wind farm layout optimization using approximate inference in graphical models · 2018-04-24 ·...

71
WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS by Aditya Dhoot A thesis submitted in conformity with the requirements for the degree of Master of Applied Science Graduate Department of Mechanical and Industrial Engineering University of Toronto © Copyright 2016 by Aditya Dhoot

Upload: others

Post on 10-May-2020

8 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

WIND FARM LAYOUT OPTIMIZATION USING

APPROXIMATE INFERENCE IN GRAPHICAL MODELS

by

Aditya Dhoot

A thesis submitted in conformity with the requirements for the degree of Master of Applied Science

Graduate Department of Mechanical and Industrial Engineering University of Toronto

© Copyright 2016 by Aditya Dhoot

Page 2: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

ii

Wind farm layout optimization using approximate inference in

graphical models

Aditya Dhoot

Master of Applied Science

Graduate Department of Mechanical and Industrial Engineering

University of Toronto

2016

Abstract

Wind farm layout optimization (WFLO) determines the optimal location of wind turbines

within a fixed geographical area to maximize the total power capacity of the wind farm,

under stochastic wind conditions and non-linear aerodynamic interferences between

turbines. This thesis develops optimization approaches to fast approximate (sub-optimal)

turbine layouts to aide engineers make design decisions. Building on previous work in

discrete quadratic WFLO models, we recast the program as a probabilistic graphical

model incorporating spatial dependencies (i.e., aerodynamic interferences, proximity

constraints, and maximum number of turbines) between the variables. Turbine layouts

are estimated using message passing inference (BP, TRW-S), which exploit the

problem’s graph-theoretic structure using decomposition and factorization. We perform

an exhaustive computational study comparing TRW-S with branch-and-cut algorithms

under varying wind-regime complexity and problem resolutions. We demonstrate the

broad applicability of techniques we develop by solving a suite of benchmark quadratic

knapsack problems, a general class of problems that arise in many settings.

Page 3: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

iii

Acknowledgements

I sincerely thank Professor Cristina Amon for believing in my academic abilities and giving me an

opportunity to pursue my research interests. Her support and guidance have helped me to grow

as a researcher, and have made the last two years academically and personally enriching. I am

particularly grateful to her for being incredibly supportive of my decisions when I chose to defer

my graduate studies to pursue entrepreneurial projects.

I owe a great debt of gratitude to Dr. David Romero for being a supportive research advisor.

His open-mindedness and patience towards exploring new ideas and concepts encouraged me

to think outside the box and find inspiration when it is least expected. I’ve become a better

researcher and engineer because of David.

I am thankful to Professor Daniel Frances and Professor Roy Kwon for serving on my thesis

committee. Their insights into optimization theory and engineering practices helped me improve

this thesis. I would also like to acknowledge Professor Brendan Frey, his graduate class and

research partially formed the inspiration to conduct this research.

I am grateful to my undergraduate thesis advisor, Professor Timothy Chan for giving me the

opportunity to explore the world of operations research. I really enjoyed his LP class, and his clear

and organized teaching helped me understand and conceptualize ideas in the formative stages

of my research.

I am thankful to the many friends I’ve made while sharing an often overcrowded office space

– Juan Stockle, Enrico Antonini, Jim Kuo, Carlos Da Silva, Sami Sorkhabi, Matthew Doyle, David

Guirguis, Armin Taheri, Fernan Saiz, Francisco Contreras, and Sean Crawford. In particular, I am

thankful to Jim for helping me navigate through the many research ideas. I would also like

acknowledge Peter Zhang for introducing me to the research group and taking the time to explain

his work.

I am grateful to my parents for their support and encouragement, and to whom this thesis is

dedicated to. I would also like to acknowledge my sister for her help and from whom I’ve learned

the art of rigorously applying logical fallacies to get your way. And now, last but not least, I would

like to thank my furry friend, Frodo for always sticking by my side through thick and thin. His

determination and dedication to fetching the ball has always been a source of inspiration.

Page 4: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

iv

Contents

1 Introduction ................................................................................................................ 1

1.1 Motivation ............................................................................................................. 1

1.2 Contribution .......................................................................................................... 1

1.3 Organization ......................................................................................................... 3

2 Literature Review ....................................................................................................... 4

2.1 Probabilistic Graphical Models ............................................................................. 5

2.2 Literature Review ................................................................................................. 6

2.2.1 Evolutionary Algorithms................................................................................ 7

2.2.2 Mathematical Programs ............................................................................... 8

2.2.3 Message-Passing Algorithms ..................................................................... 11

2.2.4 Loopy Belief Propagation ........................................................................... 12

2.2.5 Max-Product Linear Programming ............................................................. 13

2.2.6 Tree-Reweighted Message-Passing .......................................................... 14

2.2.7 Tightening LP Relaxations using Message-Passing Algorithms ................. 15

2.3 Discussion .......................................................................................................... 16

3 Quadratic Knapsack Problem ................................................................................. 18

3.1 Introduction ........................................................................................................ 19

3.1.1 Quadratic Knapsack Problem ..................................................................... 19

3.2 Related Work ..................................................................................................... 20

3.2.1 Message Passing Algorithms ..................................................................... 21

3.3 Method ............................................................................................................... 23

3.3.1 Penalty Method .......................................................................................... 23

3.4 Experiments and Discussion .............................................................................. 24

3.4.1 Benchmark Experiments ............................................................................ 25

3.5 Conclusion ......................................................................................................... 27

4 Wind Farm Layout Optimization ............................................................................. 28

4.1 Introduction ........................................................................................................ 29

4.2 Literature Review ............................................................................................... 30

Page 5: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

v

4.3 Background ........................................................................................................ 32

4.3.1 Wake Modelling .......................................................................................... 32

4.3.2 Wind Farm Layout Optimization Mathematical Program ............................ 33

4.3.3 WFLO as a Markov Random Field ............................................................. 35

4.3.4 Message-Passing Algorithms ..................................................................... 38

4.4 Experimental Results and Discussion ................................................................ 40

4.4.1 Case 1 – 100 cells ...................................................................................... 42

4.4.2 Case 2 – 400 cells ...................................................................................... 47

4.4.3 Case 3 – 2,500 cells ................................................................................... 50

4.4.4 Discreteness Analysis ................................................................................ 53

4.5 Conclusions........................................................................................................ 55

5 Concluding Remarks ............................................................................................... 58

5.1 Conclusions ........................................................................................................ 58

5.2 Future work ........................................................................................................ 58

5.2.1 Wind Farm Layout Optimization ................................................................. 59

5.2.2 Message passing algorithms for inference and optimization ...................... 60

Bibliography ................................................................................................................ 62

Page 6: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

1

Chapter 1

Introduction

1.1 Motivation

Increasing levels of greenhouse emissions threaten to irreversibly damage the climate

and eco-system. There is an immediate need to produce sustainable energy to meet the

world’s growing energy demand. In 2014, a total of $270 billion was invested in renewable

energy technologies that resulted in an addition of 128GW power capacity, of which 37%

was due to new wind farm installations [1]. Therefore, capturing wind energy will play a

major role in our shift to sustainable energy generation. A common problem encountered

while designing a wind farm is to determine the optimal placement of wind turbines within

a fixed geographical area to maximize the total power capacity of the wind farm under

stochastic wind conditions and non-linear aerodynamic interferences between the

turbines. This is known as the wind farm layout optimization (WFLO) problem. In this

thesis, we investigate mathematical models and optimization techniques that can fast

approximate turbine layouts to aide engineers to quickly assess and plan infrastructure

design decisions. While the principal goal of this thesis is to build optimization methods

for the WFLO problem, we demonstrate that the methods and techniques we develop are

widely applicable to optimal resource allocation problems constrained with a limited

budget capacity that are encountered in many fields, such as computer vision, natural

language processing, computational biology, telecommunications, and finance, among

others.

1.2 Contribution

In this thesis, we discretize the wind farm and use a quadratic formulation of the WFLO

problem developed by Turner et al. [2] to model a probabilistic undirected graphical

model, also known as Markov random fields (MRFs), which are an useful abstraction that

Page 7: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

2

succinctly capture the spatial dependencies between the variables such as aerodynamic

interferences caused by wind regimes, and turbine proximity constraint. Furthermore, the

modelled WFLO problem has a single linear inequality constraint (budget constraint) that

restricts the number of turbines that can be placed on the terrain. We use Jensen’s wake

model [3] to determine the pairwise wake interactions between turbines. However, in

contrast with previous works that use Jensen’s wake model, we do not simplify the shape

and distance of the wake interactions.

We demonstrate that the modelled WFLO problem is structurally similar to the

quadratic knapsack problem (QKP), which is a well-studied problem with NP-Hard time

complexity and for which some algorithms that seek an exact solution have been

developed [4]. Rader et al. [5] established that the QKP has graph-theoretic properties

that can be exploited to better understand the availability of polynomial approximability,

or lack thereof under various problem conditions. Using this insight, we review message

passing algorithms, a class of decentralized and asynchronous probabilistic inference and

optimization methods that operate on graphical models by iteratively passing beliefs

locally along its edges to determine the maximum-a-posteriori (optimal) configuration of

the graph’s random variables. Belief propagation (BP) [6], Sequential Tree-Reweighted

Message Passing (TRW-S) [7], and Max-Product Linear Programming (MPLP) [8] are all

types of message-passing algorithms that have been successful in practice at

determining a good configuration (sub-optimal) of dense and large graphs in problems

that arise in computer vision [9] – [10], machine learning [11], and computational biology

[12]. While much remains to be investigated about the algorithms’ theoretical guarantees,

their deep rooted relationship with linear programming, duality theory, and probabilistic

graphical models forms the basis for our motivation to apply these algorithms to the WFLO

problem.

As message passing algorithms typically operate on an unconstrained objective

function, we derive an unconstrained form of QKP using Lagrangian augmentation and

penalty functions to incorporate the budget constraint into the graphical model. As a

number of number of important sub-routines in many fields can be described as a QKP,

we apply BP and TRW-S on benchmark QKP datasets of varying graph density and

Page 8: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

3

number of random variables, and then compare the obtained results with the ones from

exact optimization methods (e.g., branch-and-bound, cutting planes, etc.). We also

conduct a thorough computational study to compare the effectiveness of turbine layouts

produced using TRW-S with computationally exhaustive branch-and-cut algorithm under

varying wind-regime complexity and discrete resolution. Additionally, to the best of our

knowledge, we generate results for the WFLO problem at a higher discrete resolution

than any existing works in literature.

As QKPs and the modelled WFLO problem are computationally intractable and in

certain cases in-approximable, we don’t intend to find optimal solutions using BP and

TRW-S. We want to assess the empirical performance of these algorithms to generate

good wind farm layouts in comparison with exact algorithms within a given time budget.

1.3 Organization

In Chapter 2, we review the literature on WFLO models and the variety of optimization

methods used to solve them. We also review probabilistic graphical models, exact

inference, and message passing algorithms such as BP, TRW-S, and MPLP. In Chapter

3, we build an undirected graphical model to represent QKPs and solve them using TRW-

S. In Chapter 4, we perform a thorough computational study comparison between TRW-

S and branch-and-cut methodology to solve the WFLO problem using varying wind-

regime complexity and problem resolution. Finally, in Chapter 5 we provide concluding

remarks and directions for future work.

Page 9: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

4

Chapter 2

Background

As wind farm layout optimization (WFLO) is a problem that consists of optimal turbine

placement based on an accurate model of the non-linear aerodynamic interferences

(wake effects) between the turbines, existing work focuses primarily on two key areas –

constructing accurate wake models and improving layout optimization algorithms.

However, sophisticated models that try to accurately portray physical and mechanical

properties make the problem non-convex, non-linear, and computationally intractable. In

this thesis, we follow previous work and use a simple analytical wake model (Jensen’s

wake model [3]) to calculate the non-linear aerodynamic interferences, and work on

developing novel optimization approaches to solve the WFLO problem. In this chapter,

we thoroughly review the various heuristic and exact optimization techniques that have

been used to solve the WFLO problem.

In this thesis, we make an extensive use of message passing algorithms, graphical

models, and probabilistic inference to solve the WFLO problem and quadratic knapsack

problems (QKPs). Therefore, we initially derive the underlying mathematical framework

describing probabilistic graphical models. Subsequently, we compare the various

message passing algorithms by discussing their advancements, benefits, and limitations.

Page 10: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

5

2.1 Probabilistic Graphical Models

A graph G = (𝒱, ℰ) is a set of nodes or vertices 𝒱 connected by edges ℰ, where the edges

between the nodes represent an underlying mathematical relationship between the

nodes. A probabilistic graphical model is a graphical representation of a joint probability

distribution of a set of random variables [13]. Therefore, in a probabilistic graphical model,

the nodes represent random variables in the probability distribution and the edges

connecting the nodes capture the dependencies between the random variables.

Probabilistic graphical models can be directed or undirected. Directed graphs are

acyclic and the edges have a directionality that encapsulates further conditional

dependency between the random variables. In undirected graphs or Markov random

fields (MRF), the edges have no directionality and maybe cyclic. MRF can be

decomposed as a product of factors composed of a set of random variables XC over the

set of maximal cliques C of graph G [13].

Given a set of random variables 𝐗 = (X1, X2, … , Xn), the joint distribution of a pairwise

MRF, shown in Eq. (2.1) is proportional to the product of potential functions θC(𝐗𝐂) over

maximal cliques C, where Z is the partition function denoted by Eq. (2.2). The partition

function normalizes the joint distribution.

p(𝐗) =1

Z∏ θC(𝐗𝐂)C (2.1)

Z = ∑ ∏ θC(𝐗𝐂)C𝐗 (2.2)

The positive potential functions θC(𝐗𝐂) encode the problem specific relationship

between the random variables. As many combinatorial optimization problem formulations

can be represented as a MRF, the potential function can be thought of as encoding the

optimization objective function [14]. Furthermore, the potential function is generally

expressed as an exponential function as denoted in Eq. (2.3), where E(𝐗𝐜) represents the

energy function. Thus, random variable configuration with a higher probability have a

lower overall energy [15].

θC(𝐗𝐂) = exp(−E(𝐗𝐜)) (2.3)

Page 11: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

6

Furthermore, the product of the exponential functions results in the sum of factors of

energy functions as shown by Eq. (2.4), where φs(Xs) represent all unary potentials over

the model vertices 𝒱 and φst(Xs, Xt) represent the pairwise potentials over the edges ℰ.

E(𝐗) = ∑ φs(Xs)s∈𝒱 + ∑ φst(Xs, Xt)(s,t)∈ℰ (2.4)

Then, the joint distribution of all the random variables is given by Eq. (2.5), which is

generally known as the Gibbs distribution [15].

p(𝐗) =1

Zexp(− ∑ φs(Xs)s∈𝒱 − ∑ φst(Xs, Xt)(s,t)∈ℰ ) (2.5)

The problem of determining the configuration of random variables in a graphical model

that yields the maximum probability over the joint distribution is known as maximum-a-

posteriori (MAP) inference problem as formulated in Eq. (2.6). This is equivalent to

minimizing the energy function as shown in Eq. (2.7).

𝐗∗ = argmax𝐗

p(𝐗) (2.6)

𝐗∗ = argmin𝐗

E(𝐗) = argmin𝐗

(∑ φs(Xs)s∈𝒱 + ∑ φst(Xs, Xt)(s,t)∈ℰ ) (2.7)

In many cases, the decomposable factors within the graphical model can be exploited

to efficiently compute the optimization problem.

We propose to use a MRF based representation to model the relationship between all

pairwise discrete locations on the wind farm by encoding the appropriate wake

interactions for any given wind speed and directions within the pairwise edge potentials.

An optimal wind farm configuration can be obtained by determining the MAP estimation

of the resulting graphical model.

