River Flow Forecasting Using Long Short-term Memory

Abstract

In this paper we propose an LSTM application to hydrological problem, the flow estimation in Jinhe river in China. The LSTM is a very powerful prediction tool that considers temporal dependencies. The model's performance highly depends on the so-called hyperparameters values. To find the best combination of hyperparameters we have used the random search method, that is, testing combinations of randomized hyperparameters until we get the best one. We then compare the built model with a watershed model called GWLF. Using two performances indicators that are the RMSE and Nash coefficient, we found that our model is much better in both calibration and validation period.