With the rapid development of technologies such as 5G and the Internet of Things, dynamic heterogeneous networks have become a key network architecture to support diversified services with their characteristics of coexistence of multiple types of nodes and integration of multiple access technologies. However, the dynamic and heterogeneous nature of network traffic and the differences in quality of service (QoS) requirements make it difficult for traditional traffic scheduling methods to achieve efficient resource allocation and traffic balancing, which in turn causes problems such as network congestion and transmission delay. To this end, this paper proposes an intelligent scheduling algorithm for dynamic heterogeneous network traffic based on deep reinforcement learning. The algorithm constructs a deep neural network model, takes network status information as input, and outputs the optimal traffic scheduling strategy; it uses the reinforcement learning mechanism to iteratively optimize the strategy with the goal of maximizing network throughput and minimizing latency. Experimental results show that in dynamic heterogeneous network tests, DRL-Scheduler shows significant performance advantages: the network throughput in low/medium/high load scenarios is increased by 3.0%, 7.1%, and 15.7% respectively compared with the SPF algorithm, especially in a high-load environment where SPF reaches its processing limit, it still achieves a 16.7% bandwidth utilization optimization. For latency-sensitive services, the algorithm reduces the end-to-end latency of real-time video streaming in high-load scenarios from 175 to 82 ms (a decrease of 53.1%), and file transfer latency from 280 to 195 ms (an improvement of 30.4%), while reducing the packet loss rate from 0.8 to 0.15%. In the test of dynamic adjustment of burst traffic, DRL-Scheduler only takes 120 ms to complete policy reconstruction, showing millisecond-level response capability compared to the rigid routing mechanism of SPF.