Abstract:Short-term precipitation forecasting plays a critical role in the areas of meteorology and hydrology. To address the limitations of traditional forecasting methods in dealing with complex meteorological phenomena and the problem of cumulative errors in image sequence prediction by recurrent neural networks, a short-term precipitation forecasting method called multi-scale attention encoding-dynamic decoding network (MAEDDN) has been presented. It predicts future precipitation by learning the spatiotemporal features of the input data. Within the encoding process, convolutional blocks with spatial and channel attention are utilized for encoding, and a multi-scale fusion module is employed to address the challenge of capturing both small-scale and large-scale information in precipitation distribution simultaneously. In short-term precipitation forecasting, the method effectively addresses weather systems' generation and dissipation. During the decoding process, a dynamic decoding network is put forward to flexibly choose the decoding process according to the learned intensity distribution and change trends derived from past input data. Experiments are carried out by utilizing the precipitation data from the open-source SEVIR dataset, and comparisons are made with the currently best methods reported. The experimental results reveal that: (1) The CSI evaluation metrics for MAEDDN in high-intensity precipitation areas have shown significant improvement, with CSI-160 at 0.3645, CSI-181 at 0.3077, and CSI-219 at 0.2330. These results indicate that the predictive capability of MAEDDN in high-intensity precipitation areas has been significantly enhanced. (2) MAEDDN outperforms other models in terms of the resolution of predicted image sequences. The constructed multi-scale attention encoding captures the complex relationships in meteorological data more effectively, while the dynamic decoding adapts the decoding process based on different scenarios, resulting in more accurate prediction outcomes.

A multi-scale attention encoding and dynamic decoding network designed for short-term precipitation forecasting