optimization with variable energy...
TRANSCRIPT
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Optimization with Variable Energy Prices
Helmut Simonis
Cork Constraint Computation CentreComputer Science Department
University College CorkIreland
CPAIOR 2012, Nantes
Helmut Simonis Optimization with Variable Energy Prices 1
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Take away message
Electricity suppliers move to time variable pricesScheduling should take this into accountA Family of resource cost aware constraintsProblem generatorWhere do forecast prices come from?Improved forecast does not imply improved scheduleInteresting extension to building energy management
Helmut Simonis Optimization with Variable Energy Prices 2
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Working with
4CTarik HadzicDiarmuid GrimesGeorgiana IfrimBarry O’Sullivan
IndustryCharles Sheridan, IntelAnnabelle Pratt, Intel
Campus21 FP7 projectUCC, TU Wien, Cartif, BAM, NEC, UTRC-I, Sirus,HSG-Zander, City of Valladolid, ESB
Helmut Simonis Optimization with Variable Energy Prices 3
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Funding Sources
SFI TIDA Grant: Energy Cost Aware SchedulingIRCSET Post Doctoral Fellowship (Intel)SFI PI Grant (O’Sullivan)FP7 Campus21 Project
Helmut Simonis Optimization with Variable Energy Prices 4
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Outline
1 Motivation
2 Scheduling with Energy Costs
3 Predicting Prices
4 Extension to Building Management
Helmut Simonis Optimization with Variable Energy Prices 5
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
A Changing World for Electricity Supply
Many countries moving away from single, nationalelectricity suppliersAllow customers to switch between suppliersStill use shared distribution networksIncreasing use of renewable sources (wind, solar PV)Political aims to reduce CO2 productionPrimary energy cost increasing rapidlySecurity of supply not guaranteed
Helmut Simonis Optimization with Variable Energy Prices 6
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
The Rising Cost of Electricity (Source: Eurostat)
Helmut Simonis Optimization with Variable Energy Prices 7
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Auction Based Electricity Markets
National/Regional price finding mechanismNew price produced every 30 minutes (Ireland)Generators bid for productionBids accepted in merit orderFill predicted/real demand for time slotPrice defined by last accepted bidAdditional uplift price, start-up/shut-down/reserve costsConsumer price includes distribution costs, margins andtaxes
Helmut Simonis Optimization with Variable Energy Prices 8
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Demand and Prices (Source: SEMO)
Demand Price
Figure: Electricity Demand and Price, ROI, Comparing January andJune data 2010
Helmut Simonis Optimization with Variable Energy Prices 9
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Why Time Variable Prices?
Determine “fair” price for variable demand/productionCapacity of renewable resources variable
“Zero” marginal costCapacity profile not linked to demand
Penalize CO2 intensive generatorsMove demand away from peak periods
Helmut Simonis Optimization with Variable Energy Prices 10
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Consumer Prices
Most consumers (residential and industrial) still pay fixedprice per kWhSome customers use simple day/night tariffs (2 meters)Smart-meters required to allow more flexibilityTime-of-Use (ToU) tariffs offered by some utilitiesConsumer real-time prices still experimental
Helmut Simonis Optimization with Variable Energy Prices 11
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Time-of-Use Tariffs
Price changes during day for certain periodsWeekend prices are differentSummer/Winter prices possibleCost at each time period known well in advance (6months?)Not really linked to market price
Helmut Simonis Optimization with Variable Energy Prices 12
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Summer Tariff
Helmut Simonis Optimization with Variable Energy Prices 13
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Winter Tariff
Helmut Simonis Optimization with Variable Energy Prices 14
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Why are real-time prices difficult to sell?
Customers face uncertainty about costLet supplier worry about problem
Pay premium for fixed prices
How to exploit changing prices?Opportunity to reduce costs by intelligent use
But impact on other cost factors (social, environmental)
Helmut Simonis Optimization with Variable Energy Prices 15
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Further Complication: Feed-in
Customers may have their own electricity sourcesWindCHP (Combined heat and power)Solar PV
These sources are cheaper than grid (sunk investmentcost)Feed-in prices can be low, fair, subsidized
Helmut Simonis Optimization with Variable Energy Prices 16
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Outline
1 Motivation
2 Scheduling with Energy CostsUse CaseCost Aware CumulativeIntegrating Energy Cost into Other SchedulingConstraintsAn Instance Generator
3 Predicting Prices
4 Extension to Building Management
Helmut Simonis Optimization with Variable Energy Prices 17
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Scheduling with variable energy cost
Extend existing scheduling constraints with a variableresource cost componentAllow prices to vary in time and in volumeAt each time, use cheapest source firstAdd-on to existing constraint-based schedulingMainly affects cost objective
Helmut Simonis Optimization with Variable Energy Prices 18
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Surely somebody is doing this already?
