This paper explores some variations of a hier-archical control framework that has been recently proposed. This framework is dedicated to the control of a network of interconnected subsystems such as those describing cryogenic processes or power plants. Recent studies have shown that handling constraints and non-linearities could challenge the real-time feasibility of the approach. This paper investigates and combines two successful directions, namely the use of truncated fast gradient and deep-neural-network-based controller modeling, to reduce the computational time of the most critical subsystem. It is also shown that by doing so, the control update period can be significantly reduced and the closed-loop performance is greatly improved. This paper can therefore be seen as a concrete implementation and validation of some key ideas in the design of real-time distributed NMPCs. All concepts are validated using the realistic and challenging example of a real cryogenic refrigerator.

This paper investigates the use of fixed-point Anderson method (AM) to a recently proposed hierarchical control framework [1]. In the previous works, the synthesis of a filter that ensures the convergence of fixed-point iteration is made by using the mathematical knowledge of subsystems, which violates the privacy-preserving requirement for the coordinator at the upper layer. Due to its model-free property, the AM-based resulting hierarchical framework becomes more generic since no mathematical model of the subsystems at the lower layer is required at the upper coordinator layer. Numerical results are proposed to evaluate the effectiveness of this approach in showing that the AM could converge faster in some cases where the filter is not compatible with the model.