M. Illunga (S. Africa)
Infilling rainfall, Neural Network, Generalised Backpropagation
Long time series of hydrological data (e.g. rainfall, river flow) are required for water resource planning and management as well as for hydrological studies. However, it happens that hydrological data have missing values or are incomplete. In this paper, a feedforward artificial neural network (ANN) technique is used for rainfall data infilling, specifically for annual total rainfall data. The Generalized backpropagation (BP) technique is used with a sigmoid activation function. The effect of the generalization parameter in the Generalized BP on the accuracy of missing annual rainfall (totals) is preliminarily investigated. The root mean square error of predictions (RMSEp) is used to evaluate the performance of the generalized BP technique. A preliminary case study in South Africa was conducted using rainfall data of the following rainfall stations: the Touwsrivier (SAWS gauge no. 0044050) (target) rainfall station and the Jan deboers (SAWS gauge no. 0044286) (control) rainfall station of the primary drainage System Rivers (J) in the Western Cape as well as the Isidenge (SAWS gauge no. 0079490) (target) rainfall station and the Izeleni (SAWS gauge no. 0079730) (control) rainfall station of the primary drainage system rivers (S) in the Eastern Cape. Generally, it was shown that the accuracy in the estimation of missing values becomes less with increasing values of generalization parameter when the generalized BP technique is applied to annual total rainfall data. Recommendations for further work on the generalization parameter effect include the application of the Generalized BP to other rainfall data (e.g. annual maximum series, etc) and to rainfall data from other climatic regions. The use of other activations functions other than sigmoid function needs to be investigated.
Important Links:
Go Back
