Understanding the effects of natural hazards on chemical emission incidents using machine learning techniques

Abstract

Natural hazard-triggered technological accidents (Natechs) pose significant risks to industrial safety, particularly in regions vulnerable to extreme weather conditions. This study explores the impact of various climate variables on the frequency of chemical emission incidents in Houston, TX, aiming to understand the major contributors of Natechs from a data-driven perspective and enhance predictive capabilities for process safety management. Machine learning models, including XGBoost, Random Forest, k-nearest neighbor (kNN), and support vector machine (SVM), were developed to predict high-risk days for chemical emission incidents, using local climate data as inputs. Conformal prediction techniques were employed to control error rates and optimize the balance between sensitivity and specificity. The results demonstrate that XGBoost and Random Forest models outperformed the others, achieving ROC AUC scores exceeding 0.8. Furthermore, the conformal wrapper indicated XGBoost as the more promising model, particularly under higher specificity requirements: at controlled specificity values of 0.75 and 0.80, its guaranteed sensitivity values were 0.765 and 0.750, compared to Random Forest’s 0.649 and 0.610, respectively. Notably, precipitation and lightning were identified as the most significant contributors to chemical emission incidents. Overall, this study provides a framework for using climate data in predictive models for Natechs with novel conformal error control strategies, offering valuable insights for proactive risk assessment and facilitating process safety protocols.