In narrowband power line networks, it is necessary for intermediate nodes to forward packets when source node transmits data to a distant destination node due to the great signal attenuation. Statically configured routing can not meet this demand without any consideration of dynamic network topology. This article makes efforts to find an effective and reliable routing solution to work in the power line communication (PLC) networks which are not only of Master-Slave structure but also composed of large quantities of nodes. We utilize on-demand multipath routing to provide high degree of fault-tolerance. As conventional multipath protocols do not perform well in large-scale scenarios, extra back off time in routing layer and random packet forwarding mechanism are introduced in our work to provide fairness in routing discovery. Broadcasting controls based on location learning is used to cut down the broadcasting overhead as well. Simulation results show that the proposed protocol can achieve a better performance in finding maximally disjoint routes in PLC networks.