Improving prediction methods to estimate power curtailment

Improving prediction methods to
estimate power curtailment
Saki Kinney
Professor Eric Suess
Department of Statistics, CSUH
CSU Annual Student Research Competition
May 3-4, 2002
Background: Curtailment Programs
• Energy demand management has long
been of interest
– Maintain power reliability
– Reduce costs
• Programs provide incentive to reduce
power when electricity is in peak
demand
– 2001: “Use your appliances after 7pm”
Measuring Demand Reduction
• Key component of incentive program is to
determine participant performance
– Requires estimating what power would
have been if a curtailment order had not
been issued
– Curtailment = Predicted Power- Actual Power
• Typically averaging methods are used
Average Method
Prediction = Average of the same hour
over the previous ten business days
789 kW
Curtailment
Average Method
• Turns out to be a poor prediction
method
• Does not account for variables affecting
power consumption
– Time (past consumption)
– Temperature
Temperature Variation
Daily Load Profile 6/98 - 5/99
San Francisco Office (100,000 sf)
600
Average Demand (kW)
500
Maximum
400
3rd Quartile
Median
300
1st Quartile
Minimum
200
100
0
0
4
8
12
Hour of Day
16
20
Temperature Variation
• Curtailments typically called for on hottest
days
• New participants complained this put
them at a disadvantage
– Performance measured against cooler days
• Past participants typically were not
concerned with temperature variability.
Analysis Overview
• Compare overall prediction accuracy for
different models to ISO average method
– “Goodness of fit”
• Compare curtailment estimates during
actual power emergency
Other Prediction Models
• Regression Model
– Power vs. Temperature
• Autoregression Model
– Predict from previous values
Regression Method
• Prediction = a + b*Temperature
– Use previous ten days power and temperature data
to get model for each hour
Power Use 2-3pm (MW)
7.5
7.0
6.5
R2 = 0.76
1534 kW
Curtailment
6.0
5.5
5.0
60
65
70
75
80
85
Maximum Daily Temperature (Deg F)
90
Autoregression Method
• Past power usage used to predict future
power usage
– Use several days data, including current
day, up to curtailment time
• Temperature variation is implicit
– Temperature data not required
• Time correlation is accounted for
Autoregression Method
8
7
1032 kW
Curtailment
Power (MW)
6
5
4
3
Actual
Predicted
2
1
0
1.00
Hourly Time Series Data
281.00
Which Model is Better?
• Regression and Autoregression models
appear better than the Average model
– Comparable to each other
• Not conclusive – small sample size
Model Comparison
9000
Predicted Value (kW )
8000
7000
6000
A verage
5000
R egres s ion
4000
A u toregres s ion
3000
2000
1000
0
0
1000
2000
3000
4000
5000
6000
A c tu a l V a lu e (k W )
7000
8000
9000
Summer 2001 Results
• July 3rd Power Emergency
– Illustrate difference in curtailment
estimates using different models
– Example: Results from two Federal
Buildings that participated in ISO program
(Source: US GSA, LBNL)
July 3rd Power Emergency
Impact on Performance
• A higher baseline leads to higher
performance estimate
– Performance= Estimated Curtailment
Promised Curtailment
• Averaging model tends to understate
actual performance
Further Considerations
• Analyze data from different types of
buildings in different climate zones
• Consider model variations
– Evaluate models used by other states and
programs
Conclusions
• Better predictive models are needed in
demand reduction programs
– Fairness
– Reliability
– Decision making and planning
• Accounting for temperature variability
and time dependence provides more
accurate results
Acknowledgments
Cal State Hayward
Statistics Department
Lawrence Berkeley Lab
Environmental Energy
Technologies Division