Development of a PM2.5 prediction model using a recurrent neural network algorithm for the Seoul metropolitan area, Republic of Korea

Abstract

Listen

Translate article

The National Institute of Environmental Research, the Ministry of Environment, has been forecasting the concentrations of particulate matter (PM) with a diameter ≤ 2.5 μm (PM2.5) over Seoul, Republic of Korea, in terms of four PM2.5 concentration categories (low, moderate, high, and very high) since August 31, 2013. The current model, the Community Multiscale Air Quality (CMAQ) model, is run four times a day to forecast air quality for up to two days in 6-h intervals. In 2018, the hit ratio (i.e., accuracy) of the model was 60%, with an additional increase of 10% with the involvement of a forecaster. The CMAQ was improved in this study by incorporating a recurrent neural network (RNN) algorithm for the Seoul Metropolitan Area. Input datasets to the RNN model—PM values, meteorological parameters, and back-trajectory tracks obtained from both observations and model forecasts—were sorted according to time as the RNN algorithm learns time sequence series information, unlike typical neural network algorithms. To reflect the seasonality of the meteorological parameters that influence the PM2.5 concentrations in the region, one year was divided into 36 sets of three-month periods (i.e., there are three sets for July: May–June–July, June–July–August, and July–August–September). Several indices representing the accuracy of the forecast were calculated based on the RNN model results for 2018 after training the model for the previous three years (2015–2017). The accuracy of the RNN model is 74–81% for forecast lead times up to two days, about 20% higher than the CMAQ-only forecasts and ~10% higher than the combined CMAQ-forecaster forecast. The RNN model probabilities of detection for both high and very high PM2.5 events are comparable to those of the CMAQ model; however, the RNN model notably reduces the false alarm rate. Overall, the RNN model yields higher performance than the current forecast methods. Hence, this model can be adopted as an operational forecast model in Korea.