Modern dynamic systems, such as transportation networks and IoT infrastructures, generate massive volumes of interrelated temporal data represented as temporal graphs. Conventional methods – like clustering, statistical thresholds, and classical time series analysis – often fail to account for the spatial-temporal dependencies inherent in these systems, leading to high false positive rates or missed complex anomalies. In this paper, we propose a novel anomaly detection approach that combines Temporal Graph Neural Networks (TGNN) with Autoencoders. The method utilizes TGNN to extract robust node representations by capturing both local connectivity and temporal evolution, while an autoencoder is trained to reconstruct normal node behavior. Anomalies are subsequently identified through significant reconstruction errors, which serve as indicators of deviations from typical patterns. Experimental evaluations on the real-world PeMSD7 dataset demonstrate that the proposed TGNN + Autoencoder method improves detection accuracy by 17.33% compared to traditional methods, reduces false positives by 4.71%, and achieves a 6.02% higher F1-score relative to using TGNN or autoencoder individually. These results underline the practical relevance of our approach for real-time monitoring of transportation networks, while also contributing theoretically to the integration of spatial and temporal features in anomaly detection.