Ensemble re-forecasting methods for enhanced power load prediction

Electric load forecasting is a key element for management and operation of the electric grid. In this study we introduce ensemble re-forecast methods that take an initial forecast and produce a better prediction by extracting information from the structured errors. The models in the ensemble rely upon the real-time information obtained from load measurements and estimates over a state-wide domain. The weights in the ensemble are optimized in three different ways based on global, hourly, and weekly performance of the models. The proposed methodology is applied to predict hour-ahead market (HAM) and day-ahead market (DAM) load for California Independent System Operator (CAISO) and Electric Reliability Council of Texas (ERCOT) respectively. Proposed models showed consistent performance enhancements for all the cases. HAM predictions show an improvement of 47% and 36% in terms of Mean Absolute Percentage Error over the forecasts provided by CAISO and ERCOT. For DAM, the improvements are 34% for CAISO and 47% for ERCOT. Temporal analysis comparing the internal forecast produced by the ISOs and re-forecasts shows significant improvement during off-peak hours and small improvement for on-peak hours. Results validate the potential of the proposed methodology to enhance the forecast accuracy, independent of load profile or forecast horizon.