Author(s): Vincent Wolfs; Patrick Willems
Linked Author(s): Patrick Willems
Keywords: Rating curve; Neural Network; Model Tree; Regression
Abstract: Reliable discharge data from rivers forms a vital source of information in hydrological practices. To obviate the need of expensive and difficult to conduct discharge measurements, hydrologists often turn to the use of rating curves in combination with stage measurements. Usually, regression techniques are employed to determine the rating curve, based on historical flow and stage data. Because of the problem complexity, these techniques do not always provide an unambiguous relationship between stage and discharge. This study aims to overcome this issue by comparing multiple approaches on an hourly set of stage-discharge couples, obtained via simulations with a detailed full one-dimensional hydrodynamic model of the Belgian river Demer. Firstly, the conventional rating curve methodology is applied. Next, the use of expert systems is examined, allowing for a more flexible model structure identification and calibration. The performance of artificial neural networks is compared to that of M5’ model trees. Results show that due to the expert systems’ flexibility, they manage to outperform the conventional rating curve, albeit only slightly. However, as demonstrated in this paper, this kind of models is prone to overfitting. Even with a carefully calibration procedure, the models can become overparameterized if the amount of data presented during training is too small, or if the data does not cover all possible conditions. Secondly, close inspection of the appropriate input space is required. This study presents autoregressive models that perform excellent in one-step ahead prediction, but fail when previously calculated values are used as input.
Year: 2013