Traditional PM2.5 point value prediction would lose the fluctuation information of concentration value, and thus could not adequately represent and estimate the range of its fluctuation and change. A multiscale combination prediction method for PM2.5 range is proposed based on long shortterm memory (LSTM), hybrid fuzzy time series (HFTS) and error correction (EC). Based on deep learning and interval multiscale decomposition method, the combined prediction model of interval time series is established by further considering the effective information hidden in the prediction error. This model can extract complex data features from time series with large randomness, and solve the problems of lag existing in traditional forecasting methods and insufficient use of error information. Finally, the empirical analysis shows that this method is suitable for the prediction of PM2.5 range with large fluctuation, and has higher accuracy and good applicability by comparing with the existing methods.