A method for short-term photovoltaic power prediction integrating long short-term memory network, differential transformer, and multi-objective escape algorithm

Abstract

With the rapid development of renewable energy, photovoltaic power generation has become a key part of the global energy transition. Short-term photovoltaic prediction is critical for intra-day real-time power grid dispatching, and enhancing its accuracy is a key research focus. However, existing methods still have limitations in handling complex nonlinear relationships in photovoltaic temporal data. To tackle this, this paper proposes a new model combining Long Short-Term Memory (LSTM), Differential Transformer (DiffTransformer), and Multi-Objective Escape Algorithm (MOESC) for short-term photovoltaic power prediction optimization: Preprocessed data is input into the LSTM-Differential Transformer model, with the Differential Transformer encoder capturing fine-grained temporal changes via optimized multi-head attention and rotary positional encoding, and the LSTM decoder integrating local temporal information for power prediction. Subsequently, Pareto-improved MOESC performs multi-objective optimization on the model’s key parameters (balancing RMSE, MAE, and R²), with the optimal parameters selected from the Pareto frontier. Experiments based on the Guoneng Rixin photovoltaic dataset show that, with user-defined weights (RMSE: 30%, MAE: 30%, R²: 40%), this method outperforms XGBoost, LightGBM, SVR, LSTM, GRU and the unoptimized LSTM-Differential Transformer model in photovoltaic power prediction. It not only can effectively improve prediction accuracy but also exhibits better stability compared with the unoptimized LSTM-Differential Transformer model.