2024 – A Machine Learning Approach to Identifying Key Factors Influencing PLL Estimates

Yanni Wang, Albert Shen, David Stephens, Ze Jiang

Potential Loss of Life (PLL). The complex and dynamic nature of human response to hazard presents challenges in determining driving factors from input variables, hence where investment is most needed to reduce PLL or modelling efforts should focus to better understand uncertainty. In this work, we explored the pioneering capability of machine learning to navigate the nonlinear and multilayered relationships in LifeSim. The machine learning model, which employs a histogram-based gradient boosting algorithm, demonstrated superior performance and robustness compared to the baseline linear regression model. An interpretative tool, called SHapley Additive exPlanations (SHAP), was employed to rank the relative importance of input variables. For the selected LifeSim scenario, hydraulic variables like maximum depth and depth-velocity, structural variables such as number of stories and evacuation related variables like mobilisation time, were identified as key factors. An interesting interaction between
mobilisation and warning time was also observed using the SHAP dependence analysis. The machine learning approach, together with SHAP, offers a data-driven framework to transform sophisticated dynamics into quantifiable insights. Meanwhile, it helps to build an understanding of machine learning within the dam community and ensure such models and their results are used in fair and ethical ways.