期刊
JOURNAL OF WATER RESOURCES PLANNING AND MANAGEMENT
卷 147, 期 10, 页码 -出版社
ASCE-AMER SOC CIVIL ENGINEERS
DOI: 10.1061/(ASCE)WR.1943-5452.0001414
关键词
Dam hazard classification; Machine learning; Tuning; Multiobjective
资金
- National Science Foundation [ACI-1532235, ACI-1532236]
- University of Colorado Boulder
- Colorado State University
- US Department of Education's Graduate Assistance in Areas of National Need (GAANN) program [P200A180024]
- Baker research group
- Kasprzyk research group
This study introduces a geospatial and machine learning model for classifying dam hazard potential, highlighting the significance of utilizing a multiobjective approach for tuning model parameters.
Within the United States' National Inventory of Dams, 15,000 dams have been classified as having a high hazard potential, meaning failure or misoperation would lead to probable loss of human life. However, state dam officials evaluate dam hazard potential on a case-by-case basis, ultimately relying on human judgement. Such a process lacks objectivity and consistency across state boundaries and can be time-consuming. Here, the authors present a parameterized geospatial and machine learning dam hazard potential classification model to overcome these limitations. The parameters of this model can be tuned for optimal performance. However, for this classification problem, the regulatory and physical implications of the types of model misclassifications are best captured through multiple objectives. Therefore, this research additionally contributes a novel multiobjective approach to machine learning parameter tuning. This research demonstrates the utility of this approach for dams in Massachusetts, United States, using a multiobjective evolutionary algorithm to explore different model parameterizations and identify analyst-relevant tradeoffs among objectives describing model performance. Such an approach allows for greater justification of model parameters as well as greater insights into the complexities of the dam hazard potential classification problem.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据