Development of penetration rate prediction models for hard rock TBM in construction phase by deep learning and block model techniques: A case study in Mae Tang-Mae Ngad Tunnel, Northern Thailand

The prediction models are proposed to utilize through multi stages in a project such as, preliminary and feasibility state, design, tender, and construction. As a result, its low accuracy exceeds the expected accuracy range (The association for the advancement of cost engineering, -10% to +15%). For example, when entering the input data from Mae Tang - Mae Ngad Project through existing models, the values of root mean square error that are evaluated by comparing between the actual and the predicted rates of penetration have its results between 0.356 and 0.893. The errors show that efficiency of prediction model remains as same level as the penetration rates of standard deviations of actual rate of penetration which equal to 0.323 and 1.238 m/h. This research aims to create a model for predicting tunnel boring machine performance through deep learning in construction phase by Deep Feed Forward and Long-Short Term Memory and applied block models with mining technique. The prediction model is created and validated using data gathered from Mae Tang - Mae Ngad project over 300 datasets from ring number 2350 to 2649. Input parameters for the model are selected through the correlation examination. Parameters that correlated with 5 rate of penetration variables are Thrust, Torque, Cutterhead Speed, Rock Mass Type, and Uniaxial Compressive Strength. UCS value is determined by Indirect test from Schmidt hammer at Tail Shield and Invert Distance technique filling missing value. Rate of penetration in validation set resulted its root mean square errors as 0.162 and 0.216 from Deep Feed Forward and Long-Short Term Memory techniques respectively. Deep Feed Forword is more accurate whereas Long-Short Term Memory is more adaptive to fluctuate geological conditions and nonassessable input parameters in advance as it is recurrent neural network of deep learning, which also usable to predict previous datasets. Results from the prediction model development can lead to further Rate of penetration prediction framework development for other situation of tunnel excavation in the future.