Prior-Guided gated convolutional networks for rainstorm forecasting

Abstract

Accurate rainstorm forecasting is crucial for the sustainable development of human society. Recently, machine learning-based rainstorm prediction methods have shown promising results. However, these methods often fail to adequately consider the prior knowledge of rainstorms and do not explicitly account for the dynamic spatio-temporal patterns of rainstorm events. This study introduces a novel end-to-end prior-informed rainstorm forecasting model that incorporates both fundamental physical priors and the spatio-temporal development patterns of rainstorms. The model utilizes a gated convolutional encoder-decoder network to effectively represent the spatio-temporal patterns of rainstorm events. A key component of the representation network is the Substantial Derivative-GuIded gated convolutional Unit (SDGiU), which updates latent states under the constraints of physical priors. Additionally, an integrated loss function is designed to minimize reconstruction errors on multiple scales and facilitate the generation of forecasts that reproduce the actual spatio-temporal patterns of rainstorm formation, development and dissipation. Experimental results on two reanalysis datasets show that the proposed forecasting model outperforms competing state-of-the-art baselines by at least 19.7% (15.0%) in overall Critical Success Index (Heidke Skill Score). Qualitative analysis indicates that the proposed model can generate predictions that are both physically consistent and spatially-temporally coherent.