Long-Term Retrospective Predicted Concentration of PM2.5 in Upper Northern Thailand Using Machine Learning Models.

Abstract

This study aims to build, for the first time, a model that uses a machine learning (ML) approach to predict long-term retrospective PM_2.5 concentrations in upper northern Thailand, a region impacted by biomass burning and transboundary pollution. The dataset includes PM₁₀ levels, fire hotspots, and critical meteorological data from 1 January 2011 to 31 December 2020. ML techniques, namely multi-layer perceptron neural network (MLP), support vector machine (SVM), multiple linear regression (MLR), decision tree (DT), and random forests (RF), were used to construct the prediction models. The best ML prediction model was selected considering root mean square error (RMSE), mean prediction error (MPE), relative prediction error (RPE) (the lower, the better), and coefficient of determination (R²) (the bigger, the better). Our study found that the ML model-based RF technique using PM₁₀, CO₂, O₃, fire hotspots, air pressure, rainfall, relative humidity, temperature, wind direction, and wind speed performs the best when predicting the concentration of PM_2.5 with an RMSE of 6.82 µg/m³, MPE of 4.33 µg/m³, RPE of 22.50%, and R² of 0.93. The RF prediction model of PM_2.5 used in this research could support further studies of the long-term effects of PM_2.5 concentration on human health and related issues.