Downstream Water Level Forecasting Using LSTM Model for Flood Warning in Pak Phanang District, Nakhon Si Thammarat Province

Abstract

This study aims to develop a deep learning model based on Long Short-Term Memory (LSTM) networks to forecast downstream water levels for flood warning purposes in Pak Phanang District, Nakhon Si Thammarat Province, southern Thailand. The study area is located downstream of the Uthokvibhajaprasid Watergate, a key structure in the Pak Phanang Basin Development Project. During the monsoon season, water releases from the upstream catchment, combined with rising sea levels, often cause flooding in low-lying areas along both sides of the lower Pak Phanang River, which lies downstream of the gate and is vulnerable to recurring flood events.

The model was developed using input variables including upstream water level (Upper_WL), reservoir outflow (Outflow_MCM), daily rainfall (DailyRain_mm), and a one-day lag of the downstream water level (Lower_WLT1). The data were arranged as sequences with a look-back period of 30 days and trained using a three-layer LSTM network with dropout regularization and early stopping to prevent overfitting.

Model evaluation results showed high accuracy, with the coefficient of determination (R²) reaching 0.8368 for the training set and 0.7264 for the testing set. The root mean squared error (RMSE) was 0.1171 and 0.1331, respectively. The comparison between observed and predicted values demonstrated that the model effectively captured trends and fluctuations in water levels, especially during periods of rapid change.

This study highlights the potential of LSTM models to serve as a decision-support tool for real-time flood early warning systems in coastal lowland areas. Specifically, the model can support municipal-level flood preparedness in Pak Phanang Municipality, which is situated alongside the lower Pak Phanang River and lies directly downstream of the Uthokvibhajaprasid Watergate. The integration of this model together with 7-day water forecast data and real-time water monitoring system will help to enhance the 7-day advance warning capability can significantly enhance proactive flood management and reduce the impacts of flood disasters.