This paper presents a training algorithm for regularized fuzzy neural networks which is able to generate consistent and accurate models while adding some level of interpretation to applied problems to act in the prediction of time series. Learning is achieved through extreme learning machines to estimate the parameters and a technique of selection of characteristics using regularization concept and resampling, which is able to perform the definition of the network topology through the selection of subsets of fuzzy neuron more significant to the problem. Numerical experiments are presented for time series problems using benchmark bases on machine learning problems. The results obtained are compared to other techniques of prediction of reference series in the literature. The model made rough estimates of the responses obtained by the models of fuzzy neural networks for time series forecasting with fewer fuzzy rules.