GENETIC PROGRAMMING AND NEURAL NETWORKS FORECASTING OF MONTHLY SUNSPOT NUMBERS

Abstract

A three-stage computational intelligence strategy is used to forecast the unsmoothed monthly sunspot number. The strategy employs agents that use two computational techniques, genetic programming (GP) and neural networks (NN), in a sequence of three stages. In the first, two agents fit the same set of observed monthly data. One employs GP, while the other employs NN. In the second, residuals (= differences between observed and solution values) from the first stage are fitted employing a different technique. The NN fitted-residuals are added to the GP first-stage solution while the GP fitted-residuals are added to the NN first-stage solution. In the third, outputs from the first and second stages become inputs to use in producing two new solutions that reconcile differences. The fittest third stage solution is then used to forecast 48 monthly sunspot numbers (September 2009 through August 2013). This modeling scheme delivered lower estimation errors at each stage. The next sunspot number peak is predicted to be around the middle of 2012.