Analysis of Factor Importance of PM<SUB>2.5</SUB> High Concentration Case Using DNN and Layer-wise Relevance Propagation

Abstract

In this study, we used Layer-wise Relevance Propagation (LRP) to analyze the level of contribution of input factors to the predictive results of the PM2.5 predictive model. First, we trained the DNN prediction model using data from 2015 to 2020, and then evaluated it using data from 2021. Next, we performed LRP on the evaluation data to analyze the importance of input factors in the prediction results. As a result, factors with consistently high importance regardless of concentration were O_TA, O_TD, O_RH, O_U, O_V, and O_PA, whereas PMSUB10/SUB and O_RN_ACC were observed to have lower importance. Furthermore, to analyze the characteristics of high-concentration data that are generally difficult to predict compared to low-concentration data, we divided the data by concentration and analyzed the importance of input factors. As a result, the importance of O_PMSUB2.5/SUB was high in the high concentration pattern and the importance of O_radiation was low, while the opposite trend was observed in the low concentration pattern. In particular, for high-concentration patterns that started suddenly and lasted more than three days, we analyzed the importance of input factors by time and factor. These high-concentration patterns with these characteristics showed significantly increased importance in the O_PMSUB2.5/SUB factor in the T12 interval closest to the prediction time, and it was observed that the importance of the F_PMSUB2.5/SUB factor increased slightly. Applying the factor importance results analyzed in this study to the PMSUB2.5/SUB prediction model is expected to improve prediction accuracy for high concentration patterns that are difficult to predict compared to general patterns.