Application of the ResNet-Transformer Model for Runoff Prediction Based on Multi-source Data Fusion

Accurate runoff prediction is a crucial component of flood control and water resource management, both of which are vital for ensuring regional safety and sustainable development. However, obtaining reliable forecasts faces numerous challenges due to complex terrain, variable topography, climate-change impacts, and the nonlinear coupling of ecological processes. Additionally, previous studies often rely exclusively on a single type of hydrological data, for example, datasets that include only point-scale rainfall measurements or only a single downstream discharge series, which fails to fully reflect the watershed’s complex hydrological processes. Moreover, traditional physical models have inherent limitations in capturing nonlinear relationships and complex spatiotemporal features, hindering their ability to precisely model runoff variations. To address these issues, this study focuses on the Jinsha River Basin in China, exploring the application of multi-source data fusion combined with a hybrid deep-learning model for runoff prediction. The study integrates meteorological station observations, digital elevation model (DEM) data, and runoff measurements from multiple hydrological stations across the basin and innovatively applies the ResNet-Transformer model that merges convolutional residual networks (ResNet) with self-attention mechanisms (Transformer) to achieve high-accuracy runoff forecasts. During the research process, historical runoff data from upstream hydrological stations were incorporated into downstream station datasets through correlation analysis, significantly improving the model’s predictive accuracy. The experimental results demonstrate that this approach delivers superior performance across multiple forecast horizons, particularly in capturing extreme runoff values, and consistently outperforms traditional models. Specifically, the ResNet-Transformer attained a mean Nash–Sutcliffe efficiency (NSE) of 0.992 in the Jinsha River Basin—approximately 13.8% higher than the 0.872 NSE achieved by a benchmark LSTM model. These findings not only open new pathways for runoff prediction in large river basins but also provide robust scientific support for water resource management.