Streamflow Prediction with Long Short-Term Memory Neural Networks Using High-resolution Meteorological Estimates. A Case Study in Southern Mexico

Streamflow simulation is an important aspect for water management and risk assessment. This task requires good-quality meteorological data; however, observations are not easy to obtain. This study evaluates the use of Long Short-Term Memory Recurrent Neural Networks (LSTM) in streamflow simulation, and its results were compared with the conceptual hydrological model GR4J. The study site is the Papagayos River basin, located in southern Mexico. Meteorological data from five ground stations as well as three high-resolution gridded meteorological datasets were selected. Results show that the estimates of precipitation and temperature have important differences in terms of correlation, spread, and mean values. The streamflow simulations were compared in terms of the Nash Sutcliffe Efficiency Coefficient, and results show a better performance of LSTM over GR4J. In addition, LSTM better simulate low and medium flows which are overestimated by GR4J. All in all, the use of LSTM provides reliable discharge simulations and can be used as a tool for risk evaluation in basins with scarce observed meteorological data.