With the digital development of the steel industry, the orders become multiple species and random change, which puts forward new requirements for the adaptability and flexibility of the hot-rolling scheduling model. For the hot rolling scheduling problem, the current mainstream method is a heuristic algorithm, which has two problems: one is that it does not consider the organizational representation of data; the other is that this kind of algorithm has strong pertinence. When the problem changes very little, it needs complex parameter adjustment. Compared with machine learning, it has better adaptability and flexibility. Therefore, this paper uses ontology to represent the organization of human-cyber-physical data and proposes a hot rolling scheduling solution method of pointer network $+$ reinforcement learning for the first time. The pointer network is used to learn the mapping from sequence to sequence. In order to solve the problems of the pointer network training difficulty and low performance, the actor-critical network is used to improve the accuracy and convergence speed of the model. Finally, the effectiveness and performance of the algorithm are simulated by designing the corresponding experimental scheme and compared with LK-H's local search algorithm to further verify the effectiveness of the proposed method.