| In this study three data-driven water level forecasting models are presented
and discussed. One is based on the artificial neural networks approach,
while the other two are based on the Mamdani and the Takagi-Sugeno fuzzy
logic approaches, respectively.
All of them are parameterised with reference to flood events alone, where
water levels are higher than a selected threshold. The analysis of the three
models is performed by using the same input and output variables. However, in order to evaluate their
capability to deal with different levels of information, two different input
sets are considered. The former is characterized by significant spatial and
time aggregated rainfall information, while the latter considers rainfall
information more distributed in space and time.
The analysis is made with great attention to the reliability and accuracy of
each model, with reference to the Reno river at Casalecchio di Reno
(Bologna, Italy). It is shown that the two models based on the fuzzy logic
approaches perform better when the physical phenomena considered are
synthesised by both a limited number of variables and IF-THEN logic
statements, while the ANN approach increases its performance when more
detailed information is used. As regards the reliability aspect, it is shown
that the models based on the fuzzy logic approaches may fail unexpectedly to
forecast the water levels, in the sense that in the testing phase, some
input combinations are not recognised by the rule system and thus no
forecasting is performed. This problem does not occur in the ANN approach. |