Model-based learning of coordinators of the decentralized multi-zone objects control systems

Decentralized control systems are gaining more and more expansion, which is due to the increase in the availability and power of microcontrollers. Decentralized control of multi-zone objects is associated with the need to coordinate the local control systems of zones state. Learning systems are preferred for implementation of the coordination methods, as they are able for flexibly adjust to the specifics of control of each zone. However, the training of coordinators is complicated task by the absence at the stage of a system creating of marked datasets for controlled multi-zonal objects. This article considers the creation of a dataset based on a simulation of a decentralized system and four scenarios for training neural coordinators. A model for simulation of a decentralized system was been created on the Scilab/Xcos platform using a pre-built library of blocks for simulating decentralized systems. The scenarios differ depending on the structure of the neural coordinators: a segmented network according to the structure of the coordinator simulation model or an integrated one, as well as on the training strategy: train all the coordinators of the decentralized system in parallel or only one coordinator and then clone the results. Experimental studies of the proposed method of training neural network coordinators, implemented on Python TensorFlow, were conducted. The study showed greater effectiveness of segmented coordinators parallel training. However, in the course of the study, the last step of the scenarios – fine tuning on a real physical object, was not performed. A preliminary evaluation suggests that after such additional training, the advantages of mono-neural coordinators will become more visible, since such additional training will correct the shortcomings of imitation.