Enhancing Air Quality Prediction Through Holt–Winters Smoothing and Transformer-BiGRU With Bayesian Optimization

Abstract

This paper proposes a hybrid forecasting framework, HWES-TBiGRU-BO, designed for accurate and efficient air pollution prediction by seamlessly integrating statistical smoothing techniques with advanced Deep Learning (DL) architectures. The model begins with Holt–Winters Exponential Smoothing (HWES) to decompose and smooth the time series, effectively capturing level, trend, and seasonal components inherent in environmental data. These seasonally-adjusted signals are then passed through a Transformer-based Bidirectional Gated Recurrent Unit (TBiGRU) architecture, which leverages the Transformer’s Attention Mechanism (AM) to model long-range temporal dependencies while the BiGRU simultaneously learns bidirectional sequence patterns to preserve both past and future contextual information. To further enhance performance, Bayesian Optimization (BO) is employed to automatically tune critical hyperparameters, ensuring optimal predictive capability without manual intervention. Experiments conducted on a real-world air pollution dataset demonstrate that the proposed HWES-TBiGRU-BO model consistently outperforms conventional statistical approaches and state-of-the-art deep learning baselines across multiple evaluation metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Coefficient of Determination (R2) , Directional Accuracy (DA), Win/Loss Ratio (WLR), and Theil’s U-Statistic (TUS). Moreover, the model achieves competitive computational efficiency and robust generalization, making it a scalable and reliable solution for real-time environmental monitoring and air quality forecasting.