Targeted residual analysis for improving electric load forecasting

Abstract

Management and operation of the electrical grid in the US is handled in large part by regional authorities called Independent System Operators (ISO's). One of the key activities of an ISO is load forecasting which is critical to short-term energy trading markets and effective operation of the power grid. In this paper, we analyze load forecasts and develop methods for improving forecasts that can be used directly by ISO's or third parties. Specifically, we assess the hourly electrical load forecasts against actual load data provided by Midwest ISO over a two-year period. Residual analysis shows systematic inaccuracies in hourly forecasts that can be caused by a variety of factors including modeling errors and pumped storage in the grid. We utilize machine learning-based methods to improve forecasts over short time horizons. Our methods reduce the mean squared error of forecasts over the entire year by roughly 20%. By shortening the forecast horizon to 1 to 32 hours, we are able to improve by over 90%. These improvements can be important in operational energy market contexts, where even small differences in forecasts can lead to large swings in transmission behavior and market activity.