Air Quality Index Forecasting Based on a CEEMDAN-KPCA-TCN-BiGRU-AMRC Hybrid Model
TL;DR
Experimental results show that the proposed model outperforms the comparative model in terms of RMSE, MAE, MAPE, and R², and exhibits better tracking capabilities during local peaks, drastic fluctuations, and medium-to-high pollution stages.
Abstract
: To address the nonlinear, non-stationary, and abrupt fluctuations in the Air Quality Index (AQI) series, influenced by pollutant emissions, meteorological conditions, and random disturbances, this paper proposes a hybrid prediction model integrating Adaptive Noisy Complete Ensemble Empirical Mode Decomposition (CEEMDAN), Kernel Principal Component Analysis (KPCA), Temporal Convolutional Network (TCN), Bidirectional Gated Recurrent Unit (BiGRU), and Adaptive Markov Residual Correction (AMRC). First, CEEMDAN is used to decompose the original AQI series at multiple scales to separate high-frequency disturbances, periodic fluctuations, and long-term trend information. Second, KPCA is employed to extract nonlinear features from pollutants and meteorological variables, reducing input redundancy and enhancing the representation of key features. Subsequently, a TCN-BiGRU prediction network is constructed, extracting local temporal patterns through a temporal convolutional structure and learning sequence dependencies using a bidirectional recurrent structure. Finally, based on the state transition characteristics of the prediction residuals, an AMRC mechanism is introduced to dynamically correct the initial prediction results. Experimental results show that the proposed model outperforms the comparative model in terms of RMSE, MAE, MAPE, and R², and exhibits better tracking capabilities during local peaks, drastic fluctuations, and medium-to-high pollution stages. Residual distribution, autocorrelation analysis, and state transition heatmaps further validate the corrective effect of AMRC on systematic errors.