China Differentiable Learning Distributed Hydrologic Model (CDLM) 全国可微分分布式水文模型（CDLM）

The national distributed differentiable hydrologic model consists of three coupled components: a 3-km gridded GR4J differentiable runoff-generation operator, a 1-km gridded Muskingum differentiable routing operator, and a neural-network (NN)-based spatiotemporal dynamic parameter adapter. The differentiable runoff-generation operator converts daily precipitation, air temperature, and potential evapotranspiration into gridded runoff; the differentiable routing operator propagates runoff through channels following the topology of the 1-km river network; and the spatiotemporal dynamic parameter adapter generates parameter fields that vary in space and time from grid attributes and daily hydrometeorological states. Embedding all three components within a single end-to-end neural-network framework allows streamflow errors at basin outlets to be backpropagated layer by layer to the parameter adapter. This enables spatiotemporally adaptive learning of runoff-generation and routing parameters, overcoming the homogenized and static limitations of conventional parameter-calibration methods.

全国分布式可微分水文模型由 3 km 网格 GR4J 可微分产流算子、1 km 网格马斯京根可微分汇流算子和基于神经网络（NN）的参数时空动态适配器等 3 个相互耦合的部分组成。可微分产流算子负责将逐日降水、气温和潜在蒸散发转化为网格产流；可微分汇流算子按照 1 km 河网拓扑对产流量进行河道演进；参数时空动态适配器则根据网格属性以及逐日水文气象状态生成时空变化的参数场。将三者嵌入同一个端到端的神经网络框架内，可实现流域出口断面流量误差向参数适配器的逐级反向传播，进而实现模型产汇流参数的时空自适应学习，破除以往参数率定方法“均一化、静态化”的局限。