Abstract
Precipitation plays a critical role in the global water cycle and significantly impacts various aspects of human life. Although deep learning models have advanced the development of precipitation forecasting, most existing models rely on a single data source, neglecting the multifactorial nature of precipitation formation. To deal with this issue, a novel spatiotemporal transformer network with multivariate fusion for short-term precipitation forecasting abbreviated as ST-MFTransNet is proposed. Specifically, firstly, a multivariate fusion module based on Omni-dimensional Dynamic Convolution is designed to efficiently combine information from diverse meteorological variables such as temperature, humidity, and wind speed. Secondly, based on the multivariate fusion module, an encoder-decoder framework comprising a Transformer and a multi-scale convolution module is constructed to extract spatiotemporal features from the fused data. To validate the effectiveness of ST-MFTransNet, Experiments were conducted using multiple meteorological variables, accumulated over the past 12 and 24 hours, to forecast accumulated precipitation for the subsequent 12-hour and 24-hour periods. Experimental results demonstrate that ST-MFTransNet achieves enhancements in POD and CSI of 21.2% and 18.4%, respectively, relative to VIT, for the 12-hour accumulated precipitation forecast at a threshold of 15. For the 24-hour forecast, the improvements in POD and CSI are 10.3% and 10.7%, respectively, at the threshold of 25.