Predicting atmospheric dispersion of industrial chemicals using machine learning approaches

Abstract

This study presents an intelligent framework for assessing atmospheric dispersion in industrial accident scenarios involving chemical substances. The research focuses on modeling the dispersion of key chemicals, such as chlorine, methanol, and propane, under various accident conditions, including leaks, fires, and explosions. Atmospheric and contextual variables, such as wind speed, air temperature, tank specifications, and chemical release parameters, were thoroughly characterized to construct a robust database using experimental data and software simulations. Machine learning techniques were rigorously trained and tested to predict atmospheric dispersion, emphasizing hyperparameter optimization to enhance model performance. Dimensionality reduction methods, such as principal component analysis and correlation-based dimensionality reduction, were implemented to improve computational efficiency, reduce data noise, and maintain essential information. Results demonstrate the effectiveness of the proposed approach, with satisfactory predictions across all evaluated risk areas. Key contributions include the development of a replicable framework adaptable to diverse industrial scenarios, applying hyperparameter tuning to optimize model accuracy, and integrating dimensionality reduction techniques to streamline data processing. These advancements establish a foundation for future studies to incorporate additional chemicals and accident scenarios, improving the flexibility and reliability of atmospheric dispersion modeling. Future work will explore hybrid machine learning models and advanced dimensionality reduction methods to enhance the system’s applicability to complex industrial environments.