One reference found for production schedulingP.M. Castro, I. Harjunkoski and I. E. Grossmann. A New
Continuous-Time Scheduling Formulation for Continuous
Plants under Variable Electricity Cost, IND ENG CHEMRES , vol. 48, no. 14, pp. 6701-6714, 2009.Continuous processTime of use tariff
Large body of work on energy efficient scheduling insideprocessors, not time variableDatacentre scheduling (follow the moon)
Helmut Simonis Optimization with Variable Energy Prices 19
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Use Case: Feed Mill Scheduling (Simonis 2006)
Animal feed production in UKFeed in different sizesFor different speciesHuman health risk
ContaminationBSE
Strict regulationsConstraints
Avoid contamination risksMachine setup timesMachine choice (quality/speed)Limited storage of finished productsVery short lead times (8-48 hours)Factory structure given as data
Status (COSYTEC Product)Operational since Nov 96
Helmut Simonis Optimization with Variable Energy Prices 20
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Product Made: Feed Pellets (Image: Wikipedia)
Helmut Simonis Optimization with Variable Energy Prices 21
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Feed Mill Schematic (www.feedmachinery.com)
Pellet Press: Rating 375kWYearly UK production (2009-2010): 10M tonnes (Defra)
Helmut Simonis Optimization with Variable Energy Prices 22
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Why this use case?
Day-by-Day scheduleMake during night for delivery tomorrow morningOnly need to know prices for 24/36 h aheadEnergy use depends on recipeCurrent day/night tariff limits production capacityMargins extremely tight
Helmut Simonis Optimization with Variable Energy Prices 23
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Reminder: Cumulative
Aggoun, Beldiceanu 1993Core global constraint for constraint-based schedulingLarge number of algorithmic developments, few changes ofbasic constraintTime/volume dependent resource cost not considered sofar
Cumulative([s1, s2, ...sn], [d1, d2, ...dn], [r1, r2, ...rn], l , p),
Helmut Simonis Optimization with Variable Energy Prices 24
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
New Constraint Variant: CumulativeCost
Add cost elementPer unit cost expressed with areasIntersection of resource use profile with areas defines costGlobal reasoning required
Helmut Simonis Optimization with Variable Energy Prices 25
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Formally: CumulativeCost
∀ 0 ≤ t < p : prt :=�
{i|si≤t<si+di}
ri ≤ l
∀ 1 ≤ i ≤ n : 0 ≤ si < si + di ≤ p
ov(t , prt ,Aj) :=
�max(0,min(yj + hj , prt)− yj) xj ≤ t < xj + wj
0 otherwise
∀ 1 ≤ j ≤ m : aj =�
0≤t<p
ov(t , prt ,Aj)
cost =m�
j=1
ajcj
Helmut Simonis Optimization with Variable Energy Prices 26
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Toy Example
12
1
23
1
45
1
34
1
01
1
3
← j
← cj
← aj
2 52
1t1
1 52
2t2
0 5
3
1
t3
Helmut Simonis Optimization with Variable Energy Prices 27
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Toy Example: Optimal Solution
12
1
23
1
45
1
34
1
01
1
3
← j
← cj
← aj
cost = 15
2 52
1t1
1 52
2t2
0 5
3
1
t3
23 2 02
Helmut Simonis Optimization with Variable Energy Prices 28
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
How to implement this?
element Element constraints linking start and costA greedy assignmentB greedy assignment
Flow flow from tasks to areasLP1 Flow + extra inequalitiesLP2 Flow + extra inequalitiesDLP time indexed LP model
DMIP MIP variant of DLP
Helmut Simonis Optimization with Variable Energy Prices 29
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Comparative Power of Algorithms
Element
Alg A
Flow
Alg B
LP 1 LP 2 DLP DMIP
Helmut Simonis Optimization with Variable Energy Prices 30
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Clear Winner: DLP
Based on LP relaxation of Cumulative by HookerMost others can be quite weakMIP is too expensive to solveDLP has very good cost boundScalable for 100s (not 1000s) of tasksCost based pruning: Reduced Cost Filtering
Helmut Simonis Optimization with Variable Energy Prices 31
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
LP Lower Bound
lb = minq�
j=1
aj cj
prt ∈ [0, l] yit ∈ {0, 1} zjt ∈ [0, hj ]
∀ 1 ≤ j ≤ q : 0 ≤ aj ≤ wj hj
∀ 1 ≤ i ≤ n : si =p−1�
t=0
tyit
∀ 1 ≤ i ≤ n :p−1�
t=0
yit = 1
∀ 0 ≤ t < p : prt =�
1≤i≤n
�
t�≤t<t�+di
yit� ri =q�
j=1
zjt
∀ 1 ≤ j ≤ q : aj =
xj+wj−1�
t=xj
zjt
Helmut Simonis Optimization with Variable Energy Prices 32
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Toy Example: DLP and DMIP
Direct LP Direct MIP
12
1
23
1
45
1
34
1
01
1
3
← j
← cj
← aj
lb = 1233 1 11
12
1
23
1
45
1
34
1
01
1
3
← j
← cj
← aj
cost = 1523 2 02
Helmut Simonis Optimization with Variable Energy Prices 33
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Can we use this for something else?
