College of Mining Engineering, Taiyuan University of Technology

由于在重介分选过程中各个传感器位置不同，导致重介分选主要工艺参数与灰分存在时间滞后，影响了精煤灰分结果。基于回归模型的灰分预测方法缺乏对时间序列信息的利用，无法捕捉重介生产过程随时间变化的动态特性；基于时间序列的灰分预测方法未能充分考虑灰分和重介分选主要工艺参数之间的时间依赖关系。针对上述问题，提出了一种基于时间序列对齐和TCNformer的重介精煤灰分多步预测方法。通过滞后相关性分析来量化灰分与重介分选主要工艺参数之间的滞后步长，依此对重介分选主要工艺参数在时间维度上进行移动，使得灰分和重介分选主要工艺参数时间序列对齐，消除灰分和重介分选主要工艺参数之间的时间滞后。在Transformer模型的基础上，引入时间卷积网络（TCN）提取特征，并将单向编码器扩展为双向编码器，构建了TCNformer模型来实现精煤灰分多步预测。将时间序列对齐得到的与未来时刻灰分数据对应的过程变量序列作为解码器的输入，以提升模型预测精度。实验结果表明：该方法的平均绝对误差为0.157 9%，均方根误差为0.215 2%，平均皮尔逊相关系数为0.505 1，能有效提升精煤灰分预测精度。

Due to the different positions of various sensors during the dense medium separation process, there is a time lag between the main process parameters of dense medium separation and ash content, which affects the results of clean coal ash content. The grey prediction method based on regression models lacks the utilization of time series information and cannot capture the dynamic features of the dense medium production process over time. The time series based ash prediction method fails to fully consider the time dependence relationship between the main process parameters of ash content and dense medium separation. In order to solve the above problems, a multi step prediction method for dense medium clean coal ash content based on time series alignment and TCNformer is proposed. The method quantifies the lag step between the main process parameters of ash content and dense medium separation through lag correlation analysis. The method moves the main process parameters of dense medium separation in the time dimension accordingly, aligning the time series of the main process parameters of ash content and dense medium separation, and eliminating the time lag between the main process parameters of ash content and dense medium separation. On the basis of the Transformer model, a time convolutional network (TCN) is introduced to extract features, and the unidirectional encoder is extended to a bidirectional encoder to construct the TCNformer model for multi-step prediction of clean coal ash content. The sequence of process variables corresponding to the grey data at future moments obtained from the time series alignment is used as an input to the decoder to improve the model prediction precision. The experimental results show that the average absolute error of this method is 0.157 9%, the root mean square error is 0.215 2%, and the average Pearson correlation coefficient is 0.505 1, which can effectively improve the precision of predicting clean coal ash content.

国家自然科学基金项目（52274157）；内蒙古自治区重点专项项目（2022EEDSKJXM010）；山西省重点研发计划项目（202102100401015）。

王珺，王然风，魏凯，等. 基于时间序列对齐和TCNformer的重介精煤灰分多步预测[J]. 工矿自动化，2024，50（5）：60-66.

WANG Jun, WANG Ranfeng, WEI Kai, et al. Multi step prediction of dense medium clean coal ash content based on time series alignment and TCNformer[J]. Journal of Mine Automation，2024，50（5）：60-66.