Résumé:
In the management of water resources in different hydro- systems it is important to evaluate and
predict the sediment load in rivers. It is difficult to obtain an effective and fast estimation of
sediment load by Artificial Neural Network without avoiding over-fitting of the training data. The
presented thesis comprises of three steps in order to obtain an Artificial Neural Network model. In
the first step the study comprises the comparison of a multi-layer perception network one with
non-regularized network and the other with regularized network using the Early Stopping
technique to estimate and forecast suspended sediment load in the Isser River, upstream of Beni
Amran reservoir, northern Algeria. The study was carried out on daily sediment discharge and
water discharge data of 30 years (1971–2001). In the second step, the author used the same
Artificial Neural Network model once again, using non regularized and then regularized model to
forecast suspended sediment in the Sebaou Wadi, in the Great Kabyle watershed, northern
Algeria. The study was conducted on daily water and sediment discharge data of 7 years between
(1978 and 1989). Both studies on different valleys were compared using the regularized and non
regularized neural networks. The models were evaluated in terms of the Coefficient of
Determination (R²) and the Root Mean Square Error (RMSE). The comparison results indicated
that the regularizing neural network using the Early Stopping criterion to avoid over fitting
performs better than the non regularized networks in both studied areas, with a priority of a better
performance values to the application of the Isser Wadi. The results show that the overtraining in
the back propagation occurs because of the complexity of the data introduced to the network.
In the third step authors tried to confirm the efficiency of their neural network model using the
Early Stopping technique, the application of the neural network model was the prediction of
suspended sediment discharge in un-gauged river. The study was applied on two different sites,
firstly, we used the input data of the Isser Wadi to forecast the suspended sediment in the Sebaou
Wadi, carried on daily water and sediment discharge in a period of 7 years (9 years using training
inputs from the Isser Wadi, and two years for validation and testing depending on the data of the
Sebaou Wadi). Secondly, we used the input data of the Sebaou Wadi to forecast the sediment
discharge of the Isser Wadi during the period of 7 years with the same divided data sets as the
previous application on the Sebaou Wadi. The comparison of the results indicated that the overfitting occurred often in our models, and the Early Stopping technique showed acceptable values
but still further from the applications using real river data that were shown in the first and second
7
steps. The use of the early stopping technique in forecasting sediment discharge is very effective
and robust especially to avoid the over-fitting that occurred often in our models.
Keywords: Artificial neural network, Suspended sediment, Back propagation, Water discharge,
Erosion, Early stopping, Isser, Sebaou, Algeria, Wadi, Watershed, Dam, Ungauged river, Beni
Amran, Taksebt.
