Multiple step ahead river flow modelling for south east tasmanian aquaculture

We develop multiple step ahead prediction models of river flow for locations in Tasmania (Australia) for decision support in aquaculture. In predicting river flows for multiple days ahead, we first statistically determine the maximum input lags of rainfall and river flow. We then use machine learning techniques in building models. In multiple step ahead prediction, we consider both static and dynamic approaches. In dynamic approach, one day prediction is served as input to two days ahead prediction. The experimental results demonstrate that, in general, a dynamic approach provides better accuracy in multiple day's ahead prediction. For Duck Bay location using dynamic approach, support vector regression performs best over linear regression, M5P and multilayer perceptron. However, at Montagu Bay location, we find that M5P performs best over methods. We find that multiple step ahead prediction of river flow for each location requires modelling of lags with associated machine learning techniques.