With the development of urbanization and industrialization, the severity of air pollution has increased, among which fine particulate matter (PM2.5) is one of the most concerned air pollution factors. However, the strong nonlinearity and non-stationarity of PM2.5 series makes accurate prediction challenging. To improve the prediction accuracy of PM2.5 concentration, we propose a long-short-term memory neural network (LSTM) model based on residual connection and Transformer encoder improvement, referred to as RT-LSTM. Firstly, we employ a time-series multiscale data augmentation (AUG) strategy to enhance the time-domain diversity and convert the 1D time-series into 2D tensor ensembles based on multiple cycles. New dimensions are provided for the time series. Secondly, the RT-LSTM model proposed in this paper is used to predict PM2.5 concentration. Finally, the coati optimization algorithm (COA) was used to optimize the parameters of the model. The experimental results show that the proposed model significantly outperforms the existing comparative models, which fully validates the effective enhancement of the multidimensional approach to the representation of sequence features, and also demonstrates the significant role of the RT-LSTM model in improving prediction accuracy. The proposed model exhibits excellent accuracy, provides reliable air quality information for the society.