Machine Learning for Water Level Prediction in the Chao Phraya River Basin

Abstract

High precision of hydrological prediction is crucial for real–time operation of flood and drought risk mitigation and strategic planning. This study assessed the predictive performances of three machine learning algorithms; Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Deep Neural Networks (DNNs) for water level prediction. Accordingly, the one–day and one–week water level prediction models for six key gauged stations along the Chao Phraya River and its major tributaries were developed. Selecting input features was carried out based on the physical river–reservoir system using past water level, rainfall, controlled reservoir outflow, and upstream discharges with different travel times. The statistical evaluation indicated that both XGBoost and RF with rainfall input robustly outperformed than DNNs, as it strongly achieved higher R² from 0.937 to 0.999 for model training and from 0.743 to 0.995 for model testing and lower MAE, MSE, and RMSE values for all daily prediction scenarios. Among these algorithms, RF demonstrated the superior performance for low water level prediction exhibiting the smallest percentage error of overestimating lying between +0.0088% and +0.9380%. XGBoost, RF, and DNNs algorithms exhibited small average percentage errors for high water level prediction ranging from –2.2696% to +1.1587%. Additionally, daily model can capture the entire testing dataset with high precision than weekly model. Daily predictions provide valuable real–time insights for forecasting water levels during critical flood and drought periods. In contrast, weekly predictions assist in strategic water resource planning to address challenges in diverse hydrological environments.