PM2.5 presents substantial health hazards to individuals, underscoring the critical importance of precise PM2.5 concentration predictions. This paper proposed a groundbreaking hybrid deep learning model designed for long-term PM2.5 forecasting. This model utilizes singular spectrum analysis and kernel principal component analysis to effectively remove noise, extract essential features, and reduce the dimensionality of air quality data. These techniques are seamlessly integrated with the Informer model to enable accurate long-term forecasting of PM2.5 concentrations. To evaluate the effectiveness of our prediction model, we employ three key metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination. Upon analyzing experimental data for PM2.5 levels over the next 24 hours, it becomes clear that our model surpasses other models noticeably. Specifically, our model achieves a coefficient of determination value of at least 0.2142 higher than that of the comparison models, reduces RMSE by at least 2.1991 compared to the comparison models, and lowers MAE by at least 1.4868 compared to the comparison models.