Runoff simulation in Shiyang River Basin based on differentiable parameter learning
摘要
纯数据驱动的深度学习模型在降雨径流模拟中具有良好性能,但同时也存在可解释性差等问题.基于可微分参数学习的水文模拟有机融合了深度学习和机理过程两类模型的优点,可兼顾模拟精度和可解释性,但该方法是否适用于我国西北寒区流域尚有待验证.本研究选择石羊河流域三个典型寒区流域,基于可微分的HBV(Hydrologiska Byråns Vattenbalansavdelning)水文模型,对可微分参数学习方法在西北寒区流域的适用性进行评估.该方法把参数化网络与可微分HBV模型串联在一起,前者根据流域静态属性和动态气象因子推测HBV模型的参数.在利用径流数据优化模型参数时,损失函数的梯度可通过可微分HBV水文模型反向传播至参数化网络,从而可以利用径流数据优化参数获取网络的权重.在石羊河三个流域的径流模拟案例中,纳什效率系数均超过0.63,整体略高于HBV模型和纯数据驱动模型.空间外推实验进一步表明,dPL_HBV模型在以两个流域训练并外推至剩余流域的实验中整体优于HBV模型和LSTM模型,显示出较好的跨流域泛化能力.模拟的蒸散发量与GLEAM数据产品的相关性超过0.8,表明可微分参数学习在我国西北寒区流域具有良好的适用性和应用潜力.
Purely data-driven deep learning models have demonstrated good performance in rainfall-runoff simu-lation because of their strong ability to capture complex and nonlinear relationships between meteorological driv-ing factors and hydrological responses.However,such models generally lack explicit representations of hydro-logical processes and have weak physical interpretability,which may limit their reliability and spatial transfer-ability when applied to basins with hydroclimatic conditions significantly different from those represented in the training region.Differentiable parameter learning-based hydrological simulation provides a new approach to ad-dressing the above problems.By integrating the strong data learning capability of deep neural networks with the physically interpretable structure of process-based hydrological models,this method is expected to improve phys-ical consistency,interpretability,and generalization ability of models while ensuring simulation accuracy.Nev-ertheless,whether this framework can be effectively applied to cold-region basins in Northwest China,particu-larly under the combined effects of snow accumulation and melting,seasonal temperature variations,evapotrans-piration,soil moisture dynamics,and complex water balance processes,remains insufficiently investigated.In this study,three representative cold-region basins in the Shiyang River Basin of Northwest China were selected as the study areas.A differentiable parameter learning hydrological simulation framework based on the differen-tiable HBV(Hydrologiska Byråns Vattenbalansavdelning)hydrological model,referred to as dPL_HBV,was developed by coupling a neural-network-based parameterization module with the differentiable HBV model.Ad-ditionally,the applicability of dPL_HBV to runoff simulation in cold-region basins of Northwest China was sys-tematically evaluated.The parameterization network estimated HBV model parameters from basin static attri-butes and dynamic meteorological driving factors,while the differentiable HBV model explicitly represented ma-jor hydrological processes and generated simulated runoff.During model training,the discrepancy between sim-ulated and observed runoff was used to define the objective function,and the gradients of the loss function were propagated backward through the differentiable HBV model to the parameterization network.Consequently,ob-served runoff data directly constrained the neural network and optimized its weights through gradient-based learn-ing,enabling data-driven learning of hydrological model parameters while maintaining the original physical pro-cess structure of the HBV model.The performance of dPL_HBV was systematically compared with that of the conventional HBV model and a purely data-driven long short-term memory(LSTM)model.The results showed that the Nash-Sutcliffe efficiency coefficients(NSE)of dPL_HBV exceeded 0.63 in all three study basins,with its overall performance slightly better than that of both the traditional HBV and LSTM models.To further exam-ine the spatial transferability of the models,spatial extrapolation experiments were conducted by training the models using data from two basins and transferring them to the remaining basin that was not included in the train-ing.Under these conditions,dPL_HBV generally outperformed both HBV and LSTM models,indicating stron-ger cross-basin transferability and generalization ability.In addition to runoff simulation,the Global Land Evap-oration Amsterdam Model(GLEAM)evapotranspiration product was used to assess the physical consistency of internal hydrological variables simulated by the model.The correlation coefficients between the simulated evapo-transpiration from dPL_HBV and the GLEAM product exceeded 0.80,indicating that the model can reproduce not only runoff dynamics but also important components of the basin water balance.Overall,the findings dem-onstrate that differentiable parameter learning can effectively combine the predictive capability of deep learning models with the process interpretability of conceptual hydrological models.It demonstrates considerable applica-bility in cold-region basins of Northwest China and has great potential for applications in cold-region hydrologi-cal simulation,runoff prediction in ungauged basins,and regionalization of hydrological models across basins.