Modelling of the river flowrate: the influence of the training set selection

Abstract

A study of the influence of the training set selection, the modelling technique, and the number of objects in the training set was performed on a data set of 2 years' daily measurements of atmospheric precipitation and river flowrates. Twenty-five different data sets were prepared by the following selection methods: random selection (RND), Kohonen neural network selection, and Kennard–Stone selection (K–S). On these data sets, 125 models (regressions) were generated using the following five methods: multiple linear regression (MLR), partial least squares regression (PLSR) and two feed-forward neural networks with the error back-propagation and Levenberg–Marquardt learning algorithm (LM). The models were tested using a single set separated from the rest of the data. Additionally, a bottleneck neural network with back-propagation learning was tested, since it combines both modelling and mapping capabilities. The results are discussed and two examples of model use are presented. The best method for sample division into the training/monitoring sets was the Kohonen map (KOH) division. The best model obtained among those generated yields the test set RMS error of 0.0845 and was achieved by L–M neural net model on 400 randomly selected training objects.