This position paper describes a novel approach for dynamic flexible job-shop scheduling. The proposed approach enables batched and continuous scheduling, taking into account setup times, transportation times, new job arrivals and machine breakdowns. This is achieved by using a heterogeneous graph with a graph neural network and deep reinforcement learning. The heterogeneous graph represents the relationships between operations, machines and vehicles. Reinforcement learning and the graph neural network are used to learn the embedding and ranking of nodes and sets of nodes and therefore enable the scheduling of operations. The graph representation considers that one operation can possibly be completed in different ways, and that machines and vehicles have different capabilities that can be mapped to the required capabilities of the operation and transportation requirements. The jobs are represented in a way that enables parallel execution and partially ordered operations on top of the usual sequential and totally ordered execution. The training uses multi-objective optimization to adapt the policy to the improvement of the makespan of the entire batch, the duration of a single workpiece and the cost.