This paper proposes an innovative and lightweight deep learning model called Attention-Based Deep Convolutional Autoencoding Prediction Network (AT-DCAEP) for unsupervised anomaly detection in industrial multivariate time series data. The model consists of a characterization network based on convolutional autoencoders and a prediction network based on attention mechanisms. The characterization network extracts spatial features from multivariate time-series data and calculates reconstruction errors, while the prediction network captures time-dependent relationships in the reconstruction errors. Both subnetworks are jointly optimized to minimize reconstruction and prediction errors, significantly enhancing the model's anomaly detection performance. The AT-DCAEP model is tested on six publicly available datasets, demonstrating superior performance compared to current state-of-the-art methods. It achieves the optimal balance between anomaly detection performance and computational cost, with a 5.68% improvement in F1 score. The model's effectiveness is validated through extensive experiments, showing its ability to accurately detect anomalies in multivariate time series data. The model's contributions include the introduction of AT-DCAEP, which characterizes spatiotemporal patterns through simultaneous reconstruction and prediction analysis. It also incorporates multi-head attention between the convolutional encoder and decoder to focus on crucial information in low-dimensional space, enhancing the characterization network's reconstruction capabilities. Additionally, the model is benchmarked against state-of-the-art anomaly detection methods, demonstrating its superior performance. The model's performance is evaluated using precision, recall, ROC/AUC, and F1 scores. The results show that the AT-DCAEP model achieves a high F1 score across all datasets, with an average F1 score of 0.9191. The model's effectiveness is further validated through ablation studies, overhead analysis, and sensitivity analysis, demonstrating its robustness and efficiency in detecting anomalies in multivariate time series data. The model's ability to accurately detect anomalies in real-world scenarios is also illustrated through visualizations of the results on different datasets. The study concludes that the AT-DCAEP model is a significant advancement in the field of unsupervised anomaly detection for industrial multivariate time series data.This paper proposes an innovative and lightweight deep learning model called Attention-Based Deep Convolutional Autoencoding Prediction Network (AT-DCAEP) for unsupervised anomaly detection in industrial multivariate time series data. The model consists of a characterization network based on convolutional autoencoders and a prediction network based on attention mechanisms. The characterization network extracts spatial features from multivariate time-series data and calculates reconstruction errors, while the prediction network captures time-dependent relationships in the reconstruction errors. Both subnetworks are jointly optimized to minimize reconstruction and prediction errors, significantly enhancing the model's anomaly detection performance. The AT-DCAEP model is tested on six publicly available datasets, demonstrating superior performance compared to current state-of-the-art methods. It achieves the optimal balance between anomaly detection performance and computational cost, with a 5.68% improvement in F1 score. The model's effectiveness is validated through extensive experiments, showing its ability to accurately detect anomalies in multivariate time series data. The model's contributions include the introduction of AT-DCAEP, which characterizes spatiotemporal patterns through simultaneous reconstruction and prediction analysis. It also incorporates multi-head attention between the convolutional encoder and decoder to focus on crucial information in low-dimensional space, enhancing the characterization network's reconstruction capabilities. Additionally, the model is benchmarked against state-of-the-art anomaly detection methods, demonstrating its superior performance. The model's performance is evaluated using precision, recall, ROC/AUC, and F1 scores. The results show that the AT-DCAEP model achieves a high F1 score across all datasets, with an average F1 score of 0.9191. The model's effectiveness is further validated through ablation studies, overhead analysis, and sensitivity analysis, demonstrating its robustness and efficiency in detecting anomalies in multivariate time series data. The model's ability to accurately detect anomalies in real-world scenarios is also illustrated through visualizations of the results on different datasets. The study concludes that the AT-DCAEP model is a significant advancement in the field of unsupervised anomaly detection for industrial multivariate time series data.