Abstract: In software-defined networks (SDN), traffic scheduling is crucial for enhancing network performance, resource utilization, energy efficiency, and load balancing.Traditional multi-objective optimization algorithms significantly affect the convergence speed of the algorithms under high traffic and increased network dynamics, making it difficult to meet the diverse demands of complex network environments.To address this problem, a deep reinforcement learning-based traffic prediction on-the-line routing algorithm, OTPR-DRL, is proposed: predicting critical and ordinary flows based on traffic characteristics and building a linear programming problem to obtain the optimal solution for critical flow routing by combining network state and traffic information.In order to meet the different QoS (quality of service) demands of ordinary flows, a general utility function is introduced to achieve multi-objective optimization, and routes are selected for ordinary flows through multi-intelligentsia and priority experience playback mechanisms.Experimental results show that OTPR-DRL improves the convergence speed, reduces at least 10.26% of network transmission delay, reduces 3.09% of packet loss, and improves 1.7% of throughput compared to existing algorithms under high traffic intensity.