Analysis of Electric Load Forecasts Using Machine Learning Techniques

Abstract—Load forecasting forms the basis of demand response planning in energy trading markets where smart grid operators look to achieve effective resource planning in servicing customers’ electricity needs. To this end, historical data on customer electricity usage patterns have been deployed extensively in literature, applying state-of-the-art approaches in building predictive models for various use cases. In this research paper, a comparative study of different approaches for modeling time series load data is presented. The aim is to measure the impact and performance of classical statistical approaches, and more recent machine learning approaches on different lengths of historical data. To achieve this, 10 years of electric load data from PJM transmission and weather data were used to deploy day ahead forecasts using ARIMA, SARIMA, Extra Trees regressor and XGBoost. Performance was compared using evaluation metrics such as RMSE, MAPE and MAE. Upon successful completion of experiments, results show that regression-based machine learning models generally showed better results for modeling with lengthier historical data (more than two years). This is especially so as Extra Trees regressor delivered the best RMSE of 0.425 for 6 years data overall. For shorter time series length between 4 weeks to 6 months, SARIMA delivered an RMSE of 0.129, which is the best result overall. Keywords: Electric Load Forecasting, Time Series Forecasting and Machine Learning