hems: a home energy market simulator
TRANSCRIPT
Andrea Monacchi, Sergii Zhevzhyk, Wilfried ElmenreichInstitute of Networked and Embedded Systems / Lakeside Labs
Alpen-Adria-Universität Klagenfurt
HEMS: A Home Energy Market Simulator
3rd Energy Informatics Conference - November 2014 - Zurich
Motivation
● Home Energy Management Systems○ Unifying solution for storage, generation, consumption○ Rationalization of local energy resources○ Residents in the loop: awareness and cost control○ Especially interesting for rural off-grid communities (uGrids)
● Goal: aid smart controller design○ Energy allocation as a market
Market-based control
● Distributed self-interested agents in competition○ Preferences as private price model (local view)○ Better reflects needs of multiple users in a community○ Price proportional to demand and inverse to supply○ Demand and production are not known in advance
● Double-sided auctions○ BID and ASK offers ordered by price in shared orderbook○ Shared best BID (H), best ASK (L)○ NYSE spread reduction rule○ k-pricing: p = k * bid.p + (1 - k ) * ask.p, k ∈ [0,1]
Agent-based modeling
● Scenario description format○ Weather model: cloudiness timeserie○ Power grid connections: truth-telling agents
■ price models: energy tariff & feed-in tariff■ power availability & power capability models
○ Prosumers: profit-oriented agents
A. Monacchi, S. Zhevzhyk, and W. Elmenreich. Home energy market simulator: The scenario definition format. Alpen-Adria Universität Klagenfurt, Technical Report, September 2014.
unit price
Production model
Demand model
Flexible or Inflexible service?
The generation model
● Available production over timea. External time series of measured power datab. Simple battery modelsc. Exploiting a generation model dependent on current weather
M. Pöchacker, T. Khatib, and W. Elmenreich. The microgrid simulation tool RAPSim: description and case study. In Proc. of the 2014 IEEE Innovative Smart Grid Technologies (ISGT) Conference, Kuala Lumpur, Malaysia, May 2014.
Photovoltaic plant:● peak power● efficiency● latitude● longitude● height● size
getCurrentProduction(current_date, current_weather)
● Device signature■ sequence of n states σ, with σi = ❬Pi, di, χ
s❭■ device start delay sensitivity χb (secs.)■ intra-state interruption sensitivity χi
The operation model
Example power profile: dishwasher
Available for smart appliances, mined for legacy devices
Currently none
User discomfort
Correct operation
● Usage model○ Time of use probability Puse
■ Puse = 1 − (Phold)N , over N instants
● Puse = 1 − (1 - ω*)N then ω* = 1 - root(1 - Puse , N)■ device starting willingness ω* ∈ [0, 1] for specific time instant
○ Willingness decay λ to update Puse■ Puse = Puse(1 − λ)n , with n completed operations■ simple but not expressive enough
○ Can be extracted from consumption data!
The operation model
Extracting appliance usage models
● GREEND dataset○ P @ 1Hz in 8 selected households in AT and IT for 1 year
A. Monacchi, D. Egarter, W. Elmenreich, S. D’Alessandro, and A.M. Tonello. GREEND: an energy consumption dataset of households in Italy and Austria. in Proc. of IEEE Int. Conf. on Smart Grid Communications (SmartGridComm), Venice, Italy, Nov 2014.
Example: Coffee machine
Extracting appliance usage models
● Appliance Usage Model Manager (UMMA)
A. Monacchi, S. Zhevzhyk, and W. Elmenreich. Home energy market simulator: The appliance usage model manager. Alpen-Adria Universität Klagenfurt, Technical Report, September 2014.
Device usage model interface Model extraction Model assessment and exploitation
Example model: BN
Learning Smart controllers(ψs/pmax)
fully-meshedANN
(amin/pmax) posASK mtcdASK
Hour Month Weekday
τImportance
χlb /χb
(ψb/pmax) (bmax/pmax) posBID mtcdBID
Seller’s
Context
Trader’s
Buyer’s
p = 2 ∗ pmax ∗ poutput − pmax
Offer Type
ASK p < -pth
NOOP -pth <= p <= pth
BID p > pth
Offer price
trading tendency: τ ∈ [-1,+1], with -1.0 sell and +1.0 buy
Learning Smart controllers
● Objective: minimizing costs and discomfort
● Heuristic fitness function○ F = R + (δg ∗Igrid) − C
I.Fehervari and W.Elmenreich. Evolving neural network controllers for a team of self-organizing robots. Journal of Robotics, 2010.
Demand qjt = qj
t,s + qjt,o , ∀bj ∈ B
Rationality ptj ≤ ψbj,∀bj ∈ B; pt
i ≥ ψsi,∀si ∈ S
Brokerage qjt − qi
t ≥ 0, with bj = si , ∀bj ∈ B,∀si ∈ S
Self-trading bθ ≠ sθ , ∀θ ∈ Θ
Reward: utility delivered to users upon operation completion
Income from fed into Grid
Running costs
Similar approach
Learning Smart controllers
● Running costs Expenses for energy imported from the Grid
Discomfort from delayed start of operated flexible devices
Discomfort from delayed start of intermediate states in operated multi-state flexible devices
Discomfort from state interruption of operated flexible devices
Penalty from violating the market policy: NYSE
Inflexible demand/supply not allocated
Penalty resulting from irrational trading (sensitivity & reservation price violations)
Penalty from price not reflecting needs: according to operation and generation model
δg, δb, δs, δi, δm, δl, δn, etc Penalty selection assigns objective importance and drives evolution: cumbersome!
*Values are normalised on device parameters
Evolving smart controllers
● Evolutionary computing: selection of the fittest
● FREVO: framework for evolutionary computinghttp://frevo.sourceforge.net
A. E. Eiben and J. E. Smith, Introduction to evolutionary computing. Springer, 2003.
Rank candidates by fitness
Explore solution space
Combine successful solutions
For g generations or until a fitness f is reached
Early experiments
● Uniform-price double auction with 1 sec alloc.● 250 generations NNGA with 80 candidates each
Results
● Learned controllers○ Well perform given objective fitness function
● Selected market mechanism:○ Perform well, although might yield suboptimal allocations
i. Commodity divisibilityii. Prevention of offers exceeding available supply
○ Bundling problem: I only want B if I also get Ai. Resource re-allocation to solve “conflicts”ii. Conflict avoidance (self-org.) through additional sensory inputiii. Combinatorial auction (WDP is NP-Complete and centralised)
Conclusions & Future work● HEMS Modeling & Simulation tool
○ Assisted design of energy trading behavior○ Part of the FREVO evolutionary framework○ Built-in generation/load models
○ Testbed for market mechanisms & policies■ Market solver uncoupled from abstract simulator loop■ Easy to add agents for specific markets
● Implementing variant of clock auction
Questions?
Andrea Monacchi
Smart Grid groupAlpen-Adria Universität Klagenfurt
E: [email protected]: http://wwwu.aau.at/amonacch
Double-side auctions
● More balanced than monopsonistic and monopolistic○ Convergence to price equilibrium even with a few traders
● BID and ASK offers○ Ordered by price in a shared order book○ shared market status: best BID (H), best ASK (L)○ Offer accepting policy: NYSE spread reduction rule
● Pricing mechanism○ k-pricing: p = k * bid.p + (1 - k ) * ask.p, k ∈ [0,1]○ DDA vs CDA vs UCDA