Forecasting hourly PM2.5 concentration with an optimized LSTM model

Machine learning has become a powerful tool in air quality assessment which can provide timely and predictable information, alert the public, and take timely measures to prevent deteriorating air quality. The study proposed a deep learning-based long -short term memory (LSTM) model to predict hourly PM2.5 in one of the most polluted areas in Taiwan. A series of sensitivity assessments with model settings was conducted to optimize the performance of the LSTM model. Regarding the model input parameters, aerosol optical depth, pressure, and PM2.5 concentrations from the three nearby stations were used and later showed significant improvement in the forecast results. As a result of the 1-24 h forecast in 2021, the root-mean-square error (RMSE) shows a range from 6.3 to 13.1 mu g m-3, and the Pearson correlation coefficient (r) varies from 0.92 to 0.59, as compared with the observed PM2.5. The model's predictability decreases as time increases-a strong correlation (r higher than 0.7) within a 9-h PM2.5 forecast. The seasonal variation showed that the highest RMSE, about 16.2 mu g m-3, was observed during the winter, which is the high-polluted season in the area. Additionally, the spatial representation of the model was examined. The model can perform an efficient and satisfied forecast in the radius of 15 km from the training station. We further compared several deep learning-based algorithms in forecasting PM2.5, and our model performs better prediction results. The deep learning-based model investigated in this study can be implemented for routine air quality monitoring in urban areas and air-quality alarms associated with public health.