2.2 Literature Review

The overall goal of WFLO research community is to generate turbine layouts that produce

maximum energy, while adhering to several project specific constraints such noise,

Page 12: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

7

infrastructure costs, etc. WFLO is a challenging problem because it is not only difficult to

accurately model wind turbine behaviour, but also a difficult discrete optimization task.

Therefore, existing work on wind turbine layout optimization focuses primarily on two key

areas – improving the model accuracy, and improving optimization algorithms.

In order to improve the model accuracy researchers are studying the physical and

mechanical relationships between turbine properties and environmental factors.

Sophisticated models that try to accurately portray physical and mechanical properties

make the problem non-convex, non-linear, and generally intractable. In this review, we

initially focus on the success and limitations of the optimization algorithms that have been

applied to various formulations of the WFLO problem. We then review a class of message

passing algorithms that have been widely applied within the field of computer vision and

computational biology with much success.

2.2.1 Evolutionary Algorithms

Two decades ago when the WFLO problem started to garner the interest of the

engineering research community, evolutionary algorithms were the first set of methods

applied to solve the optimization problem. Evolutionary algorithms are biologically

inspired metaheuristic optimization methods that generate populations of candidate

solutions and evolve them until a convergence criterion is met, in the hope that the final

population of solutions contains the optimal solution to the problem. Depending on the

search space, problem representation, and evolution operators used (e.g., the crossover

strategy used to generate new candidate solutions), evolutionary algorithms can in many

cases find the globally optimal solution for problems that are difficult to solve using other

methods.

In 1994, Mosetti et al. [16] were the first group of researchers to formulate the WFLO

problem and use a genetic search algorithm to find optimal turbine placements on a

simple 10 x 10 square cell grid that maximizes energy, while minimizing installation costs

and wake interactions for single directional and multi-directional wind scenarios. Even

though simple assumptions regarding the turbine type, wake modeling, and terrain data

are made, the promising results encouraged the entire wind engineering community to

Page 13: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

8

broadly investigate the wide applicability of using evolutionary algorithms to solve realistic

WFLO problems.

In 2004, Grady et al. [17] also used genetic search algorithm and modeled a similar

objective function that was presented by Mosetti et al. in [16]. Grady et al. obtained better

layouts with higher power output and lower installation costs compared to the results

published in [16] by increasing the simulation population size and number of generations.

This shows that by fine tuning the parameters involved in genetic algorithms such as

crossover type, diversity percentage, and population size the overall results of the

algorithm can be drastically improved.

Huang et al. [18] further enhanced earlier approaches by applying a distributed genetic

algorithm to decompose a large wind farm terrain search space into local search spaces

to maximize annual wind farm profits. It is shown that a distributed approach yields better

solution quality and is computationally faster than traditional genetic algorithms. Huang

et al. also suggests the use of local search techniques to further enhance the distributed

genetic algorithm.

The evolutionary methods described here used to solve the WFLO problem work well

with small-scale non-linear and non-convex problems. These methods, however, do not

provide any optimality guarantees; therefore, solutions generated maybe sub-optimal and

further methods need to be investigated to achieve globally optimal solutions.

2.2.2 Mathematical Programs

By discretizing the wind farm terrain into grid cells and representing each cell as a binary

decision variable, mathematical programming models can be developed for the WFLO

problem. Individual decision variables take on a binary state and determine whether at

every location a turbine is placed or not placed. The relationship between the decision

variables (e.g., wake interactions, infrastructure costs, etc.) can be encoded within the

objective function and hard constraints can be developed for budget, spacing, proximity,

terrain usage, noise, and other infrastructure constraints. Existing research on formulating

the WFLO problem as a mixed integer program (MIP) and quadratic integer program

Page 14: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

9

(QIP) has been conducted with promising results. Integer programs can be solved exactly

using well-designed algorithms [19] that are generally implemented in an off-the-shelf

commercial solver such as CPLEX (IBM Corp., Armonk, NY), and have several theoretical

guarantees, which is one of the biggest advantages to model WFLO problem as a

mathematical program. However, in many cases, models need to be simplified to make

the problem tractable and convex.

Donovan [20] [21] used a MIP model to formulate the WFLO problem and applied

branch-and-bound method to solve the problem. Donavan exploits the inherent graph

based structure of the problem to model the problem as a vertex packing problem. Edges

in the graph represent turbine proximity and wake interference relationship between

potential turbine placements. Furthermore, maximal cliques are used limit local placement

of turbines and satisfy turbine proximity constraints. Donovan notes the limitations of the

constraints used within the model, and suggests that a hybrid heuristic algorithm and MIP

can be applied together to handle more realistic non-linear constraints and improve non-

optimal solutions generated by the MIP.

Donovan in [21] explored varying branching strategies to reduce the time taken to

reach optimality. The paper dynamically generates violated constraints as needed to

significantly reduce the time required to find the optimal solution. A brief analysis on the

implications of linear programming (LP) relaxation of the integer constraints is conducted.

Donovan also claims that the branching strategy will garner benefits up to a maximum

placement of 50 turbines beyond which more efficient techniques need to be developed.

Nevertheless, this approach illustrates that good solutions can be generated for simplified

models that can be used along with hybrid heuristic algorithms with non-linear constraints.

Similarly, Archer et al. [22] formulated a WFLO MIP model that minimizes the wake

interactions between turbines. Wake interactions are modelled using the PARK model

[23] and a Weibull distribution is implemented to handle variability in the wind velocity.

The authors were able to place up to 25 wind turbines and the resulting solution quality

outperforms those generated using other heuristic algorithms.

Turner et al. [2] formulate the discrete wind farm layout problem as a quadratic integer

program (QIP). Jensen’s wake model is used to calculate the wake interactions between

Page 15: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

10

turbines, and the overall formulation determines turbine placement that minimizes the

overall kinetic energy deficit. Furthermore, the paper also uses Adams-Sherali zero-one

quadratic program linearization technique to formulate a MIP. CPLEX 12.1 (IBM Corp.,

Armonk, NY) was used to find optimal turbine placements for a multi-directional wind

problem on a 10 x 10 (2km by 2km) grid. Solution quality for placing 10 to 40 turbines

was better than solutions from Mosetti et al. [16] and Grady et al [17]. Nevertheless, it

could take up to 24 hours to solve the problem, which is non ideal as a realistic WFLO

problem would consist of further constraints and a higher dimensional objective function.

Thus, there is a need for developing algorithms that quickly produce good solutions

(approximate) for mathematical program formulations of single and multi-objective WFLO

problems.

Recognizing the limitation of existing WFLO MIP models to incorporate realistic

nonlinearities in the problem, in 2014, Zhang et al [24] proposed constraint programming

(CP) and MIP models that takes into account the inherent non-linear aerodynamic

behaviour of the wake interactions among multiple wind turbines. The paper showed that

the MIP based decomposition models outperforms CP models. Additional models with

noise and land-owner participation constraints were also solved. The results from [24] are

promising and further research in mathematical programs for the WFLO problem needs

to be conducted in this direction.

In general, the application of MIP models and constraint programming to WFLO

problem has been recent but very promising. These models offer exact solutions and

optimality guarantees as opposed to evolutionary algorithms. Nevertheless, further

research in this area is required to develop models that consider the full non-linearity of

the wake interactions, stochastic wind conditions, difficult constraints, and are also

capable of encoding multiple objectives, while solving the problems in a reasonable

amount of time. Furthermore, a hybrid approach using mathematical models alongside

with evolutionary algorithms is also an area of research worth exploring.

Page 16: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

11

2.2.3 Message-Passing Algorithms

A large class of discrete optimization problems can be posed as a maximum-a-posteriori

(MAP) inference problem in a probabilistic graphical model by recasting the objective cost

function as a graph structured probability distribution known as Markov random fields

(MRFs). Vertex cover (e.g. [25]), facility location (e.g. [26]), quadratic assignment (e.g.

[27]), and resource allocation (e.g. [28]) are some examples of discrete optimization

problems that have been reformulated and solved as MAP inference problems.

A class of algorithms known as message passing algorithms approximate the MAP

assignment by iteratively passing beliefs along the edges of the graphical model in a

distributed, decentralized, and asynchronous manner. Message passing algorithms

exploit the decomposability of the graphical model to find the optimal configuration. MIP,

QIP, and CP WFLO models with a graph-theoretic structure can be posed as a MAP

inference problem. Therefore, it is hypothesized that message passing algorithms can be

used to efficiently approximate layout configurations of discretized WFLO model with a

graph based structure.

Pearl [6] introduced a type of message passing algorithm called the max-product belief

propagation (BP) that is guaranteed to converge to the optimal MAP configuration for

graphical models that contain only trees. BP is not guaranteed to converge or find optimal

configurations for graphical models that include cycles. Nevertheless, empirical evidence

has shown that BP can produce good results even for graphs with cycles in the fields of

computer vision (e.g. [9] and [10]) and computational biology (e.g. [12]). Even though BP

cannot be directly applied to graphical models with cycles, understanding their

convergence properties has helped researchers develop several variants of BP

algorithms that have been empirically successful at finding optimal MAP configurations.

As performing exact inference in graphical models is an NP-Hard problem, Wainwright

et al. [29] argue that fixed points, the stationary points of the energy function Eq. (2.7), of

message passing algorithms do not generally yield globally optimal solutions;

nevertheless, these algorithms in practice yield good sub-optimal MAP configurations.

Page 17: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

12

2.2.4 Loopy Belief Propagation

Loopy Belief Propagation (LBP) is simply BP applied over graphical models with cycles

(“loops”) to estimate the MAP configuration. In [30] Sanghavi et al. used Loopy Belief

Propagation to determine the MAP estimation of the weighted matching problems on

general graphs. In general, weighted matching problems can be formulated as integer

programs; however, the distributed and asynchronous characteristics of message

passing algorithms have been used to empirically solve the problem efficiently. The paper

proves that LBP converges to the optimal weighted matching if the linear programming

relaxation is tight, and does not converge when the linear programming relaxation is not

tight for a given weighted matching problem. Thus, any WFLO problem modelled as a

weighted matching problem can be optimally and efficiently solved using LBP if the

resulting LP relaxation of the problem is tight. However, this is generally not the case.

Therefore, further variants of message passing algorithms need to be generated that work

well with cycles.

Additionally, as many real-world objective functions are non-convex it is always ideal

to develop message passing algorithms that are able to solve these problems to

optimality. For instance, Turner et al. [2] formulated the WFLO problem as a QIP model,

which is generally non-convex and thus linear approximations of the problem yield

solutions that are not globally optimal. In [31] Kumar et al. used the concave-convex

procedure (CCCP) with message-passing algorithm to estimate the MAP inference of the

relaxed quadratic integer formulations of pairwise MRF. The CCCP technique

decomposes a non-convex quadratic objective function into a difference of two convex

functions. Using CCCP with message passing algorithms is guaranteed to converge to a

local minima or a global minimum for a non-convex and convex quadratic programs

respectively. This methodology results in richer solution quality than using standard LP

formulations of max-product for Ising graphs and protein design problems. This technique

can be applied to the QIP model generated in [2], however, the relaxed QIP is not always

tight.

Similarly, Ravikumar et al. [32] formulated a relaxed QIP to estimate the MAP

inference. They show that a quadratic program solves the MAP solution exactly as it

Page 18: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

13

accurately represents the energies from a pairwise Markov random field compared to a

more commonly used linear formulation of the MAP problem. Furthermore, simple convex

approximations are made to make non-convex quadratic formulations convex. Standard

solvers and iterative projected conjugate gradient methods are used to solve small and

large problems respectively. It is shown that computing a quadratic relaxation program

exactly is sometimes faster and yields better solutions than LP Relaxation, iterative

conditional modes, and tree-reweighted max product algorithms.

Even though traditional integer, quadratic, and linear programming algorithms from

[33] can be applied to solve MAP problems, it is sometimes more efficient to exploit clique

factors within the graphical model to efficiently compute solutions. Furthermore, in many

cases very large graphs are cannot be solved using techniques in [33] in a limited amount

of time. Thus, it is important to know when message-passing algorithms can be applied

to best exploit graphical model structure and solve for its MAP configuration.

Since this thesis will be using message passing to tackle WFLO MIP models, we need

to further investigate message passing algorithms that are able to efficiently compute

integer programs that are not tight when relaxed. Two such variants of belief propagation

that efficiently compute MAP configuration for MRFs with cycles are Max Product Linear

Programming algorithm (MPLP) and Tree-Reweighted Message Passing (TRW).

However, both these methods relax the integer constraints to solve the problem.

Therefore, we also need to investigate methodologies that can be used to dynamically

generate constraints to make the solution space results tight.

2.2.5 Max-Product Linear Programming

Globerson et al. [8] introduced a new type of max-product message passing algorithm

known as Max Product Linear Programming algorithm (MPLP). MPLP uses block

coordinate descent in the dual of the LP Relaxation of the MAP problem to compute

beliefs between nodes. MPLP is guaranteed to converge and computes the exact solution

to the relaxed LP problem for binary variables. MPLP explicitly exploits the graph structure

of graphical models that traditional LP solvers fail to exploit.

Page 19: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

14

2.2.6 Tree-Reweighted Message-Passing

In [29], Wainwright et al. show that by decomposing an arbitrary graphical model with

cycles into a convex combination of tree-structured distributions, an optimal upper bound

can be calculated for the MAP estimation problem. The authors show that an optimal MAP

configuration and a tight lower bound can be obtained if and only if the combination of

tree distributions share a common optimum (known as tree agreement). This property is

derived by initially formulating a LP by relaxing the integer polytope into loose tree-based

consistency conditions and then solving its Lagrangian dual. Furthermore, Wainwright et

al. in [29] apply an iterative tree-reweighted message passing algorithm that

reparametrizes the tree-structured distributions to determine fixed points that specify the

solutions to the Lagrangian dual. Lastly, it is shown that when the relaxed integer

constraints yield fractional values, the relaxation can be tightened by forming additional

clusters of random variables within the graphical model.

As demonstrated in [29], TRW works by maximizing the lower bound of the free energy

objective. However, Kolmogorov in [7] shows that TRW does not guarantee to increase

this lower bound, and in certain situations may actually decrease the lower bound value

during the iterative algorithm. Furthermore, TRW is not guaranteed to converge. In [7],

Kolmogorov develops sequential tree-reweighted message passing algorithm (TRW-S)

that guarantees that the lower bound will never decrease.

Messages are also updated sequentially rather than in parallel. Tested on benchmark

computer vision problems, TRW-S yields better results and is memory efficient compared

to TRW and MPLP. However, it is important to note that TRW-S does not work well when

the resulting relaxed LP formulation is not tight.

In [34], Kolmogorov et al. derive key optimality conditions of TRW-S for problems with

binary variables. The derivation decomposes an arbitrary graph into a convex

combination of trees to illustrate that any fixed point (stationary points of the energy

function) that satisfy the weak tree agreement can be used to specify a subset of a

globally optimum solution [34]. Similar to MPLP, it is shown that for problems with binary

variables a weak tree agreement fixed point always determines the optimum of the

relaxed linear program of the MAP estimation problem. Lastly, the properties derived in

Page 20: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

15

[34] are similar to the results found in integer programming literature of Rhys relaxation,

roof duality, and weak persistency.

Thus, given MPLP’s and TRW’s optimality guarantees and computation efficiency this

paper encourages researchers to solve discretized WFLO problems using message

passing algorithms. However, it is important to note that both MPLP and TRW are not

general purpose LP solvers, they both work well when the graph can be decomposed into

factors and sub-problems.

