170 likes | 297 Views
Load Forecasting Process Review Calvin Opheim Generation Adequacy Task Force October 7, 2013. Outline. Long-Term Load Forecast Process Review Previous Model Approach What we’ve learned New Modeling Approach Approach Weather Normalization 4 Coincident Peak Analysis Questions.
E N D
Load Forecasting Process Review Calvin Opheim Generation Adequacy Task Force October 7, 2013
Outline • Long-Term Load Forecast Process Review • Previous Model • Approach • What we’ve learned • New Modeling Approach • Approach • Weather Normalization • 4 Coincident Peak Analysis • Questions
Previous Model • 2-3 weather stations per weather zone • Used non-farm employment to capture future growth • Weather Zone Forecasts • Daily Energy Per Job = f(weather, season, day type, daylight) • Hourly Demand = f(temperatures, previous hour’s load) • ERCOT Forecast • ∑ eight weather zone loads
Previous Model – What we’ve learned • Historical values of economic data are subject to significant revision for two years • During the first quarter of 2013, the Bureau of Labor Statistics increased non-farm employment values by 1% in 2011 and 2% in 2012. • While values may seem small, relative impacts are significant. • Changing historical data compromises the accuracy of the model as “historical” relationships are subject to change. • Model was based on the assumption that non-farm employment values were stable.
Previous Model – What we’ve learned • Historical revisions impact forecast years • Moody’s forecast for CY2013 was increased by 2% in order to align with the revised historical values for CY 2012. • Did load suddenly increase by 2% due to these revisions? • Economic forecasts have been trending high, resulting in forecasts that reflect overly optimistic growth scenarios.
What’s new? • Daily energy forecasts are now based on Neural Network Models. • Growth is determined by multiple factors (premise growth rates, weather variables, day types, and their interactions). • A single economic variable has less influence on forecast outcome. • Benefits • ERCOT can determine/account for variable interactions more robustly, compared to linear regression models. • All predictor variables are used as inputs in each network node. • This approach produces more detailed/precise model formulation.
What’s new? • Forecasts will now be based on many model simulations instead of being based on a single linear model. • Neural Network models were developed with 33% of the historical data being withheld from model development. • The data being withheld was determined randomly. • Randomly withholding data mitigates over-fitting of the data. • The model’s accuracy was determined based on how well it predicted the sample holdout data. • Process was repeated hundreds of times (model convergence). • Benefits • In statistics, repeated sampling gives a more accurate estimate than a single sample. • The result is a more robust forecast.
What’s new? • Historical energy relationships will now be based on premise counts by customer class (residential, commercial and industrial). • Historical energy relationships will no longer be based on non-farm employment values. • Benefits • Historical premise accounts will be very stable and will not be subject to the significant changes exhibited by non-farm employment revisions. • “Historical values are actually historical.”
What’s new? • The determination of 15-year normal forecast will now be based on model output using the most recent 15 years of historical weather data. • Will no longer create a synthetic weather file for use in the model • Will no longer time align weather conditions for time of peak • Benefits • More accurately reflects historical weather patterns • More accurately reflects load diversity at time of peak
2011 Summer Peak – Impact of 4 CP load reduction • 4 CP impact shown is based on aggregated transmission load values for ~430 premises. • Estimate is not based on an analysis of individual premises. • Difference represents the 4 CP impact of ~600 MW on an aggregated basis. • 4 CP impact would likely be greater if analysis were performed on individual premises. 4 CP impact
2012 Summer Peak – 4CP Impact • 4 CP impact shown is based on aggregated transmission load values for ~430 premises. • Estimate is not based on an analysis of individual premises. • Difference represents the 4 CP impact of ~900 MW on an aggregated basis. • 4 CP impact would likely be greater if analysis were performed on individual premises. 4 CP impact
2013 Summer Peak - 4CP Impact • 4 CP impact shown is based on aggregated transmission load values for ~440 premises. • Estimate is not based on an analysis of individual premises. 4 CP Impact • Difference represents the 4 CP impact of ~500 MW on an aggregated basis. • 4 CP impact would likely be greater if analysis were performed on individual premises.
ON OFF Questions