Abstract: With the development of high-resolution numerical weather models and new-generation Earth observation systems, the amount of data in the field of meteorology is rapidly increasing. The rapid increase of meteorological data provides rich information for weather and climate theoretical research and operational applications, but also brings new challenges to traditional data processing methods and weather analysis and forecasting techniques. Deep learning could extract complex spatiotemporal features from a large amount of high-dimensional spatiotemporal distributed meteorological data with high computational efficiency and transferability, and excellent synergy and flexibility. Deep learning has been widely applied in convective nowcasting, extreme event detection, and improving numerical weather models and their prediction. In this paper, we firstly introduce the main deep learning methods and models currently applied in meteorology, and then discuss the applications of the data-driven deep learning to the theory-driven numerical weather prediction models in detail. Deep learning presents significant impacts on improving data assimilation, parameterization of sub-grid physical processes, and numerical weather model output postprocessing. Meanwhile, the interpretability and uncertainty quantification of the deep learning models are highly desirable and challenging. Developing hybrid models of data-driven deep learning and theory-driven numerical weather models to exploit the synergy between deep learning and numerical weather models present a new way to further improve numerical weather models.