Multi-step Prediction of Photovoltaic Power Based on Multi-view Features Extraction and Multi-task Learning

Accurate multi-step prediction results of photovoltaic (PV) power have important guiding significance for the scheduling optimization of power grid. To solve the problems of insufficient prediction accuracy caused by insufficient feature extraction of historical data and ignoring the correlation between multi-step prediction values, a multi-step prediction method of photovoltaic power based on multi-view feature extraction and multi-task learning was proposed. Firstly, in order to obtain rich and comprehensive feature information, feature extraction of input data is carried out from time series, local, and global viewpoints. Subsequently, the multi-step PV power prediction task is transformed into multiple single-step PV power prediction sub-tasks, and multi-step PV power prediction is carried out by using the multi-task learning model based on the attention mechanism and expert network to realize the full utilization of the correlation between multi-step prediction values. Lastly, an improved dynamic weight average method is proposed to adaptively optimize the loss weight to further improve the performance of the model. Experimental results show that the proposed method can be adopted to effectively improve the accuracy of photovoltaic power multi-step prediction. © 2024 Science Press. All rights reserved.