Since the 1960s, waste generated by power plants has been a major environmental concern. In Chile, the regulation of emissions from the power industry remains limited, leading to significant negative impacts on both the environment and public health. International organizations have developed models and software to monitor emissions, yet their accuracy in assessing the true impact of pollutants remains unsatisfactory, as they rely solely on past observations to simulate pollutant concentrations without incorporating learning from historical behavior. This study explores key concepts related to air quality, atmospheric emission modeling, and the critical parameters involved in pollutant dispersion analysis. The objective of this work is to evaluate the efficacy of a feedforward-backpropagation neural network (NN) as an alternative to CALPUFF for predicting SO 2 concentrations in environmental impact assessments. Emission, meteorological, and stack data were collected from official, private, and public sources. We compare the performance of the CALPUFF/CALPOST workflow with a neural-network-based approach for predicting SO 2 concentrations in the Industrial Bay of Mejillones, Chile. The NN substantially reduces prediction error at the training site (Ferrocarriles (FFEE); Root Mean Square Error (RMSE) 4.75 and Mean Absolute Error (MAE) 2.90) compared with CALPOST (RMSE 14.74 and MAE 12.76). However, transfer to an independent station (Cactus Gaviotín (CV)) remains more challenging (NN RMSE 6.76 and MAE 5.57), highlighting both the potential and limitations of data-driven surrogates in environmental impact assessment contexts.
Concha et al. (Mon,) studied this question.