At present, most PM_2.5 concentration prediction models only use time series data from a single station for concentration prediction, but do not take into account the regional correlation among air quality monitoring stations. This will lead to a certain one-sidedness of the prediction. In this paper, the KNN algorithm was used to select the relevant spatial factors in the area where the target site is located. Combined with the LSTM model, a KNN-LSTM PM_2.5 concentration prediction model based on spatiotemporal features was proposed. The simulation experiments were performed on pollutant data from 10 air quality monitoring stations in Harbin, and the KNN-LSTM model was also compared with other prediction models. The results show that the model compared with the BP neural network model, Mean Absolute Error (MAE), Mean Square Root Error (RMSE) decrease by 19.25% and 13.23% respectively; compared with the LSTM model, MAE and RMSE decreased by 4.29% and 6.99% respectively. It shows that the KNN-LSTM model proposed in this study can effectively improve the prediction accuracy of LSTM model.