From Local to Regional: Deep Learning Models for Daily Water Discharge Forecasting in a Data-Scarce Basin and Engineered River

Accurate prediction of water discharge (Q) is crucial for effective water resource management and meeting rapidly growing human demands. This paper presents novel data-driven frameworks: local, regional, and regional-local approaches, for daily water discharge estimation in the data-scarce, transboundary Red River basin, spanning China, Laos, and Vietnam. We utilized approximately 60 years of observational data from 11 hydrological gauging and climatic stations located within the Vietnamese portion of the Red River system. Additionally, climate parameters (air temperature, rainfall, and evaporation) from the ERA5 dataset were used to supplement data gaps in areas where direct measurements are unavailable, particularly within China’s territory. Accordingly, we proposed a regional-local modeling framework employing an ensemble Deep Learning (DL) model that combines Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM) networks (CNN-LSTM). This framework achieved the best statistical performance for one-day-ahead forecasting with a 3-day time lag. The water discharge predicted by the CNN-LSTM model exhibited lower variability compared to other standalone and ensemble models, resulting in forecasts that closely matched observed values. Notably, using input data from three days prior yielded the highest prediction accuracy. Our findings also indicated a strong agreement between observed and forecasted water discharge, especially during peak flow periods, with optimal results at 1-day and 2-day lead times. However, prediction accuracy significantly diminished for lead times beyond three days. The proposed model offers significant practical implications, serving as a valuable tool for improving operational water resource management, optimizing reservoir operations, and enhancing early warning systems for flood mitigation across the Red River basin.