Atmospheric dispersion prediction for different toxic gas clouds by using a machine learning approach

Abstract

Several tools are available to predict atmospheric dispersion considering climatic and chemical parameters, event scenarios, and the frequency of industrial accidents caused by chemicals. However, some tools, such as the Areal Locations of Hazardous Atmospheres (ALOHA) software, cannot integrate these parameters owing to scalability issues and other limitations. To overcome this, a new intelligent approach is developed in this thesis to forecast the atmospheric dispersion of common industrial chemicals. The proposed approach involves a dynamic system in which several machine learning models are tested, and the model with the best evaluation metrics for each type of accident is selected. This thesis includes a comprehensive analysis of correlation studies and the application of Principal Component Analysis (PCA) as a dimensionality reduction method. The results demonstrate that for chlorine scenarios, the best model metrics were achieved without applying PCA. In contrast, for methanol and propane scenarios, utilizing PCA led to more favorable outcomes. Evaluation metrics such as RMSE, Accuracy, and MAE were used to assess the model performance across the different scenarios. The results show that there is no single type of model that performs adequately across all the studied scenarios; instead, models such as Random Forest, K Nearest Neighbors, Multilayer Perceptron, and Support Vector Regressor tend to excel in specific scenarios.