Time series data is widely used in the Internet of Things (IoT) domain, such as equipment condition monitoring, environmental sensing, and industrial sensor data analysis. As a key task in time series analysis, time series anomaly detection is of great significance for achieving intelligent operation and maintenance, fault early warning, etc. Reconstruction-based methods using deep neural networks are one of the

current mainstream approaches. However, they still face several challenges in IoT scenarios: firstly, the fixed and limited input window size hinders the adequate capture of global contextual information; secondly, the model's insufficient ability to represent and memorize normal temporal features limits its anomaly detection capability; thirdly, relying solely on data reconstruction struggles to effectively improve the model's detection performance. To address these issues, we propose TPC-AD, an anomaly detection method based on Temporal Patch Embedding, enhanced Prototype Memory, and Contrastive Fusion. This approach enhances the model's ability to capture long-range dependencies through temporal patch embedding; explicitly learns and stores normal temporal patterns via a prototype memory mechanism, improving the modeling and memorization of normal behaviors; and actively separates normal and abnormal patterns in the feature space by incorporating contrastive learning, thereby significantly boosting detection performance. Experiments on four public IoT time series datasets demonstrate that our method achieves superior detection accuracy compared to state-of-the-art methods, with an average improvement of 2.93% in F1-score and 3.19% in AUC.