A Two-Fold Machine Learning Approach for Efficient Day-Ahead Load Prediction at Hourly Granularity for NYC

Abstract

Recent surge in electricity requirements has propelled a need for accurate load forecasting methods. Several peak load demand forecasting methods exist which predict the highest load requirement of the day. However, Short Term Load Forecasting (STLF) takes precedence owing to the constant load fluctuation over the day, especially in developed cities, and therefore finds more practical and economical use. While statistical methods have largely been used for STLF, contemporary works involving Machine Learning (ML) have seen more success. Such ML methods have made use of several years of data, focused on testing only for a short duration (few weeks), disregarded federal and public holidays when the load demands are erratic, or utilized simulated and not real-time data. This provokes the need for a solution that is capable of forecasting real-time load accurately for all days of the year. The authors of this paper propose a unique two-fold approach to model the training data used for accurate day-ahead hourly load prediction, which also predicts suitably well for federal and public holidays. The New York Independent System Operator’s (NYISO) electrical load dataset is used to evaluate the model for the year 2017 with a Mean Absolute Percentage Error (MAPE) of 3.596.