To improve the ability of echo state network for feature extraction and prediction on chaotic time series, a hierarchical plasticity echo state network (HPESN) model is proposed. In this model, multiple reservoirs are connected in sequence, and the ability of nonlinear multi-scale dynamic feature extraction is enhanced through layer-by-layer feature transformation. Meanwhile, the intrinsic plasticity mechanism in neuroscience is introduced to simulate the firing rate distribution of real biological neurons, and the reservoir is pre-trained with the goal of maximizing neuronal information transmission. HPESN not only increases the capacity of the model and reduces the instability caused by random projection, but also provides a new idea for understanding the representation, processing, memory and storage operations of the reservoir. The simulation results show that compared with the other seven improved echo state network models, the model proposed in this paper achieves the best prediction accuracy in the chaotic time series prediction task composed of synthetic data and real-world data.