The well-established occurrence of exceptionally high floods on the Red River prior to the record of annual peak flows at Winnipeg is an important factor in the flood risk assessment for that city and for the entire Red River valley. But the weight given to this occurrence is quite dependent on the autocorrelation structure assumed for the spring peak time series. It is therefore important to decide whether the clustering of high peak flows, which can be observed in the record, is a mere chance phenomenon or indeed a characteristic of the runoff process. In earlier studies this clustering was found to be significant in a statistical sense. The present study aims at finding a physical explanation for this particular type of correlation structure. It presents the accumulated basin storage (ABS) as a physically based parameter that measures average soil moisture conditions in the drainage basin. The reconstructed record of ABS values just prior to the spring runoff shows a very high first-order autocorrelation coefficient. Relatively wet and relatively dry soil conditions therefore tend to persist over long periods. Since the magnitude of the spring peak is significantly affected by soil moisture conditions prior to snowmelt, the structure of the annual ABS time series can be expected to be reflected in the peak flow time series. This was found to be the case. The study thus supports earlier conclusions based on statistical evidence that the conventional assumption of serially independent spring peak floods seriously underestimates the flood risk for the City of Winnipeg and the Red River valley. Key words: accumulated basin storage, Red River floods, simulation, time series, clustering, streamflow persistence, serial correlation, flood risk.
ARTIFICIAL NEURAL NETWORK MODELING FOR ANNUAL PEAK FLOWS: A CASE STUDY
