Abstract: Existing wind power prediction based on deep learning methods is an indirect prediction using meteorological data as input, and its prediction accuracy depends on the accuracy of meteorological forecasts. However, existing meteorological forecast data generally suffer from low-resolution and unstable models. At the same time, deep learning models are completely data-driven and need guidance from physical laws, making it difficult to improve prediction accuracy further. Therefore, this paper proposes a method that combines refined meteorological factors with physical deep learning. Firstly, numerical weather prediction data is processed using downscaling and multi-model integration techniques to improve meteorological forecast products' low resolution and accuracy issues. Secondly, two physical models are introduced based on wind turbine wake effects and power curves. On the one hand, the physical models are embedded in the neural network loss function as regularization terms to introduce physical constraints in the learning process and construct a physical deep learning network. On the other hand, the physical models are used to generate pre-training samples to address the insufficient observational data, obtaining a pre-training model that supports subsequent supervised learning tasks. Finally, the effectiveness and superiority of the proposed method are verified through simulation analysis of actual data from a coastal wind farm in a certain city.