A novel ensemble electricity load forecasting system based on a decomposition-selection-optimization strategy

Electricity load forecasting exhibits an irreplaceable role in enhancing the dispatching and management efficiency of power systems. However, the majority of existing research neglected the feature extraction of original series as well as interval prediction, which leads to inevitable forecasting bias and insufficient information. To fill the gaps, a novel ensemble system is proposed to realize both point and interval multi-step forecasting results. Specifically, the original series is divided and reconstructed into multi-scale sub-series by the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and the adaptive Lempel-Ziv complexity (ALZC) algorithms, while the input variables are determined by a two-stage feature selection method. Then a sub-predictor selection strategy is employed to select the effective forecasting model for each sub-series, and the improved multi-objective salp swarm algorithm (IMSSA) is utilized to optimize the ensemble point predictions. Further, the point error-based interval forecasting fitted by the forecasted fluctuation properties is conducted for uncertainty analysis. To testify the efficiency of the system, three 30-min electricity load datasets from Australia are employed for experiments. Based on the empirical results, the average mean absolute percentage errors of one-step, two-step, and three-step point predictions on three datasets are 0.2621 %, 0.4512 %, and 0.6305 %, respectively, and the average interval coverage probabilities are 0.9867, 0.9900, and 0.9767 for three datasets of one-step, two-step, and three-step interval predictions under the significance level of 95 %. The results indicated that the system has superior forecasting ability and can supply scientific and comprehensive references for power systems.