Performance Improvement of LSTM-based Deep Learning Model for Streamflow Forecasting Using Kalman Filtering

Prediction of streamflow as a crucial source of hydrological information plays a central role in various fields of water resources projects. While accurate daily streamflow forecasts are essential, predicting streamflow based on limited data is useful for minimizing computational time and supporting flood early warning systems. This study aims to improve Long Short-Term Memory (LSTM) performance by Kalman filter (KF) for streamflow forecasting. For this goal, the simulation has been specified according to the daily streamflow series for 60 years of Dez Dam, located in Iran. We compared the results of simulating the LSTM, LSTM-KF, LSTM-UKF, LSTM-KFS, and LSTM-UKFS models. In addition, the accuracy of the proposed method is evaluated with statistical analysis including Nash–Sutcliffe efficiency (NSE), root-mean-squared error (RMSE), average absolute relative error (AARE) and mean relative error (MAE). The results demonstrate that the LSTM-UKFS model is highly effective in flood forecasting, indicating the promising potential of a simple architecture deep-learning approach for predicting floods, even in the presence of dams in the study area.