Machine learning for better prediction of seepage flow through embankment dams: Gaussian process regression versus SVR and RVM.

In the present study, three machine learning methods were applied for predicting seepage flow through embankment dams, namely (i) support vector regression (SVR), relevance vector machine (RVM), and Gaussian process regression (GPR). The three models were developed using seepage flow (Q: L/mn) and piezometer level (Z:m) measured at several piezometers placed in the corps body of the dam. The proposed models were calibrated and validated using a separate subset. Models evaluation and comparison was successfully achieved using various performances metrics, i.e., coefficient of correlation (R), Nash-Sutcliffe efficiency (NSE), root mean square error (RMSE), and mean absolute error (MAE). Experimental results showed that the proposed models are a good alternative to the in situ measured and contributed significantly in overcoming the case of missing measured seepage flow. The best performances were obtained using the RVM model with R and NSE values of ≈0.909 and ≈0.823, followed by the GPR model with R and NSE values of ≈0.891 and ≈0.767, while the SVR model was ranked as the poorest one exhibiting R and NSE values of ≈0.780 and ≈0.600, respectively. While, a growing number of investigations have focused on testing machine learning in terms of their feasibilities to accurately describe seepage flow, as well as providing important support to our understanding of the factors affecting its fluctuation, the present work was demonstrated that the combination of a wide range of variables can help in simulating seepage flow, and enhance their sensitivity which has help in developing new algorithms.