融合机理启发式时间特征工程的城市内涝水位预测研究

Research on Urban Flooding Water Level Prediction Integrating Mechanism-Heuristic Temporal Feature Engineering

摘要: 随着全球气候变化的加剧及城市化进程的迅猛推进，城市内涝已成为制约我国中心城区可持续发展的核心瓶颈之一。传统基于物理机制的水文学模型(如SWMM)虽然在机理描述上具有显著优势，但在实时预警场景中往往面临计算效率低、参数获取难及模型维护成本高等现实挑战。针对这一痛点，研究提出了一种融合机理启发式时间特征工程的异构数理模型预测框架。该框架的核心创新在于将水文物理循环机制转化为可量化的时间特征空间，包括：利用二十四节气标识捕捉地球公转驱动的季节性降雨背景、采用三角函数变换映射日内排水周期，以及通过防泄露滚动统计(Shift-Rolling)和多步滞后构造长短期历史记忆。以某典型山前滨河城市为研究对象，在中心城区关键水位站进行了多模型验证。实验对比了包括集成学习(LightGBM、随机森林)、支持向量机(SVM)及深度时序网络(LSTM、CNNLSTM、Transformer)在内的异构模型库。结果表明，基于机理启发式特征工程的LightGBM模型在测试集上取得了0.9887的纳什效率系数(NSE)，其预测精度和计算实时性显著优于传统物理模型。研究证实，引入蕴含物理规律的时间特征能显著增强数理模型的泛化能力与稳健性，为智慧排水系统的动态风险评估与科学防涝决策提供了高效的技术支撑。

Abstract: With the exacerbation of global climate change and the rapid advancement of urbanization, urban waterlogging has become a major bottleneck restricting the sustainable development of central urban areas in China. Although traditional physics-based hydrological models (e.g., SWMM) possess significant advantages in mechanistic representation, they frequently face practical challenges in real-time early warning scenarios, such as low computational efficiency, difficulties in parameter acquisition, and high model maintenance costs. To address these limitations, this study proposes a heterogeneous predictive modeling framework integrating mechanism-heuristic temporal feature engineering. The core innovation of this framework lies in translating physical hydrological cycle mechanisms into a quantifiable temporal feature space. Specifically, this includes: utilizing the Twenty-Four Solar Terms to capture the seasonal rainfall background driven by the Earth’s revolution; employing trigonometric transformations to map intra-day drainage cycles; and constructing long- and short-term historical memory via data leakage-free shift-rolling statistics and multi-step lags. Taking a typical piedmont riverine city as the case study, multi-model validation was conducted at key water level stations within the central urban area. The experiments compared a diverse library of heterogeneous models, including ensemble learning algorithms (LightGBM, Random Forest), Support Vector Machines (SVM), and deep time-series networks (LSTM, CNN-LSTM, Transformer). The results indicate that the LightGBM model, enhanced by mechanism-heuristic feature engineering, achieved a Nash-Sutcliffe Efficiency (NSE) coefficient of 0.9887 on the test set. Its prediction accuracy and real-time computational efficiency significantly outperformed traditional physics-based models. This study confirms that incorporating temporal features embedded with physical principles can significantly enhance the generalization capability and robustness of data-driven models, providing highly efficient technical support for dynamic risk assessment and scientific waterlogging mitigation decisions in smart drainage systems.