Yanover et al. [35] conducted a thorough empirical study to determine the similarities

and differences in using off-the-shelf LP solvers versus TRW to estimate the MAP

problem for common graphical models generally found in computational biology and

computer vision. The paper demonstrates that a MAP problem can be represented as a

type of combinatorial optimization problem using an integer program with binary indicator

variables. The paper shows that even when the integer constraints are relaxed, barrier

and dual LP solvers in CPLEX (IBM Corp., Armonk, NY) fail to solve problems that have

over 106 variables on standard desktop computers. The paper demonstrates that due to

the sparse nature of the indicator variable constraints, TRW exploits the relaxed LP

structure to obtain good MAP configurations faster than CPLEX (IBM Corp., Armonk, NY).

However, only when the LP solution is non-fractional (tight) that the TRW beliefs had a

unique MAP configuration.

2.2.7 Tightening LP Relaxations using Message-Passing Algorithms

Many message passing algorithms rely on the underlying assumption that the resulting

linear program is tight when the integer constraints are relaxed to continuous constraints.

However, this is hardly the case for general real-world MAP estimation problems. More

specifically, none of the discretized WFLO problems discussed earlier are tight when their

integer constraints are relaxed to continuous constraints. Thus, in order to use the

efficient message passing algorithms we need to further investigate ways in which integer

constraints can be systematically enforced.

Traditionally, cutting-plane algorithms have been employed within the combinatorial

optimization community to enforce violated integer constraints. For instance, Gomory cuts

Page 21: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

16

are widely used in traditional integer program solvers to enforce integer constraints.

However, these methods are generic and not problem specific, which results in inefficient

[14].

In [36] and [37] Sontag et al. demonstrate that tighter relaxations can be achieved by

implementing higher order cluster-based LP relaxations that enforce local consistency.

Adding clusters is a lift-and-project method where the goal is to develop a relaxed

polytope that approaches the boundaries of the original integer polytope. This is achieved

by dynamically enforcing relaxations until the relaxed polytope is equal to the integer

polytope. It is expected that by developing a hierarchy of relaxations that the integer

polytope can be approximated without having to enforce all the constraints. Therefore, in

[36] clusters are incrementally added to the graphical model only when they are

guaranteed to improve the approximation, thus, as more clusters are added the relaxation

becomes tighter.

Therefore, it is hypothesized in this thesis that by applying message passing

algorithms to the relaxed WFLO MIP model while enforcing key constraints using cluster

based approach that good sub-optimal can be generated.

2.3 Discussion

WFLO is in general an intractable problem due to stochastic wind conditions, non-linear

aerodynamic wake interferences, and other difficult constraints. Simplified wake models

and linear constraints are used to make the problem tractable. In this review we have

shown that evolutionary algorithms can be used to develop heuristics that enable

researchers to solve the non-convex WFLO problems with nonlinearities. Nevertheless,

these algorithms are extremely slow and generally do not offer any optimality guarantees.

Further simplifications can be performed to formulate the WFLO problem as discretized

MIP, QIP, and CP models. Off-the-shelf exact solvers guarantee optimality but once again

are extremely slow and they fail to solve problem instances with a large number of

variables, for instance, Zhang et. al [24] solved the 20 x 20 WFLO instance, the largest

known instance, using mathematical programming with limited success. Therefore, there

is a need for an optimization methodology that is able to quickly approximate wind farm

Page 22: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

17

layouts for a general class of problem instances, considering stochastic wind conditions,

non-linear aerodynamic wake interferences, and other difficult constraints for a given

formulation.

In this chapter, we have illustrated through existing literature that many combinatorial

optimization problems can be modeled as a probabilistic graphical model and MRF, and

then MAP configuration can be found by applying message passing algorithms. Message

passing algorithms work best when the graphical model can be decomposed into factors

of small cliques. More specifically, we show that MPLP and TRW algorithms work well

with MRF with cycles; however, these algorithms relax the underlying integer constraints

to solve the resulting LP. Thus, MPLP and TRW need to be applied alongside with lift-

and-project methods that enforce a series of hierarchical relaxations until the relaxed

polytope approximates the integer polytope near the corner point with optimal solution.

The underlying rationale of the method I propose is to frame the existing WFLO integer

problem formulation from [2] as a maximum-a-posteriori (MAP) inference problem in a

probabilistic graphical model by recasting the objective cost function as a graph structured

probability. Then, I will apply an approximate inference method using message passing

algorithm augmented with cluster based LP relaxations to sub-optimally approximate the

WFLO.

Page 23: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

18

Chapter 3

Quadratic Knapsack Problem

Quadratic knapsack problem (QKP) is a binary integer program with a quadratic objective

function bounded by a linear capacity (budget) constraint. QKPs arise in many fields that

require maximizing resource allocation under a constrained budget capacity. While

several techniques that derive tight upper bounds have been extensively studied and are

effective at solving super-modular QKPs, they are computationally slow and memory

intensive to solve general QKPs without making any a priori assumptions. Additionally,

although QKPs have strong graph-theoretic properties, they have not been studied in the

context of approximate inference in graphical models using message passing algorithms

to generate approximate (sub-optimal) solutions to otherwise difficult QKPs.

In this research, we use the graphical model abstraction along with Lagrangian

augmentation to represent a QKP as an undirected graph. We develop a method to solve

a QKP using sequential tree-reweighted message passing (TRW-S) augmented with

triplet clusters to tighten the results. In this chapter, we perform an experimental

comparison between the effectiveness of solving benchmark QKPs using TRW-S and an

exact solver (branch-and-bound) over a varying number of variables and graph densities.

The subsequent chapter performs a similar computational study to solve the WFLO

problem.

Page 24: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

19

3.1 Introduction

As discussed in earlier sections, undirected graphical models are a widely used

framework in machine learning to succinctly model pairwise relationships between a set

of random variables. This relationship is expressed in the form of an energy function (Eq.

(3.1)) with unary (φs(xs)) potentials over vertices 𝒱, and pairwise (φst(xs, xt)) potentials

over edges ℰ in a graph G = (𝒱, ℰ). In general, optimal inference over graphical models

is computationally intractable (NP-Hard), and using state-of-the art exact solvers (i.e.

CPLEX) to conduct exact optimization such as branch-and-bound and cutting planes as

a subroutine task is ineffective and time consuming. Recent advancements in messaging

passing algorithms, a class of decentralized and asynchronous optimization methods that

exploit graph factorization, have shown to well approximate MAP configurations with

some optimality guarantees.

E(x, φ) = ∑ φs(xs)s∈𝒱 + ∑ φst(xs, xt)(s,t)∈ℰ (3.1)

In this chapter, we model quadratic knapsack problem, a type of quadratic integer

program, as a graphical model and apply approximate inference with an expectation of

generating good sub-optimal solutions in comparison to computationally exhaustive exact

methods.

3.1.1 Quadratic Knapsack Problem

Despite major advances in message passing algorithms, their behaviour in maximizing

Eq. (3.1) under a global constraint has not been studied. The binary quadratic knapsack

problem (QKP) is a class of constrained combinatorial integer maximization problem with

a quadratic objective function (Eq. (3.2a)) and a global linear inequality constraint (Eq.

(3.2b)). Formally, QKP assumes that there are N items each with a positive weight ai, and

an integral profit di. Furthermore, for each selected item i {i|i ∈ N}, an integral profit cij can

be derived if item j, {j|j ∈ N, j > i} is also selected. Therefore, the problem involves

choosing a subset of items from N items such that the weight of chosen items does not

exceed b, known as the knapsack capacity, while maximizing the overall profit. The

objective function (Eq. (3.2a)) is an undirected graph with unary and pairwise potentials,

which is constrained by a linear capacity constraint (Eq. (3.2b)). The linear constraint,

Page 25: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

20

also known as a knapsack polytope, appears in optimization formulations with a limited

budget capacity.

maximize𝐱 𝑓(𝑥) = ∑ ∑ cijxixjj∈Ni∈N + ∑ dixii∈N (3.2a)

subject to ∑ aixii∈N ≤ b (3.2b)

xi ∈ {0,1} ∀i ∈ {1, … , N} (3.2c)

The special case when profits cij = 0, the problem reduces to the standard knapsack

problem, which even though is a NP-hard problem can be solved using dynamic

programming in pseudo-polynomial time complexity O(bN) [38]. QKP is NP-hard in the

strong sense as shown by its reduction from the Clique problem in [39]. Contrary to

standard knapsack problem, no such pseudo-polynomial algorithm exists that can solve

all QKPs.

We motivate our study of QKPs as they commonly appear as sub-routines within a

wide variety of fields such as transportation planning [40], electronic circuit design [41],

finance [42], and wind farm layout optimization (WFLO) [2]. In this thesis, we design

methods that exploit the effectiveness of message passing algorithms on undirected

graphs bounded by a knapsack polytope to approximate a variety of QKPs.

3.2 Related Work

Existing work to solve QKPs mostly focusses on advancing exact algorithms such as

branch-and-bound by developing methods to generate better bounds and branching

strategies that yield a lower duality gap and a tighter integrality gap. Therefore, techniques

such as Lagrangian decomposition [43] [44], Lagrangian relaxation [39], semidefinite

programming [45], upper planes [46], linearization [47] and heuristic methods [48] have

been widely studied in the context of solving super-modular QKPs. A QKP is called super-

modular when all quadratic profits cij ≥ 0. Although these methods yield solutions with

optimality guarantees, they are ineffective at solving QKPs with a very large number of

variables, arbitrary profits, and in cases where QKPs appear as a subroutine as part of a

much larger optimization problem.

Page 26: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

21

Therefore, we need to investigate approximate and heuristic algorithms that perform

reasonably well with general QKPs without making any a priori assumptions and when

the size of the problem scales exponentially. More specifically, for practical purposes we

need to test the viability of approximate algorithms that seek good solutions without

exhaustively searching for optimality. Rader et al. [5] illustrated that the graphical

representation of the problem’s objective function gives insight into the problem’s

solvability, that are otherwise difficult to understand. Rader et al. prove that graph’s that

are edge series-parallel (ESP) can be solved in pseudo-polynomial time in O(b2N) time

complexity using dynamic programming. Furthermore, a fully polynomial time

approximation scheme (FPTAS) is available when the graph is ESP and all profits, cij

and di, are non-negative (super-modular). Rader el al. [5] also show that graphs that are

vertex series-parallel (VSP) are strongly NP-Complete without a pseudo-polynomial time

solution unless P = NP. QKPs with positive and negative profits are in-approximable

unless P = NP. Similarly, Rader el al. [5] derive that tree-based graphs with arbitrary

profits do not have polynomial time approximation unless P = NP. It is important to note

that these results have only been studied under conditions with integral costs, cij and di.

3.2.1 Message Passing Algorithms

As Rader el al. [5] shows that the underlying graph-structure of the objective function

yields significant insight into the difficulty of QKPs, it is natural to investigate the

application of message passing algorithms to solve QKPs, as these algorithms have long

been used to factorize graphs to best estimate its MAP assignment. Pearl’s Max-product

Belief Propagation (BP) [6] is a common message-passing scheme that estimates the

MAP assignment of random variables by decomposing the graph G into factors and

iteratively passing beliefs along adjacent nodes over edges in a distributed, decentralized,

and asynchronous manner. Beliefs represent the potential value of the random variable

as determined by the unary and pair-wise potential function of adjacent nodes. BP is

guaranteed to converge to the optimal MAP configuration for acyclic MRFs [10]; although

no such convergence and optimality guarantees exist for cyclical (loopy) graphs, BP has

still shown to produce very good results [10].

Page 27: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

22

Variants of BP such as Max-Product Linear Programming (MPLP) [8] and Sequential

Tree-Reweighted Message Passing (TRW-S) [7] conduct MAP estimation by linearizing

Eq. (3.1) into an integer program (MAP-IP) using auxiliary marginal variables {μs(xs)}s∈𝒱

and {μst(xs, xt)}(s,t)∈ℰ for every vertex and edge in the undirected graph G respectively.

These auxiliary variables also satisfy the conditions ∑ μs(xs)Xs= 1 and ∑ μst(xs, xt)Xs,Xt

=

1 [29]. The MAP-IP can be defined by Eq. (3.3) over the binary marginal polytope M(G)

from Eq. (3.4), where p(x, φ) is the graph G’s joint distribution characterizing its

factorization [29].

TRW and MPLP work by relaxing the marginal polytope in Eq. (3.4) and iteratively

minimizing the dual to approximate the lower bound using graph tree decompositions and

re-parameterization. These variants are guaranteed to produce optimal configurations for

marginal polytopes that are tight when relaxed.

MAP − IP: argmax𝛍∈M(G)

𝛗𝐓𝛍 = argmax𝛍∈M(G)

(∑ ∑ μs(xs)φs(xs)Xss∈𝒱 +

∑ ∑ μst(xs, xt)φst(xs, xt)Xs,Xt(s,t)∈ℰ )

(3.3)

M(G) = {𝛍 ∈ ℝd|∃p(x, φ) s. t μs(xs) = ∑ p(x, φ)x𝒱\s

μst(xs, xt) = ∑ p(x, φ)x𝒱\s,t

} (3.4)

Existing literature in the context of graphical models and probabilistic inference

remains inconclusive and uncertain regarding methods to incorporate linear inequality

constraints such as the knapsack polytope (Eq. (3.2a)) into graphical models. Therefore,

it is important to develop procedures that are able to effectively incorporate global linear

inequalities within graphical models so that message passing algorithms can become an

attractive and robust subroutine to generate good approximate feasible solutions within

large-scale optimization problems.

The purpose of this study is to determine methods to solve a variety of QKPs with

positive, negative, and real cost coefficients using message passing algorithm. We

investigate methods to incorporate knapsack inequality constraint within graphical models

to generate good approximate feasible solutions. Lastly, we test the effectiveness of these

Page 28: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

23

methods to solve QKPs against exact methods solved using state-of-the-art solvers (i.e.

CPLEX).

3.3 Method

In the context of MRFs, the QKP objective function 𝑓(𝑥) (Eq. (3.2a)) can be reformulated

as an energy function shown in Eq. (3.1), defining the unary potentials (φs) using linear

costs and pairwise potentials (φst) using quadratic costs as illustrated in Eq. (3.5a) and

Eq. (3.5b) respectively.

φs(xs) = dsxs, ∀ s ∈ 𝒱 (3.5a)

φst(Xs, Xt) = cstxsxt, ∀ (s, t) ∈ ℰ (3.5b)

There are two methods that can be used to augment MRFs to incorporate knapsack

polytope feasibility. These two methods are based on penalty methods and auxiliary

variables.

3.3.1 Penalty Method

QKP can be transformed into an unconstrained quadratic integer program (QIP) using

penalty methods based on the augmented Lagrangian function as described in [49]. It is

hypothesized that the unconstrained QIP is generally easier to solve compared to the

constrained QKP. We use a quadratic penalty function to incorporate the knapsack

capacity constraint into the quadratic objective function, and the resulting augmented

Lagrangian function with the penalty parameter, μ is shown in Eq. (3.6).

𝐿(𝑥, 𝜇) = 𝑓(𝑥) + μ [max {0,1 −1

b∑ aixij∈N }]

2

(3.6)

The undirected graph, G𝐿 representing 𝐿(𝑥, 𝜇) can have a higher or equal graph

density compared to the constrained undirected graph G. Nevertheless, both BP and

TRW-S have been applied to complete graphs with some success in [7] .

maximize𝐱 𝐿𝑘(𝑥, 𝜇𝑘) (3.7a)

subject to xi ∈ {0,1} ∀i ∈ {1, … , N} (3.7b)

Page 29: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

24

The unconstrained QIP is shown in Eq. (3.7a), where {𝜇𝑘} is a sequence of penalty

parameters such that 0 < μk < μk+1. Therefore, the unconstrained QIP in Eq. (3.7a) is

successively solved for a given penalty parameter μk until Lk(x, μk) converges to a

stationary point within the feasible domain. In general, the problem may become ill-

conditioned if μk → ∞. This procedure is described in Algorithm 1.

