Data driven day-ahead electrical load forecasting through repeated wavelet transform assisted SVM model

Abstract

Electrical load forecasting is an integral tool used by the grid operator to operate the smart power network. The information related to the electrical load is a prerequisite towards the effective and optimal operation of the power network with renewable and conventional generation resources. Economical bidding in the energy markets is directing research towards a superior forecasting model. Statistical and machine learning models have been used for electrical load forecasting while considering the electrical load as a time-series signal. This paper proposes a hybrid model that combines the Wavelet Transform (WT) and Support Vector Machine (SVM) features in estimating a regression model for electrical load forecasting utilizing the historical time-series information of electrical load. The WT decomposes the electrical load time-series data into various sub-series. The error contribution in forecasting due to the individual sub-series is estimated using Mean Absolute Error (MAE) in forecasting for each sub-series. The proposed Repeated WT-based SVM model (RWT-SVM) selects the sub-series with the highest MAE for further decomposition through WT. This results in a better forecasting model for the sub-series with the highest MAE, thereby improving the overall forecasting ability of the RWT-SVM model. The superiority of the proposed Repeated WT-based SVM model (RWT-SVM) for electrical load forecasting is justified using various data sets and comparing with some of the existing forecasting models.