Short-Term Power Load Forecasting Based on Feature Selection and Optimized Extreme Learning Machine

For the difficulty in determining the feature quantity and the low output stability of the extreme learning machine (ELM) due to randomly generated initial weights and thresholds during load forecasting, it is proposed in this paper that the Gram-Schmidt orthogonalization and Pearson correlation analysis (GSO-PCA) and improved gray wolf optimization (IGWO) algorithm be used to optimize the short-term power load forecasting model of ELM (IGWO-ELM). GSO and PCA are used for optimal selection of two distinct features respectively, and the mean absolute percentage error (MAPE) is based for determination of the optimal feature set. MAPE in GSO-PCA is 1.3%, 0.55% and 0.83% lower respectively compared with that in traditional experience, maximum information coefficient and random forest for feature selection. This verifies its superiority. The Tent chaotic mapping and particle swarm optimization (PSO) algorithm are integrated into the gray wolf optimization algorithm to obtain IGWO algorithm which is tested with two typical test functions and proved with better optimization capability. The IGWO algorithm is used to dynamically optimize initial weights and threshold of the ELM and establish IGWO-ELM short-term power load forecasting model. This established model is then compared with a traditional model in terms of four evaluation indexes, i.e., goodness of fit test coefficient, mean absolute error, root mean square error and MAPE and in line with case study. The simulation results show that these four indexes are 0.997 8, 54.90 kW, 72.02 kW and 1.52% and are siqnificantly better other models. This verifies the effectiveness and superiority of the proposed model. © 2022, Editorial Office of Journal of Xi'an Jiaotong University. All right reserved.