Water levels forecasting for the river basins of Kien Giang in Quang Binh province using Transformer-based approaches

Abstract Precision and timely forecasting of water levels holds critical significance across diverse applications such as flood prediction, hydroelectric power management, and environmental surveillance. While conventional recurrent neural network (RNN)-based techniques have been extensively employed for this purpose, recent strides in long-term time-series forecasting have ushered in transformer-based models, showcasing remarkable advancements in predicting time-series data. This study delves into the utilization of transformer-based models for water level prediction, specifically concentrating on the Nhat Le River Basin. Through numerous experiments encompassing different scenarios and employing various model architectures, we offer detailed analyses of the model's predictive capabilities. Demonstrating consistent superiority over traditional RNN-based methods, the transformer-based models excel across various evaluation metrics, including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). Notably, these models exhibit outstanding precision in forecasting flood peaks, consistently maintaining errors below 0.02 meters. The resilience and scalability of transformer-based models position them as highly promising for achieving accurate water-level predictions in practical, real-world applications.