인공신경망 모델과 배경대기 측정자료를 활용한 서울시 PM2.5 농도 단기예측 및 입력변수의 기여도 분석

Abstract

Recently, Artificial Neural Network (ANN) models have been successfully applied to predict PM2.5 mass concentration. However, the complex nature of ANNs hinders understanding of the actual relationship between input variables and output PM2.5. In this study, a simple ANN model was constructed to predict the PM2.5 mass of Seoul 12 hours in advance using nine atmospheric variables routinely measured in Seoul and three Background sites. The contribution of the input variables from the four sites to the predicted PM2.5 mass was then estimated using the Connection Weight Method (CWM) and the Garson’s Algorithm (GA). The second rank of Baengnyeong Island PM2.5 after Seoul suggests the impact of transport, and the least contribution of reactive gases of Seoul including O3, NO2, SO2, and CO, indicates the relatively insignificant contribution of in situ formation to PM2.5. The ranking of meteorological variables including temperature, relative humidity, and wind direction and speed highlights the importance of synoptic meteorological conditions in determining PM2.5 levels in Seoul. It also reveals the role of stagnation in increasing PM2.5 mass.