Forecasting of turbid floods in a coastal, chalk karstic drain using an artificial neural network

Water collected at the Yport (eastern Normandy, France) Drinking Water Supp ly well, situated on a karst cavity, is affected by surface runoff-related turbidity spikes that occur mainly in winter, In order to forecast turbidit y, precipitation was measured at the center of the catchment basin over two years, while water level and turbidity were monitored at the web site. Application of the approach of Box and Jenkins (1976) leads to a linear mod el that can accurately predict major floods about eight hours in advance, p roviding an estimate of turbidity variation on the basis of precipitation a nd mater level variation over the previous 24 hours. However, this model is intrinsically unable to deal with (1) nonstationary changes in the time pr ocess caused by seasonal variations of in ground surface characteristics or tidal influence within the downstream past of the aquifer, and (2) nonline ar phenomena such as the threshold for the onset of runoff. This results in many false-positive signals of turbidity in summer. Here we present an alternative composite model combining a conceptual runof f submodel with a feedforward artificial neural network (ANN), This composi te model allows us to deal with meaningful variables, the actioneffect of w hich on turbidity is complex, nonlinear, temporally variable and often poor ly described. Predictions are markedly improved, i.e,, the variance of the target variable explained by 12-hour forward predictions increases from 28% to 74% and summer inaccuracies are considerably lowered. The ANN can adjus t itself to new hydrological conditions, provided that on-line learning is maintained.