Scheduling problems with manpower costsMan/hour cost varies over timeOffice Hours/Nights/Weekends/HolidaysExtra staff costs more: Temps/FreelanceNatural, otherwise change hire rules
Helmut Simonis Optimization with Variable Energy Prices 34
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Why don’t we just use CumulativeCost with otherresource constraints?
Combine a CumulativeCost with a Disjunctiveresource constraintCost estimate does not take non-overlapping into accountToo optimistic cost valueWorks if we have one global LP relaxation
Helmut Simonis Optimization with Variable Energy Prices 35
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
DisjunctiveCost
The DisjunctiveCost constraint allows one task to be runon a machine at any time. We can describe the constraint byadding
∀ i , j |i �= j : si + di ≤ sj ∨ sj + dj ≤ si
to constraints.In the LP/MIP model, we extend constraints with the condition
∀ 0 ≤ t < p :�
1≤i≤n
�
t �≤t<t �+di
yit � ≤ 1
Helmut Simonis Optimization with Variable Energy Prices 36
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
ParallelMachineCost
The ParallelMachineCost constraint consists of constraintsplus the constraints
∀ 1 ≤ k ≤ b, ∀ i , j |i �= j : mi �= mj ∨ si + di ≤ sj ∨ sj + dj ≤ si
We can express this condition in the LP/MIP model by addingconstraints of the form
∀ 1 ≤ k ≤ d , ∀ 0 ≤ t < p :�
{i|mi=k}
�
t �≤t<t �+di
yit � ≤ 1
to constraints.
Helmut Simonis Optimization with Variable Energy Prices 37
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
MachineChoiceCost
∀ 0 ≤ t < p : prt :=�
{i|si≤t<si+dimi}
rimi≤ l
∀ 1 ≤ i ≤ n : 0 ≤ si ≤ si < si + dimi≤ si + dimi
≤ p
ov(t , prt ,Aj) :=
�max(0,min(yj + hj , prt)− yj) xj ≤ t < xj + wj
0 otherwise
∀ 1 ≤ j ≤ q : aj =�
0≤t<p
ov(t , prt ,Aj)
cost =q�
j=1
ajcj
Helmut Simonis Optimization with Variable Energy Prices 38
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Instance Generator
An instance generator is available in the form of a Java jar fileCostInstance.jar which can be downloaded from
http://4c.ucc.ie/∼thadzic/CostInstance.jar
To create an instance, execute:
java -cp CostInstance.jar Instance <parameters>
where <parameters> are:
Helmut Simonis Optimization with Variable Energy Prices 39
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Instance Generator
-instanceType 0 - CumulativeCost, 1 - DisjunctiveCost,2 - ParallelMachineCost
-n number of required tasks-m number of required areas-d_max maximum duration of a task-r_max maximum resource consumption of a task-s_diff_portion portion of the horizon restricting the start time domain-util utilization of the total available area-cost_distr cost distribution, 0 - explicitly given, 1 - random-w width of each area-machineNo number of machines for parallel machine instances-randomSeed initial random seed-maxCost maximal random cost of an area-costFileName a valid file name containing a vector of costs (or input "no-file")
Helmut Simonis Optimization with Variable Energy Prices 40
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
XML Output Example
<?xml version = ”1.0” encoding = ”UTF − 8” standalone = ”yes”? >< instance resource − limit = ”6” horizon = ”21”
xmlns : xsi = ”http : //www.w3.org/2001/XMLSchema − instance”xsi : noNamespaceSchemaLocation = ”resourcecost.xsd”/ >
< tasks number = ”5” >< task id = ”0” start_min = ”4” start_max = ”4” duration = ”3” resource = ”2”/ >< task id = ”1” start_min = ”1” start_max = ”5” duration = ”4” resource = ”5”/ >< task id = ”2” start_min = ”8” start_max = ”8” duration = ”6” resource = ”5”/ >< task id = ”3” start_min = ”0” start_max = ”10” duration = ”3” resource = ”2”/ >< task id = ”4” start_min = ”3” start_max = ”9” duration = ”3” resource = ”3”/ >< /tasks >< areas number = ”7” >< area id = ”0” x = ”0” y = ”0” width = ”3” height = ”6” cost = ”14”/ >< area id = ”1” x = ”3” y = ”0” width = ”3” height = ”6” cost = ”11”/ >< area id = ”2” x = ”6” y = ”0” width = ”3” height = ”6” cost = ”7”/ >< area id = ”3” x = ”9” y = ”0” width = ”3” height = ”6” cost = ”7”/ >< area id = ”4” x = ”12” y = ”0” width = ”3” height = ”6” cost = ”0”/ >< area id = ”5” x = ”15” y = ”0” width = ”3” height = ”6” cost = ”7”/ >< area id = ”6” x = ”18” y = ”0” width = ”3” height = ”6” cost = ”7”/ >< /areas >< machines number = ”2” >< machine id = ”0” tasks = ”2 4 3”/ >< machine id = ”1” tasks = ”0 1”/ >< /machines >
