Scholarlink

<p>With the increasing demand for urban flood hazard mitigation systems, machine learning has become an important approach to improve the rapid-response capability for pollutant transport and dispersion at the city scale. However, pollutant monitoring data are still incomplete in many cities, making it difficult to meet the data requirements for training machine learning models. This study proposes a city-scale numerical model for pollutant transport and dispersion, which couples two-dimensional surface processes with a one-dimensional sewer network. The model is used to generate a training database for machine learning. Based on this database, an LSTM (Long Short-Term Memory) model is developed to rapidly predict pollutant concentrations on the surface and within the sewer network. The results show that the proposed numerical model achieves good accuracy under validation scenarios. Compared with observed data, the Nash–Sutcliffe efficiency (NSE) is greater than 0.8. Under different rainfall return periods, the pollutant concentration patterns differ clearly. For small return periods, rainfall intensity is low and pollutant concentrations remain high. For large return periods, intense rainfall dilutes concentrations. At the same time, the pollutant load increases and the spatial extent of pollution becomes larger. The LSTM model provides high prediction accuracy for different return periods. The NSE of predicted pollutant concentrations exceeds 0.9 for both surface cells and sewer nodes. For surface concentrations, the NSE remains above 0.96. The mean absolute error (MAE) is below 1.5&#xa0;mg/L, and the root mean square error (RMSE) is below 2&#xa0;mg/L. In terms of efficiency, the machine learning model’s average simulation time for a single rainfall event does not exceed 0.83s. This performance satisfies the requirement for rapid response in pollutant transport and dispersion. Under the same training dataset and computing environment, the LSTM model outperforms KNN (K-Nearest Neighbors) and RF (Random Forest) in both accuracy and efficiency. The 1D–2D coupled numerical model captures pollutant spatiotemporal dynamics under complex urban pipe networks and diverse underlying surfaces. Coupling the physics-based results with an LSTM model enables rapid forecasting in data-scarce cities and supports early warning and refined urban water-environment management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Simulation and Rapid Prediction of Water Quantity and Quality Processes Based on Numerical Models and Deep Learning

  • 期刊论文

  • 发表日期：

    2026年3月4日

查看全文
  • Qingyuan Guo,
  • Jingming Hou,
  • Tian Wang,
  • Xinxin Pan,
  • Guangxue Luan,
  • Rongbing Zhang

摘要

With the increasing demand for urban flood hazard mitigation systems, machine learning has become an important approach to improve the rapid-response capability for pollutant transport and dispersion at the city scale. However, pollutant monitoring data are still incomplete in many cities, making it difficult to meet the data requirements for training machine learning models. This study proposes a city-scale numerical model for pollutant transport and dispersion, which couples two-dimensional surface processes with a one-dimensional sewer network. The model is used to generate a training database for machine learning. Based on this database, an LSTM (Long Short-Term Memory) model is developed to rapidly predict pollutant concentrations on the surface and within the sewer network. The results show that the proposed numerical model achieves good accuracy under validation scenarios. Compared with observed data, the Nash–Sutcliffe efficiency (NSE) is greater than 0.8. Under different rainfall return periods, the pollutant concentration patterns differ clearly. For small return periods, rainfall intensity is low and pollutant concentrations remain high. For large return periods, intense rainfall dilutes concentrations. At the same time, the pollutant load increases and the spatial extent of pollution becomes larger. The LSTM model provides high prediction accuracy for different return periods. The NSE of predicted pollutant concentrations exceeds 0.9 for both surface cells and sewer nodes. For surface concentrations, the NSE remains above 0.96. The mean absolute error (MAE) is below 1.5 mg/L, and the root mean square error (RMSE) is below 2 mg/L. In terms of efficiency, the machine learning model’s average simulation time for a single rainfall event does not exceed 0.83s. This performance satisfies the requirement for rapid response in pollutant transport and dispersion. Under the same training dataset and computing environment, the LSTM model outperforms KNN (K-Nearest Neighbors) and RF (Random Forest) in both accuracy and efficiency. The 1D–2D coupled numerical model captures pollutant spatiotemporal dynamics under complex urban pipe networks and diverse underlying surfaces. Coupling the physics-based results with an LSTM model enables rapid forecasting in data-scarce cities and supports early warning and refined urban water-environment management.