In order to solve the contradiction that only the local data can be optimized using traditional algorithms, an abnormal data processing algorithm based on long short-term memory network (LSTM) deep learning model proposed, and then the neural network model is built based on tensor flow framework, so as to obtain alternative data sets and calibration data sets. Through the comparison of sample labels, abnormal points are determined, the data is replaced and the standard samples are updated. To verify the effectiveness of the algorithm, the K2077 section traffic data of Kunming-Shantou expressway was used for checking and analysis. The results show that the LSTM model can quickly process the original traffic time series data and detect and correct algorithm. It can also optimize the data quality and makes up for the limitations of traditional algorithms. The establishment of the working window simplifies the algorithm flow and improves the accuracy of data correction.