A Hybrid Framework for Air Pollution Forecasting using DSES, DLA, Self-Attention Transformer, and BiGRU with Bayesian Optimization

Abstract This paper presents a novel hybrid forecasting framework, DSES-DLA-Transformers-BiGRU-BO, for accurate and interpretable air pollution prediction by synergistically capturing temporal dynamics and seasonal variations within multivariate time series. To be specific, this study first incorporates Dual Seasonal Exponential Smoothing (DSES) to extract and smooth the underlying trend along with dual seasonal components, inspired by the seasonal nature of environmental signals. Second, the proposed Dual-Level Attention Mechanism (DLA) simultaneously performs feature-level attention to identify influential environmental variables and temporal-level attention to capture critical historical dependencies in air pollution sequences. These enriched representations are then processed using a Bidirectional Gated Recurrent Unit (BiGRU) architecture to learn bidirectional sequence patterns, thereby enabling the model to understand both historical context and future trend. Third, Bayesian Optimization (BO) is employed to adaptively fine-tune the model’s hyperparameters, enhancing predictive robustness and performance across different time scales. Extensive experiments conducted on real-world air pollution dataset confirm the superior forecasting accuracy of the proposed DSES-DLA-Transformers-BiGRU-BO model, outperforming state-of-the-art statistical and deep learning counterparts by notable margins across various metrics, including MAE, MSE, RMSE, MAPE, DA, WLR, and TUS. Additionally, the proposed framework demonstrates computationally feasible hybrid forecasting architecture and generalization capability, offering a scalable and effective solution for environmental monitoring systems.