The transient storage model of the transport of solutes in rivers may be used as a warning mechanism, when the values of its three parameters are known. It requires expensive and time-consuming tracer tests to be performed, as no reliable empirical formulae have been developed so far. In this paper the parameters of the transient storage model were evaluated by means of multi-layer perceptron artificial neural networks (ANN), using easily accessible hydraulic and morphometric data as inputs. The major obstacle was the scarcity of available data. For comparison purposes, the ANNs were trained by three optimization techniques, namely the Levenberg-Marquardt algorithm and two global approaches (much less popular in hydrological sciences): particle swarm optimization and differential evolution. Some minor modifications in these procedures, enabling the neural networks to avoid overfitting, were proposed. The ANNs revealed their superiority over available empirical formulae and the linear regression method applied to the same data sets, and performed similarly to the nonlinear multi-variable robust minimum covariance determinant method. Differential evolution appeared to be the most reliable among the optimization approaches investigated.
Transient storage model, longitudinal dispersion, tracer tests, artificial neural networks, particle swarm optimization, differential evolution, global optimization
