Improving peak flow estimates in artificial neural network river flow models

Abstract

In this paper, the concern of accuracy in peak estimation by the artificial neural network (ANN) river flow models is discussed and a suitable statistical procedure to get better estimates from these models is presented. The possible cause for underestimation of peak flow values has been attributed to the local variations in the function being mapped due to varying skewness in the data series, and theoretical considerations of the network functioning confirm this. It is envisaged that an appropriate data transformation will reduce the local variations in the function being mapped, and thus any ANN model built on the transformed series should perform better. This heuristic is illustrated and confirmed by many case studies and the results suggest that the model performance is significantly improved by data transformation. The model built on transformed data outperforms the model built on raw data in terms of various statistical performance indices. The peak estimates are improved significantly by data transformation.

HydroTest: A web-based toolbox of evaluation metrics for the standardised assessment of hydrological forecasts

“…PEP would thus appear to be more appropriate for single-event modelling as opposed to continuous modelling and care must be taken when applying this metric to continuous hydrographs since the two maximums that are being compared might derive from different storm events in the observed and modelled datasets The test could also be modified to perform a comparison that was limited to a consideration of annual maximums such that numerous computations would be performed on temporal subsets of the full record e.g. Sudheer et al (2003). Equation 11 is in fact a modified version of the original simple "percent error in peak" that was recommended for the comparison of single-event models in the major review of evaluation criteria that was conducted by Green and Stephenson (1986) and thereafter endorsed by ASCE (1993).…”

An Hourly Streamflow Forecasting Model Coupled with an Enforced Learning Strategy

“…During the training of NNs, the weights are adjusted iteratively to minimize the overall error between the desired and the actual outputs. Owing to the fact that the data of peak flow are insufficient in size, NN-based models are usually unable to yield satisfactory solutions of extreme values in the river flow [43]. The idea to increase the rate of the data with specific characteristics in the entirety of the learning data is applied.…”

“…For example, the peak flows, which are the most valuable part in constructing flow forecasting models, are always rare. Due to the insufficient data of peak streamflow in size, NN-based models are usually unable to yield satisfactory solutions of extreme values in the streamflow [43]. To overcome this problem, studies that are attempted to improve the quality and the quantity of training data of NN-based models are available in the literature (e.g., [28,40,[44][45][46]).…”

Complexity selection of a neural network model for karst flood forecasting: The case of the Lez Basin (southern France)

“…Extra care had to be taken when normalizing the input and output datasets, e.g. by choosing a specific transformation to reduce these irregularities (Sudheer et al, 2003). Inputs must also be chosen with precaution: data on rainfall, outflow at previous times and sometimes temperature and evapotranspiration are typically available.…”