Development of a deep neural network for predicting 6-hour average PM_2.5 concentrations up to two subsequent days using various training data

Abstract

Abstract. This study aims to develop a deep neural network (DNN) model as an artificial neural network (ANN) for the prediction of 6-hour average fine particulate matter (PM2.5) concentrations for a three-day period—the day of prediction (D+0), one day after prediction (D+1) and two days after prediction (D+2)—using observation data and forecast data obtained via numerical models. The performance of the DNN model was comparatively evaluated against that of the currently operational Community Multiscale Air Quality (CMAQ) modelling system for air quality forecasting in South Korea. In addition, the effect on predictive performance of the DNN model on using different training data was analyzed. For the D+0 forecast, the DNN model performance was superior to that of the CMAQ model, and there was no significant dependence on the training data. For the D+1 and D+2 forecasts, the DNN model that used the observation and forecast data (DNN-ALL) outperformed the CMAQ model. The root-mean-squared error (RMSE) of DNN-ALL was lower than that of the CMAQ model by 2.2 μgm−3, and 3.0 μgm−3 for the D+1 and D+2 forecasts, respectively, because the overprediction of higher concentrations was curtailed. An IOA increase of 0.46 for D+1 prediction and 0.59 for the D+2 prediction was observed in case of the DNN-ALL model compared to the IOA of the DNN model that used only observation data (DNN-OBS). In additionally, An RMSE decrease of 7.2 μgm−3 for the D+1 prediction and 6.3 μgm−3 for the D+2 prediction was observed in case of the DNN-ALL model, compared to the RMSE of DNN-OBS, indicating that the inclusion of forecast data in the training data greatly affected the DNN model performance. Considering the prediction of the 6-hour average PM2.5 concentration, the 8.8 μgm−3 RMSE of the DNN-ALL model was 2.7 μgm−3 lower than that of the CMAQ model, indicating the superior prediction performance of the former. These results suggest that the DNN model could be utilized as a better-performing air quality forecasting model than the CMAQ, and that observation data plays an important role in determining the prediction performance of the DNN model for D+0 forecasting, while prediction data does the same for D+1 and D+2 forecasting. The use of the proposed DNN model as a forecasting model may result in a reduction in the economic losses caused by pollution-mitigation policies and aid better protection of public health.

Development of a deep neural network for predicting 6 h average PM2.5 concentrations up to 2 subsequent days using various training data

“…Code and data availability. The code and data used in this study can be found at https://doi.org/10.5281/zenodo.5652289 (Lee et al, 2021) or https://github.com/GercLJB/GMD (last access: 28 January 2022).…”

“…To overcome the uncertainty and limitations of the atmospheric CTM, a model for predicting air quality using artificial neural networks (ANNs) based on statistical data has recently been developed (Cabaneros et al, 2019;Ditsuhi et al, 2020). Studies using ANNs, such as the recurrent neural network (RNN) algorithm which is advantageous for timeseries data training (Biancofiore et al, 2017;Kim et al, 2019;Zhang et al, 2020;Huang et al, 2021) and the deep neural network (DNN) algorithm which is advantageous for extracting complex and non-linear features, are underway (Schmidhuber et al, 2015;LeCun et al, 2015;Lightstone et al, 2017;Cho et al, 2019;Eslami et al, 2020;Chen et al, 2021;Lightstone et al, 2021). Kim et al (2019) developed an RNN model to predict PM 2.5 concentrations after 24 h periods at two observation points in Seoul.…”

“…The evaluation of the prediction performance of the RNN model for the May to June 2016 period yielded an index of agreement (IOA) range between 0.62 and 0.76, which constituted a 0.12 to 0.25 IOA improvement compared with the CMAQ model. Lightstone et al (2021) developed a DNN model that predicted 24 h PM 2.5 concentrations based on aerosol optical depth (AOD) data and Kriging PM 2.5 . The DNN-model predictions for the January to December 2016 period yielded a root-meansquared error (RMSE) of 2.67 µg m −3 , thereby demonstrating a prediction-performance improvement of 2.1 µg m −3 compared with the CMAQ model.…”