Success and challenges in predicting TBM penetration rate using recurrent neural networks. (December 2022)

• Record Type: Journal Article
• Authors: Shan, Feng; He, Xuzhen; Jahed Armaghani, Danial; Zhang, Pin; Sheng, Daichao
• Is Part Of: Tunnelling and underground space technology. Volume 130(2022)
• Publication Date: 2022-12
• Subjects: Tunnel boring machine -- Penetration rate -- Time series forecasting -- Recurrent neural network -- Long short-term memory -- Random walk; TBM tunnel boring machine -- RNN recurrent neural network -- LSTM long short-term memory -- R2 coefficient of determination -- MSE mean squared error -- RMSE root mean squared error -- MAPE mean average percentage error; Tunneling -- Periodicals; Underground construction -- Periodicals; Tunnels -- Periodicals; Underground areas -- Periodicals; 624.193
• DOI: 10.1016/j.tust.2022.104728
• Languages: English
• ISSNs: 0886-7798

Abstract

Highlights: A framework for forecasting tunnel boring machine performance is presented. 23 tests are conducted to study-one-step forecast and multi-step forecast. Recurrent neural networks perform well in time series forecasting. The time lag problem and effects of time horizon are discussed. Abstract: Tunnel Boring Machines (TBMs) have been increasingly used in tunnelling projects. Forecasting future TBM performance would be desirable for project time management and cost control. We aim to use recurrent neural networks to predict the near future TBM penetration rate from historical data. Our datasets are composed of Changsha and Zhengzhou metro lines, with totally different geological conditions. In the experiments, the one-step forecast of TBM penetration rate by the traditional recurrent neural network (RNN) or long short-term memory (LSTM) is relatively accurate, irrespective of the different geological conditions used in training and evaluation. Predicting the next N th step penetration rate proves to be more challenging and depends on the time to the future or the distance ahead of the TBM cutterhead. There are generally time lags between measured and predicted results. The recursive RNN is then developed to address the lag problems, but to no avail. Alternative methods for predicting future penetration rates are studied, including the penetration rate at the N th step in the future and the average penetration rate of the next N steps, with the latter being trained by long-input or short-input methods. The average N -step forecast using short inputs provides the best results, and its performance over other alternatives becomes more distinct as the number N increases. We also discuss the possibility of the forecast problem as a quasi-random walk, which means that forecasting penetration rate cannot easily be achieved using low-frequency data with RNNs, and that the accuracy depends on the correlation between the last and predicted steps in the data.