<p>The job shop scheduling problem (JSSP) is the core of industrial production and manufacturing with strongly NP-hard complexity. Recently, deep reinforcement learning (DRL) has been shown to be an ideal technique for learning priority dispatching rules (PDRs) to solve complex scheduling problems. However, In real-world environments, scheduling is complex, stochastic, and dynamic, with stochastic job arrivals and random machine breakdowns, which is known as dynamic JSSP (DJSSP). Most DRL methods focus on JSSP problem. In this paper, we propose a framework which extends the DRL method based on graph neural networks and deep reinforcement learning to addresse DJSSP problem of dynamic jobs with random processing time. Specifically, we extend the graph attention network (GAT) as a common encoder between JSSP and DJSSP, encoding job and machine information into nodes and arcs represented by a disjunctive graph. Then we exploit the encoder to learn high-quality PDRs via an deep reinforcement learning agent with proximal policy optimization. Moreover, by the communality of the GAT encoder, the embeddings of nodes and arcs pretrained in JSSP can be transferred into DJSSP as a good prior knowledge. Thus, the learning framework can pay more attention to learn the multilayer perceptron (MLP) predictor of the policy, which combines GAT as a policy network and produces better performance, rather than randomly initializing the GAT, trained from the original environment. Experiments show that the framework outperforms the DRL methods without pretrained embeddings and is computationally efficient, even on instances of larger scales and different properties unseen in training.</p>