Comprehensive Data-Driven Modeling for Water Level Forecasting in the Nan River Basin, Thailand

Accurate water level forecasting plays a critical role in flood management and water resource planning, particularly in river basins prone to extreme hydrological events. This study presents a comparative analysis of three predictive data-driven modeling approaches, Artificial Neural Networks (ANN), Extreme Gradient Boosting (XGBoost), and Linear Regression (LR), to test the forecast accuracy of water levels at the N. 67 gauging station on the Nan River, Thailand. Using a 10-year dataset of upstream water levels and dam discharge data, multiple input scenarios were developed based on different time lag structures, incorporating daily data and the inclusion or exclusion of target station water level data. Among the models, the ANN consistently demonstrated superior performance, achieving an R² of 0.91 under optimal conditions, particularly for a 3-day ahead forecast. The 7-day model performed well, and the 15-day ahead model decreased performance. The study also revealed that while ANN and XGBoost models effectively learn from complex temporal patterns, simpler models, such as LR, can also perform well. Nevertheless, LR was less effective in predicting peak water levels. These findings underscore the value of selecting appropriate input features and model architectures to maximize forecasting reliability and highlight the strategic role of local station data in enhancing predictive accuracy for flood-prone regions.