LSTM变体模型在径流预测中的性能及其可解释性

基于湘江流域1999—2013年实测水文气象数据，采用LSTM模型和其变体模型研究多个预见期下不同输入变量和不同模型结构对径流预测结果的影响，评估LSTM模型及其变体模型在短期径流预测中的性能，基于排列重要性法和积分梯度法探究了LSTM模型对流域径流预测的可解释性。结果表明：在历史径流输入数据的基础上增加有效的水文气象变量输入，可以明显改善模型的预测效果，输入变量的改变比模型结构的差异对预测结果的影响更大；随着预见期的增大，降水数据的加入对预测效果表现出不同程度的提升，预见期为1 d时，预测结果的纳什效率系数(NSE)提升2.0%,预见期为2～4 d时，NSE提升可达13.6%;降水和历史径流在预测中起着重要的作用，而前期湿润条件与降水事件的共同作用是湘江流域洪水的主要诱发因素；LSTM模型可反映两种不同的输入输出关系，这两种关系对应于近期降雨和历史降雨两种洪水诱发机制。

Performance of variant LSTM models in runoff prediction and their interpretability

Based on observed hydrometeorology data in the Xiangjiang River Basin from 1999 to 2013, the long short-term memory (LSTM) model and its variant models were used to study the influences of different input variables and different model structures with different foresight periods on runoff prediction results, and performance of the LSTM model and its variant models in short-term runoff prediction was evaluated. Then, the interpretability of the LSTM model for runoff prediction of the Xiangjiang River Basin was explored using the permutation importance and integrated gradient methods. The results demonstrate that incorporating hydrometeorological variable data with historical runoff data as the model input can significantly improve the prediction performance of the model, and changes in input variables have a greater impact on prediction results than differences in model structure; with the increase of the foresight period, the addition of precipitation data leads to different degrees of improvement in the model prediction performance, and the Nash-Sutcliffe efficiency coefficient (NSE) of prediction results increased by 2.0% with a foresight period of one day, while the increase in NSE was up to 13.6% with a foresight period of two to four days. Precipitation and historical runoff were found to play critical roles in runoff prediction, while the joint effect of early humid conditions and precipitation events was identified to be the primary factor inducing floods in the Xiangjiang River Basin. The LSTM model can reveal two distinct input-output relationships, corresponding to two flood-triggering mechanisms of recent rainfall and historical rainfall.