Abstract: Given that PM2.5 concentration is not only related to local pollutants, but also affected by long-distance transmission of PM2.5 in adjacent areas, the key to improving the prediction accuracy of PM2.5 concentration is to comprehensively consider the effect of local influencing factors and the transmission effect of PM2.5 in adjacent areas. For this purpose, a novel seasonal discrete multivariable grey prediction model, encompassing spatial influencing factors, has been established. Firstly, we analyze the mechanism of spatial influencing factors and the reasonableness of using PM2.5 concentration in adjacent areas as the spatial influencing factors. Secondly, based on the spatial agglomeration characteristics of PM2.5 concentration, the K − means clustering algorithm is used to cluster adjacent cities with similar PM2.5 concentration, then the comprehensive value of PM2.5 concentration in each city cluster is calculated by weighted average combination. On this basis, a driving term of spatial influencing factors and a cosine trigonometric function term are introduced into the novel model to characterize the effect of spatial influencing factors on PM2.5 concentration and the seasonal fluctuation of itself, respectively. More importantly, the Genetic Algorithm Toolbox is employed to optimally determine the emerging parameters of this model, and the time response function of the novel model is calculated by mathematical induction method. Lastly, the new model is deemed valid through testing its PM2.5 concentration predictions for the cities of Beijing, Tianjin, and Baoding in the Beijing-TianjinHebei region. Based on the original observations from 2018Q1 to 2023Q2, the novel model is built for PM2.5 concentration prediction in 2023Q3 to 2024Q2 for the three cities. The findings imply that the newly developed model outperforms its competitors significantly and has the potential to serve as a robust tool for predicting PM2.5 concentration.

Key words: Grey prediction, Spatial influencing factors, PM2.5 concentration, Spatial autocorrelation analysis