Modeling of Seepage Flow Through Concrete Face Rockfill and Embankment Dams Using Three Heuristic Artificial Intelligence Approaches: a Comparative Study

The goal was to predict seepage flow (Q) through concrete face rockfill and embankment dams, using three artificial intelligence models, i.e., multivariate adaptive regression splines (MARS), least squares support vector machine (LSSVM), and M5 model tree (M5Tree). The three models were constructed exclusively using in situ measured data from two dams: El Agrem dam located at Jijel province, and Fontaine Gazelles dam located at Biskra province. The obtained results using artificial intelligence models were compared to those obtained using the multiple linear regression (MLR) models. We used two different input variables for developing the models: (i) the daily reservoir water level (WL) and the piezometer elevation (PL) measured at seven different piezometers (PZ1 to PZ7). The results show that the estimation accuracy for Fontaine Gazelles dam is much better than those obtained for El Agrem. All the models performed reasonably well, but the LSSVM was the most consistent predictor of seepage flow for the two data sets. The validation results showed that the LSSVM model has showed significantly better accuracy of seepage flow prediction with root mean square error (RMSE) of 0.432 L/s, mean absolute error (MAE) of 0.302 L/s and correlation coefficient R of 0.952 for Fontaine Gazelles, and RMSE of 0.544 L/s, MAE of 0.344 L/s and correlation coefficient R of 0.731 for El Agrem dam. From this study we conclude that, seepage flow is likely to vary considerably, depending on the reservoir water level, and that the proposed model can be very helpful in estimation of seepage flow, while limitations of the prediction using a standard regression model are illustrated.