Air pollution directly affects people’s life and work and is an important factor affecting public health. An accurate prediction of air pollution can provide a credible foundation for determining the social activities of individuals. Scholars have, thus, proposed a variety of models and techniques for predicting air pollution. However, most of these studies are focused on the prediction of individual pollution factors and perform poorly when multiple pollutants need to be predicted. This paper offers a DW-CAE model that may strike a balance between overall accuracy and local univariate prediction accuracy in order to observe the trend of air pollution more comprehensively. The model combines deep learning and signal processing techniques by employing discrete wavelet transform to obtain the high and low-frequency features of the target sequence, designing a feature extraction module to capture the relationship between the variables, and feeding the resulting feature matrix to an LSTM-based autoencoder for prediction. The DW-CAE model was used to make predictions on the Beijing PM2.5 dataset and the Yining air pollution dataset, and its prediction accuracy was compared to that of eight baseline models, such as LSTM, IMV-Full, and DARNN. The evaluation results indicate that the proposed DW-CAE model is more accurate than other baseline models at predicting single and multiple pollution factors, and the R2 of each variable is all higher than 93% for the overall prediction of the six air pollutants. This demonstrates the efficacy of the DW-CAE model, which can give technical and theoretical assistance for the forecast, prevention, and control of overall air pollution.