Algorithm 1: Penalty Method

1: Inputs: QKP with model parameters: set of cost coefficients ({d}i, {c}ij), set of knapsack

capacity coefficients {𝑎}𝑖, knapsack capacity 𝑏, and problem size 𝑁

2: Result: Approximate MAP: argmax𝒙 (𝐿(𝑥, 𝜇) = 𝑓(𝑥) + μ [max {0,1 −1

b∑ aixij∈N }]

2)

3: Choose step-size sequence {γ(i)}0∞ in descending order

4: Initialize penalty parameter 𝜇(𝑘=0) = 𝛾𝑖=0, iteration 𝑘 = 0

5: while ‖𝑓(𝒙𝒌+𝟏∗ ) − 𝑓(𝒙𝒌

∗ )‖ is below some threshold (convergence-check) do

4: Compute 𝒙𝒌

∗ ← argmax𝒙 𝐿(𝑥, 𝜇𝑘) = 𝑓(𝑥) + μk [max {0,1 −1

b∑ aixij∈N }]

2 using TRW-S

5: if 𝒙𝒌∗ is feasible (∑ ai𝑥𝑘𝑖

∗j∈N ≤ b) then

6:

Update step-size μk+1 ← μk + γi

Update iteration count k ← k + 1

7: else if 𝒙𝒌∗ is not feasible (∑ ai𝑥𝑘𝑖

∗j∈N > b) then

8:

Update step-size by reverting to previous feasible penalty parameter and incrementing

μk ← μk − γi + γi+1 Update step-size

i ← i + 1 end if

9: end while

10: Return Approximate MAP result 𝑓(𝑥𝑘) and arguments 𝑥𝑘

3.4 Experiments and Discussion

We use Algorithm 1 over benchmark dataset to establish a comparison between the

solutions generated using TRW-S and an exact solving algorithm. Results generated by

Algorithm 1 over benchmark dataset gives us insight into the benefits and limitations of

TRW-S as applied to QKPs of varying difficulties. In the subsequent chapter, we perform

a similar computational study in the context of wind farm layout optimization.

Page 30: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

25

3.4.1 Benchmark Experiments

Experiments are conducted on benchmark dataset to measure the effectiveness of

message passing algorithms with existing methods. The benchmark dataset as shown in

[46], [44], [47], [39], [50], and others use test cases that randomly generate 10 test

problems for 100, 200, and 300 variables for a range of graph densities (∆d). The set of

cost coefficients c and d for every test case are uniformly distributed integers between 0

and 100, while the set of knapsack capacity coefficients a are uniformly distributed

integers between 0 and 50, and the capacity constraint b is randomly selected integer

between 50 and max(50, ∑ aii∈N ). Billionnet et al. [50] showed these test cases to be

reasonable difficult by measuring the gap of random feasible solutions and optimal

solutions to be between 40% – 60%.

For each set of 10 test problems with a specific number of variables and graph density,

we apply TRW-S via Algorithm 1 to measure the average error (Eμ) and the standard

deviation in the average error (Eσ). We also measure the average time (Tμ) taken to

compute the MAP configuration over a convergence criteria, and the standard deviation

in the average time (Tσ) is also calculated.

In general, MAP configurations generated using TRW-S are not tight since TRW-S

computes fixed-points (extreme points) over a relaxed LP, which is transformed from a

QIP using a linearization technique. Therefore, tighter relaxations are generated by

iteratively generating triplet clusters that enforce local consistency between cluster

marginals based on cutting-plane methods and Sherali-Adams hierarchy as illustrated by

Sontag et al. [36] and [37]. Clusters are iteratively searched and added after termination

of Algorithm 1. Thus, measurements using this approach is also computed, and it is

expected that MAP configurations are tighter and yield better value at convergence.

Table 3.1 shows the results generated using the two approaches and an exact method

that uses Lagrangian decomposition to generate upper bounds [50]. We compare our

results to ones from [50], as generating upper bounds to compute exact solution is

representative of the existing direction of research in generating optimal solutions to

QKPs.

Page 31: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

26

As shown in Table 3.1, we can see that applying TRW-S using Algorithm 1 without

clusters yields sub-optimal solutions in time that is several orders of magnitude less than

the time it takes exact method to generate optimal solution. Nevertheless, these solutions

have an average error of 2-4%. Depending on the application type and solution reliability

constraints, a higher threshold of average error maybe acceptable. As searching for

triplet-clusters is a computationally exhaustive task [37], we see that generating a tighter

solution is time consuming, however, it yields better MAP assignments compared to

applying TRW-S without clusters. While both methods are unable to generate optimal

solutions for all test cases as illustrated in Figure 3.1 and Figure 3.2, generating triplet-

clusters consistently generates solutions with same or better MAP assignments as

quantified by a lower error rate.

TRW-S performs well in comparison with an exact method. Applying TRW-S with

clusters over test cases with 200 variables in a fully-connected graph produces an error

margin of less than 1% using 40 times less the amount of time as exact method.

Performance with 100 and 300 variables is also computationally comparable to exact

method. We believe that performance of TRW-S can be further enhanced by exploiting

the distributed and asynchronous properties of message passing algorithms by

implementing parallel processes.

Test Case TRW-S (Via Algorithm 1) TRW-S (Via Algorithm 1) with

triplet-cluster tightening

Exact Solution using Lagrangian

Decomposition [50]

𝐍 ∆d(%) Eμ(%) Eσ Tμ(s) Tσ Eμ(%) Eσ Tμ(s) Tσ Tμ(s) Tσ

100 25 4.00 2.75 7 5 2.71 1.52 394 175 117 122

200 100 2.19 3.34 64 37 0.66 0.93 1,656 811 70,633 167,326

300 25 3.87 3.25 219 120 3.31 3.35 7,252 1,151 7,765 7,600

50 2.41 2.12 321 127 0.69 0.63 6,814 1,291 4,786 4,591 Table 3.1: A comparison of results generated using penalty method, penalty method applied with tightening technique, and Lagrangian decomposition to generate upper bounds.

Page 32: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

27

Figure 3.1: Resulting error of MAP assignments generated using TRW with and without triplet-clusters in comparison to exact method. Test cases labelled 100:1-10 correspond to a QKP with 100 variables with 25% graph density. Test cases labelled 200:1-10 correspond to a QKP with 200 variables with 100% graph density.

Figure 3.2: Resulting error of MAP assignments generated using TRW with and without triplet-clusters in comparison to exact method. Test cases labelled 300:1-10 correspond to a QKP with 300 variables with 25% graph density. Test cases labelled 300:11-20 correspond to a QKP with 300 variables with 50% graph density.

3.5 Conclusion

In this chapter, we derived a method to solve QKPs using approximate inference in

graphical models using message passing algorithms. While unconstrained quadratic

objective function have been represented as graphical models and solved using a variety

of message passing algorithms such as BP, MPLP, and TRW-S, they have not been

studied when the quadratic objective function is constrained by a linear capacity

constraint. We incorporate this constraint into the objective function using Lagrangian

augmentation and a quadratic penalty function, thereby making the problem an

unconstrained QIP that can be represented as a graphical model. By applying TRW-S

over a variety of benchmark test cases, we showed that TRW-S can be used as an

alternative method to generate good feasible results to QKPs.

Page 33: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

28

Chapter 4

Wind Farm Layout Optimization

As rising greenhouse emissions threaten to harmfully disrupt the climate and eco-system,

there is an immediate need to produce sustainable energy to meet the world’s growing

energy demand. Capturing wind energy will play a major role in our shift to sustainable

energy generation. A common problem encountered while designing a wind farm is to

determine the optimal location of wind turbines within a fixed geographical area to

maximize the total power capacity of the wind farm under stochastic wind conditions and

non-linear aerodynamic interferences between the turbines. This is known as the Wind

Farm Layout Optimization (WFLO) problem. This thesis investigates optimization

techniques that can fast approximate turbine layouts to aide engineers to quickly assess

and plan infrastructure design decisions.

We discretize the wind farm and use a quadratic integer formulation of the WFLO

problem to model an undirected graph that succinctly captures the spatial dependencies

between the variables due to aerodynamic interferences caused by wind regimes. By

performing probabilistic inference using sequential tree-reweighted message passing

(TRW-S) on the undirected graphical model the most probable placement of the turbines

is approximated. Turbine placements are also determined using branch-and-cut algorithm

using CPLEX, a state-of-the-art exact solver. This research conducts a computational

study to compare the effectiveness of turbine layouts produced using approximate

probabilistic inference with computationally exhaustive branch-and-cut algorithm under

varying wind-regime complexity and discrete resolutions. We demonstrate that

probabilistic inference can produce sub-optimal turbine layouts that are within 5% of the

power capacity of optimal layouts for low resolution. Our results also suggest that

probabilistic inference can produce approximate layouts that are equivalent or better than

branch-and-cut in a shorter period of time for higher resolutions.

Page 34: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

29

4.1 Introduction

In 2014, a total of $270 billion was invested in renewable energy technologies that

resulted in an addition of 128GW power capacity, of which 37% was due to new wind

farm installations [1]. As the world’s energy demand continues to rise, governments

continue to invest heavily in wind farms to harness wind energy to limit their greenhouse

gas emissions to curtail the threat of climate change and achieve energy security.

Therefore, an important stepping stone in achieving this goal is to build fast computational

tools that help design efficient large-scale wind farms that maximize power generation

and minimize infrastructure costs, while adhering to local land-use, environmental, and

mechanical constraints.

Wind Farm Layout Optimization (WFLO) is a problem that consists of determining the

optimal location of wind turbines within a fixed geographical area to maximize the total

power capacity of the wind farm under stochastic wind conditions and non-linear

aerodynamic interferences (wake effects) between the turbines. Wake effects created by

turbines effectively reduce the wind speed directly downstream of the placed turbines,

which decreases the expected power capacity of any turbines placed in their wakes.

Additionally, wakes can overlap with each other to further decrease the effective wind

speeds. Thus, turbines need to be placed optimally to minimize the total wake effects

caused by complex wind regimes.

Existing work focuses primarily on two key areas – constructing accurate wake models

and improving layout optimization algorithms. Standard wake models such as Jensen [3]

and Eddy Viscosity [51] quickly calculate the wake decay constant through momentum

mixing rate simplifications and ignoring the turbine effects on the planetary boundary

layers (PBL) [52]. Computationally expensive models (i.e. Deep-Array Wake Model,

Fuga, and WindModeller) capturing detailed wake interactions can be generated using

complex Reynolds-averaged Navier-Stokes equations and large eddy simulations [52]. In

this resarch, we use Jensen’s wake model and focus on developing fast optimization

algorithms for WFLO problem with discretized wind farms.

Page 35: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

30

In developing better optimization algorithms, the two most important factors of any

wind farm design objective function are minimizing wake effects and power capacity. A

more comprehensive objective function maximizes the wind farm’s net-present-value

profitability by incorporating installation (i.e. civil and electrical), maintenance, and

operations costs in addition to wake effects and power capacity as illustrated by Gonzalez

et al. [53]. Furthermore, design constraints such as land-use availability, noise generation,

turbine proximity, and other infrastructure constraints generally accompany objective

functions. In this thesis, we focus on developing fast approximate optimization algorithms

by taking into account wake effects and power capacity for small and large wind farms.

4.2 Literature Review

Early research used evolutionary algorithms (EA) to maximize power capacity and

minimize infrastructure costs in small wind farms. In 1994, Mosetti et al. [16] formulated

the WFLO problem as a small discretized wind farm by modelling the wake interactions

using Jensen’s wake model and using a simplified turbine cost model. Genetic algorithm

(GA) was used to find sub-optimal layouts with completely random initial configuration for

varying complexity of wind regimes. By fine tuning GA parameters such as crossover

type, diversity percentage, population size and generation count, Grady et al. [17]

obtained better layouts with higher power capacity and lower installation costs. Huang et

al. [18] further enhanced earlier approaches by applying a distributed genetic algorithm

to decompose a large wind farm terrain search space into local search spaces to

maximize power capacity. Thus, a distributed approach yields better solution quality and

is computationally faster than traditional genetic algorithms. By incorporating problem

specific heuristics and fine tuning optimization parameters, EAs can find good solutions

to non-linear, non-convex, and non-differentiable functions that are difficult to solve using

other methods. Nevertheless, EAs don’t offer optimality guarantees and can be

computationally expensive.

Particle swarm optimization (PSO) is applied in [54], [55] by Chowdhury et al. to

varying WFLO objective functions to handle discrete and continuous variables with non-

convex objective functions. Similarly, Wan et al. [56] apply Gaussian PSO while also

Page 36: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

31

incorporating a local search strategy based on differential evolution to enhance the

optimization results. Due to the varying objective functions and varying convergence

criteria, it is unclear whether PSO offers any computational time complexity advantage

over GA methods.

By discretizing the wind farm terrain into grid cells and representing each cell as a

mathematical decision variable, mathematical programming models can be developed for

the WFLO problem. Individual decision variables take on a binary state and determine

whether at every location a turbine is placed or not placed. The relationship between the

decision variables (i.e. wake interactions, infrastructure costs, etc.) can be encoded within

the objective function and hard constraints can be developed for budget, spacing,

proximity, terrain usage, noise, and other infrastructure constraints. Integer programs can

be solved exactly using well-designed algorithms [19] that are generally implemented in

an off-the-shelf commercial solver such as CPLEX (IBM Corp., Armonk, NY), and have

several theoretical guarantees. However, in many cases, models need to be simplified to

make the problem tractable and convex.

Donovan [20], [21] and Archer et al. [22] used a mixed integer problem (MIP) model

to minimize the wake interactions between the layout turbines. Donovan exploits the

inherent graph based structure of the problem to model the problem as a vertex packing

problem, and further explores varying branching strategies within the branch and bound

algorithm to reduce optimization time. Turner et al. [2] formulated a binary quadratic

integer program (QIP) by minimizing the wind speed deficit at each turbine as derived

from Jensen’s wake model. In [2], Adams-Sherali zero-one quadratic program

linearization technique [57] is applied to convert the QIP to a MIP. It is shown that MIP

solutions in [2] outperforms EA for complex wind regimes; nevertheless, it can take up to

24 hours to obtain good solutions for standard small wind farms. Thus, there is a need for

finding an algorithm that quickly produces good solutions for mathematical program

formulations of WFLO.

In this thesis, we pose the binary QIP formulation from [2] as an undirected graph

known as a Markov Random Field (MRF) where the pairwise wake interactions and global

Page 37: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

32

constraints are encoded within the graph’s edges. Probabilistic inference known as

maximum-a-posteriori (MAP) is conducted on the resulting MRFs’ joint log-likelihood

function to determine the optimal turbine placements. The overall goal of this method is

to approximate turbine layouts for small to large wind farms much faster than solving the

mathematical model using exact solvers to aide engineers quickly assess and plan

infrastructure design decisions.

4.3 Background

The power capacity of each turbine is proportional to the cube of the effective local wind

speeds at the turbine hub height. Similar to [16], [17], and [2], the total power capacity of

the wind farm is calculated using Eq. (4.1), where ui is the effective wind speed of the ith

turbine of a wind farm with a total of K turbines, and the coefficient is developed from

empirical observations.

Ptotal = ∑ 0.3uj3K

i=1 (4.1)

4.3.1 Wake Modelling

By assuming that momentum is conserved in the turbine wakes, the effective wind speed

can be calculated using Jensen’s wake model (Eq. (4.2)). The effective wind speed u at

the turbine hub height is directly related to the distance x from a turbine upstream.

