Wu Hongshi, Shi Peng, Qu Simin, Yang Xiaoqiang, Zhang Hongxue, Wang Le, Ding Song, Li Zichun, Lu Meixia, Qiu Chao
College of Hydrology and Water Resources, Hohai University, Nanjing 210098, China.
College of Hydrology and Water Resources, Hohai University, Nanjing 210098, China; Cooperative Innovation Center for Water Safety & Hydro Science, Nanjing 210098, China.
Sci Total Environ. 2024 Jan 10;907:167767. doi: 10.1016/j.scitotenv.2023.167767. Epub 2023 Oct 11.
Conventional hydrological modeling is usually based on the assumption that parameters are statically time-invariant. However, recent studies suggest that the influences of climate change and human interventions have made this hypothesis questionable. Meanwhile, machine learning techniques are increasingly used to extract patterns and insights from the ever-increasing hydrometeorological data. Here, we proposed a hybrid framework (HSPDM) to improve the precision of real-time flood forecasting in response to nonstationary conditions. It utilizes multiple machine learning techniques to dynamically retrieve calibrated hydrological model parameters from historical similar floods, thus continuously obtaining hourly time-variant parameters in real-time flood forecasting operations. Using the Quzhou Basin in China as a case study, the effectiveness and advancement of HSPDM framework was examined. Three schemes, including traditional time-invariant parameters (scheme 1), hourly time-variant parameters (scheme 2), and probabilistic forecasting scheme (scheme 3), were built for comparison purpose. The results were summarized as follows: (1) The proposed framework can successfully identify continuous flood subsequence with a high retrieval accuracy (1.74) and acceptable time consumption (175.05 s) by adopting k-means, K-Nearest Neighbor (KNN), and embedding-based subsequence matching (EBSM) method. (2) Compared to scheme 1, scheme 2 provided more reliable forecasting results with higher accuracies, in terms of the general goodness-of-fit (higher NSE value) and reproducing flood peak and flood process. (3) The streamflow hydrographs forecasted by scheme 2 fell exactly in the predictive uncertainty bounds of scheme 3 and even showed superiority compared to a preferred deterministic forecast (Q50) of scheme 3. The major scientific contribution of this study lies in advancing the technique of real-time flood forecasting based on the hourly time-variant model parameters, thereby strengthening our understanding of model behaviors under changing conditions. The proposed framework can also act as a new alternative for flood control and ultimately contribute to the mitigation of flooding disasters.
传统水文模型通常基于参数在时间上静态不变的假设。然而，最近的研究表明，气候变化和人类干预的影响使这一假设受到质疑。与此同时，机器学习技术越来越多地用于从不断增加的水文气象数据中提取模式和见解。在此，我们提出了一种混合框架（HSPDM），以提高在非平稳条件下实时洪水预报的精度。它利用多种机器学习技术从历史相似洪水中动态检索校准后的水文模型参数，从而在实时洪水预报操作中持续获取每小时随时间变化的参数。以中国的衢州盆地为例，检验了HSPDM框架的有效性和先进性。为了进行比较，构建了三种方案，包括传统的时不变参数（方案1）、每小时随时间变化的参数（方案2）和概率预报方案（方案3）。结果总结如下：（1）所提出的框架通过采用k均值、K近邻（KNN）和基于嵌入的子序列匹配（EBSM）方法，能够成功识别连续洪水子序列，具有较高的检索精度（1.74）和可接受的时间消耗（175.05秒）。（2）与方案1相比，方案2在总体拟合优度（更高的NSE值）以及重现洪峰和洪水过程方面提供了更可靠、精度更高的预报结果。（3）方案2预测的流量过程线恰好落在方案3的预测不确定性范围内，甚至与方案3的首选确定性预报（Q50）相比表现出优势。本研究的主要科学贡献在于推进了基于每小时随时间变化的模型参数的实时洪水预报技术，从而加强了我们对变化条件下模型行为的理解。所提出的框架还可以作为防洪的新选择，最终有助于减轻洪水灾害。
