A study on ultra-short-term photovoltaic power prediction model based on signal decomposition-driven feature reconstruction and Mamba architecture

To address the challenges posed by climate change and mitigate negative impacts on the natural environment, promoting the efficient utilization of renewable energy has become a critical pathway. Among various renewable energy sources, accurate prediction of photovoltaic power generation is of great significance for maximizing the utilization of solar energy resources and continuously improving the efficiency of photovoltaic systems. In view of this, this paper proposes a method that combines signal decomposition-driven feature reconstruction with the Mamba architecture. To verify the superiority of the proposed method, comparative experiments with traditional models are conducted using the photovoltaic dataset from Alice Springs, Australia. The coefficient of determination (R2), mean squared error (MSE), mean absolute error (MAE), and root mean squared error (RMSE) are adopted as evaluation metrics to comprehensively assess the prediction accuracy and inference efficiency of the model. Experimental results demonstrate that the CEEMDAN-Mamba model achieves an average improvement of 1.38% in R2, 2.5% in MSE, 5.07% in MAE, and 5.13% in RMSE compared with the contrast models, while the training speed per batch is increased by 6 times.