Decomposition integration and error correction method for photovoltaic power forecasting

Abstract

Photovoltaic power generation has remarkable environmental benefit, and it is one of the effective means to fundamentally solve environmental problem. An accurate photovoltaic power forecasting system is critical for short-term photovoltaic power station scheduling and power generation plan operation. Aiming at the problems of high randomness and low prediction accuracy of photovoltaic power, a multi-factor forecasting model for photovoltaic power based on fuzzy C-means (FCM), improved variational mode decomposition (VMD) by white shark optimizer (WSO), long short-term memory (LSTM) optimized by artificial gorilla troops optimizer (GTO), error correction (EC), named WVMD-GTO-LSTM-EC, is proposed. First, screen out the main meteorological factors affecting photovoltaic power by Spearman correlation coefficient, and use these factors as clustering feature vector of FCM, divide photovoltaic power by FCM, and obtain similar daily samples. Second, improved VMD by WSO, named WVMD, is proposed to decompose similar daily samples of different meteorological models to obtain intrinsic mode functions (IMFs). Third, LSTM optimized by GTO, named GTO-LSTM, is proposed, and the IMFs combined with meteorological factors data are input into GTO-LSTM for training and testing. Final, EC is established to further increase the forecasting precision and the evaluation index, correlation coefficient R and DM test are introduced to comprehensively evaluate the model. In this paper, to demonstrate the superiority of proposed model, two groups of photovoltaic power data gathered by an Australian solar power generation system are used as experimental data, and the prediction performance is compared with that of eight comparison models. The experimental results show that forecasting value of this model accords with true value to the greatest extent, R = 0.99, and MAPE, MAE, RMSE and R2 can reach 0.02, 0.09, 0.12 and 0.99, respectively, which indicates that its forecasting ability is superior to all comparison models.