Spatio-Temporal Prediction of PM1.0 Concentration in South Korea Using a Machine Learning Algorithm

Since smaller particles can get through deeper into the human body, a model that predicts PM_1.0 concentration temporally and spatially is important. Despite their importance, there are significantly fewer PM_1.0 measurement stations than PM_2.5 and PM₁₀ in South Korea. Therefore, in this study, PM_1.0 prediction models were constructed using a machine learning algorithm to predict the spatiotemporal concentration of PM_1.0 in a place where the PM_1.0 measurement was not available.

From January to December 2021, hourly data for the concentration of particulate matters(PM₁₀ and PM_2.5), the composition of PM_2.5, the concentration of gaseous pollutants(SO₂, O₃, CO, NO, NO_y, NH₃), and weather conditions(wind direction, wind speed, temperature, relative humidity, press, precipitation, cloud) measured at three different Air Quality Measurement Systems were collected. PM_1.0 concentrations were collected at two of these sites which are Seoul and Ansan(Gyeonggi-do), and no PM_1.0 concentration was measured at the other site which is Seosan(Chungcheongnam-do). Since the three measurement stations were located close to each other and had similar sources and characteristics, the concentration of PM_1.0 in Seosan was predicted by using a model trained based on Seoul and Ansan data.

Based on collected data, the importance of variables in the model was identified and variables that are important in predicting PM_1.0 concentration were selected. Ensemble models (Random Forest, gradient boosting) and sequence models (RNN, LSTM, GRU) were compared to find a better model. Each model was evaluated by calculating the coefficient of determination and the proportion of impossible concentrations. Finally, based on the model with the best prediction result, the PM_1.0 concentration was predicted at the target location (Ansan) where the PM_1.0 concentration was unknown.

Our results showed that PM_1.0 concentration can be predicted with high accuracy considering both the spatial distribution and temporal variability of the concentration. The results of this study can be used as data for selecting models in air quality prediction studies using machine learning. In addition, economical and efficient choices can be made in the construction of new monitoring stations in the future.

 

Acknowledgments

This study was supported by the National Research Foundation of Korea (grant number NRF-2021R1C1C1013350) and by a grant from the National Institute of Environmental Research (NIER), funded by the Ministry of Environment (ME) of the Republic of Korea (NIER-2022-04-02-087).