An investigation into improving the accuracy of real-time flood forecasting techniques for the Onkaparinga River catchment, South Australia

Abstract

Three aspects of real-time flood forecasting were investigated on the Onkaparinga River catchment in the Mount Lofty Ranges, South Australia, firstly antecedent soil moisture indices, secondly artificial neural networks (ANNs) and finally, the application of the RORB model. The aim was to investigate whether the subjectivity associated with flood forecasting can be eliminated, particularly real-time updating and understanding the hydrologic processes occurring in order to achieve an improvement in forecast accuracy. Antecedent precipitation index and pre-storm baseflow were used to estimate initial loss, however poor correlations were obtained. Antecedent precipitation index performed marginally better than pre-storm baseflow in estimating initial loss, although the best relationship containing antecedent precipitation index had an R2 of only 0.68. A backpropagation artificial neural network was used as a real-time runoff forecasting model. Both 1 and 5 hour runoff forecasts were made with input combinations of previous runoff and rainfall. The ability for the ANN model to forecast in real-time was not as encouraging as first hoped, in fact most forecasts were no better than what a 'trained forecaster' could achieve. Best results were obtained when the training sets contained the least amount of hydrologic variability and the most accurate representation of the storms. It was concluded that perhaps the limiting factor in the development of an ANN model of this nature is the accuracy of the model data itself, particularly the rainfall input. The RORB model was automated using a PASCAL program and a DOS keystroke faker for 5 hour real-time forecasts. Forecasts were performed with different rainfall scenarios, continuing loss values and model parameters kc and m. In one case, the model parameter kc was adjusted in real-time. Best estimates were made using an accurate continuing loss value and the exact time distribution of rainfall. Again, the forecasts were probably no improvement to what could be achieved by a 'trained forecaster'. Altering the continuing loss as the storm proceeds to achieve a more accurate forecast, is probably more desirable than adjusting kc alone. The main factor that prevented a reduction in the subjectivity and an associated improvement in the accuracy of the real-time flood forecasts for the Onkaparinga River catchment, was due to the availability of only a small historical data set. This data set contained floods with vastly different characteristics, combined with an inadequate spatial representation of rainfall throughout the catchment which meant that relationships were unable to be successfully developed.