Abstract: This paper proposes a novel energy consumption prediction system based on digital twin technology, with the objective of addressing the issues of low visualization and high energy consumption in spinning. The system aims to enhance the accuracy of energy consumption predictions in spinning workshops through intelligent means. Given the multitude of factors that influence energy consumption in the spinning process, including the type of raw cotton and the spindle speed, precise prediction of energy consumption not only facilitates the optimization of production processes but also enables significant cost savings for enterprises. The system architecture proposed in this text is comprised of four layers: the physical layer, the perception layer, the twin layer, and the application layer. The physical layer refers to the actual physical equipment and environment, which serves as the foundation for all data collection. The perception layer involves various sensors and other data acquisition devices responsible for real-time collection of data from the physical layer. The twin layer establishes virtual models to simulate various variables and conditions in the actual production process, thereby creating a "digital twin" corresponding to the physical entity. The application layer is concerned with the process of supporting decision-making through data analysis and processing. 
In this system, the combination of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) is utilized to explore the potential correlations between different data dimensions in the spinning process. CNN is particularly adept at processing spatial features, such as pattern recognition in images, while BiLSTM is well-suited to the task of capturing dynamic changes in timeseries data. Through this combination, the model becomes more flexible and efficient in handling complex and variable production data. To further enhance the model's capacity for generalization, the system also introduces the Attention Mechanism. The mechanism enables the model to focus on the most salient aspects of the input data, thereby better understanding and processing complex data structures. Finally, the Fully Connected Layer is responsible for integrating the features extracted from the previous layers to output the final prediction result. The results demonstrate that this CNN-BiLSTM-Attention model exhibits extremely high accuracy in energy consumption prediction, reaching 98.21%, with prediction errors all controlled within 2.5%. In comparison to models such as CNN-BiLSTM, the new model demonstrates superior performance in both prediction accuracy and error control, suggesting that the incorporation of the Attention Mechanism can indeed significantly improve model performance. 
In conclusion, the energy consumption prediction system based on digital twin technology offers a novel approach to addressing the energy consumption prediction challenge in the conventional spinning process. It not only enhances the accuracy of energy consumption predictions but also offers powerful technical support for future production management, energy conservation, and emission reduction. This achievement is of great significance for promoting the development of the textile industry in a more intelligent and green direction.

Key words: digital twin, energy consumption prediction, convolutional neural network, bidirectional long short-term memory, attention mechanism

摘要： 针对纺纱生产过程可视化程度低、纺纱生产能耗较高的问题，以及能耗受原棉种类、锭子转速等多种因素影响的问题，提出一种基于数字孪生的纺纱车间能耗预测系统的构建方法，以提升能耗预测的精度。该系统涵盖数字孪生物理层、感知层、孪生层和应用层4部分，融合了卷积神经网络和双向长短时记忆网络，提取纺纱生产过程数据维度之间的特征关联，并增加注意力机制提升泛化能力，最后由全连接层输出预测结果。结果表明：CNN-BiLSTM-Attention模型的预测精度达到98.21%；与CNN-BiLSTM等模型相比，该模型的预测精度更高，误差更小，表现出一定的优越性和可靠性。

关键词: 数字孪生, 能耗预测, 卷积神经网络, 双向长短时记忆网络, 注意力机制