Time-series Anomaly Detection has important applications, such as credit card fraud detection and machine fault detection. Anomaly detection based on the generative model generally detect samples with high reconstruction errors as anomalies. However, some anomalies may get low reconstruction errors, as they can also be well reconstructed due to the strong generalization ability of the model. To ensure the high reconstruction error of anomalies, we propose a novel anomaly detection algorithm named RAN (Reconstruct Anomalies to Normal) based on the Autoencoder. We try to force the reconstruction samples of both normal samples and anomaly samples obey the distribution of normal samples, then the difference between normal sample and its reconstruction sample is small while the difference between anomaly sample and its reconstruction sample is large, and higher reconstruction error for anomaly samples is guaranteed. The Autoencoder constructed by 1D-FCN with different kernel sizes is utilized to extract richer features of time-series data. Imitated anomaly samples are feed to the model to provide more information about anomalies. Then, constraints in the latent space and original data space are added to control the reconstruction process. Extensive experiments on real-life time-series datasets also show that RAN outperforms some state-of-art algorithms.