In order to tackle the dynamic scheduling problem of tow trains in mixed-model assembly lines, a scheduling approach is proposed based on the knowledge base and neural network. Firstly, the dynamic scheduling problem of material delivery in the automotive assembly line is formally described. The throughput of the assembly line and the total delivery distances are selected as components of the objective function. After that，the sample data of mixed-model assembly lines are generated by the Plant Simulation software and are used to train the neural network model offline. Finally, the trained neural network model and the knowledge base are adopted in the real-time scheduling process to select the optimal scheduling rule for tow trains. The experimental results indicate that the scheduling rules selected by the selection method proposed in the paper are mostly the optimal ones. The lower computational complexity of scheduling rules ensures the real-time performance of scheduling. It can cope well with changes in the dynamic environment, thus effectively improving the dynamic scheduling of tow trains.