A comparative study between time series and soft computing models for river discharge forecasting

Abstract River discharge forecasting plays a critical role in sustainable water resource management. This study evaluated six predictive approaches: three machine learning models—Support Vector Regression (SVR), Random Forest (RF), and K-Nearest Neighbors (KNN); a deep learning model—Long Short-Term Memory (LSTM); and two time-series models: Contemporaneous Autoregressive Moving Average (CARMA) and CARMA with Generalized Autoregressive Conditional Heteroskedasticity (CARMA-GARCH). Eight hydrometric stations on the Kashkan River Basin in western Iran were selected as case studies to evaluate model performance. Two scenarios were evaluated: the first involved predicting river discharge while accounting for a lag time within the data itself, and the second focused on forecasting river discharge at any station based on upstream hydrometric station data. Error metrics included Root Mean Square Error (RMSE), Coefficient of Determination (R 2 ), and the Nash–Sutcliffe Efficiency (NSE). Results indicated that time series models and machine learning methods effectively forecast river discharge. The best results in machine learning models were achieved with RF models for both scenarios, with average RMSE values of 4.6 m 3 /s and 4.1 m 3 /s, while the CARMA-GARCH was the best time series model and had RMSE values of 2.5 m 3 /s and 2.48 m 3 /s.