River flow prediction using artificial neural networks: generalisation beyond the calibration range

Artificial neural networks (ANNs) provide a quick and flexible means of creating models for river flow prediction, and have been shown to perform well in comparison with conventional methods. However, if the models are trained using a dataset that contains a limited range of values, they may perform poorly when encountering events containing previously unobserved values. This failure to generalise limits their use as a tool in applications where the data available for calibration is unlikely to cover all possible scenarios. This paper presents a method for improved generalisation during training by adding a guidance system to the cascade-correlation learning architecture. Two case studies from catchments in the UK are prepared so that the validation data contains values that are greater or less than any included in the calibration data. The ability of the developed algorithm to generalise on new data is compared with that of the standard error backpropagation algorithm. The ability of ANNs trained with different output activation functions to extrapolate beyond the calibration data is assessed.