Additionally, the effective wind speed is dependent upon the axial induction factor a,

entrainment or wake decay constant α, downstream rotor radius r1 that is the wake radius

immediately after the upstream turbine, and freestream wind velocity uo. A schematic

illustration of Jensen’s wake model for a turbine, T, is shown in Figure 4.1.

u = uo [1 −2𝑎

(1+𝛼(𝑥

𝑟1))

2] (4.2)

Page 38: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

33

Figure 4.1: Jensen’s wake model schematic illustration

The turbine induction factor, wake decay constant, and downstream rotor radius are

calculated using Betz’s theory [58] as shown in Eq. (4.3) – (4.5) where RR is the turbine

radius, CT is the turbine thrust coefficient, z is the turbine hub height, and zo is the surface

roughness constant. The rate of expansion of the wake downstream is given by Eq. (4.6),

where rx is the radius of the wake distance x downstream of the turbine. Finally, to avoid

structural damage to the turbines due to wake interactions, the proximity between any

two turbines need to be at least more than 5 times their rotor radius.

CT = 4a(1 − a) (4.3)

r1 = Rr√1−𝑎

1−2𝑎 (4.4)

α =0.5

ln(𝑧

𝑧𝑜) (4.5)

rx = r1 + αx (4.6)

4.3.2 Wind Farm Layout Optimization Mathematical Program

In order to simplify wind farm modeling and optimization, the land being surveyed for

turbine siting can be discretized into smaller areas. More particularly, the wind farm can

be divided into cells, in which each cell can only hold a single turbine, thereby limiting the

number of possible locations that a turbine can be placed within the wind farm. Thus, the

problem space is discretized into N cells, where each cell is represented with a binary

variable xi such that i ∈ I, I = {1, … , N}. If a turbine is placed at a given location i then the

xi takes on the value of 1 and 0 otherwise. In order to calculate the effective wind speed

at any location i, wakes caused by turbines upstream need to be considered (Figure 4.2).

Page 39: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

34

Figure 4.2: Wakes effecting Turbine 3 (T3) due to the wake region produced by Turbine 1 and 2 (T1, T2)

The effective wind speed at a given location j, j ∈ I due to multiple wakes is calculated

by assuming that the kinetic energy deficit at j is the linear sum of all the kinetic energy

deficits caused by the individual wakes produced by turbines at locations i upstream of j

(Eq. (4.7)). Thus, the effective wind speed uj of a downstream turbine j is calculated using

Eq. (4.8).

(1 −uj

uo)

2

= ∑ (1 −ui

uo)

2N

i=1,i≠j

xi (4.7)

uj = uo [1 − √∑ (1 −ui

uo)

2

xiNi=1,i≠j ] (4.8)

As illustrated by Turner et al. [2] maximizing wind farm power capacity is equivalent to

maximizing the effective wind speed at individual turbine hub heights. Eq. (4.9) maximizes

the effective wind speed by only placing turbines at optimal locations j that have an overall

smaller kinetic energy deficits and higher effective wind speeds. The freestream constant

speed uo is omitted as it does not affect the maximization.

max ∑ xjNj=1 [1 − √∑ (1 −

ui

uo)

2Ni=1,i≠j xi] (4.9)

Similarly, a quadratic integer program (QIP) is derived in [2] by illustrating that

maximizing effective wind speed is approximately equivalent to minimizing the kinetic

energy deficit (Eq. (4.10)).

Page 40: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

35

min ∑ ∑ xj (1 −ui

uo)

2

xiNi=1.i≠j

Nj=1 (4.10)

An interaction matrix W ∈ ℝNxN can be generated by calculating the kinetic energy

deficits at every location j ∈ I caused by a turbine at all locations i ∈ I, i ≠ j prior to the

optimization (Eq. (4.11)). The sparsity of W depends upon wake parameters and wind

regime complexity.

wij = {∑ ∑ (1 −ui

uo)

2Ni=1

Nj=1

0

, i ≠ j, i = j

(4.11)

The QIP formulation can be written as Eq. (4.12a), where X = (x1, x2, … , xN) is a binary

decision vector such that X ∈ {0,1}N. Due to budget constraints, land feasibility,

government regulations, and grid capacity the number of turbines to be placed on the

wind farm is determined prior to the farm development. Thus, the turbine constraint in

Eq. (4.12a) enforces that only a set number of turbines K will be placed on the farm, where

𝐞 is a vector {1, … ,1}N. Depending on the Jensen model’s parameter values, complexity

of the wind regime, and the problem state space, QIP can be highly non-convex and non-

tractable.

minimize XTWX (4.12a)

subject to ∑ xjNj=1 = K or 𝐞T𝐗 = K (4.12b)

xj ∈ {0,1} (4.12c)

4.3.3 WFLO as a Markov Random Field

The binary QIP (Eq. (4.12a)) is posed as an undirected graph G = (𝒱, ℰ) known as a

Markov Random Field (MRF) with a set of vertices 𝒱 represented by decision vector X ∈

{0,1}N (also known as random variables) connected by edges ℰ that encode the pairwise

wake interactions from W and the turbine constraints. To encode the turbine constraint

from Eq. (4.12b), we dualize the constraint using an augmented Lagrangian method by

applying a quadratic penalty function (Eq. (4.13)), which can be simplified into Eq. (4.14).

Page 41: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

36

f(X, β) = XTWX + β(𝐞TX − K)T(𝐞TX − K) (4.13)

f(X, β) = XTWX + β𝐗T [

00⋮

2 ⋯ 20 ⋱ ⋮⋮ ⋱ 2

0 0 ⋯ 0

] X + β(−2K + 1)𝐞TX + K2 (4.14)

Eq. (4.15) shows the unconstrained binary QIP where parameter β affect the

smoothness of the penalty contour and the K2 constant term is omitted as it does not

affect the optimization.

argminX∈ {0,1}N,β

XT(W + βE)X + (−2Kβ + β)𝐞TX (4.15)

The relationship between the random variables from Eq. (4.15) can be concisely

captured using a graphical model. Thus, the resulting undirected graph G (Eq. (4.15)) can

be formulated as a probabilistic graphical model, which is a graphical representation of

the joint probability distribution of the set of random variables [13]. Probabilistic inference

known as maximum-a-posteriori (MAP) is conducted on the resulting MRFs’ joint log-

likelihood function to determine the optimal turbine placements. The overall goal of this

method is to approximate optimal turbine layouts much faster than solving the

mathematical model using exact solvers to aide engineers quickly assess and plan aide

infrastructure design decisions.

MRFs can be decomposed as a product of factors composed of a set of random

variables Xc over the set of maximal cliques C of graph G [13], and such that the set of

maximal cliques satisfy Xc ⊂ X ∀ c ∈ C and ⋃ Xcc∈C = X. Depending on the Jensen

model’s parameter values and the wind regime, the interactions between the random

variables maybe localized and form maximal cliques. Thus, given a set of binary random

variables X ∈ {0,1}N, the joint distribution of a pairwise MRF based on the QIP model (Eq.

(4.15)) is shown in Eq. (4.16), which is proportional to the product of potential

functions θC(XC) over maximal cliques C, where Z is the partition function (Eq. (4.17)). The

partition function normalizes the joint distribution.

p(X) =1

Z∏ θC(XC)C (4.16)

Page 42: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

37

Z = ∑ ∏ θC(XC)CX (4.17)

The positive potential functions θC(XC) encodes the problem specific relationships

between the random variables, which in this case are the pairwise wake interactions and

turbines constraints. Furthermore, the potential function is generally expressed as an

exponential function, Eq. (4.18), where E(Xc) represents the energy function. Thus, a

random variable assignment with a higher probability has a lower overall energy [15].

θC(XC) = exp(−E(Xc)) (4.18)

The product of the exponential functions results in the sum of factors of energy

functions (see Eq. (4.19)), where φs(Xs) represent all unary potentials over the model

vertices 𝒱, and φst(Xs, Xt) represent the pairwise potentials over the edges ℰ. The term

(W + β1E) and (−2Kβ + β) from Eq. (4.15) represent the coefficients of the quadratic

pairwise and linear unary potentials respectively.

E(X, φ) = ∑ φs(Xs)s∈𝒱 + ∑ φst(Xs, Xt)(s,t)∈ℰ (4.19)

Then the joint distribution (i.e. Gibbs distribution) [15] is given in Eq. (4.20).

p(X, φ) =1

Zexp(− ∑ φs(Xs)s∈𝒱 − ∑ φst(Xs, Xt)(s,t)∈ℰ ) (4.20)

Determining the configuration of random variables XM in a graphical model that yields

the maximum probability over the joint distribution (Eq. (4.20)) is known as maximum-a-

posteriori (MAP) inference problem as formulated in Eq. (4.21). This is equivalent to

minimizing the energy function (Eq. (4.22)).

XM = argmaxX

p(X, φ) (4.21)

XM = argminX∈ {0,1}N

E(X, φ) = argminX∈ {0,1}N

(∑ φs(Xs)s∈𝒱 + ∑ φst(Xs, Xt)(s,t)∈ℰ ) (4.22)

The MAP problem can be linearized by introducing auxiliary marginal variables over

the random variables associated with the vertices {μs(Xs)}s∈𝒱 such that ∑ μs(Xs)Xs= 1,

and for every edge {μst(Xs, Xt)}(s,t)∈ℰ within the graphical model such that

Page 43: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

38

∑ μst(Xs, Xt)Xs,Xt= 1 [29]. Thus, the linear program can be defined by Eq. (4.23) over the

binary marginal polytope M(G) (Eq. (4.24)) where the dimension d of 𝛍 is equivalent to

2|𝒱| + 22|ℰ| and the vertices of M(G) are always integral. In general, MAP problem is an

NP-hard problem [15].

argmin𝛍∈M(G)

𝛗𝐓𝛍 = argmin𝛍∈M(G)

(∑ ∑ μs(Xs)φs(Xs)Xss∈𝒱 + ∑ ∑ μst(Xs, Xt)φst(Xs, Xt)Xs,Xt(s,t)∈ℰ )

(4.23)

M(G) = {𝛍 ∈ ℝd|∃p(𝐗, 𝛗) s. t μs(Xs) = ∑ p(𝐗, 𝛗)X𝒱\s

μst(Xs, Xt) = ∑ p(𝐗, 𝛗)X𝒱\s,t

} (4.24)

The IP in Eq. (4.23) can be solved using cutting plane, branch-and-bound, and branch-

and-cut algorithms by solving a partially relaxed IP and iteratively applying appropriate

integer constraints to ensure feasibility and satisfactory optimal or sub-optimal solutions.

However, depending on the number of variables and state-space, general purpose

solvers can take a very long time to solve this problem. For example, the branch-and-cut

method, when applied to linearized QIP from Eq. (4.12c) for complex wind regimes, can

take up to 24 hours to solve discretized problem space with 100 potential turbine locations

[2].

4.3.4 Message-Passing Algorithms

It is hypothesized that approximate solutions can be generated quickly for large MRFs by

class of algorithms known as message-passing algorithms that works by exploiting the

decomposable factors within the graphical model. Message-passing algorithms

approximate MAP assignments of relaxed IPs by iteratively passing beliefs locally along

the edges of the graphical model in a distributed, decentralized, and asynchronous

manner.

In 1988, Pearl [6] developed the max-product belief propagation (BP), which is

guaranteed to converge to the optimal MAP configuration for graphical models with only

trees. BP is not guaranteed to converge or find optimal configurations for graphical

models with cycles. Nevertheless, BP has still produced good empirical results as shown

Page 44: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

39

in the fields of computer vision (e.g. [9] and [10]) and computational biology (e.g. [12])

with problem structures similar to Eq. (4.15).

Multi-directional wind regimes create cycles within the MRF; thus, it is important to

study variants of BP that properly handle graphs with cycles to optimally estimate the

MAP assignment. Two such variants of BP that efficiently compute MAP configuration for

MRFs with cycles are Max Product Linear Programming algorithm (MPLP) and Tree-

Reweighted Message Passing (TRW).

MPLP [8] uses block coordinate descent in the dual of the LP Relaxation of the MAP

problem to compute beliefs between nodes. MPLP is guaranteed to converge and

computes the exact solution to the relaxed LP problem for binary variables. Similarly, in

[29] Wainwright et al. developed the TRW algorithm by decomposing an arbitrary

graphical model with cycles into a convex combination of tree-structured distributions to

calculate the optimal upper bound. The authors show that an optimal MAP configuration

and a tight lower bound can be obtained if and only if the combination of tree distributions

share a common optimum when the integer constraints are relaxed. TRW works by

maximizing the lower bound of the free energy objective. However, Kolmogorov [7]

showed that TRW does not guarantee to increase this lower bound, and in certain

situations may actually decrease the lower bound value during the iterative algorithm.

Furthermore, TRW is not guaranteed to converge. In [7], Kolmogorov developed the

sequential tree-reweighted message passing algorithm (TRW-S) that guarantees that the

lower bound will never decrease. Due to its convergence properties, in this work, we use

TRW-S to minimize the objective function in Eq. (4.23) over the binary marginal polytope

to determine the approximate turbine layout configurations.

Both MPLP and TRW-S rely on the assumption that the resulting linear program is

tight when the integer constraints are relaxed by forming a local polytope L(G), where

M(G) ⊂ L(G) and min𝛍∈L(G)

𝛗𝐓𝛍 ≤ min𝛍∈M(G)

𝛗𝐓𝛍. The relaxed constraints are hardly tight, which

requires us to investigate methodologies that can be used to dynamically generate

constraints within the message passing paradigm to make the solution space results tight.

Page 45: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

40

Similar to cutting-plane algorithms, in [36] and [37] Sontag et al. demonstrated that

tighter relaxations can be acquired by iteratively enforcing edge consistency over a small

subset of triplet clusters c ⊂ C to generate the polytope P. Similar to first lifting of the

Sherali-Adams hierarchy, clusters are chosen at every iteration that improve the dual LP

bound to create a sequence of polytopes P0 ⊂ P1 ⊂ P2 … ⊂ M(G) in which the relaxations

are continually tighter and approach the marginal polytope [14]. Thus, by iteratively

approximating the MIP solution using TRW-S and generating tighter polytopes, it is

expected that good turbine placements can be acquired in a relatively shorter amount of

time compared to other exact solver techniques.

It is hypothesized that by applying message passing algorithms, more specifically

TRW-S, to the relaxed WFLO MIP model, while enforcing key constraints using cluster

based approach, an approximate and sub-optimal turbine layouts can be generated faster

than traditional exact approaches. In this research, we conduct a thorough computational

study to measure the effectiveness of TRW-S in comparison with branch-and-cut

algorithm to generate turbine layouts under varying wind regime complexity and problem

dimensionality. Please note that in the following sections, to preserve simplicity, we refer

to TRW-S as message passing algorithm (MP).

4.4 Experimental Results and Discussion

We calculate the effectiveness of message-passing algorithm (MP) over two benchmark

wind resources – WR-1 and WR-36. Wind regime WR-1 has uniform wind speed blowing

from a single direction, and WR-36 is a complex wind regime with varying wind speeds

blowing from 36 uniformly distributed directions with varying probabilities. The wind rose

of WR-1 and WR-36 are illustrated in Figure 4.3 and Figure 4.4 respectively. Wake

modelling, wind farm, and simulation parameters are listed in Table 4.1; the parameter

values and wind regime distributions are chosen to establish a fair comparison with

respect to results from [16], [17], and [2]. Additionally, we compare results obtained by

solving the linearized MIP using exact state-of-the-art solver CPLEX 12.1 (IBM Corp.,

Armonk, NY) by applying branch-and-cut algorithm, and solving the unconstrained

objective function (Eq. (4.23)) over the relaxed marginal polytope (Eq. (4.24)) using the

Page 46: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

41

developed approximate solver based on MP method on varying discrete resolutions (100,

400, and 2,500 square cells), while keeping the square farm area to a constant at 4km2.

