Machine Learning for Daily Streamflow Forecasting in the Rhine River Basin: Modeling and Predictive Insights

Accurate prediction of streamflow is crucial for various purposes, such as flood control, dam design and operation, water supply systems, and hydropower generation. Estimating streamflow in a catchment presents challenges due to factors such as chaotic distribution, periodicity in streamflow patterns, and intricate/nonlinear relationships among catchment elements. The limitations of traditional models and the growing availability of time series data on flow rates and relevant weather and climate variables are leading to an increased use of Machine Learning-based models. Among these, neural networks have proven to be highly effective for making accurate predictions. In this research, three different types of Machine Learning (ML) algorithm, Multilayer Perceptron (MLP), Support Vector Regression (SVR), and Random Forest (RF), were employed to forecast the daily streamflow of the Rhine River (Worms catchment). The predictive features at Worms gauging station (Rhine River) encompassed lagged values of streamflow (Qs) from the previous 1, 2, and 3 days, flow rate at the Maxau station (Rhine River) with a single lag-time (Q_m-1), and daily precipitation (P). In this study, the data from 1 January 2013 to 31 August 2021 was employed for building models (training), and the data from 1 September 2021 to 31 October 2023 was used for model validation. The performance of the proposed models in predicting streamflow were investigated using some quantitative metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Nash-Sutcliffe efficiency (NSE), and Percent bias (PBIAS). The results showed that the Maxau flow rate (Q_m-1) and daily streamflow with one day lag (Q_s-1) are the most effective input variables for forecasting streamflow at Worms gaugin station. According to the NSE metric, all models have very good predictive power, but the RF algorithm outperformed the others.