Use of an improved ANN model to predict collapse depth of thin and extremely thin layered rock strata during tunnelling

• A neural network model for the prediction of collapse depth was developed.
• A quantitative criterion and degradation rules of support strength was presented.
• A small initial weight range of ANNs which should be given priority was proposed.
• A principle for determination of termination conditions of ANNs was put forward.

Numerous collapses have occurred during the excavation of diversion tunnels in the thin and extremely thin layered rock strata at Wudongde Hydropower Station in China. Hence, a reliable method is required to predict the risk and the depth of collapse. However, both theory and practice indicate that one single criterion methods cannot satisfactorily predict the collapse depth accurately. In this study, using an artificial neural network (ANN), an intelligent prediction method has been investigated. Through theoretical and statistical analyses, six input parameters (i.e., cover depth, minor–major principal stress ratio, geological strength index, excavation method, support strength and attitude of rock), have been selected and used in the model. Obtained from three diversion tunnels at Wudongde Hydropower Station, forty-five learning samples and six testing samples were used in the training of the model. The structural parameters and the initial weights of the ANN have been optimized by a genetic algorithm (GA). The trained model was then used to predict the collapse depth of another six excavation sites. The predictions show good agreement with the measurements at the sites. The absolute errors between the predicted and the measured collapse depths are all less than 0.35 m, and the relative errors are all less than 15%. Application of the improved ANN method to the tunnel collapse analysis at Wudongde Hydropower Station confirms its effectiveness in predicting collapse depth during tunnelling.