Figure 4.3: Wind rose of unidirectional wind regime

Figure 4.4: Wind rose of multidirectional complex wind regime

Page 47: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

42

Parameter Value

Turbine rotor radius (𝐑𝐑) 20 m Turbine hub height (𝐳) 60 m Thrust coefficient (𝐂𝐓) 0.88 Surface roughness constant (𝐳𝐨) 0.3 Axial induction factor (𝒂) 0.3268 Entrainment/wake-decay constant (𝜶) 0.1

Farm size 4km2 (2,000m x 2,000m)

Workstation Intel® Xeon® Processor 16 cores @ E5-2450 2.10GHz 256 GB Memory

Exact solver CPLEX 12.1 (IBM Corp., Armonk, NY) Table 4.1: Table of parameters

4.4.1 Case 1 – 100 cells

Discretizing the 4km2 wind farm area into 100 cells yields a square cell size

of 200m x 200m, which automatically satisfies the turbine proximity constraint. Wind farm

layout power capacity results for varying number of turbines for WR-1 and WR-36 are

shown in Table 4.2 and Table 4.3 respectively. The corresponding values in the columns

are calculated using branch-and-cut algorithm and MP methods. Initially, the branch-and-

cut algorithm is conducted using CPLEX using 1 thread with a cut-off of 1 hour (E1);

however, if relative optimality is not guaranteed as indicated by the relative gap, further

branch-and-cut iterations are conducted based on 4 threads with a cut-off of 1 hour (E2)

and 12 threads with a cut-off of 24 hours (E3). An asterisk (*) expresses that an optimal

solution was found, while a dash (-) indicates a 100% gap. MP simulations are run using

source-code in [59] by applying a total of 5,000 triplet clusters with a cut-off of 1 hour on

a single thread. Triplet clusters used to tighten the relaxed marginal polytope (Eq. (4.24))

are generated using modified source-code in [60] developed from [37].

Page 48: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

43

MIP – 1 Thread – 1 hour

cut-off (E1) MP

K Time (s)

Gap (%)

Power (W)

Time (s)

Power (W)

10 1 * 5,184 3,303 5,184

20 21 * 10,157 3,414 9,974

30 2 * 14,410 3,824 13,972

40 1 * 17,833 3,953 17,122

50 1 * 20,256 3,119 19,951

60 1 * 21,927 3,944 21,850

70 1 * 22,929 3,570 22,795

80 1 * 23,761 3,269 23,706

90 1 * 24,466 3,093 24,443

100 1 * 25,121 50 25,121

Table 4.2: Power capacity results under WR-1 with 100 cells

MIP – 1 Thread – 1 hour

cut-off (E1) MIP – 4 Threads – 1 hour

cut-off (E2) MIP – 12 Threads – 24

hour cut-off (E3) MP

K Time (s)

Gap (%)

Power (W)

Time (s)

Gap (%)

Power (W)

Time (s) Gap (%)

Power (W)

Time (s)

Power (W)

10 3,570 11 9,302 5,455 0 9,302 8,654 0 9,302 3,046 9,096

20 3,590 83 17,862 11,516 75 17,862 583,138 41 17,920 3,304 17,447

30 3,591 89 25,834 11,787 79 25,891 788,843 57 25,900 3,523 25,334

40 3,597 50 33,546 13,257 43 33,548 847,190 33 33,576 3,604 32,575

50 3,598 15 40,900 12,390 11 40,900 670,572 1 40,937 3,399 40,270

60 3,590 47 47,726 13,251 5 47,763 201,977 1 47,809 3,700 47,050

70 3,576 4 54,125 12,856 3 54,196 657,841 1 54,197 3,369 53,741

80 3,590 2 60,341 13,132 2 60,347 237,185 1 60,347 3,180 59,934

90 296 * 66,249 219 * 66,249 227 * 66,249 3,141 66,068

100 1 * 71,923 1 * 71,923 1 * 71,923 3,311 71,923

Table 4.3: Power capacity results under WR-36 with 100 cells

All turbine placements for WR-1 reach optimality using E1 in a few seconds due to the

low sparsity of the interaction matrix and the relatively small state space. Under WR-36,

the highly dense interaction matrix causes turbine placements to have a larger gap when

applying E1, thus, needing application of E2 and E3 to generate better bounds and cuts,

requiring up to several days in CPU time to compute optimal solutions. However, a large

gap still occurs while placing 20 – 40 turbines due to the relatively large complexity of the

turbine constraint polytope as characterized by the difficulty to generate effective cuts and

better bounds that well approximate the integer polytope.

As shown in Table 4.2 – 4.3 many layouts generated using MP capture a lower power

capacity compared to E1 and E1,2,3 for WR-1 and WR-36 respectively. The percentage

difference in the power capacities between MP and various branch-and-cut instances are

Page 49: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

44

illustrated in Figure 4.5 for both wind regimes. Under WR-1, MP generates layouts with a

power capacity that are 0 – 4% less than the optimal values captured by layouts from E1,

with the largest difference occurring between layouts with demanding turbine constraint

polytopes (i.e. 20 – 50 turbines).

To better compare MP with E1, E2, and E3 for WR-36, we compute the percentage

difference between the three pairs and plot the range of values as indicated by the

minimum and maximum bars for a given number of turbines. The minimum of the range

occurs between E1 and MP, while the maximum occurs between E3 and MP because E3

has a higher cut-off and uses more computing resources to generate similar or better

layouts compared to E1. Power capacity values generated using MP are less than 0 – 3%

compared to E1, E2, and E3 for placement of all turbine values, while varying less than

0.5% between E1, E2, and E3 for any given number of turbines.

Figure 4.5: Percent difference in power capacity values between CPLEX and MP for varying turbines under WR-1 and WR-36

Even though MP generates layouts that are sub-optimal, they are consistently within

5% of the power capacity generated using state-of-the-art branch-and-cut algorithm for a

wide variety of cases with varying gaps and cut-off periods. We attribute sub-optimality to

the generation of lack of effective triplet clusters, choice of penalty constant, and local

Page 50: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

45

numerical instability caused by locally similar wake decay values. In latter sections, we

test whether these inefficiencies of MP perform worse in comparison to branch-and-cut

when generation of cuts and bounds is further challenged when the problem’s state space

is increased by increasing discrete resolution.

Table 4.4 – 4.5 compare power capacity results from MP with EAs in [16] and [17],

and mathematical programming in [2]. MP outperforms Mosetti et al. [16] for placing 26

and 15 turbines under WR-1 and WR-36 respectively; contrarily, MP performs worse than

Grady et al. [17] for placing 30 and 39 turbines under WR-1 and WR-36 respectively.

Even though methods employed by both [16] and [17] are similar, it is postulated that

Grady et al. [17] outperforms MP due to fine-tuned GA parameters and longer run time.

Interestingly, results from branch-and-cut either outperform or match Turner et al. [2] even

when both methods have similar mathematical formulation; we speculate that this may

be due to failure to account for an appropriate angle of wake spread by Turner et al. [2].

K Model Power (W)

26

Mosetti et al. [16] 12,474

Turner et al. [2] 12,686

MILP (E1) 12,709

MP 12,486

30

Grady et al. [17] 14,410

Turner et al. [2] 14,410

MILP (E1) 14,410

MP 13,972

Table 4.4: Results comparison under WR-1 with 100 cells

K Model Power (W)

15

Mosetti et al. [16] 13,374

Turner et al. [2] 13,671

MIP (E3) 13,679

MP 13,395

39

Grady et al. [17] 32,377

Turner et al. [2] 31,947

MIP (E3) 32,818

MP 32,142

Table 4.5: Results comparison under WR-36 with 100 cells

Figure 4.6 – Figure 4.9 show the various layouts for the results presented in Table 4.4 –

4.5. In general, we observe that layouts generated using mathematical programming in

Page 51: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

46

Turner et al. [2] and MIP (E1,2,3) yield uniformly distributed layouts, while EAs in [16] and

[17], and MP suffer from local non-uniformity.

Optimizing over several other parameters (e.g. land-use constraints, noise generation,

infrastructure costs) significantly adds to the computational complexity of the problem.

Therefore, design engineers optimize over few key variables such as wind stochasticity,

and then heuristically modify determined turbine placements to adhere to further set of

constraints and objectives. Consequently, the benefits of uniformity and non-uniformity is

highly dependent upon project specifics. For instance, while uniform layouts generated

using branch-and-cut are wake optimal, they may increase electrical cabling costs due to

the increased distance between turbines. Similarly, non-uniform layouts that are wake

sub-optimal with closer turbine proximity may have lower maintenance costs, yet increase

the concentration of noise level in certain parts of the wind farm.

Mosetti et al. Turner et al. MIP (E1) MP

Figure 4.6: Comparison of layouts for WR-1 scenario with 26 Turbines

Grady et al. Turner et al. MIP (E1) MP

Figure 4.7: Comparison of layouts for WR-1 scenario with 30 turbines

Mosetti et al. Turner et al. MIP (E3) MP

Figure 4.8: Comparison of layouts for WR-36 scenario with 15 turbines

Page 52: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

47

Grady et al. Turner et al. MIP (E3) MP

Figure 4.9: Comparison of layouts for WR-36 scenario with 39 turbines

4.4.2 Case 2 – 400 cells

Resolution is increased by four-fold from 100 cells to 400 cells, while keeping the wind

farm area to a constant 4km2 and decreasing the square cell size to 100m x 100m.

Turbine proximity constraint is met without an addition of further constraints due to the

adequate distance between centroids of the neighbouring cells. Higher resolutions further

increases the computational complexity of finding good solutions due to an exponential

increase in the state space of the integer program. To the best of our knowledge, no other

studies were found in the literature within the realm of mathematical programming and

heuristic methods for discrete formulations that assess the effects of increasing resolution

to power capacities, layout design, and run-time for varying number of turbines.

Table 4.6 – 4.7 show the wind farm layout power capacity results for varying number

of turbines for WR-1 and WR-36 respectively. For both wind regimes we apply E1, E2,

and E3. As running branch-and-cut under using E3 requires significant computational and

memory resources, in some cases, in sufficient memory (INM) due to large problem size

causes the program to terminate without generating a layout.

Generation of triplet clusters for 400 cells is a computationally expensive task and can

take up to 3 – 4 hours. The graphical model structure based on the marginal polytope

generated using the unconstrained objective function (Eq. (4.23)) hardly changes while

optimizing for varying number of turbines using MP; thus, we pre-calculate 5,000 triplet

clusters based on 10 turbines to save time on layout generation for additional turbines.

Thus, all results presented on Table 4.6 – 4.7 using MP are based on pre-generated

clusters for individual wind regimes.

Additionally, due to the unconstrained nature of the formulation, MP simulations

sometimes produce layouts that contain one more or less turbine than needed. This

Page 53: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

48

aberration is fixed by greedily adding or subtracting a turbine from the layout that results

in the highest power capacity.

Under WR-1, the layout power capacity increases for each set of added turbines until

it peaks for a placement of 220 turbines for both methods. Power capacity decreases at

the addition of more turbines as the densely packed turbines couple together to increase

the wake interactions between turbines and significantly decrease the effective wind

speed in the narrow horizontal wake columns. Power capacity does not peak under WR-

36 as the stronger winds allows more energy to be harnessed than the energy lost due

to wake interactions. However, under both wind regimes addition of each turbine over 20

turbines results in a lower power capacity yield per turbine compared to power capacity

yield for a lower number of turbines.

Figure 4.10 shows the percentage difference in power capacity results obtained from

the three instances of branch and cut, and MP for varying number of turbines under both

wind regimes. A positive percentage demonstrates that MP produces a higher power

capacity, while a negative percentage a lower power capacity compared to a given

instances of branch and cut and number of turbines. Under WR-1, similar to the

formulation with 100 cells, MP consistently captures layouts with a 0 – 5% lower power

capacity than E1, E2, and E3 and with under 1% variation between the three instances.

Furthermore, the largest difference occurs in formulations with 30 – 140 turbines due to

their difficult turbine constraint polytope.

Under WR-36, placing 50 – 240 turbines using MP produces a higher power capacity

than E1 and E2. Applying E3 yields tighter bounds and cuts, hence, reducing the gap

within this turbine range, which results in E3’s power capacity results that are better than

MP. The performance of MP is favorably comparable to instances where there exists a

large optimality gap in results computed using branch-and-cut. Furthermore, MP

consistently performs well as an approximate solution for instances, commonly E3, in

which branch-and-cut produces results that are optimal or have a small gap.

Page 54: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

49

Figure 4.10: Percent difference in power capacity values between CPLEX and MP for varying turbines under WR-1 and WR-36

MIP – 1 Thread – 1 hour

cut-off (E1) MIP – 4 Threads – 1 hour

cut-off (E2) MIP – 12 Threads – 24

hour cut-off (E3) MP

K Time (s)

Gap (%)

Power (W)

Time (s)

Gap (%)

Power (W)

Time (s) Gap (%)

Power (W)

Time (s)

Power (W)

10 1 * 5,184 1 * 5,184 1 * 5,184 646 5,184

20 1 * 10,368 1 * 10,368 1 * 10,368 1,002 10,368

30 3,595 - 15,339 10,671 - 15,341 586,310 - 15,341 1,024 15,259

40 3,596 - 20,101 10,646 - 20,054 564,897 - 20,106 1,029 19,245

50 3,596 - 24,417 10,518 - 24,429 559,054 - 24,503 1,052 23,804

60 3,597 - 28,325 10,531 - 28,398 597,192 - 28,454 1,060 27,697

70 3,596 - 31,858 9,843 - 31,944 560,929 - 32,065 1,073 30,768

80 3,597 - 35,472 10,330 - 35,605 620,235 - 35,697 1,069 33,729

90 3,597 - 37,960 10,633 - 38,124 650,517 - 38,188 1,063 36,665

100 3,598 - 40,377 8,845 - 40,439 641,507 - 40,727 1,062 39,178

120 3,598 - 44,667 8,951 - 44,954 INM - - 1,085 42,852

140 3,593 19 47,484 12,385 12 47,486 INM - - 1,050 45,960

160 3,594 3 49,357 12,809 1 49,402 INM - - 1,058 47,806

180 3,594 2 50,472 12,949 1 50,428 26,962 1 50,428 1,065 48,962

200 148 1 51,282 642 1 51,288 685 1 51,308 1,089 50,118

220 378 1 52,025 2,252 1 51,250 5,192 1 51,250 1,078 51,292

240 3,595 1 50,906 3,039 1 51,116 4,033 1 50,882 1,068 50,927

260 3,597 1 50, 082 12,975 1 50, 165 23,153 1 50,207 1,063 50,150

280 3,581 1 49, 261 12,221 1 49, 310 INM - - 1,009 49,240

300 3,582 1 48,320 12,618 1 48,325 INM - - 1,033 48,326

320 3,588 1 47,277 10,946 1 47,248 INM - - 1,050 47,169

340 3,590 1 46,151 11,661 1 46,151 INM - - 983 46,156

360 3,569 1 45,013 11,339 1 45,013 618,898 1 45,013 980 45,086

380 2 1 43,820 7 1 43,820 7 1 43,820 972 43,941

400 1 * 42,660 1 * 42,660 1 * 42,660 4,506 42,660

Table 4.6: Power capacity results under WR-1 with 400 cells

Page 55: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

50

MIP – 1 Thread – 1 hour

cut-off (E1) MIP – 4 Threads – 1 hour

cut-off (E2) MIP – 12 Threads – 24

hour cut-off (E3) MP

K Time (s)

Gap (%)

Power (W)

Time (s)

Gap (%)

Power (W)

Time (s) Gap (%)

Power (W)

Time (s)

Power (W)

