Multi-step-ahead electricity load forecasting using a novel hybrid architecture with decomposition-based error correction strategy

Abstract

In this study, a novel architecture combining a hybrid learning paradigm and an error correction strategy is presented for multi-step-ahead electricity load forecasting. The detail of the proposed architecture is provided as follows: (1) a novel hybrid learning paradigm based on complementary ensemble empirical mode decomposition (CEEMD) and backpropagation (BP) neural network improved by particle swarm optimization (PSO-BP) is developed for preliminary prediction of the electricity load; (2) an error prediction approach based on variational mode decomposition (VMD) and PSO-BP is established for prediction of the subsequent error; (3) calibrate the preliminary prediction values using the forecast results of the error prediction model. Specifically, in the error correction process, the original data series is separated into three subsets to generate a reasonable historical error series used for establishing the error prediction model. Two case studies based on the data of PJM and Ontario electricity markets are presented and investigated to assess the effectiveness of the proposed architecture. The evaluation results demonstrate that the proposed architecture can yield results in higher accuracy than other benchmark models considered in this study.