Toxic heavy gas sulfur dioxide (SO2) is a specific life and environmental hazard. Predicting the diffusion of SO2 has become a research focus in fields such as environmental and safety studies. However, traditional methods, such as kinetic models, cannot balance precision and time. Thus, they do not meet the needs of emergency decision-making. Deep learning (DL) models are emerging as a highly regarded solution, providing faster and more accurate predictions of gas concentrations. To this end, this study proposes an innovative hybrid DL model, the parallel-connected convolutional neural network-gated recurrent unit (PC CNN-GRU). This model utilizes two CNNs connected in parallel to process gas release and meteorological datasets, enabling the automatic extraction of high-dimensional data features and handling of long-term temporal dependencies through the GRU. The proposed model demonstrates good performance (RMSE, MAE, and R2 of 20.1658, 10.9158, and 0.9288, respectively) with real data from the Project Prairie Grass (PPG) case. Meanwhile, to address the issue of limited availability of raw data, in this study, time series generative adversarial network (TimeGAN) are introduced for SO2 diffusion studies for the first time, and their effectiveness is verified. To enhance the practicality of the research, the contribution of drivers to SO2 diffusion is quantified through the utilization of the permutation importance (PIMP) and Sobol\' method. Additionally, the maximum safe distance downwind under various conditions is visualized based on the SO2 toxicity endpoint concentration. The results of the analyses can provide a scientific basis for relevant decisions and measures.

Deep learning models are tools for data analysis suitable for approximating (non-linear) relationships among variables for the best prediction of an outcome. While these models can be used to answer many important questions, their utility is still harshly criticized, being extremely challenging to identify which data descriptors are the most adequate to represent a given specific phenomenon of interest. With a recent experience in the development of a deep learning model designed to detect failures in mechanical water meter devices, we have learnt that a sensible deterioration of the prediction accuracy can occur if one tries to train a deep learning model by adding specific device descriptors, based on categorical data. This can happen because of an excessive increase in the dimensions of the data, with a correspondent loss of statistical significance. After several unsuccessful experiments conducted with alternative methodologies that either permit to reduce the data space dimensionality or employ more traditional machine learning algorithms, we changed the training strategy, reconsidering that categorical data, in the light of a Pareto analysis. In essence, we used those categorical descriptors, not as an input on which to train our deep learning model, but as a tool to give a new shape to the dataset, based on the Pareto rule. With this data adjustment, we trained a more performative deep learning model able to detect defective water meter devices with a prediction accuracy in the range 87-90%, even in the presence of categorical descriptors.