Data-driven multi-step prediction and analysis of monthly rainfall using explainable deep learning

Monthly rainfall prediction is a crucial topic for the management of water resources and prevention of hydrological disasters. To make a multi-step monthly rainfall prediction and discover the primary factors affecting rainfall, this study presents an explainable deep learning approach that integrates four consecutive modules, namely the Gated-Recurrent-Unit-based (GRU-based) encoder module, attention mechanism module, GRU-based decoder module, and expected-gradient-based explanation module, respectively. The first three modules constitute an encoder-decoder with an attention mechanism which could predict monthly rainfall in multiple continuous months in the future, while the explanation module computes attribution values for the input weather and climate features to quantify their importance. A case study is conducted on monthly rainfall data collected from Darwin and Perth, Australia, from Jan 1921 to Dec 2020. The results mainly indicate that: (1) Compared with the baseline methods, the multi-step monthly rainfall predictions made by the encoder-decoder with an attention mechanism show better agreement with the ground truth. (2) For January rainfall in Darwin, the most important weather factors are last January and February rainfall, while the most critical climate index is last February's Southern Oscillation Index, with its high values inhibit while low values promote January rainfall in Darwin. (3) The most significant feature for June rainfall in Perth is last June's solar radiation whose feature value is weakly negatively correlated with its attribution value. The study's significance lies in enhancing the accuracy of multi-step rainfall prediction and providing interpretability through identification of influential factors, which facilitates long-term planning of water resources and deeper understanding of complex weather systems.