The large-scale flexible job shop dynamic scheduling problem (LSFJSDSP) has a more complex solution space than the original job shop problem because of the increase in the number of jobs and machines, which makes the traditional solution algorithm unable to meet the actual production requirements in terms of the solution quality and time. To address this problem, we develop a dynamic scheduling model of a large-scale flexible job shop based on noisynet-double deep Q-networks (N-DDQNs), which takes the minimum expected completion time as the optimization objective and thoroughly takes into account the two dynamic factors (the new job arrival and the stochastic processing time). Firstly, a Markov decision process model is constructed for dynamic scheduling of a large-scale flexible workshop, and the corresponding reasonable state space, action space and reward function are designed. To address the problems (of solution stability and unsatisfactory scheduling strategy selection) in the conventional exploration method of DDQNs, learnable noise parameters are added to the DDQNs to create the N-DDQN algorithm framework, where the uncertainty weight is added. Secondly, the learnable noise parameters are added to the DDQNs to form the N-DDQN algorithm framework, and the uncertainty weight is added to realize automatic exploration. Hence, the issue is solved that the traditional DDQN exploration method may result in unsatisfactory solution stability and scheduling strategy selection. The proposed method, which has significant flexibility and efficacy, is demonstrated (by experimental findings) to be superior to the conventional method based on compound scheduling rules in tackling large-scale flexible job shop dynamic scheduling problems.