The combination of immune evolution and neural network for nonlinear time series forecasting

Forecasting and dynamic modeling have common applications in science and engineering. Time series data are often found from different sources, astrophysical, biological, economical, etc. It is then very important to predict future values of these data series from the existing data. The immune system is a very complex system with several mechanisms to defense against pathogenic organisms. Inspired by the principles of immune system and biology evolution, a novel algorithm based on the combination of immune evolution and neural network is proposed to forecast nonlinear time series, which imitates the cellular clonal selection theory of biology immune system and the mutation ideas of biology evolution process. Then, the mutation intensity of each antibody is decided by its objective function value; similar antibodies are suppressed by computing the affinity of antibodies and new antibodies are produced dynamically to maintain the diversity. Application of the proposed algorithm to nonlinear time series of sunspots number modelling and prediction is investigated. The experimental results by different methods confirm that the proposed method has better generalization performance than that of the Fuzzy genetic algorithm (FGA), Genetic programming (GP), Automatic Regression Model (AR) and Automatic Regression Moving Average Model (ARMA)