Multi-Source Domain Transfer Network for Short-Term Photovoltaic Power Forecasting

To address the challenges of low prediction accuracy and weak generalization ability due to insufficient historical data in newly built photovoltaic power plants, this paper proposes a novel multi-source domain transfer network with dynamic weighting adjustment and physics-informed constraints. Firstly, a multi-layer long short-term memory network is utilized to extract common temporal features across multi-source and target domains. Secondly, a dynamic weight adjustment module based on Wasserstein distance is designed, combining a time decay factor and an abnormal domain filtering strategy to adaptively allocate source domain contributions. Finally, the physical equations of PV power generation are embedded into the loss function to constrain the physical reasonableness of predictions, enhancing robustness in extreme scenarios. Experiments on multiregional datasets demonstrate that under limited data conditions, the proposed method achieves significant improvements over traditional deep neural networks and single-source transfer models. The study provides a novel high-precision solution for PV power forecasting in data-scarce scenarios.