High-precision and accurate runoff simulation is crucial for the management and allocation of water resources, the operation of hydraulic engineering, and the prevention of flood and drought disasters. However, consensus remains elusive regarding effective methods to filter and reshape the impact of numerous external factors on runoff, and theoretical foundations for such processes are also inadequately established. To maximize the accuracy of runoff simulation metrics and better capture the intrinsic hydrological characteristics of runoff, the concept of granular computing from artificial intelligence was drawn on, terrain factors were extracted and their attribute features were optimal-selected based on granulation rules, and a Long Short-Term Memory (LSTM) model incorporating the climate characteristic index (LSTM-new) was developed based on delineated sub-region areas in this study. Finally, a unidirectional feedback framework was proposed, combining process-driven method based on the Variable Infiltration Capacity (VIC) model with a data-driven method using the established LSTM (CouplingVIC-new), to enhance the hydrological process characteristics of the simulated runoff and improve simulation accuracy. The results showed that the average NSE, R2, KGE, and RMSE of CouplingVIC-new during training, validation, and testing periods achieved 0.93, 0.92, 0.91, and 334.86 m3/s, respectively, which increased by 7.29 %、2.97 %、9.73 %、-19.41 % and 13.41 %, 12.19 %, 19.73 %, −46.95 % compared to uncoupled LSTM and VIC. Additionally, the proposed framework effectively captured the interannual variation trend of runoff in all seasons except late spring and summer, though it also overestimated the risk of the occurrence of annual maximum daily peak flow (AMDPF) and total flood volume of annual continuous maximum 5-day (TFAM5D) and their joint variables. The overall results indicated that the scheme of introducing climate characteristic index, based on sub-region division, can more accurately capture extreme runoff in the study area, as well as the variation of seasonal runoff on both intra-annual and interannual scales. Although CouplingVIC-new still had limited ability to capture extreme flow, the structure of extreme value of the output runoff became more robust after unidirectional coupling. This research advances the application of machine learning in hydrological modelling and provide a useful reference for related studies.

中文翻译：

---

通过引入地形因子构建气候特征指数，基于粒度计算的径流模拟

高精度和准确的径流模拟对于水资源的管理和分配、水利工程的运行以及预防洪水和干旱灾害至关重要。然而，对于过滤和重塑众多外部因素对径流影响的有效方法，人们仍然难以达成共识，而且此类过程的理论基础也未充分建立。为了最大限度地提高径流模拟指标的准确性，更好地捕捉径流的内在水文特征，利用人工智能的粒度计算概念，提取地形因子，根据粒化规则对其属性特征进行最优选择，并基于划定的子区域开发了包含气候特征指数的长短期记忆 （LSTM） 模型 （LSTM-new）。最后，提出了一种单向反馈框架，将基于可变渗透能力 （VIC） 模型的过程驱动方法与基于已建立的 LSTM （CouplingVIC-new） 的数据驱动方法相结合，以增强模拟径流的水文过程特征并提高模拟精度。结果表明，CouplingVIC-new 在训练、验证和测试期间的平均 NSE、R2、KGE 和 RMSE 分别达到 0.93、0.92、0.91 和 334.86 m3/s，与未耦合的 LSTM 和 VIC 相比，提高了 7.29 %、2.97 %、9.73 %、-19.41 % 和 13.41 %、12.19 %、19.73 %、-46.95 %。 此外，所提出的框架有效地捕捉了除春末和夏季外所有季节径流的年际变化趋势，但也高估了年最大日峰值流量 （AMDPF） 和年连续最大 5 天总洪水量 （TFAM5D） 及其联合变量的发生风险。总体结果表明，基于次区域划分的气候特征指数引入方案能够更准确地捕捉研究区的极端径流，以及年内和年际尺度上季节性径流的变化。尽管 CouplingVIC-new 捕获极端流量的能力仍然有限，但单向耦合后，输出径流的极值结构变得更加稳健。本研究推动了机器学习在水文建模中的应用，并为相关研究提供了有益的参考。