Distributed Photovoltaic Lightweight Power Prediction Method Based on Transfer Learning and Meta-Learning

This paper proposes a distributed PV power prediction method that integrates transfer learning, meta-learning, and knowledge distillation to address the difficulties of small sample learning and rapid adaptation in distributed photovoltaic output prediction. Firstly, by calculating the similarity between sites and constructing the adjacency matrix of the graph neural network, the target data is pre trained using transfer learning methods to achieve the transfer and initialization of model parameters. Secondly, model agnostic meta-learning (MAML) is introduced to quickly adapt to changes in photovoltaic station output under different weather conditions through weather scene partitioning, thereby improving the model's generalization ability. Finally, using knowledge distillation, the complex MAML teacher model transfers its knowledge to a lightweight LSTM student model, combined with distributed photovoltaic station output samples, to construct an efficient power prediction model. The case analysis shows that the proposed method can significantly reduce computational costs while maintaining model prediction accuracy, providing an efficient and robust solution for distributed PV power prediction.