Improving multi-scale short-term precipitation forecasting through frequency domain analysis and attention mechanisms

Existing data-driven models still exhibit shortcomings during short-duration precipitation events, with forecasts lacking multi-scale characteristics and overall intensity predictions tending to be low. Concurrently, improving forecasting accuracy often requires more computational resources. This study quantifies the multi-scale distribution characteristics of precipitation in the frequency domain. Integrating a cross-channel multi-scale attention mechanism with the TransUNet hybrid architecture further enables the capture of nonlinear relationships in precipitation's temporal evolution. This has led to the development of a novel short-term precipitation forecasting model. It reduces the number of parameters while enabling the rapid and precise prediction of hourly precipitation distributions over the next 0-24 hours. Experimental results demonstrate that, compared to four commonly used deep learning models, WE_TransUNet achieves optimal forecasting performance with an average threat scores (TS) improvement of 12 % to 0.447 and an average RMSE of just 0.281. The EMA demonstrates superior capability to capture nonlinear relationships in time-series evolution across different scales compared to conventional CBAM modules, thereby extending the effective forecasting horizon. The visualisation of the results indicates that the wavelet transform facilitates the preservation of high-frequency, localised severe convective phenomena in forecast outcomes. The proposed model can effectively extract multi-scale information from historical precipitation data using a small number of parameters, thereby mitigating the ambiguity inherent in deep learning models. This design offers a novel optimisation approach for short-term precipitation forecasting.