Research on short-term photovoltaic power forecasting method

Owing to the significant time variability, randomness, and volatility inherent in photovoltaic power generation, ensuring the stable operation of photovoltaic power plants presents considerable challenges. Therefore, the accurate prediction of photovoltaic power generation is an important technical means to improve the absorption capacity of photovoltaic power generation and ensure the safety of the power system. This paper introduces a short-term photovoltaic power prediction method based on the Long Short-Term Memory (LSTM) network. The Pearson correlation coefficient is utilized to identify input features with strong correlations. The Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) technique is employed to mitigate the fluctuation of the original sequence, thereby extracting data features at various frequencies. Permutation entropy is used to quantify the complexity of each component, and sequences with similar entropy values are reconstructed to reduce computational complexity. By comparing the prediction outcomes of the LSTM model with those of the Backpropagation (BP) neural network model and the decision tree model, this study employs metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination (R²) to demonstrate the superior dynamic capture and prediction accuracy of the LSTM network for time series data. This approach provides accurate and reliable predictions, effectively supporting electrical scheduling decisions.