The dynamic reservoir of the randomly generated echo state network(ESN) contains a significant amount of redundant neurons, which leads to collinearity in the high-dimensional state space matrix of the network and subsequently affects its prediction performance. In order to address this issue, this paper proposes a self-organizing choice ESN(SC-ESN) structure optimization model based on reinforcement learning. The essence of the SC-ESN model lies in its construction of multiple initial reserve pools, which is founded upon the idea of ensemble learning.The contribution of each neuron in the reservoir pool to the network performance is then measured using mutual information, and the decision mechanism of reinforcement learning is utilized to screen out effective neurons for network output. The purpose of this optimization is to improve the network's structure and prediction performance. Results of experiments conducted on both manual and actual datasets show that the SC-ESN model proposed has a more streamlined structure while still maintaining superior prediction performance compared to other prediction models.