Cloud Layer and Precipitation Forecasting via Multi-Scale Gated Temporal and Spatial Attention Network

Abstract

Cloud layer and precipitation forecasting play a crucial role in daily life and decision-making. Most existing deep learning models extract features at a single scale and ignore the correlation between features at different scales in the cloud layer and precipitation data. This hinders the ability to extract multi-scale cloud layer features and precipitation features and further constrains the predictive accuracy of the model. To address these challenges, we propose the multi-scale gated temporal and spatial attention network (MGTSA-Net). This network is designed to capture multi-scale spatiotemporal features in the cloud layer and precipitation data more effectively. As a result, it can improve the accuracy of cloud layer and precipitation forecasting. The core component is the multi-scale temporal gated (MTG) module, which integrates multi-scale convolutions and gated recurrent unit (GRU). To further enhance the model's capability of spatial feature extraction, we integrate a multi-scale spatial attention (MSA) module into the encoder. Experimental evaluations on the WeatherBench dataset demonstrate that the MGTSA-Net outperforms state-of-the-art models in predictive accuracy and computational efficiency.