Runoff simulation of the upper Jinsha River Basin based on LSTM driven by elevation dependent climatic forcing

Abstract

The upper Jinsha River has seen increased variability of stream runoff under the global warming since 1990 and extreme flood events with a 100-year recurrence period have occurred in recent years with flood peaks double or triple that of its normal annual mean flow, which has led to challenges to the utilization of water resources and reservoir operation in the basin. The upper Jinsha River Basin covers a large area but has few observation stations. Compared to the traditional models, the single objective runoff simulation based on machine learning (ML) model has shown advantages in forecasting floods, but the research on runoff prediction of ML model for large rivers originated in alpine mountains is insufficient. In this study, the long short-term memory (LSTM) neural network model was used to simulate the annual runoff process in the upper Jinsha River and the model was driven by daily precipitation, mean temperature, and snow cover area extracted from the 500 m elevation bands of GPM, ERA5-Land, and MODIS snow cover products. The model was built with the runoff observation data as the objective. An ensemble model driven by daily means of all above parameters of the whole basin was also built and compared with the LSTM model. Both models used data from 2001-2014 for training and 2015-2019 for validation. The results show that the Nash-Sutcliffe efficiency (NSE) of the two models was greater than or equal to 0.80 within 15 days lead-time, the models had similar NSE in adjacent lead-times, and the NSE decreased to about 0.70 for the lead-times of 25 and 30 days, which indicates that the runoff simulation results of the two models are reasonable at the 30 days and shorter lead-times. Better results of runoff simulation were generated by the LSTM model driven by the vertical zonation data for the 1-5 days lead-times as compared to the ensemble model. The advantage of the vertical zonation data-driven model reduced for the 7-13 days lead-time and the simulation results are equivalent for the 15-30 days lead-time. The vertical zonation data-driven model was superior to the ensemble model in simulating flood season runoff. In general, the runoff simulation accuracy of the vertical zonation data-driven model is the highest at the 3 days lead-time, especially for spring and summer floods. We conclude that the developed model driven by the elevation zonation data can provide reliable prediction of floods, which can provide a reference for the operation of the downstream cascade hydropower stations of the middle Jinsha River. However, the improvement of the ML model for extreme spring floods should still be an important direction in future research.