This paper introduces the Attention-Based Deep Convolutional Autoencoding Prediction Network (AT-DCAEP), a novel and lightweight deep learning model for unsupervised anomaly detection in multivariate time series data. The AT-DCAEP consists of two main components: a characterization network based on convolutional autoencoders and a prediction network based on attention mechanisms. The characterization network extracts low-dimensional spatial features from the time series data, while the prediction network captures time-dependent relationships using external attention. The model is designed to handle the challenges of large-scale data annotation and complex variable relationships in industrial settings. Extensive experiments on six publicly available datasets demonstrate that the AT-DCAEP outperforms state-of-the-art methods in terms of both anomaly detection performance and computational efficiency, achieving a 5.68% improvement in F1 score. The paper also includes ablation studies, overhead analysis, and sensitivity analysis to validate the effectiveness and robustness of the proposed model. Future work will focus on enhancing spatiotemporal feature coupling and developing methods suitable for edge devices.This paper introduces the Attention-Based Deep Convolutional Autoencoding Prediction Network (AT-DCAEP), a novel and lightweight deep learning model for unsupervised anomaly detection in multivariate time series data. The AT-DCAEP consists of two main components: a characterization network based on convolutional autoencoders and a prediction network based on attention mechanisms. The characterization network extracts low-dimensional spatial features from the time series data, while the prediction network captures time-dependent relationships using external attention. The model is designed to handle the challenges of large-scale data annotation and complex variable relationships in industrial settings. Extensive experiments on six publicly available datasets demonstrate that the AT-DCAEP outperforms state-of-the-art methods in terms of both anomaly detection performance and computational efficiency, achieving a 5.68% improvement in F1 score. The paper also includes ablation studies, overhead analysis, and sensitivity analysis to validate the effectiveness and robustness of the proposed model. Future work will focus on enhancing spatiotemporal feature coupling and developing methods suitable for edge devices.