Efficient calculation of distributed photovoltaic power generation power prediction via deep learning

Abstract

Distributed photovoltaic (PV) power generation has gained significant support from national policies and has seen rapid development due to its ability to adapt to local conditions, its cleanliness and efficiency, as well as its notable environmental and economic benefits. However, PV power generation is highly susceptible to fluctuations and unpredictability caused by varying weather conditions. Accurate prediction of PV power generation is essential for maintaining grid stability and efficient operation. To improve prediction accuracy, we propose a novel model, PerfCNN-LSTM, which combines a convolutional neural network (CNN) and a long short-term memory (LSTM) network with the Performer self-attention mechanism. This model aims to enhance PV power generation forecasting. By extracting local features from the data, the model further captures global features through the integration of the Performer self-attention mechanism layer. This layer introduces linear random feature mapping, transforming the originally nonlinear attention weight calculation into linear attention, which simplifies the attention process and reduces the model's computational complexity. The output from the Performer layer is directly fed into the LSTM model to generate the final PV power generation prediction. We evaluated the performance of the model across three different datasets using key metrics such as MAE, RMSE, MSE, and R². When compared with six other deep learning models, the PerfCNN-LSTM demonstrates superior prediction accuracy.