Study of Rough Set Based Grey Relational BP Neural Network on Grain Yield Forecasting

BP neural network is a nonlinear dynamical system widely-used in forecasting target values influenced by diverse variables. Its redundant input nodes often generate imprecision in the predict results. By introducing rough set and grey relation theories, condition attributes were considered as correlation sequences and decision attributes as reference sequences. The grey correlation coefficients represented the weights upon which the condition attributes were reduced and the initial decision table was renewed with the remaining core factors. As a result of training the network by the reduced condition attributes, the prediction precision was improved prominently. In the model application case to forecast the grain yields of China in 2001 and 2002, the results show great improvement of prediction precision as 0.83% and 1.93%, respectively. And the fitting precision of the grain yields in the other 11 years (1990-2000) are all above 99%. The redundancy elimination also increases the network training rate by reducing the input and hidden nodes. Thus, this optimized model was promising in reducing dimension during knowledge discovery and data dining from a large-scale information pool