Development of a deep neural network for predicting 6 h average PM<sub>2.5</sub> concentrations up to 2 subsequent days using various training data

Abstract. Despite recent progress of numerical air quality models, accurate prediction fine particulate matter (PM2.5) is still challenging because uncertainties in physical and chemical parameterizations, meteorological data, emission inventory databases. Recent advances artificial neural networks can be used to overcome limitations models. In this study, a deep network (DNN) model was developed for 3 d forecasting 6 h average PM2.5 concentrations: the day (D+0), 1 after (D+1), 2 (D+2). The DNN evaluated against currently operational Community Multiscale Air Quality (CMAQ) modeling system South Korea. Our study demonstrated that outperformed CMAQ results. provided better skills by reducing root-mean-squared error (RMSE) 4.1, 2.2, 3.0 µg m−3 consecutive days, respectively, compared with CMAQ. Also, false-alarm rate (FAR) decreased 16.9 %p 7.5 7.6 (D+2), indicating substantially mitigated overprediction high concentrations. These results showed when it simultaneously trained using observation data from Notably, more benefits as days increased. suggest our data-driven machine learning approach useful tool implemented models together model-oriented systematic biases.