Model-Oriented Training of Coordinators of the Decentralized Control System of Technological Facilities With Resource Interaction

The control process of technological facilities with resource interaction in a decentralized system requires coordination of local systems for control of the state of objects. For the implementation of coordination methods, learning systems have an advantage since they can flexibly adapt to the specifics of each facility control. However, the coordinators’ training process is complicated by the lack of labelled datasets for technological facilities. In decentralized control systems, the problem is complicated by the need to train all coordinators, with the outcome depending on the coordinator’s position within the structure of the distributed control system. This article explores the prospects of model-based learning for solving the problem of missing datasets used for coordinators’ training. An approach to determining the optimal statistics of the training dataset for the coordination control of nonlinear technological facilities with resource interaction is proposed. A combined three-stage process of coordinator training for the decentralized system is proposed. In the first stage, one coordinator is trained on the basis of a distributed system simulation. In the second stage, the settings of the trained coordinator are applied to other coordinators, which are retrained in parallel on the basis of simulation models of local control systems of the relevant parts of the technological facilities. In the third stage, coordinators are fine-tuned to real conditions using Bayesian random search. Conducted experimental studies of the proposed method of training neural network coordinators, implemented on Python TensorFlow, showed greater effectiveness of Collaborative Federated Learning compared to independent training of coordinators or direct transfer of learning outcomes between coordinators.