Association of COVID-19 outbreak with meteorological and air pollution parameters in Thailand

Abstract

On January 13, 2020, the first infected case was identified in Thailand. Since then, the virus has spread throughout the country. Thailand’s COVID-19 situation is still ongoing and widespread, especially in the country's central region. In order to forecast the number of cases in Thailand's central region, we investigated the relationships between meteorological, and air quality factors with COVID-19 cases. The data for the air quality were obtained from the World Air Quality Index (WAQI), while the data for the meteorological variables were collected from the Global Surface Summary of the Day (GSOD). The air quality variables comprise the concentrations of PM 2.5, PM 10, ozone, and nitrogen dioxide whereas the meteorological variables include temperature, rainfall, and relative humidity. In order to model, we first smoothen the data by doing a 7-day moving average and then applying the feature selection method. To forecast the COVID-19 daily infection in Thailand's central region, we used a multivariate Generalized Linear Model (GLM) and univariate, multivariate Long short-term memory (LSTM) frameworks. For the Generalized Linear Model (GLM), we used a Dendrogram correlation in the feature selection method. Due to data fluctuations, the multivariate GLM prediction does not fit the reported cases very well. Then, we applied Long short-term memory (LSTM) for both univariate and multivariate models. Taken’s Theorem is used to choose the lag time in the Univariate model, and Principal Component Analysis (PCA) and XG-Boost are used to select the suitable parameters and lag time in the Multivariate Long short-term memory model. The results show that the best model is the multivariate LSTM model where the data were split with a ratio of 85%:15% into training and testing datasets. The best model was selected out of 100 experiments by comparing root mean square error (RMSE) and mean absolute error (MAE). The prediction performance for one step ahead was presented. We found that forecasting the number of cases can be done using our model with a relative humidity factor.

-