Abstract Precise forecasting of the Air Quality Index (AQI) is essential for environmental management and public health protection. However, the non-linear and non-stationary nature of AQI time series presents a significant challenge for traditional predictive models. Most current deep learning approaches still face limitations in feature extraction and rely on inefficient manual hyperparameter tuning. To address these constraints, this study proposes an integrated forecasting framework combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Convolutional Neural Networks (CNN), Bidirectional Gated Recurrent Units (BiGRU), and an Attention mechanism. The methodology begins by using CEEMDAN to decompose the complex AQI signals into multiple stable frequency components, which effectively reduces the impact of data noise. Each component is then processed by a hybrid sub-network where CNNs extract local features and BiGRU units capture long-term temporal dependencies. An attention layer is incorporated to dynamically assign weights to critical time steps. Furthermore, an Improved Grey Wolf Optimizer (IGWO) is introduced to automate the hyperparameter search, ensuring optimal network performance without manual intervention. Experimental results on a long-term dataset from Guangzhou (2014–2024) show that the proposed model achieves an MSE of 10.2456 and a coefficient of determination ( R 2 ) of 0.9615. These findings, supported by detailed ablation studies and cross-city generalization tests, demonstrate that the model is both robust and accurate for real-world air quality early-warning systems.
Fang et al. (Fri,) studied this question.