Exploring the role of artificial intelligence in predicting piezometric levels in homogeneous earth-fill dams

Seepage flow through the body of an earth-fill dam, directly reflected by the piezometric level, is the most important parameter in the structure's design and during the operational period. Moreover, if an abnormal seepage is not detected and addressed on time, it will probably result in dam failure. As an alternative to physical-based models, artificial intelligence (AI) tools, known as soft computing techniques, have proven, in the recent decades, their effectiveness in dam engineering. This study explores the application of three AI techniques to predict piozometric levels in the body of a homogeneous earth-fill dam. The developed models are feed-forward neural network (FFNN), radial basis function (RBF) network and support vector machine (SVM) model. The monitoring system of the studied earth-fill dam has provided 528 data instances wich are used for model calibration and validation. The delayed effects of the reservoir water level fluctuations on the seepage phenomenon were investigated to achieve the best FFNN architecture and the optimal input parameters were then used to feed the RBF network and the SVM model. Statistical fitting methods such as average relative variance (ARV), correlation coefficient (R), scatter plots, and Taylor diagram were used to evaluate the results. The findings indicated that the FFNN model is more accurate than the RBF and the SVM models. Specifically, during the test set, the FFNN has an ARV of 0.094, while the RBF and the SVM have respectively 0.462 and 0.243. This study highlights the potential of such models to allow continuous simulation of dam's behaviour and to prevent dam failures.