Optimising tunnel support design with machine learning models

Abstract

The Q-system is one of the broad techniques used in tunnel design and aids in the determination of tunnel support – a crucial aspect for safety and stability in tunnel engineering. It is complex, costly, and time-consuming to acquire all the necessary Q-system characteristics. This study predicts the Q value using parameters that have the largest coefficient of relevance in the value of Q and determines the most important Q-system parameters. The predictions of the models are correlated with the actual Q values using the marginal histograms for training, testing and validation of the datasets. The histogram of the ANN and GB models is closer to that of the measured Q for training, while the RF model is a bit different from the actual Q. The imposed normal distribution curves of the ANN and GB are also closer to that of the actual Q, while RF shows a much more curved cone. These observations account for the high R² values of 0.9992 and 0.9998 obtained for the ANN and GB models for training, while an R² of 0.9716 is observed for the RF model. The histograms of the ANN models are the closest resemblance to the actual histogram of Q, followed by those of the GB and then RF for testing and validation. The ANN models have the highest R² values for the testing and validation of the dataset, which can be attributed to the closeness of their histograms to the actual Q. The ANN model performs better than the other ensemble models, demonstrating its superiority in predicting rockmass quality. The Taylor diagram displays the prediction efficacy of the three proposed models by using the testing and validation datasets, and confirms that the ANN predictive models are the closest to the actual Q values.