Forecasting river daily discharge using decision tree and time series methods

Abstract

River floods disrupt communication and transportation networks, damage buildings and infrastructure, destroy agricultural products and livestock, cause capital losses and endanger human life. Accurate and proper flood prediction and forecasting are major challenges in hydrology and water resources management. The aim of this study was to forecast and estimate the daily flows of three rivers in Iran using four tree-based data-mining methods, two ensemble bagging methods and the stochastic time series model Arima (auto-regressive integrated moving average). A comparison of these different methodologies is the main contribution of this work. Five statistical measures were used to evaluate the accuracy of these models based on 4 years of daily discharge flow data. The hold-out method was used to divide the data into training (70%) and testing (30%) sets. It was found that the ensemble tree-based chi-square automatic interaction detector provided the most precise forecasts. The overall results indicate that the data-mining methods of ensemble models and tree-based models improved the average accuracy of the models by 25.0% and 15.5% compared with the stochastic Arima model, respectively, indicating the superiority of their potential in capturing the non-linear behaviour of flow discharges.