AF-SRNet: Quantitative Precipitation Forecasting Model Based on Attention Fusion Mechanism and Residual Spatiotemporal Feature Extraction

Abstract

Reliable quantitative precipitation forecasting is essential to society. At present, quantitative precipitation forecasting based on weather radar represents an urgently needed, yet rather challenging. However, because the Z-R relation between radar and rainfall has several parameters in different areas, and because rainfall varies with seasons, traditional methods cannot capture high-resolution spatiotemporal features. Therefore, we propose an attention fusion spatiotemporal residual network (AF-SRNet) to forecast rainfall precisely for the weak continuity of convective precipitation. Specifically, the spatiotemporal residual network is designed to extract the deep spatiotemporal features of radar echo and precipitation data. Then, we combine the radar echo feature and precipitation feature as the input of the decoder through the attention fusion block; after that, the decoder forecasts the rainfall for the next two hours. We train and evaluate our approaches on the historical data from the Jiangsu Meteorological Observatory. The experimental results show that AF-SRNet can effectively utilize multiple inputs and provides more precise nowcasting of convective precipitation.

Keywords

• quantitative precipitation forecasting
• attention mechanism
• multimodal fusion
• spatiotemporal prediction