Research on physically guided deep learning for spatiotemporal streamflow prediction in data-scarce river basins

High-precision streamflow forecasting is fundamental to water resources management and flood disaster mitigation. However, in data-scarce basins, traditional hydrological models are difficult to calibrate, limiting forecasting accuracy; deep learning methods focus on single-station time-series modeling and do not fully capture basin-wide spatiotemporal correlations. This paper proposes a physics-constrained CNN-GAN （PCG） model for basin-scale daily runoff forecasting. The model employs a convolutional neural network （CNN） to extract spatial features and a generative adversarial network （GAN） to constrain and optimize the runoff field, while embedding physical constraints into the loss function to enhance physical rationality and robustness. In a case study of the data-scarce Nujiang-Irrawaddy River Basin, the PCG model demonstrates better spatial consistency on the basin scale. On the station scale, the Nash-Sutcliffe efficiency （E_NS） and other metrics remain at relatively high levels, with stable performance even at sites under complex hydrological conditions （E_NS > 0.6）. Its overall performance outperforms that of the benchmark models and GloFAS. This study provides a novel method for streamflow forecasting in data-scarce regions that combines high accuracy with physical interpretability.