Short-Term Photovoltaic Power Generation Forecasting with Dynamic Decomposition and Digital-Analog Fusion

Accurately predicting photovoltaic (PV) power generation is crucial for energy companies to plan and conduct safety assessments. However, existing model-based methods are limited by low prediction accuracy, while data-driven methods also suffer from low accuracy due to the difficulty in extracting time series features from raw data. To address these issues, we developed a method based on improved moving average decomposition, which achieves short-term PV power generation prediction by integrating Unscented Kalman Filtering and an extended long short-term memory network (xLSTM). First, we used an exponentially weighted moving average method to decompose the data into seasonal and trend components, specifically modeling the trend component using a double exponential model. Second, the xLSTM consists of two improved LSTM modules, sLSTM and mLSTM, which enhance storage decision-making capabilities and storage capacity, respectively, and enable parallelization. Finally, through PV power generation prediction experiments, the results demonstrate that our method significantly outperforms existing methods in prediction accuracy, with an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> prediction result reaching 97%.