Comparative Analysis of LSTM and HEC-HMS Models for Streamflow Prediction: A Case Study of the Upper Turtle River

Streamflow predictions are important in planning water resources and effective hydrological management. Physically based models such as Hydrologic Engineering Center-Hydrologic Modeling System (HEC-HMS) have been used in the past for this purpose, but the use of machine learning models such as long short-term memory (LSTM) has been proven to complement the use of the traditional models. This study compares the reliability performances of HEC-HMS and LSTM models to predict streamflow in an Upper Turtle River catchment. Twenty-nine years daily meteorological data from two gauging stations and discharge data from United States Geological Survey (USGS) station 05082625 were used to conduct this study. Calibration of the HMS model was done using the Soil Conservation Service (SCS) Curve Number loss method and the SCS Unit Hydrograph transform method. Routing was done using the Muskingum method. The LSTM model was pre-processed using hyperparameter tuning, the train–test split ratio was 0.8:0.2, epoch=100 but there was an early-stop function which was added to avoid overfitting. The performance of both models was evaluated using NSE, RMSE, PBias, and MAE matrices. The result shows a better performance by the LSTM model with R2 of 0.79, RMSE = 49.96.