Data-driven Versus Expertise-based AI Prediction of Industrial Air Pollutants

With the development of industrialization, air pollution problems are rapidly accelerated. Industrial air pollutants can deteriorate human lives while accelerating global warming. Thus it is important to figure out the variables that affect air pollutants in an industry. With the growth of artificial intelligence, many researches on the prediction of industrial air pollutants have been conducted to prove high performance. Yet the prediction of the air pollutants with the data-driven selected input was not evaluated compared to the expertise-based input. Herein, we predicted emissions of nitrogen oxides (NOx), sulfur oxides (SOx), and total suspended particles (TSP) at once in a heat recovery steam generator system by constructing four different multivariate AI models; a random forest regressor, a shallow long-short term memory (LSTM), a shallow bidirectional LSTM (BiLSTM), and a BiLSTM based autoencoder (BiLSTM-AE). The input of a prediction model was selected by combining the results of three univariate random forest regressors where one model predicts each air pollutant and a multivariate random forest regressor. Through average one-minute predictions to averaged 30-minute predictions, we compared the performances of the AI models. Among all of them, the random forest regressor showed the best performance for predicting NOx and SOx, and the BiLSTM-AE for predicting TSP with respect to the mean absolute error. We also compared the sensitivity by differentiating input variables of the BiLSTM-AE, the data-driven and the expertise-based selection. We constructed a multivariate random forest to examine the importance of each variable in the prediction of three air pollutants. Both the data-driven input and the expertise-based input include the gas turbine variables and some thermal variables as important variables. As a result, the expertise-based input may be good standards, but the data-driven input can be complementary to the expertise-based input for generalization and ease of selection. This study enables self-diagnosis and proactive action for each industry to regulate its air pollutants in advance of the law regulation.