A dynamic ensemble method for residential short-term load forecasting

Abstract

In this paper, a dynamic ensemble method is proposed to forecast the residential short-term load accurately. The main idea is to utilize the state-space approaches to dynamically adjust the weight coefficients used to combine the base models. The dynamic ensemble method is divided into two stages. In the first stage, the least square method is utilized to estimate the weight coefficients. In the second stage, the particle filter is employed to exclude the estimation error in the first stage and update the weight coefficients dynamically to improve the ensemble accuracy. Then, three heterogeneous models (i.e., vector autoregression model, Gaussian process regression model, and the long short-term memory neural network) are employed as the base models and combined based on the weight coefficients to forecast the residential load. The numerical tests are conducted considering two scenarios in two public datasets to investigate the performance of the proposed method. The first scenario assumes the base models are accurate and stable, while the second scenario considers the dynamics of the base models. Root mean square error, mean absolute error, and mean absolute percentage error are used to evaluate the performances. These indicators show that the proposed method outperforms in both scenarios and datasets.