PM2.5 is an important air pollution index, which has been widely concerned. An excellent PM2.5 prediction system can effectively help people protect their respiratory tract from injury. However, due to the strong uncertainty of PM2.5 data, the accuracy of traditional point prediction and interval prediction method is not satisfactory, especially for interval prediction, which is usually difficult to achieve the expected interval coverage (PINC). In order to solve the above problems, a new hybrid PM2.5 prediction system is proposed, which can quantify the certainty and uncertainty of future PM2.5 at the same time. For point prediction, a multi-strategy improved multi-objective crystal algorithm (IMOCRY) is proposed; the chaotic mapping and screening operator are added to make the algorithm more suitable for practical application. At the same time, the combined neural network based on unconstrained weighting method further improves the point prediction accuracy. For interval prediction, a new strategy is proposed, which uses the combination of fuzzy information granulation and variational mode decomposition to process the data. The high-frequency components are extracted by the VMD method, and then quantified by FIG method. By this way, the fuzzy interval prediction results with high coverage and low interval width are obtained. Through 4 groups of experiments and 2 groups of discussions, the advanced nature, accuracy, generalization, and fuzzy prediction ability of the prediction system are all satisfactory, which verified the effect of the system in practical application.