Helmut Simonis Optimization with Variable Energy Prices 41
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Use CaseCost Aware CumulativeIntegrating Energy Cost into Other Scheduling ConstraintsAn Instance Generator
Integration into JSR-331 API
We have integrated our experimental code into the JSR-331:
Constraint Programming API. We extended a standardResource class with ResourceWithCost class that supports acost-declaring method:
Var setCost ( i n t x1 , i n t x2 , i n t y1 , i n t y2 , i n t cost )
Helmut Simonis Optimization with Variable Energy Prices 42
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Outline
1 Motivation
2 Scheduling with Energy Costs
3 Predicting Prices
4 Extension to Building Management
Helmut Simonis Optimization with Variable Energy Prices 43
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Where do the prices come from?
We need to know prices in advance, for completescheduling horizonWhen we publish schedule, we are commitedSEMO publishes a 24h day-ahead priceThis is a forecast, not the actual priceIt is not known how this forecast is computedCan we do better?
Helmut Simonis Optimization with Variable Energy Prices 44
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Price Linked to Load
Helmut Simonis Optimization with Variable Energy Prices 45
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Surprises Happen
Helmut Simonis Optimization with Variable Energy Prices 46
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
SMP (System Marginal Price) in 2009-2011
Year Min Median Mean Max2009 4.12 38.47 43.53 580.532010 -88.12 46.40 53.85 766.352011 0 54.45 63.18 649.48
Helmut Simonis Optimization with Variable Energy Prices 47
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Features Used in Forecast
Historical SMPShadow priceLoadNational/Local Weather ForecastWind ForecastCalendar InformationScheduled OutagesSupply Bids
Helmut Simonis Optimization with Variable Energy Prices 48
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Two Forecasting Models
FR1 Predict the SMP using historical and forecastedSMP, shadow price, load and supply.
FR2 Predicting the SMP using the local average-SMPand a learned difference-from-average model.Average price in each time period is quite stable,predict difference from average price.
Helmut Simonis Optimization with Variable Energy Prices 49
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Details
Using SVM (Support Vector Machine) modelsNot all features are useful/required for forecast
e.g. load forecast already based on calendar
Helmut Simonis Optimization with Variable Energy Prices 50
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Evaluation (Errors and paired t-tests)
Model MAE MSESEMO 12.64 1086.25FR1 11.14 821.01FR2 11.21 781.72Baseline Price SEMO FR1 FR2Actual L 761.8 513.5 486.9
U 1410.7 1128.4 1076.4SEMO L - 172.4 209.7
U - 358.0 399.3FR1 L - - 11.5
U - - 66.9
Helmut Simonis Optimization with Variable Energy Prices 51
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Apply this to Schedule
Our forecast prices are significantly betterFor many applications this is enoughDoes this mean we produce better schedule?Apply to feedmill problem, but abstract detail to solve tooptimalityOnly consider task assignment to machines beforedue-dateOptimize energy cost with forecast, evaluate with actualpriceIgnore product sequence, set-up costs, etc.
