Evolving and evaluation of 3LP FFBP deterministic ANN model for district level long range monsoon rainfall prediction.

In order to develop a forecasting model for monsoon rainfall over the districts of Chhattisgarh, 3LP (Three Layer Perception) FFBP (Feed Forward Back Propagation) Deterministic ANN models have been proposed. In the proposed model, eleven neurons in input layer, one hidden layer with eleven neurons, a single neuron in output layer, 132 trainable weights in three layers, transfer function sigmoid 1/(1+e -deltax+eta)) with slope ä = 1, threshold ç = 0 have been used to evolve networks. Training of the network is continued till the mean square error becomes less than a pre-assigned value ranging from 0.0005 to 0.001. Data for the years 1945-2006 have been used, out of which the data of first 51 years, i.e., (-deltax + eta) 1945-1995, are used for training the network and data for the remaining period, i.e. during the period 1996-2006 are used independently for validation. In the present study, it has been observed that the mean absolute deviation (% of mean) values for the independent period (1996-2006) are less than and half of the standard deviation (% of mean) for all the districts. The performances of these 3LP FFBP Deterministic ANN models have been found to be extremely good and better evaluated over statistical trend models also. The models developed and their evaluations have been presented in this paper.