Dynamic training of ANN for its application In real time flood forecasting

Models explaining why catchments respond to rainfall in a certain way, and predicting how they will respond to the future events, have been built for many different catchments around the world. These efforts can be usefully examined with reference to a spectrum. At one end of the spectrum, there are empirical models which are based on the black box school of thought (pioneered by Sherman, 1932) and the other end of the spectrum is the physically-based models which are based on the white box school of thought (pioneered by Horton, 1933). In recent years, Artificial Neural Network models (ANN), which are categorized as nonlinear empirical models, is thought to be an appropriate instrument for dynamic and nonlinear system modeling. In this study, an empirical ANN-based model for Chamriz sub-basin outflow prediction is developed. The most important feature of this developed model is that the model parameters are updated recursively at each time step before streamflow being forecasted which is named dynamic training. Since, dynamic training was the most important consideration in the model development, the RBF paradigm of neural network was used, which is famous due to its rapid training, generality, simplicity, and capability in solving the multidimensional interpolation problems. Due to the nature of the problem under consideration and its relationship to the time series analysis context, ANN model was used in an imperfect autoregressive format. The model was tested on five experiments using the data gained from the hydrometric and raingauge stations distributed in Chamriz sub-basin. The predicted streamflow hydrograph will imitate and capture the variability in observed hydrograph, if an imperfect autoregressive ANN model with dynamic calibration process is used. The shorter the lead time is, the stronger the effect of hydrometric station situated in the outlet region will be. We consider the developed network structure to be very suitable for online real-time flood forecasting.

Keywords:

RBF neural networks , Dynamic training , Streamflow forecasting , Autoregressive ANN , Certainty factor