A temporal–image parallel hybrid solar radiation–wind speed–green hydrogen production potential prediction model based on federated learning and rolling real-time decomposition

Green hydrogen, essential for carbon neutrality, is primarily produced through water electrolysis using photovoltaic and wind energy. However, the highly fluctuating nature of solar radiation and wind speed complicates green hydrogen production forecasting. In this study, a multi-modal solar radiation–wind speed hybrid prediction model is proposed based on federated learning and rolling real-time decomposition to overcome key limitations: information leakage, single-model architecture, and inefficient parameter tuning. Jump plus AM-FM mode decomposition (JMD) with a rolling real-time mechanism is first applied to generate regular subsequences without future information leakage. To overcome the single-model architectures limitation, a multi-modal prediction framework is used to integrate temporal- and image-based prediction. A parallel dual-predictor model structure with attention-based feature fusion is employed in the temporal prediction to achieve complementary advantages. The image prediction converts data into two-dimensional images for prediction to enhance short-term variations and abrupt changes, offering complementary spatial information for temporal prediction. To address parameter tuning challenges, the Dream Optimization Algorithm (DOA) optimizes both models' parameters and output weights. The prediction results were used to calculate the green hydrogen production potential. Federated learning ensures data privacy during distributed training. Performance was validated through comparative experiments and statistical analysis across four wind–photovoltaic power stations. Compared to the random forest (RF), the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) decreased by 78.39 % and 76.88 % on average, demonstrating the model's superior predictive accuracy. This approach offers an innovative, scalable and privacy-preserving forecasting paradigm for optimizing wind–photovoltaic systems and hydrogen production.