Study on the Effect of LSTM Hyper-Parameter Tuning for Prediction of River Discharge

This study employs the long short-term memory (LSTM), one of the widely-used models in the machine learning to predict discharge in some sub-watersheds of the Hanjiang River Basin, to evaluate the accuracy of the prediction with the Nash-Sutcliffe efficiency (NSE). The results show that the hyper-parameters in the LSTM have a significant effect on the prediction of discharge, and the appropriate values of these parameters can obtain NSE of above 0.93 in the study sub-watersheds. The results also show that the accuracy of prediction is sensitive to the number of look back, the dropout rate, and the number of epochs out of the commonly used hyper-parameters. And these hyper-parameters have certain commonality in different sub-watersheds of the same basin. However, the dropout rate, a usual hyper-parameter to prevent overfitting, behaves differently from the ordinary machine learning. It should be set at the lower level to avoid under-predicting large floods.