Adaptive Weighting Ensemble Approach for High-Fidelity PM10 Forecasting via Evolutionary TCN and Bidirectional GRU

Rapid socioeconomic development has intensified air pollution, making accurate air quality forecasting essential for effective pollution management and public health protection. However, short-term PM10 prediction remains challenging because of complex temporal dependencies, nonlinear pollutant–meteorological interactions, and rapidly changing pollution episodes. To address these challenges, this study proposes the AWE-Bi-GRU-EA-TCN-LR hybrid framework, where Bi-GRU captures bidirectional temporal dependencies, EA-TCN extracts multi-scale temporal patterns through enhanced attention, and Linear Regression (LR) statistically calibrates residual relationships. The framework is evaluated using hourly air quality and meteorological observations from an industrialized urban area. As a consequence of the estimate, the RMSE value of 17.740 obtained with the comparison model, the persistence model, was decreased to 12.228 in the proposed model, while the R2 values were obtained as 0.786 and 0.899 in the same models, respectively. A number of different models were used in this research for the purpose of predicting PM10, and the Enhanced Temporal Convolutional Network (EA-TCN) model was found to have a root mean square error (RMSE) value of 12.804. Consequently, this proposed synergistic model has the capability to generate accurate results while maintaining robustness in air quality forecasting using complex datasets in regions with different environmental characteristics and meteorological variables. © 2026 The Author(s). Advanced Theory and Simulations published by Wiley-VCH GmbH.