Scholarlink

<p>Rapid prediction of storm water flooding is an important tool for mitigating current urban flooding disasters. This paper constructs a fast prediction model for urban flooding based on a machine learning approach. Firstly, MIKE numerical model simulations with high accuracy results are used as the data driver, and then sliding-window samples are constructed from the simulated inundation depth series. Four deep learning models, namely LSTM, GRU, CNN, and TCN, are developed and compared under a unified framework using Zhoukou City as an example. The results show that the MIKE model reproduced observed inundation depths with acceptable accuracy, with an MAE of 3.10 cm, an RMSE of 3.38 cm, and an R<sup>2</sup>of 0.949. Among the four models, GRU achieved the best overall performance, with mean RMSE, MAE, and R<sup>2</sup> values of 0.0188 m, 0.0138 m, and 0.9833, respectively. Therefore, the trained deep learning surrogate model can quickly predict urban inundation processes with high accuracy and can meet the needs of emergency flood control.</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

Surrogate Prediction of Urban Waterlogging for Emergency Warning: Deep Learning Comparison Based on Hydrodynamic Simulation-Generated Samples

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Jian Chen,
  • Yangyang Tian,
  • Luyao Zhang,
  • Haizhou Wang

摘要

Rapid prediction of storm water flooding is an important tool for mitigating current urban flooding disasters. This paper constructs a fast prediction model for urban flooding based on a machine learning approach. Firstly, MIKE numerical model simulations with high accuracy results are used as the data driver, and then sliding-window samples are constructed from the simulated inundation depth series. Four deep learning models, namely LSTM, GRU, CNN, and TCN, are developed and compared under a unified framework using Zhoukou City as an example. The results show that the MIKE model reproduced observed inundation depths with acceptable accuracy, with an MAE of 3.10 cm, an RMSE of 3.38 cm, and an R2of 0.949. Among the four models, GRU achieved the best overall performance, with mean RMSE, MAE, and R2 values of 0.0188 m, 0.0138 m, and 0.9833, respectively. Therefore, the trained deep learning surrogate model can quickly predict urban inundation processes with high accuracy and can meet the needs of emergency flood control.