Infilling Annual Rainfall using Pseudo Mac Laurin Generalized Feedforward Backpropagation Artificial Neural Networks

Masengo Ilunga


Generalised backpropagation, infilling hydrological data


Rainfall data play is an important element for hydrology and water resource management and development, in particular for integrated watershed management, The current study uses feedforward artificial neural networks (ANNs); i.e. pseudo Mac Laurin generalised BP order 1 and order 2 derivatives (GenerMcL1BP and GenerMcL2BP) techniques to fill in hydrological data, specifically annual rainfall data (totals). Performance of these techniques is evaluated by using the root mean square error of predictions (RMSEp). A preliminary case study in South Africa is done using the Bleskop (SAWS gauge no. 0228170) (control) rainfall station and the Luckhoff-Pol (SAWS gauge no. 0228495) (target) rainfall station, both from the Orange River drainage system (D). Results are compared with the generalized BP (GenerBP) technique which was used on the same data set in a previous study. It was confirmed that increasing gap size affects negatively the accuracy of estimated missing values (RMSEp). The two techniques (GenerMcL1BP and GenerMcL2BP) gave close results as compared with GenerBP, however their performance was generally slightly lower. Hence the GenerMcL1BP and GenerMcL2BP were generally acceptable to fill in the annual total rainfall values at Luchoff-Pol station. Further work could be recommended on other rainfall data sets of South Africa, which is a semi-arid country.

Important Links:

Go Back