10 3,435 - 9,282 10,279 - 9,300 814,609 90 9,317 2,017 9,105

20 3,598 - 17,774 10,649 - 17,906 804,187 98 17,950 2,050 16,632

30 3,590 - 24,941 10,810 - 25,220 808,427 - 26,038 2,064 25,131

40 3,598 - 32,214 10,618 - 33,040 807,081 - 33,604 2,070 32,179

50 3,598 - 38,916 10,785 - 39,201 826,492 - 40,860 2,050 39,879

60 3,598 - 46,005 11,081 - 45,682 827,122 - 47,669 2,135 46,250

70 3,598 - 51,627 10,922 - 51,838 850,897 - 54,439 2,153 54,089

80 3,594 - 58,015 10,711 - 58,157 836,268 - 61,227 2,157 59,200

90 3,596 - 64,057 10,264 - 63,905 756,351 - 65,274 2,140 65,792

100 3,596 - 69,734 10,378 - 69,730 861,872 - 72,490 2,137 71,482

120 3,596 - 80,656 11,855 - 80,618 887,169 - 86,061 2,092 83,324

140 3,594 - 90,046 11,171 - 90,384 536,166 98 92,860 2,078 95,042

160 3,597 - 101,116 7,189 - 101,116 409,573 90 104,895 2,089 105,101

180 3,593 - 91,298 7,182 - 109,116 763,095 86 114,923 2,079 114,980

200 3,591 - 119,716 6,864 97 119,716 772,600 84 124,243 2,170 123,691

220 3,587 - 109,237 10,102 79 126,958 825,105 67 132,782 2,150 131,678

240 3,591 62 135,515 10,376 61 137,218 789,276 49 141,312 2,083 139,900

260 3,582 48 143,280 10,740 41 148,203 821,311 36 148,627 2,094 147,390

280 3,586 36 150,598 9,237 27 155,605 INM - - 2,120 154,533

300 3,594 18 162,118 9,234 18 162,317 INM - - 2,038 160,593

320 3,593 11 168,355 8,537 11 168,419 INM - - 2,021 166,908

340 3,597 6 174,269 8,203 6 174,359 INM - - 2,018 173,108

360 3,588 2 179,698 6,919 2 179,703 INM - - 2,034 178,908

380 3,587 1 184,740 7,630 1 184,747 INM - - 2,016 184,414

400 1 0 189,395 1 0 189,395 1 0 189,395 4,763 189,395

Table 4.7: Power capacity results under WR-36 with 400 cells

4.4.3 Case 3 – 2,500 cells

Integer optimization using any method becomes a magnitude more challenging when the

resolution in increased twenty-five fold to 2,500 cells, while still maintaining the wind farm

area to a constant 4km2 and decreasing the square cell size to 40m x 40m. At this

resolution, placing turbines in adjacent and nearby cells violate turbine proximity

constraint, therefore, an inequality constraint is introduced such that if the distance

between the two cells is less than 5 times the rotor radius, 100m, then there can only be

a placement of 1 turbine between the two cells. Generation of effective bounds, cuts, and

triplet clusters is challenging under this case due to the demanding memory requirement

and computational complexity. Due to the large state-space we notice that we run out of

memory while iterating to tighten the relaxed marginal polytope. Therefore, to preserve

computational feasibility we decrease the cut-off time of E3 to 4 hours, and remove

Page 56: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

51

generation of triplet clusters and employ a simple rounding scheme to construct a feasible

integral solution for MP.

Table 4.8 – 4.9 show the wind farm layout power capacity results for varying number

of turbines for WR-1 and WR-36 respectively. The 100m proximity constraint and 40m

distance between adjacent cells limits the layout feasibility to no more than approximately

280 turbines. Results show that there exists large gap for many placements after

applying E1, E2, and E3 for both wind regimes, and E3 generally produces layouts with

better power capacity. Even though the lack of triplet cluster generation for MP is

expected to produce layouts with uncertain power capacity we observe that in certain

cases power produced by MP is better than E1, E2, and E3. It is conjectured that this occurs

due to the difficulty in constructing effective bounds and cuts using CPLEX given the

limited amount of resources available and the large problem size.

In Figure 4.11 we plot the percentage difference in power capacity results obtained

from the three instances of branch and cut, and MP for varying number of turbines under

both wind regimes. Under WR-1, branch and cut algorithms determines an optimal layout

configuration for the placement of 10 – 50 turbines, while MP is only able to determine

optimal configuration for 50 turbines. Furthermore, E1, E2, and E3 produce layouts with up

to 20% higher power capacity than MP for up to 100 turbine placements. Contrarily, MP

produces layouts with up to 20% higher power capacity than branch and cut instances for

turbine placements over 100. This occurs as branch and bound is able to generate better

bounds and cuts for the turbine constraint polytope with a smaller state space due to a

lower demand on computational complexity. Furthermore, the performance

between E1, E2, and E3 varies between 0 – 5%.

Under WR-36, both algorithms produce layouts with similar power capacity for

placement of over 50 turbines. Nevertheless, MP performance is approximately better

than branch and bound for less than 50 turbines. The large gap even after applying E3 for

both wind regimes highlights the need to better assess the problem and generate further

local constraints that assist in producing problem specific bounds and cuts that are

effective in solving in the problem.

Page 57: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

52

Figure 4.11: Percent difference in power capacity values between CPLEX and MP for varying turbines under WR-1 and WR-36

MIP – 1 Thread – 1 hour

cut-off (E1) MIP – 4 Threads – 1 hour

cut-off (E2) MIP – 12 Threads – 1

hour cut-off (E3) MP

K Time (s)

Gap (%)

Power (W)

Time (s)

Gap (%)

Power (W)

Time (s) Gap (%)

Power (W)

Time (s)

Power (W)

10 89 * 5,184 92 * 5,184 33 * 5,184 42 5,184

20 63 * 10,368 98 * 10,368 33 * 10,368 41 10,037

30 95 * 15,552 85 * 15,552 33 * 15,552 41 15,312

40 70 * 20,736 95 * 20,736 32 * 20,736 42 17,516

50 3,626 - 25,861 4,009 - 25,768 33 * 25,920 42 25,920

60 3,586 - 30,201 9,275 - 30,283 13,918 - 30,229 41 29,683

70 3,618 - 32,478 8,362 - 33,945 16,888 - 34,757 41 30,505

80 3,598 - 39,207 9,939 - 39,032 20,280 - 39,211 41 33,743

90 3,571 - 43,329 10,559 - 43,506 22,514 - 43,673 42 42,803

100 3,571 - 46,403 8,832 - 41,309 18,191 - 45,104 41 46,361

120 3,541 - 46,049 10,122 - 45,693 22,280 - 47,131 41 48,961

140 3,600 - 49,832 9,444 - 51,257 24,681 - 52,310 42 58,671

160 3,523 - 55,119 10,719 - 55,214 26,602 - 54,811 42 62,988

180 3,531 - 59,957 8,857 - 59,490 29,474 - 60,660 58 72,157

200 3,616 - 65,108 9,621 - 63,060 29,990 - 63,677 59 73,449

220 3,613 - 66,504 11,111 - 66,143 11,364 - 67,262 58 83,843

240 3,618 - 64,616 7,298 - 68,431 13,461 - 68,431 41 79,727

260 3,618 - 70,790 6,761 - 70,790 7,927 - 70,790 41 86,356

280 3,622 - 72,470 7,486 - 72,433 6,055 - 72,145 42 85,247

Table 4.8: Power capacity results under WR-1 with 2,500 cells

Page 58: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

53

MIP – 1 Thread – 1 hour

cut-off (E1) MIP – 4 Threads – 1 hour

cut-off (E2) MIP – 12 Threads – 1

hour cut-off (E3) MP

K Time (s)

Gap (%)

Power (W)

Time (s)

Gap (%)

Power (W)

Time (s) Gap (%)

Power (W)

Time (s)

Power (W)

10 3,600 - 8,572 10,366 - 8,572 6,831 - 8,874 121 9,184

20 3,600 - 16,974 4,609 - 16,178 INM - - 122 17,562

30 3,600 - 25,075 8,538 - 24,177 INM - - 149 25,450

40 3,600 - 31,931 6,027 - 31,931 INM - - 172 32,964

50 3,600 - 39,261 10,139 - 39,021 INM - - 171 36,983

60 3,600 - 47,158 5,529 - 46,859 16,043 - 46,859 171 47,766

70 3,600 - 53,845 5,573 - 53,845 18,107 - 53,845 172 54,628

80 3,600 - 60,501 6,095 - 60,501 16,422 - 60,501 171 60,953

90 3,600 - 67,078 4,356 - 67,139 17,817 - 67,139 171 67,920

100 3,600 - 72,698 3,972 - 72,915 INM - - 171 72,429

120 3,600 - 84,469 4,183 - 84,134 16,914 - 84,134 172 85,141

140 3,600 - 95,124 3,507 - 95,188 19,700 - 95,188 124 95,520

160 3,600 - 103,911 3,989 - 105,431 16,834 - 105,431 122 104,691

180 3,600 - 113,692 4,205 - 115,251 16,154 - 115,251 121 112,329

200 3,602 - 122,406 10,279 - 124,471 17,688 - 124,471 122 123,728

220 3,607 - 132,770 9,540 - 133,254 17,411 - 133,254 122 129,204

240 3,602 - 141,592 10,960 - 141,592 18,290 - 141,592 122 141,200

260 3,614 - 149,573 11,225 - 149,573 19,454 - 149,573 121 149,426

280 3,617 - 157,275 10,590 - 157,687 18,973 - 157,687 121 156,616

Table 4.9: Power capacity results under WR-36 with 2,500 cells

Figure 4.12 shows wind farm layouts for placing 200 turbines generated under WR-1

using MIP (E1) and MP. Similar to the visibly uniform and symmetrical layouts (Figure 4.6

– Figure 4.9) generated by branch-and-cut under 100 cells, layouts generated by MIP

under 2,500 cells are also uniformly distributed and symmetrical in comparison to layouts

generated by MP as shown in Figure 4.12. Nevertheless, in Figure 4.12 layout generated

by MP yields a higher power capacity and is more structured compared to the MIP layout.

MIP (E1) MP

Figure 4.12: Comparison of layouts for WR-1 scenario with 200 turbines

Page 59: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

54

4.4.4 Discreteness Analysis

In earlier analysis, we’ve observed that computational complexity of the integer program

increases as the wind farm resolution is increased. While choosing a lower resolution may

offer computational feasibility, the resulting turbine configurations may be highly sub-

optimal due to stricter limitations on the placement of the turbines, compared to

continuous formulations or discrete formulations with a higher resolution. Therefore,

determining the problem resolution while maintaining a handle on tractability poses a

challenge for the WFLO problem.

Figure 4.13 illustrates the power capacity for various number of turbines for three

cases (100, 400, and 2,500 cells) produced using branch-and-cut and MP algorithms

under WR-1 and WR-36 wind regimes. Please note that while most plotted power

capacity values for branch-and-cut were generated using E3, values generated using E2

were used when using E3 resulted in insufficient memory. Both algorithms, MP and

branch-and-cut, produce a higher power capacity for a higher resolution compared to a

lower resolution under WR-1. Contrarily, both algorithms produce similar power capacity

regardless of the resolution under WR-36. This illustrates the need to carefully assess the

effects of problem parameters on the effectiveness of the resolution prior to optimization.

Scalability of branch-and-cut and MP poses a challenge due to the difficulty in

generating effective cuts and triplets respectively. Nevertheless, for many turbine

placements MP’s performance is equivalent or better than E1, E2, and E3 for resolutions

with 400 and 2,500 cells under both wind regimes; interestingly, these occurrences also

produce layouts that are sub-optimal under both methods. Hence, MP provides a more

competitive and scalable approach to generating sub-optimal layouts in comparison with

exhaustive branch-and-cut while using requiring fewer computational resources.

Furthermore, sub-optimal layouts generated using MP can be augmented by feeding the

layouts to fine-tuned heuristic and evolutionary methods. Therefore, MP provides an

interesting alternative perspective to generating fast and approximate layouts to the

challenging WFLO problem.

Page 60: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

55

Figure 4.13: Power capacity generated using E3 and MP for 100, 400, and 2,500 square cells under WR-1 and WR-36 wind regimes

4.5 Conclusions

In this thesis, we model the binary QIP formulation of the WFLO problem as an undirected

graphical model by incorporating pairwise wake interactions and global turbine budget

constraint within the graph’s edges. The linearized QIP formulation is NP-hard, which

requires an exhaustive and computationally expensive branch-and-cut algorithm to

determine optimal layouts. In many cases, branch-and-cut produces sub-optimal layouts

with a large relative gap for high discrete resolution and challenging turbine budget

constraints. We’ve demonstrated that by determining the most probable assignment of

the graph’s random variables based upon local, distributed, and dynamic probabilistic

inference using message passing algorithm we can decode good sub-optimal layouts for

a given number of turbine placements. This thesis conducts a thorough computational

study to determine the effectiveness of message passing algorithm in comparison with

branch-and-cut algorithm in generating turbine layouts under varying wind regime

complexity and problem resolution.

Page 61: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

56

Branch-and-cut is effective at finding optimal layouts for a low resolution problem (100

cells) under both wind regimes, while in comparison, MP produces sub-optimal layouts

that are consistently within 5% of the power capacity. As both methods have difficulty

generating cuts to tighten the relaxed polytope at higher resolutions, the generated

layouts are sub-optimal and rely on locally optimal rounding schemes. At these higher

resolutions with 400 and 2,500 cells, the performance of MP is better or equivalent in

comparison to branch-and-cut to many turbine placements under WR-1 and WR-36.

Additionally, we show that MP can produce better approximate layouts in a shorter period

of time while using significantly less computing resources. Overall, MP offers a

competitive and scalable alternative to computationally expensive branch-and-cut

algorithm, especially when design engineers are seeking to generate approximate and

sub-optimal layouts quickly. We foresee that a population of approximate layouts

generated from MP can be further optimized for other project variables (e.g. land-use

constraints, noise generation, infrastructure costs) using fine-tuned stochastic and

evolutionary search methods.

Graphical models provide a very succinct way to capture the problem objectives’

structural variable dependencies. There are still considerable limitations for employing

probabilistic inference methods to approximate graphically modelled integer programs.

For instance, although current state-of-the-art message passing algorithms (i.e. TRW-S

and MPLP) generate good approximate solutions in practice, they don’t provide a

certificate of optimality. Furthermore, feasibility issues arise as multi-variable constraints

are incorporated using augmented Lagrangian method. While generating triplets tighten

the relaxed polytope, a method to generate and incorporate problem specific cuts within

the graphical model is required for message passing algorithms to become a practical

alternative approximate solver for problems with multiple constraints and objectives.

While we’ve demonstrated that WFLO using approximate probabilistic inference offers

a compelling alternative optimization technique, there are further challenging facets that

need to be addressed. More specifically, approximate algorithms that better handle

problem intractability and non-convexity need to be developed to solve comprehensive

Page 62: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

57

multi-objective WFLO incorporating wake effects, noise level, land-use availability,

infrastructure costs, and other geographical constraints.

Page 63: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

58

Chapter 5

Concluding Remarks

5.1 Conclusions

The principle goal of this thesis has been to explore techniques to solve the wind farm

layout optimization (WFLO) problem. We restricted the scope of the problem to optimal

placement of a limited number of wind turbines on a flat terrain in response to wake effects

generated by upstream wind turbines due to stochastic wind conditions. This restriction

allows us to model the WFLO problem as a binary discrete mathematical program using

a simple analytical wake model (Jensen’s model [3]), and motivate our study on

optimization methodologies.

In Chapter 3, we showed that the derived WFLO problem is structurally similar to the

Quadratic Knapsack Problem (QKP). Subsequently, in Chapter 4 we illustrate that while

several exact solving procedures based on linearization and generating better bounds

exist to solve quadratic integer programs, these methods don’t scale well to large-scale

optimization problems. By observing that a QKP’s objective function can be represented

as an undirected graph, we study message passing algorithms, a class of probabilistic

inference methods, that exploit the graph-theoretic structure through graph factorization

and decomposition to determine the maximum-a-posteriori (MAP) assignment of the

graph’s random variables. By devising a method to incorporate the QKP’s global linear

inequality constraint (budget constraint) into the graphical model, we show that a variant

of belief propagation (TRW-S) can be used to generate good results when applied to in-

approximable NP-Hard benchmark QKPs and the WFLO problem.

