PM2.5 구성 성분 입력자료를 이용한 DNN의 XAI 분석과 PM2.5 예측 정확도 개선

• Published in: 한국대기환경학회지(국문) Vol. 39; no. 4; pp. 411 - 426
• Main Authors: 이주용(Ju-Yong Lee), 이채연(Chae-Yeon Lee), 정민우(Min-Woo Jung), 안준영(Joon-Young Ahn), 왕경희(Kyung-Hui Wang), 최대련(Dae-Ryun Choi), 윤희영(Hui-Young Yun)
• Format: Journal Article
• Language: Korean
• Published: 한국대기환경학회 01.08.2023
• Subjects: 대기과학; XAI; component; DNN; CMAQ; 2.5< forecast; PM< SUB&gt
• ISSN: 1598-7132; 2383-5346
• DOI: 10.5572/KOSAE.2023.39.4.411

The aim of this study was to assess whether a simple Deep Neural Network (DNN) model could address the limitations of the current overestimated CMAQ model used for fine dust forecasting by NIER in Seoul, South Korea. Two DNN models, DNN-1 and DNN-2, were developed using data from 2016 to 2020. DNN-1 used urban air monitoring network data and CMAQ predictions, while DNN-2 incorporated additional PM2.5 component measurement data. The models forecasted PM2.5 concentrations for three days, and evaluation focused on the accuracy of daily average concentrations. DNN-1 and DNN-2 outperformed the CMAQ model in terms of accuracy, probability of detection (POD), and false alarm rate (FAR). DNN-2 showed lower POD but significantly improved ACC and FAR, indicating a compensatory relationship between the two metrics. Using the XAI technique called Layer-wise Relevance Propagation (LRP), the study analyzed the importance of input parameters in the models. Meteorological factors, including temperature, humidity, radiation, wind speed, and wind direction, were found to be highly important. PM2.5 component measurement data, such as NO3 - and OM (Organic Matter), showed relatively lower importance compared to meteorological factors. It was determined that the limited impact of fine dust composition was due to the XAI analysis conducted on yearly results rather than high-concentration cases in this study. KCI Citation Count: 4