Long-term exposure to air pollution will pose a serious threat to human health. Accurate prediction can help people reduce exposure risks and promote environmental pollution control. However, most previous studies have ignored the spatial spillover of air pollution, i.e., that the current region’s air quality is also correlated with that of geographically adjacent areas. Therefore, this paper proposes an innovative spatiotemporal hybrid early warning system based on adaptive feature extraction and improved fuzzy cognition maps. Firstly, a spatial spillover analysis model based on the Moran index and local gravitational clustering was proposed to capture the diffusion and concentration characteristics of air pollution between regions. Then, an adaptive feature extraction model based on an optimized Hampel filter was put forward to process and correct the outliers in the original series. Finally, a hesitant fuzzy information optimized fuzzy cognitive maps model was proposed to forecast the air quality of urban agglomeration. The experimental results show that the air quality forecasting accuracy of urban agglomerations can be significantly improved when the geographical conditions and other interactions among cities are comprehensively considered, and the proposed model outperformed other benchmarks and can be used as a powerful analytical tool during urban agglomeration air quality management.

Air pollution forecasting plays a pivotal role in environmental governance, so a large number of scholars have devoted themselves to the study of air pollution forecasting models. Although numerous studies have focused on this field, they failed to consider fully the linear feature, non-linear feature, and fuzzy features contained in the original series. To fill this gap, a new combined system is built to consider features in the original series and accurately forecast PM_2.5 concentration, which incorporates an efficient data decomposition strategy to extract the primary features of the PM_2.5 concentration series and remove the noise component, and five forecasting models selected from three types of models to obtain the preliminary forecasting results, and a multi-objective optimization algorithm to combine the prediction results to produce the final prediction values. Empirical studies results indicated that in terms of RMSE the developed combined system achieves 0.652 6%, 0.810 1%, and 0.775 0% in three study cities, respectively. Compared to other prediction models, the RMSE improved by 60% on average in the study cities.