Research on a hybrid deep learning model based on two-stage decomposition and an improved whale optimization algorithm for air quality index prediction

The Air Quality Index (AQI) is a crucial metric for assessing the severity of air pollution, making accurate AQI forecasting essential for air quality management. A hybrid deep learning model is developed for AQI prediction, incorporating two-stage decomposition and hyperparameter optimization. The two-stage decomposition method first uses Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to initially decompose the AQI sequences, followed by Singular Spectrum Analysis (SSA) applied to the subsequence with the highest Weighted Permutation Entropy (WPE), and all subsequences are then reconstructed. In the feature modelling stage, to capture the spatial correlations among monitoring stations, stations are represented as graph nodes, and a Graph Attention Network (GAT) with residual layers is employed to extract spatiotemporal features from multi-site air pollution data. In the prediction stage, a hybrid model that integrates a Gated Recurrent Unit (GRU) and a self-attention mechanism is used to forecast AQI. The model's hyperparameters are optimized by the Improved Whale Optimization Algorithm (IWOA), which improves search efficacy by including chaotic mapping, a nonlinear shrinkage factor, and a Levy flight strategy. The experimental results show that the proposed model achieves superior performance in one-step-ahead AQI prediction, with a Root Mean Squared Error (RMSE) of 3.445, a Mean Absolute Percentage Error (MAPE) of 4.737%, a Coefficient of Determination (R2) of 0.993, a Mean Absolute Error (MAE) of 2.263, a Percentage Bias (PBIAS) of −0.511%, and a Willmott Index of Agreement (WI) of 0.998, outperforming other baseline models.