Abstract: To address the challenges in power load forecasting, such as the inability of a single kernel function to adapt to complex data characteristics and the low prediction accuracy of long time series, a short-term power load forecasting method incorporating Informer’s Bayesian algorithm for optimisation (BO) and hybrid kernel extreme learning machine (HKELM) is proposed. Firstly, the random forest (random forest, RF) algorithm is used to perform correlation analysis of the load sequence, from which suitable feature matrices are selected and input into the Informer algorithm for modelling, thus improving the prediction efficiency of the load sequence. Secondly, the parameters of the HKELM model are optimized by the BO algorithm to further improve the prediction accuracy of the power load. Finally, the power load forecasting accuracy is improved by adopting a certain province of East China’s power load data from an East China. The simulation results show that the BO-HKELM network incorporating Informer outperforms other models in various error evaluation indexes, in which the MAE, RMSE, and MAPE are 32.43、49.23 and 0.19, respectively, and the prediction accuracy reaches 0.998 7.

Key words: load sequence, random forest, Bayesian optimisation algorithm, hybrid kernel extreme learning machine, Informer