Due to the complex evolution of chaotic systems, direct long-term prediction of chaotic time series is often difficult to achieve well performance. Considering this problem, the chaotic time series is transformed into a series of feature subsequences using the variational mode decomposition method. The rationality of selecting the number of subsequences is evaluated by the permutation entropy to ensure that the feature subsequences contain the long-term evolution trend of the original sequences. The modified cycle reservoir with regular jumps network is used as a predictive model for subsequences. The reservoir nodes of the modified model are connected in a uni-directional cycle with bi-directional shortcuts. This improvement avoids the problems of low prediction accuracy caused by deterministic connection structure and network instability caused by random connections in reservoir. Finally, this paper uses the proposed model to predict the time series in a long-term, and uses a variety of evaluation methods to analyze the prediction results, which show that the proposed model has a great advantage in long-term prediction.