The evaluation of RNNs-Bayesian models’ training.

Flood forecasting exhibits rapid fluctuations, water level forecasting shows great uncertainty and inaccuracy in small watersheds, and the reliability and accuracy performance of traditional probability forecasting is often unbalanced. This study combined Recurrent Neural Networks (RNN) and Bayesian to establish a comprehensive forecasting model framework of RNNs-Bayesian for the forecasting of water level confidence interval, to achieve both reasonable reliability and accuracy. In the Bayesian structure, weight training was used. In the RNNs, base RNN, Long Short-term Memory (LSTM), and Gated Recurrent Unit (GRU) are used for comparative analysis, and experiments are carried out at the point of the Qixi Reservoir in a small watershed in Zhejiang Province of China. We used the multidimensional disaster data input unit for water level forecasting, including hydrology, meteorology, and geography, and 5 days of time windows for forecasting, The comprehensive reliability of LSTM-Bayesian for 0~102 hours flood reached 92.31%, and the comprehensive accuracy reached 89.15%, and confidence interval forecasting using LSTM is the best method, and achieved reasonable balance of reliability and accuracy. Overall, compound RNN could be a good alternative for forecasting hourly streamflow and extreme water level in small watersheds.

洪水预报存在快速波动特性，小流域内的水位预报则面临不确定性强、精度不足的问题，而传统概率预报的可靠性与精度表现往往难以兼顾均衡。本研究将循环神经网络（Recurrent Neural Networks, RNN）与贝叶斯方法相结合，构建了RNN-贝叶斯综合预报模型框架，用于水位置信区间预报，以实现可靠性与精度的协同优化。在贝叶斯结构中，本研究采用权重训练策略。针对RNN系列模型，本研究选取基础循环神经网络、长短期记忆网络（Long Short-term Memory, LSTM）以及门控循环单元（Gated Recurrent Unit, GRU）开展对比分析，并以中国浙江省某小流域的七溪水库作为试验站点开展验证。本研究采用多维度灾害数据输入单元开展水位预报，涵盖水文、气象与地理数据，并设置5天的时间窗口进行预报。针对0~102小时的洪水预报场景，LSTM-贝叶斯模型的综合可靠性达92.31%，综合精度达89.15%；采用LSTM的置信区间预报方案为最优方法，实现了可靠性与精度的合理均衡。总体而言，复合循环神经网络可作为小流域逐小时径流与极端水位预报的优质替代方案。