A Novel Grey Residual Modification Model Using Neural Networks

摘要

The GM(1,1) grey prediction model is a very popular forecasting method, which uses a limited number of samples without statistical assumptions. Its prediction performance can be improved by establishing a residual model for the original model. However, several of these methods simply address the estimation of the sign for a residual. By contrast, we apply the functional-link net to construct a novel grey residual modification model by mapping the predicted value onto a range with respect to its corresponding residual. In particular, the traditional GM(1,1) models in original and residual models are replaced with the neural-network-based GM(1,1) (NN-GM(1,1)) model because NN-GM(1,1) has the advantage of being free from the dependency on the background value. Prediction accuracies of the proposed prediction models were verified using real power and energy demand cases. The experimental results verified that the proposed prediction models performed well compared with other grey residual modification models based on sign estimation.