Structural identification of multivariate neural networks for rainfall-runoff modelling

Abstract Designing neural networks predictors by pruning instead of trial and errors significantly reduces the amount of guesswork required to select the optimal architecture. Furthermore, the obtained model is partially connected and hence very parsimonious in the number of parameters, leading to relevant operational advantages in the hydrological forecast practice: in fact, removing from the model redundant measurement stations results in an improved forecast availability and in the reduction of the costs of the data acquisition system. We exploited pruning algorithms to design the network, providing also a better basin state representation in comparison to existing schemes. Thanks to this modelling improvement, the obtained pruned networks overperform some fully connected ones published in previous works on the same basins, while requiring a significantly smaller set of measurement stations.