Impact of the weighted loss function on the innovative CMAQ-CNN PM_2.5 forecasting model

Developing an accurate air quality forecasting model with a lead time longer than 72 hours is important for predicting hazardous pollution events. An air quality forecasting system (AQF) conducted with the WRF-CMAQ numerical model has been developed. This study used the predicted PM2.5 from AQF, observed PM2.5, longitude, latitude, the land use index of the corresponding grid, and indexes of synoptic weather patterns to develop a forecasting model through the CNN algorithm to predict 72 hours of PM2.5 concentrations at 75 surface air quality monitoring sites. The training dataset was from October 2019 to September 2021, and the data from October 2021 to September 2022 was divided into validation and test datasets in order of date. Since the PM2.5 varies in a wide range, the weighted loss function was applied to improve the model performance. In general, the CMAQ-CNN models developed by this study show better performance than the AQF, especially during high pollution events when the CMAQ-CNN model utilizes the weighted loss function.