Abstract: To solve adaptability problems of the reservoirs of echo state network in complicated conditions, such as suffering from random faults and deliberate attacks, a restorable echo state network with biological evolution characteristics-3DP-RESN was proposed. The 3DP-RESN was designed to be able to recover automatically from destroyed network topology based on the evolution strategies of preferentially matched duplication, newly added connection-oriented divergence and newly added connection-oriented death. In experiments, 3DP-RESN, classic ESN (CESN) and destroyed ESN (DESN) are applied to approximating three kinds of nonlinear time series, i.e., the NARMA system, Henon map and figure8. Experimental results show that, when reservoirs suffer from failure, for three kinds of time series, the prediction accuracy of 3DP-RESN significantly outperforms DESN, and is close to or even higher than that of CESN which has not suffered from failure. Especially in the experiment of figure8, compared with CESN and DESN, the prediction accuracy of 3DP-RESN is improved by 30.56% and 7.01% respectively. Besides, the short-term memory capacity of the 3DP-RESN is also close to that of CESN. Hence, 3DP-RESN can possess strongly adaptive self-recovery capacity.