Nowadays, time series data is more and more likely to appear in various real-world systems, such as power plants, medical care, etc. In these systems, time series anomaly detection is necessary, which involves predictive maintenance, intrusion detection, anti-fraud, cloud platform monitoring and management, etc. Generally, the anomaly detection of time series is regarded as an unsupervised learning problem. However, in a real scenario, in addition to a large set of unlabeled data, there is usually a small set of available labeled data, such as normal or abnormal data sets labeled by experts. Only a few methods use labeled data, and the existing semi-supervised algorithms are not yet suitable for the field of time series anomaly detection. In this work, we propose a semi-supervised time series anomaly detection model based on LSTM autoencoder. We improve the loss function of the LSTM autoencoder so that it can be affected by unlabeled data and labeled data at the same time, and learn the distribution of unlabeled data and labeled data at the same time by minimizing the loss function. In a large number of experiments on the Yahoo! Webscope S5 and NAB data sets, we compared the performance of the unsupervised model and the semi-supervised model of the same network framework to prove that the performance of the semi-supervised model is improved compared to the unsupervised model.