Machine Learning-Based Streamflow Prediction for Hydrological Applications: A Case Study with LSTM and Random Forest

Abstract

Flood forecasting plays a vital role in disaster management within water resources engineering. This study evaluates the effectiveness of two data-driven approaches: a Long Short-Term Memory (LSTM) recurrent neural network and a Random Forest (RF) regression model for predicting river discharge using a publicly available hydrological dataset. Historical streamflow data, including lagged flow and precipitation variables, serve as inputs to the models. Performance metrics such as root-mean-square error (RMSE), mean absolute error (MAE), Nash– Sutcliffe efficiency (NSE), and coefficient of determination (R²) are employed for model evaluation. Results indicate that the LSTM model exhibits superior predictive performance, with lower RMSE and MAE and higher NSE and R² values. These findings support the advantage of recurrent neural networks in modelling temporal hydrological patterns.