A novel ensemble framework based on intelligent weight optimization and multi-model fusion for air quality index prediction

Abstract

The accuracy of air quality prediction is crucial for public health and environmental management. This paper proposes a hybrid deep learning model based on TimesNet, Crossformer and Modified Honey Badger Algorithm (MHBA) for air quality prediction. First, the original air quality index (AQI) series is decomposed using Seasonal-Trend decomposition based on Loess (STL). Then, the decomposed three components are predicted separately using TimesNet and Crossformer, while the hyperparameters of TimesNet and Crossformer are optimized using the Metis algorithm. In addition, half uniform initialization and Levy flight are added to the original HBA algorithm to make up for its shortcomings of slow optimization search speed and the tendency to fall into local optimal position, and the MHBA algorithm is obtained. Finally, the MHBA algorithm is used to weight the component prediction results of the two models, and compare the advantages and disadvantages of different weighting methods, and select the optimal weighting method to get the final AQI prediction results. The experimental results show that the STL-Metis-MHBA-TC model reduces RMSE, MAE, and MAPE by 19–34 %, 22–38 %, and 22–44 %, respectively, compared to the Transformer model. Therefore, the STL-Metis-MHBA-TC hybrid model proposed in this paper can effectively improve the AQI prediction accuracy.