Abstract: According to the dynamic characteristics of ad hoc network topology, a wavelet neural network prediction (WNNP) model was used to predict the geometrical location of the nodes. Comparing the predicted total holding time with the threshold, the stabilization of a cluster in next time can be measured. If the cluster tends to be unstable in next time, a routing pre-repair mechanism can be initiated before the link failure to avoid frequent breaks of links. Thus the network performance is significantly improved. Simulation results show that compared with the lowest-identifier (lowest ID) algorithm and location-based WCA (LWCA) which has no prediction model, WNNP-LWCA can improve by 7% and 5% of the packet delivery rate, reduce by 63% and 50% of the broken routing number, and maintain the stabilization of the cluster.