Time-Series Forecasting Energy Loads: A Case Study in Texas

Abstract

Future predicted energy demand on the grid is a major factor that drives the prices of energy contracts on trading markets. Errors in forecasting are problematic for energy traders who buy and sell futures contracts on the expected price of energy: when decisions are made on inaccurate predictions, the market will be inefficient, leading to price volatility and investment losses. This paper proposes the use of an ensemble model of lasso and ridge regressions to predict energy loads. Specifically, the methodology is used to forecast hourly energy demand for up to forty-one hours in the future for the Electric Reliability Council of Texas (ERCOT). The features in the model include previous energy loads and time identifiers such as month, day, and hour of the prediction horizon. The methodology resulted in the creation of forty-one hourly models, each an ensemble of lasso and ridge regression models. The performance of the methodology is measured via out-of-sample data from ERCOT in 2020 against the ERCOT predictions for the same period.