Accurate prediction of PM2.5 concentration is crucial for public health and environmental protection. This paper develops a novel forecasting model that combines optimized signal decomposition with multi-objective feature selection techniques and error correction to enhance the accuracy of PM2.5 concentration predictions. Initially, the RIME algorithm is employed to precisely set the parameters of Variational Mode Decomposition (VMD), which decomposes the raw PM2.5 data into high, medium, and low-frequency components based on sample entropy values. Subsequently, a multi-objective feature selection approach is utilized to identify key feature subsets that significantly influence each frequency domain component. Finally, an optimized Informer model is deployed for comprehensive forecasting, complemented by an error correction mechanism to obtain the final PM2.5 concentration predictions. Experimental results indicate that the optimized decomposition effectively extracts key information from the data, reducing prediction complexity. The multi-objective feature selection approach provides superior identification of feature subsets compared to traditional single-objective methods. The enhanced Informer model, coupled with error correction, significantly improves the model’s accuracy and robustness.