Forecasting the Discharge of the River Sava in Serbia Using Seasonal Auto-Regressive Integrated Moving Average (SARIMA)

Abstract

This study evaluates Seasonal Auto-Regressive Integrated Moving Average (SARIMA) techniques for predicting river discharge up to 12 months ahead. The Sava River, Serbia, was chosen as the test basin because it plays a crucial role in regional water resource management. However, accurate long-term discharge prediction is challenging due to climatic variability and non-stationarity in hydrological series. Monthly discharge series from 1926 to 2022 at the Sremska Mitrovica station were modeled and forecasted using SARIMA and compared with Zero-Order Forecasting (ZOF) and Random Forests. The selected SARIMA model predicted discharge with a normalised Root Mean Square Error of 21% and a normalised Mean Absolute Error of 16% calculated as a percentage of mean observed discharge. This study evaluates both classical stochastic methods (SARIMA) and machine learning (Random Forest) for the Sava and underscores the potential for improved hydrological predictions in the Western Balkans. Future research should incorporate additional covariates and apply these models to diverse river systems to enhance predictive accuracy.