Daily Peak Load Forecasting Based on Sequential-parallel Ensemble Learning

Load forecasting is a hot research since the birth of the electric power industry. Daily peak load forecasting for consecutive days often plays an important role in the operation, security and stability of power grid. In this paper, consecutive daily peak load model based on sequential-parallel ensemble learning was proposed considering the forecasting error distribution of statistical learning algorithms. Firstly, the decoupling process of generalization error was analyzed, the mechanism of XGBoost sequential ensemble learning, bagging parallel ensemble learning and particle swarm optimization were introduced. Then, the load forecasting based on XGBoost algorithm under bagging framework were deployed, associated with the distribution of bias and variance at the training stage. The PSO was used to cross validate the hyperparameters of XGBoost model. Finally, the effectiveness of the algorithm was verified by Slovenia power load data. The results show that XGBoost algorithm calculates the characteristic importance to aid to select the valuable features. The particle swarm optimization algorithm effectively shorten duration of hyperparameters optimization. And the Bagging- XGBoost algorithm has better forecasting accuracy compared with several conventional models. The forecasting results indicate that sequential-parallel ensemble learning method have higher application value in engineering application for load forecasting. © 2020 Chin. Soc. for Elec. Eng.