Artificial Neural Network Rainfall-Discharge Model Assessment Under Rating Curve Uncertainty and Monthly Discharge Volume Predictions

The accuracy of rainfall-discharge volume model predictions depends on the model design and uncertainty of the available stage-discharge measurements used to fit the rating curve, which converts a time-series of recorded stage into discharge. In general, the rating curve uncertainty is the product of several combined sources. Over Algerian rivers, the extrapolation of the rating curve beyond the gauging range is the main source of this uncertainty. This study, therefore, represents a quantitative approach to reflect rigorously the impact of the rating curve uncertainty on the improvement of monthly discharge volume prediction quality by the artificial neural network (ANN) rainfall-discharge model. The rating curve uncertainty of the Fer à cheval hydrometric station in the Mazafran watershed is performed within Bayesian analysis for stationary rating curves using the BaRatin method. This allows as to build a new time series of discharge in order to assess an ANN rainfall-discharge model. To do that, Levenberg–Marquardt back propagation neuronal network has been applied over 1972-2012 time-period, for five hydrometric stations in the Algiers Coastal Basin. The model inputs were constructed in different ways, during the algorithm development, such as precipitation, antecedent precipitation with different monthly lag times and antecedent monthly discharge volume. The results indicate that training/validation of ANN rainfall-discharge volume model is widely affected by the streamflow datasets uncertainty. A large proportion of model prediction errors are significantly improved when considering the rating curve uncertainty.