融合水文模型与深度学习的青海湖流域径流模拟

Integrating Deep Learning with HydrologicalModels for Runoff Simulation in the Qinghai Lake Basin

摘要: 本文聚焦我国青海湖流域的水文过程，基于多年气象和水文动态数据，发展了一种融合概念性水文模型FLEX (FluxExchange)和门控循环单元(Gated Recurrent Unit, GRU)的混合模型对流域内最大支流布哈河的逐日径流进行了模拟和预测。在构建混合模型中，采用了三种策略提升模拟精度：引入差分进化自适应算法DREAM_(zs)反演水文参数优化FLEX模型；采用变分模态分解(VMD)提取径流数据的信息和特征；利用麻雀搜索算法(SSA)优化深度学习GRU的参数。研究将FLEX模型的模拟结果连同气象数据一起作为神经网络的输入，从而构建了FLEX-VMD-SSA-GRU混合模型。同时，探讨了不同的气象输入条件对模拟结果的影响和贡献：基于7个主要气象要素，由少及多设置了14组输入情景模拟。最后通过SHAP对深度学习方法的结果进行分析，揭示了气象变量对径流长期趋势的贡献和重要度。

Abstract: This paper focuses on the hydrological processes of the Qinghai Lake Basin in China. Based on meteorological and hydrological data, this research developed a hybrid model that integrates the conceptual hydrological model FLEX (FluxExchange) with the deep learning model GRU (Gated Recurrent Unit) to simulate and predict the daily runoff of the Buha River, the largest tributary in the basin. To enhance the simulation accuracy of the hybrid model, three strategies were employed: The DREAM_(zs) was used to invert hydrological parameters and optimize the FLEX model. Variational Mode Decomposition (VMD) was applied to extract hidden information and features from the runoff time series. The Sparrow Search Algorithm (SSA) was used to optimize the parameters of the GRU neural network. By placing the FLEX model in the first layer of the neural network, this research constructed the FLEX-VMD-SSA-GRU hybrid model. This paper also explored the influence and contribution of different meteorological input conditions on the simulation results by creating 14 input scenarios with varying numbers of the seven main meteorological elements. Finally, SHAP (Shapley Additive Explanations) was used to analyze and reveal the contribution and importance of each meteorological variable to the long-term runoff trend.