Hybrid Load Forecasting Using Gaussian Process Regression and Novel Residual Prediction

Abstract

Short-term electricity load forecasting has attracted considerable attention as a result of the crucial role that it plays in power systems and electricity markets. This paper presents a novel hybrid forecasting method that combines an autoregressive model with Gaussian process regression. Mixed-user, hourly, historical data are used to train, validate, and evaluate the model. The empirical wavelet transform was used to preprocess the data. Among the perturbing factors, the most influential predictors that were recorded were the weather factors and day type. The developed methodology is upgraded using a novel closed-loop algorithm that uses the forecasting values and influential factors to predict the residuals. Most performance indicators that are computed indicate that forecasting the residuals actually improves the method’s precision, decreasing the mean absolute percentage error from 5.04% to 4.28%. Measured data are used to validate the effectiveness of the presented approach, making it a suitable tool for use in load forecasting by utility companies.

Keywords

• energy forecasting
• Gaussian process regression
• machine learning models
• empirical wavelet transform
• residuals prediction