Aiming at the characteristics of large volume, rapid generation and concept drift of current stream data, a stream data anomaly detection method based on Long Short-Term Memory (LSTM) network and sliding window was proposed. Firstly, the LSTM network was used for data prediction, and the difference between the predicted value and the actual value was calculated. For each datum, the appropriate sliding window was selected, and the distribution modeling was performed to all the differences in the sliding window interval, then the probability of data anomaly was calculated according to the probability density of each difference in the current distribution. The LSTM network was not only able to predict data, but also able to predict and learn at the same time, as well as update and adjust the network in real time to ensure the validity of the model. The use of sliding windows was able to make the allocation of abnormal scores more reasonable. Finally, the simulation data made on the basis of real data were used for experiment. The experimental results verify that the average Area Under Curve (AUC) value of the proposed method in low-noise environment is 0.187 and 0.05 higher than that of direct difference detection and Abnormal data Distribution Modeling (ADM) method, respectively.