Anomaly detection in multivariate time series is critical for Internet of Things production line control systems, and distributed energy resource management systems. These systems generate high-dimensional sequential data characterized by intricate spatiotemporal dependencies, non-stationarity, and nonlinearity. Existing methods often fell short of expectations on the real-world data, demonstrating low anomaly detection accuracy and high false-alarm rates. To address these challenges, this study proposes Dynamic Spatio-temporal Collaborative network (DSCnet) for Multivariate Time Series Anomaly Detection. DSCnet leverages a Transformer architecture enhanced with RBF neurons to extract temporal patterns from multivariate sequences. Meanwhile, it dynamically infers the spatial dependencies between entities through graph structure learning enhanced by multi-head attention, to capture the complex feature interactions. Our collaborative computing model fuses the temporal features by the dynamic spatial dependency from the graph learning module with entity-aware normalization. Furthermore, we calculate the anomaly score by combining both the reconstruction error and the RBF-based similarity metrics from the Transformer-based temporal feature extraction architecture. Evaluations on four public multivariate time-series benchmarks demonstrate the superiority of our DSCnet with an average AUROC of 89.8%, and a 5.47% improvement over the state-of-the-art models.