Medium-term prediction of atmospheric PM_2.5 concentration based on the VG-TCABI hybrid architecture.

Abstract

To address the limitations of existing PM_2.5 concentration prediction models in extracting temporal features and integrating multi-modal information, the VMD-GWO-TwoConvAttBiLSTM-IAFF (VG-TCABI) model is proposed in this paper. This framework employs decomposition driven, collaborative feature extraction, and adaptive fusion capabilities as the core components. First, the key parameters of Variational Mode Decomposition (VMD) are optimized by the Grey Wolf Optimizer (GWO), the non-stationary original PM_2.5 sequence is decomposed into multiple stationary sub-sequences with distinct physical meanings. Secondly, a dual-stream spatio-temporal analysis architecture is constructed. One channel primarily captures the external driving relationships among multiple pollutants. The other channel focuses on the internal evolution patterns of the sub-sequences decomposed by the VMD-GWO algorithm, synergistic exploration of internal and external features is achieved. Finally, an improved attention fusion mechanism (IAFF) is designed to dynamically calibrate and adaptively fuse heterogeneous features from two channels by learning gating weights. The pollutant and meteorological data from the main urban area of Dalian City in Liaoning Province is used, The proposed model outperforms all benchmark models in predicting the next 24-h PM_2.5 concentrations. The superior performance is evidenced by an R² of 0.892, an RMSE of 5.555 µg/m³, and an MAE of 8.279 µg/m³. This validates the effectiveness of the model in analyzing complex pollution causes and the robustness in capturing long-sequence trends.