A novel deep learning model for air quality index prediction integrating time series decomposition and intelligent optimization

Air quality forecasting is indispensable for effective atmospheric pollution control and environmental management. Nevertheless, achieving high accuracy and robustness remains challenging due to numerous influencing factors and the complex, nonlinear dynamics of air quality time series. To address these issues, we introduce STL-TimesNet-iTransformer-CPO, a novel hybrid framework that fuses time-series decomposition, deep learning architectures, and intelligent optimization. First, we apply Seasonal-Trend decomposition using Loess (STL) to partition the AQI series into trend, seasonal, and residual components. We then leverage the TimesNet model to capture long-term and periodic patterns within the trend and seasonal series, while employing an improved Transformer (iTransformer) to model the short-term, high-frequency residuals. To further enhance predictive performance and generalization, we integrate the Crested Porcupine Optimizer (CPO) for adaptive tuning of key hyperparameters across all submodels, ensuring coordinated model performance. The final AQI forecast is reconstructed by aggregating the individual component predictions. We rigorously evaluate our approach through extensive experiments, including varied train–test splits, comparisons with state-of-the-art baselines, component ablation studies, and cross-city transfer tests. Empirical results demonstrate that STL-TimesNet-iTransformer-CPO consistently outperforms existing methods in both single-step and multi-step forecasting tasks, offering a robust and accurate tool for regional air quality prediction and decision support.