A Universal Prediction Model Based on Hybrid Neural Network

Single neural network is difficult in performing accurate predictions for complex model. A hybrid model, which involves a radial basis function network, a multi-layer perceptron network with back-propagation and a control module, is proposed and used for forecasting complex system. The control module serves as a linear mapping network which combines the outputs of two neural networks to gain the final output value. The prediction methods of the hybrid model are mainly discussed: Firstly taking advantage of the improved algorithm to train two networks respectively and obtain the output values; Secondly, the linear mapping network is optimized by self-adaptive genetic algorithm to gain higher prediction accuracy; Finally, this paper has carried out two experiments to compare the prediction performance of a single network and the proposed model. The experimental results show that the proposed hybrid neural network provides a superior performance in prediction accuracy than other methods and offers a common tool for complex prediction.