Advanced Prediction Method for Shield Tunneling Cutterhead Torque Based on WaveNet Network

Objective Cutterhead torque is a crucial parameter that characterizes the safety of shield tunneling and the operating status of equipment. To address the difficulties in cutterhead torque prediction and excavation parameters timely correction, an advanced prediction method for shield tunneling cutterhead torque based on WaveNet network is proposed. Method The preprocessing method for working condition data is introduced, and the basic structure and construction method of the initial static model based on WaveNet network are proposed. A training set is constructed by extracting historical data of shield construction monitoring within the initial 50 m tunneling distance, and advanced prediction of the cutterhead torque is made after five construction steps based on the shield construction monitoring data of the previous 20 construction steps. With the increase of shield excavation distance, the model is retrained and updated every five construction steps using newly generated data set, thus a long-term dynamic model for cutterhead torque advanced prediction is proposed. Taking the left-line data of the Yufuhe Sta.-Wangfuzhuang Sta. shield tunnel interval of Jinan Rail Transit Line 1 as example, the prediction effect of cutterhead torque is analyzed and verified. Result & Conclusion The cutterhead torque values by advanced prediction for the first 50 m tunneling distance show a basic consistency with the changing trend of actual values, with an average relative error of 10.07%. The initial static model exhibits relatively high prediction accuracy. As the tunneling distance increases, the relative error of the initial static model increases from 10% to about 30%, while that of the continuously updated long-term dynamic model remains stable at around 10%. The update time of the long-term dynamic model is generally distributed between 1 and 6 seconds, with an average time consumption of 3.92 s, meeting the requirement for efficient dynamic model updates.