To achieve accurate prediction of PM2.5 concentrations during heavy haze pollution events, the model of deep feature fusion (long short-term memory and multivariate linear regression, LSTM-MLR) was proposed in this paper to predict PM2.5 concentrations in 3h, 6h, 12h and 24h of Xi’an and to explain the relationship between the main meteorological factors precursors and haze concentration. In the proposed model, a series of long short-term memory (LSTM) with different hyper-parameter were used to extract the deep features of PM2.5 precursors and meteorological time series, and a full connect layer was used to adjust the dimensions of LSTM outputs. Then, a feature fusion model with the structure of multivariable linear regression (MLR) was used to obtain the predicted PM2.5 concentrations. Data from January 2015 to March 2020 in Xi’an were used to determine parameters of the proposed model, and model performances for PM2.5 concentration prediction in 3h, 6h, 12h and 24h were also evaluated. Results showed that LSTM-MLR, which was based on deep feature fusion of time series, can predict heavy haze pollution samples correctly with the accuracy of 94.12%, 85.29%, 77.57% and 51.10% in 3h, 6h, 12h and 24h, respectively. The true positive rate (TPR) of the proposed model outperforms random forest (RF), support vector regression (SVR), MLR, LSTM_PM2.5 (PM2.5 used only), multivariable LSTM (M_LSTM) and long short-term memory-random forest (LSTM-RF) which had the same input as LSTM-MLR. The fusion coefficients showed that the influence of current PM2.5 concentrations on that in the future decreased from 80.89% (t + 3) to 16.34% (t + 24), and the influence of precursor concentration increased from 5.23% (t + 3) to 29.43% (t + 24), which illustrated the importance of time for taking emergency measures to reduce the peak value and decrease the pollution duration.