A Secondary Modeling Approach for Air Quality Prediction Based on LSTM-FC Multi-model Fusion

Abstract

Aiming at the insufficient forecasting accuracy of the traditional WRF-CMAQ model due to meteorological field uncertainty and emission inventory errors, this paper proposes a secondary modeling framework that integrates multi-source data with deep learning. The multi-pollutant concentration prediction accuracy is significantly improved by introducing the pseudo-unique thermal coded wind feature reconstruction, CEEMDAN adaptive noise cancellation technique, and combining the random forest feature selection with the LSTM-FNN combined model. Experiments show that the method has a mean absolute error (MAE) of 1.76 μg/m³ in PM2.5 prediction at monitoring points A, B, and C, which is 8.3% lower than that of a single model, and its MAE is reduced to 13.53 μg/m³ for the CEEMDAN_IPSOSE_SVR model for ozone (O3), which is a 21% decrease from the baseline. The model achieves 92.7% accuracy in the identification of the top pollutants, and the relative error of AQI prediction is only 6.2%. This study provides a highly interpretable and robust technical solution for regional air quality forecasting.