基于可变时序移位Transformer-LSTM的集成学习矿压预测方法  

Ensemble learning mine pressure prediction method based on variable time series shift Transformer-LSTM

作　　者：李泽西 LI Zexi(College of Communication and Information Technology,Xi'an University of Science and Technology,Xi'an 710054,China)

机构地区：[1]西安科技大学通信与信息工程学院,陕西西安710054

出　　处：《工矿自动化》2023年第7期92-98,共7页Journal Of Mine Automation

摘　　要：现有的矿压预测模型多为依赖固定长度时序序列特征的单一预测模型,难以准确捕捉矿压时序数据的复合特征,影响矿压预测的准确度。针对该问题,提出一种基于可变时序移位Transformer-长短时记忆(LSTM)的集成学习矿压预测方法。基于拉依达准则和拉格朗日插值法,剔除矿压监测数据中的异常值,插入缺失值,并进行归一化预处理;提出可变时序移位策略,划分不同尺度的矿压时序数据,避免固定长度时序序列可能存在的数据偏移问题;在此基础上,构建基于Transformer-LSTM的集成学习矿压预测模型,通过结合注意力机制和准确的时序特征表示能力,多层次捕捉矿压变化规律的动态特征,采用集成学习的投票算法,联合预测矿压数据,克服单一预测模型的局限性。实验结果表明:采用集成学习的投票算法可降低矿压预测平均绝对误差(MAE)和均方根误差(RMSE)的波动性,有效减小不同尺度特征序列对矿压预测结果的敏感性影响;Transformer-LSTM模型在2个综采工作面顶板矿压数据集上预测结果的MAE较Transformer模型分别提高了8.9%和9.5%,RMSE分别提高了12.7%和16.5%,且高于反向传播(BP)神经网络模型和LSTM模型,有效提升了矿压预测准确度。The existing mine pressure prediction models are mostly single prediction models that rely on fixed length time series features.It is difficult to accurately capture the composite features of mine pressure time series data,which affects the accuracy of mine pressure prediction.To solve this problem,an ensemble learning mine pressure prediction method based on variable time series shift Transformer-long short-time memory(LSTM)is proposed.Based on the Laida criterion and Lagrange polynomial method,the outlier values in the mine pressure monitoring data are eliminated,and the missing values are inserted.Then normalized preprocessing is performed.The paper proposes a variable time series shift strategy to divide mine pressure time series data at different scales.It avoids potential data shift issues that may exist in fixed length time series.On this basis,the ensemble learning mine pressure prediction model based on Transformer-LSTM is constructed.By combining the attention mechanism and the accurate time series feature representation capability,the dynamic features of the mine pressure change law are captured at multiple levels.The voting algorithm of ensemble learning is used to jointly predict the mine pressure data to overcome the limitations of a single prediction model.The experimental results root mean square error(RMSE)of mine pressure prediction.It effectively reduces the sensitivity impact of different scale feature series to the mine pressure prediction results.Compared with the Transformer model,the MAE of the Transformer-LSTM model's prediction results on two roof mine pressure datasets of fully mechanized working faces improves by 8.9%and 9.5%respectively,and the RMSE has increased by 12.7%and 16.5%respectively.The above indexes are also higher than those of back propagation(BP)neural network model and LSTM model.The method proposed in the paper effectively improves the accuracy of mine pressure prediction.

关 键 词：矿压预测 可变时序移位 Transformer-LSTM模型 集成学习 投票算法 

分 类 号：TD323[矿业工程—矿井建设]