Abstract: Since the PMIPv6 can not be efficiently operated, an improved fast handover scheme based on network- performed prediction is proposed. Without changing the decision-maker, the handover experience is generated to train the learning system. The k-distance weighted nearest neighbor algorithm is employed to predict the next mobile access gateway and the soft forwarding mechanism is employed to protect the receiving packets during the handover. Simulation results indicate that the proposed scheme can achieve better performance in terms of handover latency and packet loss compared with PMIPv6.