基于注意力加强Bi-LSTM模型的TBM掘进参数预测研究

Study on Prediction of TBM Tunnelling Parameters Based on Attentionenhanced Bi-LSTM Model

使用Bi-LSTM模型预测TBM掘进参数时在模型训练用时与收敛速度方面存在不足。对传统注意力机制使用方式进行改进,提出了一种以并行融合方式形成的注意力加强的双向长短时记忆网络(Bi-LSTM+EMB_ATT)智能预测模型,并选取完整掘进循环数据预测TBM稳定段的掘进参数。采集吉林引松工程TBM3标段的运行数据并划分训练集和测试集,采用二值状态判别函数等方法预处理数据,利用皮尔逊分析结果选取完整掘进循环段21维掘进参数作为预测模型的输入,并对比分析并行连接的Bi-LSTM+EMB_ATT模型与串行连接的BiLSTM+ATT模型。结果表明,Bi-LSTM+EMB_ATT模型对TBM掘进参数的预测拟合优度均达0.91以上,平均绝对误差均小于2.7%,比Bi-LSTM+ATT模型的预测精度更高。

Prediction of TBM tunnelling parameters with Bi-LSTM model is subject to shortcomings in model training time and convergence speed. So, an intelligent prediction model of attention-enhanced bidirectional long-/shortterm memory network(Bi-LSTM+EMB_ATT) formed by parallel fusion was proposed by improving the use of traditional attention mechanisms, and the complete tunnelling cycle data was selected to predict the tunnelling parameters in the stable section of TBM. The operation data of TBM3 Section of Jilin Songhua River Diversion Project were collected and divided into a training set and a test set and preprocessed by the binary state discriminant function and other methods, and then the Pearson method was adopted to analyze the results. The 21-dimensional tunnelling parameters of the complete tunnelling cycle section was used as the input of the prediction model to compare and analyze the parallel-connected Bi-LSTM+EMB_ATT model and serial-connected Bi-LSTM+ATT model. The results show that the goodness of fit of Bi-LSTM+EMB_ATT model for predicting TBM tunnelling parameters is above0.91 with average absolute error less than 2.7%, which is higher than that of Bi-LSTM+ATT model.