Fusing multi-model hydrological data

Seven fusion strategies were used to perform an alliance of river level forecasts. The original data comprised persistence values and continuous predictions produced using a set of conventional, fuzzy logic and neural network models for two contrasting catchments: the River Ouse and the Upper River Wye. The fusion process followed a simple “data-in-data-out” architecture and each fusion operation was implemented using a backpropagation neural network. Worthwhile gains were obtained on the timing of events for a stable regime, but it proved difficult to correct for poor peak flow prediction an a flashier catchment