基于工况知识引导注意力时间卷积网络的烧结终点预测

Burning through point prediction based on working condition knowledge guided attention mechanism temporal convolutional network

中文关键词  注意力机制  时间卷积网络  工况知识  烧结终点  预测

英文关键词  attention mechanism  temporal convolutional network  condition knowledge  burn through point  prediction

中文摘要

      烧结终点位置的实时准确预测对于优化烧结工艺具有重要的意义. 针对烧结过程中强非线性和动态时变 性造成烧结终点高精度预测难的问题, 本文提出了一种基于工况知识引导注意力时间卷积网络(AM–TCN)模型. 首 先, 构建堆叠的时间卷积模块用于充分提取烧结过程数据中深层次的非线性特征; 其次, 将历史工况知识引入注意 力机制, 引导模型在保留过程数据时序特征的同时区分不同特征的重要性; 最后, 构建预测模型用于烧结终点位置 在线预测. 工业数据实验表明, 所提AM–TCN模型具有较好的烧结终点预测精度, 对提升烧结过程热状态稳定性具 有重要意义.

英文摘要

      The precise and real-time prediction of the burning through point (BTP) is essential for optimizing process operation. However, due to the strong nonlinearity and dynamic time-varying characteristics of the sintering process, highprecision prediction of the BTP has been challenging. In this paper, an attention mechanism temporal convolutional network (AM–TCN) model is proposed based on the knowledge of working conditions. First, stacked temporal convolution blocks are developed to extract deep nonlinear features in the sintering process data. Second, the attention mechanism incorporates the knowledge of historical working conditions, allowing the model to identify the significance of various extracted features while preserving the time series features of the process data. Finally, an online BTP prediction model can be established. The results of experiments using industrial data illustrate that the proposed AM–TCN model has good BTP prediction accuracy, which is critical in improving the stability of the thermal state during the sintering process.