A Comparative Study of Supervised Machine Learning Techniques for Water Leakage Prediction in Drill and Blast Excavated Water Tunnels in the Himalayan Region

Abstract

In the Himalayan region, drill and blast (DB) method excavated water tunnels often pass through complex geological rock mass conditions which were formed with frequent tectonic movements. Due to these tectonic movements, the rock mass conditions in the Himalayas are highly faulted, folded, jointed, sheared, and fractured. The geological formations are the pathway for the water ingress and leakage out in the water tunnel. This water leakage in the tunnel causes a complicated geological hazard, which significantly increases tunnel instability and can lead to a delay in completion time and finally increase the cost of the tunnel project. Therefore, an efficient water ingress/leakage prediction model is essential to mitigate these challenges. In this research, various field data such as rock mass properties, topography, and permeability data sets were collected from the Nilgiri-II Hydroelectric project water tunnel. These real-time field datasets have been used for comprehensive assessment and to predict the water ingress/leakage in the water tunnel by using four supervised machine learning (ML) approaches such as Support Vector Regression (SVR), Decision Tree (DT) regression, K-Nearest Neighbors (KNN) and Random Forest (RF) regression models. It was observed that the KNN shows the best regression performance of 93% followed by RF of 92%, DT of 86%, and SVR of 66%. Therefore, all these machine learning approaches show good performance in predicting water ingress/leakage based on real field data except the SVR model.