Deep learning model for flood probabilistic forecasting considering spatiotemporal rainfall distribution and hydrologic uncertainty

How to consider the spatiotemporal distribution of rainfall and hydrologic uncertainty is crucial to enhance flood forecasting accuracy. This study integrates a spatiotemporal dual attention (SDA) mechanism and a recurrent encoding-decoding (RED) structure into a convolutional long short-term memory (ConvLSTM) network, and proposes a novel SDA-ConvLSTM-RED deep learning model for flood probabilistic forecasting. Two probabilistic forecasting methods, quantile regression (QR) and mean-variance estimation (MVE) are incorporated to quantify uncertainty and generate flood probabilistic forecasts. The uncontrolled interval basin between Xiangjiaba and Three Gorges Reservoir (TGR) is selected as case study. The findings demonstrate that the SDA-ConvLSTM-RED model can effectively capture the spatiotemporal rainfall distribution, thereby enhance flood forecasting accuracy. Compared to the benchmark model (LSTM-RED), the SDA-ConvLSTM-RED model performs much better in flood event forecasting, with average Nash-Sutcliffe efficiency coefficients of 0.96 and 0.93 for 24 h and 120 h forecast horizons respectively, particularly in forecasting peak discharge and occurrence time. Both probabilistic forecasting methods significantly outperform than the deterministic forecasts, particularly when the probabilistic forecasts are within the 80 % confidence interval. A further comparison of the evaluation metrics for the two probabilistic forecasting methods reveals that the MVE method is more suitable for shorter-term flood probabilistic forecasts, while the QR method performs better for longer-term flood probabilistic forecasts.