A Novel Day-Ahead Probability Prediction Method for PV Station Based on Improved CNN-Autoformer Network

Accurate prediction of PV power is essential for economic dispatch in power systems. Because of the strong randomness and non-stationarity of PV power sequences, it is challenging to accurately characterize its uncertainties. This paper introduces a short-term PV power probability prediction method based on the CNN-Autoformer network. Firstly, it utilizes convolutional neural networks to extract high-dimensional meteorological features from numerical weather forecasts and establish a mapping relationship between meteorological features and PV power. Secondly, it deeply decomposes PV power sequences based on the improved Autoformer network, incorporating an autocorrelation mechanism to capture the periodicity and trend of PV power. Finally, it combines maximum likelihood estimation with gradient optimization to derive the parameters of the PV power probability distribution through a probability density estimation layer. The effectiveness of the proposed method is verified in a PV station in China.