Multivariate time series anomaly detection is of great significance in monitoring and ensuring the stable operation of complex systems. The multivariate time series generated in real scenes often have complex dependency patterns, summarized as temporal dynamics and spatial dynamics. Specifically, temporal dynamics manifested as complex dependencies between values at different timestamps, while spatial dynamics refer to uncertain similarity relationships between different sequences. In order to simultaneously model temporal and spatial dynamics, we propose a variational autoencoder based automatic clustering method for multivariate time series anomaly detection (ACVAE), which maps input sequences to latent representations using VAE and reconstructs input sequences based on the latent representations, while detecting anomalies based on reconstruction errors. Specifically, we design an encoder network that combines TCN and GRU to learn multi-scale long short-term temporal dependencies, and introduce the Dirichlet prior to automatically capture the similarity between sequences. Finally, we conduct extensive experiments on two publicly available datasets, and the results show that the ACVAE is superior to other baseline methods.