Hybrid machine learning to enhance PM2.5 forecasting performance by the WRF-Chem model

Abstract

The weather research and forecasting with chemistry (WRF-Chem) model had been widely used in PM_2.5 concentration forecasting. However, uncertainties in global emission inventories, meteorological data, and simplified chemical parameterizations continue to pose challenges. We evaluated the performance of the original WRF-Chem model and three models augmented with machine learning (ML) algorithms, i.e. Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost) and XGBoost-LSTM (Hybrid) approaches, to enhance forecasting accuracy. The ML models were trained and tested using dataset from WRF-Chem-simulated meteorological and pollutant data at four monitoring stations in southern Thailand during the year 2019–2020. The WRF-Chem-Hybrid model significantly improved all metrics in the original WRF-Chem results - with R² increasing from insignificant to 0.90 or more, RMSE decreasing from 7.00-15.17 μg/m³ to 1.34–3.47 μg/m³, and MAE decreasing from 5.52-10.55 μg/m³ to 0.79–1.49 μg/m³. The validation test during the entire 2021 performed well, with R² = 0.94–0.96, RMSE = 1.56–2.48 μg/m³ and MAE = 1.01–1.56 μg/m³. The WRF-Chem-Hybrid model forecast PM_2.5 concentrations for 72 h in advance, with R² = 0.70–0.89, RMSE = 4.64–11.99 μg/m³, and MAE = 3.07–8.38 μg/m³. Thus, the hybrid model is suggested for forecasting PM_2.5 concentrations over southern Thailand and other regions up to 72 h in advance. Overall, this study demonstrated the advantages of augmenting the WRF-Chem model to form hybrid ML models to more accurately forecast PM_2.5 levels, their distribution and evolution over time, particularly in regions where PM_2.5 levels were affected by open biomass burning from both local and cross-border emissions.