Machine Learning tools applied to the prediction and interpretation of the structural behavior of existing dams

The safety of existing dams is mainly ensured by the correct interpretation of monitoring data recorded during the whole lifetime of these structures. In this context, an increasing number of devices are being installed to provide more and more frequent measurements. Several Machine Learning tools have emerged as possible alternatives to traditional prediction approaches in recent years. Neural Networks have shown the ability to adapt to complex interactions and, therefore, to reach greater accuracy than conventional methods. However, this technique is susceptible to parameter tuning and difficult to generalize. Other recent studies have focused on Boosted Regression Trees. Less frequently used in dam engineering, they have proved to be equally accurate compared to Neural Networks, simpler to implement, and not sensitive to noisy and low relevant predictors. However, applications are limited to a few specific cases. The present contribution aims to evaluate the performances of this novel approach on dam data with a different specificity from previous research. The case study corresponds to a double-curvature arch dam introduced as a benchmark test by the International Commission on Large Dams. The input data include raw environmental variables, some derived variables, and time-related variables. Predictions of displacements under varying environmental conditions are performed, and relative influence indices are identified to determine the strength of each input-output relationship.