Water level forecasting at Hanoi station using transformer-based AI models

Abstract

Flood forecasting is a key task for mitigating flood-related damage in the Red River basin, Viet Nam. A number of reservoirs are currently operated in the Red River system to regulate floods. This study aims to develop a rapid water level forecasting method for Hanoi station under various reservoir operation scenarios and river system conditions, thereby supporting the assessment of multiple operational strategies and providing effective real-time decision-making support. A deep learning model based on the Transformer architecture was applied to forecast water levels at Hanoi station with a 24-hour lead time. The dataset was divided into three subsets: a training set (2015–2022), a validation set (2023), and a test set (2024). Results indicate that the Mean Absolute Error (MAE) remained within an acceptable range, with values of 24.1 cm, 26.1 cm, and 30.7 cm for the training, validation, and testing phases, respectively. The model demonstrates strong capability in capturing historical patterns and achieves high accuracy on the validation dataset, highlighting the effectiveness of Transformer-based architectures for water level forecasting under normal hydrological conditions. For extreme flood events, hydraulic models can be integrated to generate additional data and further enhance forecasting performance.