Infilling Maxima Annual Monthly Rainfall using Neural Networks: Effect of Scaling Parameter

Abstract

Water engineering projects and several studies in hydrology and water resources rely on hydrological data such as rainfall. The current study uses feedforward backpropagation (BP) artificial neural networks (ANNs) models; i.e. pseudo Mac Laurin BP order 1 and order 2 derivatives (McL1BP and McL2BP) as models to fill in maxima annual monthly rainfall data. The effect of the scaling parameter of the sigmoid function is investigated on the performance of the modeling process. Statistical indicators such as root mean square error of predictions (RMSEp) and correlation coefficient (r) are used to evaluate the performance. These two models are applied preliminarily to a case study in the North-West region of South Africa, particularly the Mamogaleskraal (SAWS station no. 05124812) (control) rainfall station and the De Kroon (SAWS gauge no. 05125809) (target) rainfall station. Results confirmed that increasing gap size on the target rainfall station affects negatively the accuracy of simulated missing values. For a given value of the scaling parameter, the two models under investigation gave close results. It was observed that increasing scaled parameter decreases the accuracy of the simulated values and may slightly favor one model over the other. In general these models or techniques underestimate the missing rainfall values. McL1BP and McL2BP were acceptable to fill in the annual maxima monthly rainfall values at De Kroon rainfall station. Further work may include other rainfall data sets of South Africa, as well as other hydrological data such as evaporation, wind speed, etc.