These will only reduce impact of price prediction
Helmut Simonis Optimization with Variable Energy Prices 52
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Simplified Scheduling Model
cost =�
t
pr∗t at
where pr∗t
is the profile value at time t of the optimal solution tothe following MIP problem:
min�
t
prtvt
subject to:
∀i :�
txit = 1
∀t :�
i
�t−di+1≤t �≤t
pixit � = prt ≤ lt
∀m∀t :�
i|mi=m
�t−di+1≤t �≤t
xit � ≤ 1∀i∀t |t+di>ei
: xit = 0
Helmut Simonis Optimization with Variable Energy Prices 53
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Results over 880 runs
Price Min Median Mean MaxActual 4,383,718 5,934,654 6,093,365 9,805,821SEMO 4,507,136 6,054,220 6,272,768 10,218,804FR1 4,499,811 6,058,093 6,266,800 10,070,541FR2 4,570,552 6,094,818 6,283,261 10,059,264
Helmut Simonis Optimization with Variable Energy Prices 54
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
The Good News
Given only the forecast information, we can producehigh-quality schedules5-10% off optimal solution with perfect knowledge of futurepriceThis is lower than the mark-up that suppliers require forfixed/ToU pricesFor most cost-sensitive case, side constraints will onlyreduce overhead
Helmut Simonis Optimization with Variable Energy Prices 55
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
But: t-test Comparison between forecasts
Price SEMO FR1 FR2Actual L −200, 564.9 −193, 646.7 −211, 094.4
U −158, 241.3 −153, 222.5 −168, 697.4SEMO L - −1, 506.1 −17, 262.6
U - 13, 443.1 −3, 722.9FR1 L - - −23, 968.3
U - - −8, 954.2
Helmut Simonis Optimization with Variable Energy Prices 56
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
Can we fix it?
MAE and MSE are bad predictors for quality of scheduleHow to change the forecast generation to use schedulequality?Relative ranking of time periods is important for scheduleIt is more important to predict when peaks occur, ratherthan their magnitudeWe tested this in paperHow to do this: Current research
Helmut Simonis Optimization with Variable Energy Prices 57
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
Outline
1 Motivation
2 Scheduling with Energy Costs
3 Predicting Prices
4 Extension to Building ManagementHEMSCampus21
Helmut Simonis Optimization with Variable Energy Prices 58
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
Energy Cost Aware Buildings
HEMS (home energy management system)Public spaces (campus, sports arena)Work in progress
Helmut Simonis Optimization with Variable Energy Prices 59
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
What is different?
Time of use tariff, rather than real-time pricingThermal model in addition to electricityEnergy storage (EV, thermal envelope)pre-heat, pre-coolFew scheduled resourcesData babel
Helmut Simonis Optimization with Variable Energy Prices 60
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
HEMS Architecture
SoftwareComponent
HEMS
User
Appliances
WeatherForecast
PriceForecaster
DistributionGrid HVAC
HomeElectricity
Grid
EV
EV Charger
Baseload
Zone Outside
ActivitiesTemperature RequestBattery Charge Request
start
setpoint
charge/discharge
ratePriceSignal
heating/cooling
heat-loss
electricity
thermal
command
information
Helmut Simonis Optimization with Variable Energy Prices 61
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
EV Charging with Car-to-Home Capability
Helmut Simonis Optimization with Variable Energy Prices 62
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
Pre-Heating of Rooms
Helmut Simonis Optimization with Variable Energy Prices 63
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
Huerta del Rey Sports Center, Valladolid
Helmut Simonis Optimization with Variable Energy Prices 64
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
Heating and DHW: Principle of Operations
Helmut Simonis Optimization with Variable Energy Prices 65
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
Conclusions
Electricity suppliers move to time variable pricesScheduling should take this into accountA Family of resource cost aware constraintsProblem generatorWhere do forecast prices come from?Improved forecast does not imply improved scheduleInteresting extension to building energy management
Helmut Simonis Optimization with Variable Energy Prices 66
MotivationScheduling with Energy Costs
Predicting PricesExtension to Building Management
HEMSCampus21
References
H. Simonis and T. Hadzic. Constraint-based scheduling for reducing peak
electricity use. CompSust’10, Boston, MA, June 2010.H. Simonis and T. Hadzic. A resource cost aware cumulative. ModRef 2010, StAndrews, Scotland, September 2010.H. Simonis and T. Hadzic. A Family of Resource Constraints for Energy Cost
Aware Scheduling. CROCS 2010, St Andrews, Scotland, September 2010T. Hadzic and H. Simonis. Creating Tests for a Family of Cost Aware Resource
Constraints. CSCLP 2010, Berlin, December 2010D. Grimes, H. Simonis, A. Pratt and C. Sheridan. Automated Energy Usage
Optimization for the Residential Sector: Impact of Price Tariffs. CompSust’12,Copenhagen, Denmark, July 2012.G. Ifrim, B. O’Sullivan and H. Simonis. Energy-Cost Forecasting for Scheduling.
Submitted for publication.
Helmut Simonis Optimization with Variable Energy Prices 67