Weather forecasting plays a fundamental role in the early warning of weather impacts on various aspects of human livelihood. For instance, weather forecasting provides decision making support for autonomous vehicles to reduce traffic accidents and congestions, which completely depend on the sensing and predicting of external environmental factors such as rainfall, air visibility and so on. Accurate and timely weather prediction has always been the goal of meteorological scientists. However, the conventional theory-driven numerical weather prediction (NWP) methods face many challenges, such as incomplete understanding of physical mechanisms, difficulties in obtaining useful knowledge from the deluge of observation data, and the requirement of powerful computing resources. With the successful application of data-driven deep learning method in various fields, such as computer vision, speech recognition, and time series prediction, it has been proven that deep learning method can effectively mine the temporal and spatial features from the spatio-temporal data. Meteorological data is a typical big geospatial data. Deep learning-based weather prediction (DLWP) is expected to be a strong supplement to the conventional method. At present, many researchers have tried to introduce data-driven deep learning into weather forecasting, and have achieved some preliminary results. In this paper, we survey the state-of-the-art studies of deep learning-based weather forecasting, in the aspects of the design of neural network (NN) architectures, spatial and temporal scales, as well as the datasets and benchmarks. Then we analyze the advantages and disadvantages of DLWP by comparing it with the conventional NWP, and summarize the potential future research topics of DLWP.

天气预报在天气对人类生计各个方面的影响的预警中起着基本作用。例如，天气预报为自动驾驶汽车提供决策支持，以减少交通事故和交通拥堵，而交通事故和交通拥堵完全取决于对外部环境因素（如降雨，空气能见度等）的感知和预测。准确及时的天气预报一直是气象科学家的目标。但是，传统的理论驱动的数值天气预报（NWP）方法面临许多挑战，例如对物理机制的不完全理解，从大量观测数据中获取有用知识的困难以及对强大计算资源的需求。随着数据驱动的深度学习方法在计算机视觉，语音识别和时间序列预测等各个领域的成功应用，已证明深度学习方法可以有效地挖掘时空数据中的时空特征。气象数据是典型的大地理空间数据。基于深度学习的天气预报（DLWP）有望成为传统方法的有力补充。目前，许多研究人员试图将数据驱动的深度学习引入天气预报，并取得了一些初步成果。在本文中，我们在基于神经网络（NN）的架构设计，时空尺度以及数据集和基准方面，对基于深度学习的天气预报进行了研究。