5.2 Future work

We discuss some future considerations, recommendations, and challenges of our work

within the context of WFLO, message passing inference, and graphical models.

Page 64: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

59

5.2.1 Wind Farm Layout Optimization

Other Objectives This thesis restricts the WFLO problem to a single objective function

by only considering the optimal placement of wind turbines on a flat terrain by minimizing

their wake effects on each other. There are several other variables and constraints that

significantly affect the development and planning of wind farms that need to be

considered as part of the problem objective. We list a few notable variables that can be

considered as part of the objective function:

Noise – Wind turbines are a known cause of noise pollution, especially in cases where

turbines are placed in close proximity. Several industrial guidelines exist that constraint

noise levels in residential and industrial communities such as the ISO standard [61].

As noise level at a given location is a function of its proximity to wind turbines, noise

propagation of wind turbines must be suitably modelled and included when determining

the placement of turbines.

Infrastructure Costs – The variable cost of a wind turbine installation can be broken

down into the cost of electrical cabling required to obtain grid connectivity, maintenance

infrastructure, and the type of excavation required. These costs vary depending on the

terrain type and location. For example, although placing turbines far apart reduces their

wake interactions it may result in higher infrastructure costs due to the added distance

between them. A multi-objective function considering these costs needs to be modelled

to determine project feasibility and net-present value profitability.

Land Use Availability – Rarely is the case that the entire geographical area under

consideration of wind farm installation can be used for wind turbine installations. Land-

use restrictions also need to be added into the problem as constraints.

It is important to note that the addition of further constraints and objectives makes the

optimization function highly non-linear, non-convex, and intractable. Thus, finding the

optimal solution becomes computationally challenging. Thus, the approximate inference

algorithms we study here serve as a stepping stone towards developing advanced

Page 65: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

60

approximate and hybrid evolutionary algorithms to generate good solutions to the multi-

objective WFLO problem.

Wake Modelling and Terrain Type As described in our earlier recommendation, the

WFLO problem is a multi-objective function with difficult constraints. While we’ve used a

simple analytical wake model (Jensen’s wake model) to model wake effects, more

accurate analytical wake models need to be developed that work well with complex

terrains and that can be pre-calculated to take advantage of discrete optimization models.

In general, complex wake models that accurately describe wake effects are

computationally exhaustive and memory intensive. Thus, this further adds to the time

complexity of the optimization problem.

Robustness As with any real-world systems, all measured cost vectors suffer from

uncertainty, which compounded over several factors and variables can have a significant

impact on the solution state space optimality, feasibility, and project reliability. Hence,

uncertainty within the problem space needs to be added, and an optimization algorithm

capable of performing robust optimization needs to be used appropriately. While robust

optimization algorithms are available in the context of exact discrete and continuous

optimization, they have not been widely studied in the context of message passing

inference. However, the probabilistic nature of graphical models make them an ideal

candidate to incorporate and study variable uncertainty.

5.2.2 Message passing algorithms for inference and optimization

Capacity Constraint We showed in Chapter 3 that a global linear inequality constraint

(capacity constraint) can be incorporated successfully into an undirected graphical model

using penalty functions derived from an augmented Lagrangian approach. Nevertheless,

this approach requires added iterations for convergence, adds sub-optimal stationary

points, and it can suffer from numerical ill-conditioning. Research in the direction of adding

capacity constraints within graphical models using auxiliary variables and higher-order

graph potentials needs to be explored.

Page 66: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

61

Tight Relaxation of Integral Capacity Constraint As belief propagation and its variants

operate on the relaxed marginal polytope (relaxed objective function), we apply methods

in [36] and [37] that add triplet clusters to enforce local consistency based on cutting-

plane methods and Sherali-Adams hierarchy to iteratively tighten the relaxed marginal

polytope. While this approach has been shown to work in a limited capacity when applied

to the various benchmark QKPs and the WFLO problem to reduce their integrality gap,

we recommend generating and incorporating cuts (constraints) specifically to tighten

global capacity constraint to effectively further tighten the relaxed problem polytope.

Therefore, research in the direction of generating tight cover, extended cover, and lifted

cover inequalities [62] that tighten the relaxed integral capacity constraint needs to be

conducted. Furthermore, alternative approaches to incorporating several local inequality

constraints within the graphical model needs to be explored.

Page 67: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

62

Bibliography

[1] International Energy Agency (IEA), "Energy and Climate Change," International

Energy Agency, France, 2015.

[2] S. D. O. Turner, D. A. Romero, P. Y. Zhang and T. C. Y. Chan, "A new

mathematical programming approach to optimize wind," Renewable Energy, vol.

63, pp. 674-680, 2014.

[3] N. O. Jensen, "A note on wind generator interaction," Riso National Laboratory,

Roskilde, 1983.

[4] D. Pisinger, "The quadratic knapsack problem—a survey," Discrete applied

mathematics, vol. 155, no. 5, pp. 623-648, 2007.

[5] D. J. Rader Jr and G. J. Woeginger, "The quadratic 0–1 knapsack problem with

series–parallel support," Operations Research Letters, vol. 30, no. 3, pp. 159-166,

2002.

[6] J. Pearl, Probabilistic reasoning in intelligent systems: networks of plausible

inference, Morgan Kaufmann, 2014.

[7] V. Kolmogorov, "Convergent tree-reweighted message passing for energy

minimization," Pattern Analysis and Machine Intelligence, IEEE Transactions, vol.

28.10, pp. 1568-1583, 2006.

[8] A. Globerson and T. Jaakkola, "Fixing Max-Product: Convergent Message Passing

Algorithms for MAP LP-Relaxations," in Advances in Neural Information

Processing Systems, Vancouver, 2007.

[9] J. Sun, N.-N. Zheng and H.-Y. Shum, "Stereo matching using belief propagation,"

Pattern Analysis and Machine Intelligence, IEEE Transactions, vol. 25, no. 7, pp.

787-800, 2003.

[10] P. F. Felzenszwalb and D. P. Huttenlocher, "Efficient belief propagation for early

vision," International journal of computer vision, vol. 70, no. 1, pp. 41-54, 2006.

[11] N. Lazic, Message Passing Algorithms for Facility Location Problems, University of

Toronto, 2011.

Page 68: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

63

[12] C. Yanover and Y. Weiss, "Approximate inference and protein-folding," in

Advances in neural information processing systems, 2002.

[13] C. M. Bishop, Pattern Recognition and Machine Learning, Springer, 2006.

[14] D. A. Sontag, "Approximate inference in graphical models using LP relaxations

(Doctoral dissertation)," Massachusetts Institute of Technology, 2010.

[15] D. F. N. Koller, Probabilistic Graphical Models, Cambridge: The MIT Press, 2009.

[16] G. P. C. D. B. Mosetti, "Optimization of wind turbine positioning in large windfarms

by means of genetic algorithm," Journal of Wind Engineering and Industrial

Aerodynamics, vol. 51, pp. 105-116, 1994.

[17] S. H. M. A. M. Grady, "Placement of wind turbines using genetic algorithms,"

Renewable Energy, vol. 30, pp. 259-270, 2005.

[18] H. Huang, "Distributed genetic algorithm for optimization of wind farm annual

profits," in Intelligent Systems Applications to Power Systems, IEEE, 2007.

[19] D. W. R. Bertsimas, Optimization over integers, Belmont: Dynamic Ideas, 2005.

[20] S. Donovan, "Wind farm optimization," in Proceedings of the 40th annual ORSNZ

conference, 2005.

[21] S. Donovan, "An improved mixed integer programming model for wind farm layout

optimisation," in Proceedings of the 41st annual conference of the Operations

Research Society, 2006.

[22] R. Archer, "Wind turbine interference in a wind farm layout optimization mixed

integer linear programming model," Wind Engineering, vol. 35, pp. 165-175, 2011.

[23] "GH WindFarmer Theory Manual," Garrad Hassan and Partners Ltd, 2009.

[24] P. Y. Zhang, D. A. Romero, C. J. Beck and C. H. Amon, "Solving wind farm layout

optimization with mixed integer programs and constraint programs," The

Association of European Operational Research Societies, vol. 2, pp. 195-219,

2014.

[25] M. Weigt and H. Zhou, "Message passing for vertex covers," Physical Review E,

vol. 74, no. 4, 2006.

[26] N. Lazic, F. J. Brendan and P. Aarabi, "Solving the uncapacitated facility location

problem using message passing algorithms," in International Conference on

Artificial Intelligence and Statistics, 2010.

Page 69: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

64

[27] J. Yarkony and C. Fowlkes, "Covering trees and lower-bounds on quadratic

assignment," in Computer Vision and Pattern Recognition (CVPR), IEEE

Conference, 2010.

[28] C. C. Moallemi and B. Van Roy, "Resource allocation via message passing,"

INFORMS journal on Computing, vol. 23, no. 2, pp. 205-219, 2011.

[29] M. J. Wainwright, T. S. Jaakkola and A. S. Willsky, "MAP estimation via agreement

on trees: message-passing and linear programming," Information Theory, IEEE

Transactions, vol. 11, pp. 3697-3717, 2005.

[30] S. Sanghavi, D. Malioutov and A. S. Willsky, "Linear programming analysis of

loopy belief propagation for weighted matching," in Advances in neural information

processing systems, 2008.

[31] A. Z. S. Kumar, "Message-passing algorithms for quadratic programming

formulations of MAP estimation," in arXiv preprint, 2012.

[32] P. L. J. Ravikumar, "Quadratic programming relaxations for metric labeling and

markov random field map estimation," in Proceedings of the 23rd international

conference on Machine learning, 2006.

[33] D. Bertsimas and J. . N. Tsitsiklis, ntroduction to linear optimization, Belmont:

Athena Scientific, 1997.

[34] V. W. M. Kolmogorov, "On the optimality of tree-reweighted max-product message-

passing," arXiv preprint, 2012.

[35] C. Yanover, T. Meltzer and Y. Weiss, "Linear Programming Relaxations and Belief

Propagation--An Empirical Study," The Journal of Machine Learning Research,

vol. 7, pp. 1887-1907, 2006.

[36] D. M. T. G. A. W. Y. J. T. Sontag, "Tightening LP Relaxations for MAP using

Message-Passing," in Conference in Uncertainty in Artificial Intelligence, 2008.

[37] D. C. D. K. L. Y. Sontag, "Efficiently Searching for Frustrated Cycles in MAP

Inference," in Uncertainty in Artificial Intelligence, 2012.

[38] R. E. Bellman and S. E. Dreyfus, Applied dynamic programming, 1962.

[39] A. Caprara, D. Pisinger and P. Toth, "Exact solution of the quadratic knapsack

problem," INFORMS Journal on Computing, vol. 11, no. 2, pp. 125-137, 1999.

[40] C. Witzgall, "Mathematical methods of site selection for Electronic Message

Systems (EMS)," NASA STI/Recon Technical Report, vol. 76, p. 18321, 1975.

Page 70: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

65

[41] C. E. Ferreira, A. Martin, C. C. de Souza, R. Weismantel and L. A. Wolsey,

"Formulations and valid inequalities for the node capacitated graph partitioning

problem," Mathematical Programming, vol. 74, no. 3, pp. 247-266, 1996.

[42] D. J. Laughhunn, "Quadratic binary programming with application to capital-

budgeting problems," Operations Research, vol. 18, no. 3, pp. 454-461, 1970.

[43] A. Billionnet, A. Faye and É. Soutif, "A new upper bound for the 0-1 quadratic

knapsack problem," European journal of operational research, vol. 112, no. 3, pp.

664-672, 1999.

[44] P. Michelon and L. Veilleux, "Lagrangean methods for the 0–1 quadratic knapsack

problem," European Journal of Operational Research, vol. 92, no. 2, pp. 326-341,

1996.

[45] C. Helmberg, F. Rendl and R. Weismantel, "A semidefinite programming approach

to the quadratic knapsack problem," Journal of combinatorial optimization, vol. 4,

no. 2, pp. 197-215, 2000.

[46] G. Gallo, P. l. Hammer and B. Simeone, "Quadratic knapsack problems,"

Combinatorial Optimization, pp. 132-149, 1980.

[47] A. Billionnet and F. Calmels, "Linear programming for the 0–1 quadratic knapsack

problem," European Journal of Operational Research, vol. 92, no. 2, pp. 310-325,

1996.

[48] B. A. Julstrom, "Greedy, genetic, and greedy genetic algorithms for the quadratic

knapsack problem," in Proceedings of the 7th annual conference on Genetic and

evolutionary computation, New York, 2005.

[49] D. P. Bertsekas, Constrained optimization and Lagrange multiplier methods,

Academic press, 2014.

[50] A. Billionnet and É. Soutif, "An exact method based on Lagrangian decomposition

for the 0–1 quadratic knapsack problem," European Journal of operational

research, vol. 157, no. 3, pp. 565-575, 2004.

[51] A. Jimenez, A. Crespo, E. Migoya and J. Garcia, "Advances in large-eddy

simulation of a wind turbine wake," Journal of Physics: Conference Series, vol. 75,

no. 1, p. 012041, 2007.

[52] P. Beaucage, N. Robinson, M. Brower and C. Alonge, "Overview of six commercial

and research wake models for large offshore wind farms," in Proceedings of the

European Wind Energy Associate (EWEA).

Page 71: WIND FARM LAYOUT OPTIMIZATION USING APPROXIMATE INFERENCE IN GRAPHICAL MODELS · 2018-04-24 · optimization methods that operate on graphical models by iteratively passing beliefs

66

[53] J. S. González, A. G. G. Rodríguez, J. C. Mora, M. B. Payán and J. R. Santos,

"Overall design optimization of wind farms," Renewable Energy, vol. 36, pp. 1973 -

1982, 2011.

[54] S. Chowdhury, J. Zhang, A. Messac and L. Castillo, "Unrestricted wind farm layout

optimization (UWFLO): Investigating key factors influencing the maximum power

generation," Renewable Energy, vol. 38, pp. 16 - 30, 2012.

[55] S. Chowdhury, J. Zhang, A. Messac and L. Castillo, "Optimizing the arrangement

and the selection of turbines for wind farms subject to varying wind conditions,"

Renewable Energy, vol. 52, pp. 273 - 282, 2013.

[56] C. Wan, J. Wang, G. Yang, H. Gu and X. Zhang, "Wind farm micro-siting by

Gaussian particle swarm optimization with local search strategy," Renewable

Energy, vol. 48, pp. 276 - 286, 2012.

[57] W. P. Adams and H. D. Sherali, "A Tight Linearization and an Algorithm for Zero-

One Quadratic Programming Problems," Management Science, vol. 32, no. 10, pp.

1274 - 1290, 1986.

[58] M. O. Hansen, Aerodynamics of Wind Turbines, London: Earthscan, 2008.

[59] [Online]. Available: http://pub.ist.ac.at/~vnk/software/SRMP-v1.01.zip.

[60] [Online]. Available: http://cs.nyu.edu/~dsontag/code/mplp_ver2.tgz.

[61] International Organization for Standardization, "Acoustics -- Attenuation of sound

during propagation outdoors -- Part 2: General method of calculation (ISO 9613-

2:1996)," International Organization for Standardization, 1996.

[62] Z. Gu, G. L. Nemhauser and M. W. Savelsbergh, "Lifted cover inequalities for 0-1

integer programs: Computation," INFORMS Journal on Computing, vol. 10, no. 4,

pp. 427-437, 1998.