Power Prediction of Photovoltaic Power Generation Based on EMA Decomposition and FITS‐GRU‐MKAN

The electricity generation of photovoltaic systems is influenced by many complex and changing factors, which cause large variations and uncertainties and bring difficulties to the safe and stable operation of power grids. It is important to make accurate medium‐to‐long‐term predictions for reliable grid connection and support for power system planning. In this study, a photovoltaic power prediction model based on exponential weighted moving average decomposition and a neural network structure is presented. The decomposition procedure separates the long‐term trends from the data, removes noise, and improves the forecasting accuracy. Moreover, a parallel network is used to process the decomposed signals in both time and frequency domains, so as to effectively capture the nonlinear characteristics of photovoltaic power generation. Experimental results indicate that the proposed model has excellent predictive ability. Specifically, under sunny, cloudy, and rainy conditions, our model achieves RMSE values of 0.267, 0.639, and 0.418, respectively, achieving lower RMSE and higher R ² than transformer, LSTM, and TimeNet baselines on this dataset. The R ² scores reached 0.979, 0.855, and 0.912, demonstrating superior stability and accuracy for medium‐to‐long‐term forecasting.