统计与数据科学学院知识库

流域水文过程的复杂性和多样性使得径流预测成为水资源管理与规划中的关键科学问题之一。近年来，随着对流域管理需求的不断提高，精准的径流预测逐渐成为支持水资源系统科学决策的重要技术措施。论文以黄河上游流域作为对象，从时间和空间两个维度入手，系统揭示了径流的时空演变变化特征，为后续构建预测模型提供了合理的依据，基于图神经网络与门控循环单元建立径流预测模型，为解决复杂流域管理中的径流预测问题提供了新的思路和参考。

首先，系统分析了流域内各站点径流量的时空演变特征，揭示了时间维度上的变化趋势和不同站点之间的空间依赖关系。这些结果为构建水文预测模型提供了科学依据，其中径流量的长期增加趋势为模型提供了时间变化的参考，而站点间的空间自相关性则为模型设计提供了空间依赖关系的基础。

其次，基于时空特征分析结果，设计了一种结合静态与动态空间信息的深度学习模型SDGTN。该模型利用图卷积网络（GCN）提取静态空间信息，图注意力网络（GAT）提取动态空间信息，利用多头注意力机制自适应融合两类空间特征，并结合门控循环单元（GRU）来捕获时间信息。通过与其他时序模型、空间模型和时空模型对比，SDGTN在不同预测步长下的各项评价指标上均表现出显著优势，验证了模型在预测精度的优势。

最后，通过DM检验、消融实验和不同图结构的对比分析，进一步验证了所提模型的优越性和鲁棒性。DM检验从统计学角度显著验证了模型的优势；消融实验证明了模型各模块的有效性；与其他图结构和构建方式的对比分析则证实了基于相关系数构建有向图的合理性。为水资源管理中的时空预测提供了新的思路。

The complexity and diversity of hydrological processes in river basins render runoff prediction one of the critical scientific challenges in water resource management and planning. In recent years, as the demand for effective watershed management has grown, accurate runoff prediction has increasingly become an essential technical approach for supporting scientifically sound decision-making in water resource systems. This study focuses on the upper reaches of the Yellow River, examining runoff characteristics from both temporal and spatial dimensions to systematically reveal its spatiotemporal evolution. These findings provide a robust foundation for constructing predictive models. Leveraging graph neural network (GNN) and gated recurrent unit (GRU), the study proposes a runoff prediction model, offering novel insights and references for addressing runoff prediction in complex watershed management.

First, the study systematically analyzes the spatiotemporal evolution characteristics of runoff at various monitoring stations within the basin, uncovering temporal trends and spatial dependencies among the stations. These results establish a scientific basis for constructing hydrological prediction models. Specifically, the long-term increasing trend in runoff serves as a temporal reference for model development, while the spatial autocorrelation among stations underpins the design of spatial dependencies in the model.

Second, based on the analysis of spatiotemporal characteristics, the study develops a deep learning model, Static and Dynamic Graph Temporal Network (SDGTN), which integrates static and dynamic spatial information. The model employs graph convolutional network (GCN) to extract static spatial features and graph attention network (GAT) to capture dynamic spatial features. A multi-head attention mechanism adaptively fuses these spatial features, while gated recurrent unit (GRU) are used to capture temporal information. Comparative experiments with other temporal, spatial, and spatiotemporal models demonstrate that SDGTN exhibits significant advantages across various evaluation metrics at different prediction horizons, confirming its superior predictive accuracy.

Finally, the study further validates the robustness and efficacy of the proposed model through Diebold-Mariano (DM) tests, ablation experiments, and comparative analyses of different graph structures. The DM tests statistically confirm the model's predictive superiority; the ablation experiments demonstrate the effectiveness of each module within the model; and the comparative analysis of graph structures substantiates the rationality of constructing directed graphs based on correlation coefficients. Collectively, these findings present a novel approach to spatiotemporal prediction in water resource management.