Deep reinforcement learning LQR controller design for MIMO systems applied to gas production facility

This paper addresses performance control in the synthesis of a stabilising controller for a gas production facility. The controller's performance is closely linked to the pole values defined during synthesis. However, it is highlighted that these calculated pole values may not always be applicable due to the system's physical constraints, such as the impossibility of reducing a biological chemical reaction from hours to microseconds. Initially, the controller is synthesised using an LQR controller with state estimators generated by a specific observer. An in-depth discussion of pole values is provided, referencing Hadamard's lemma, Gerschgorin discs, and the Nyquist stability criterion. To enhance stabilisation performance, Deep Reinforcement Learning is employed to modify the poles by adjusting LQR values in a learning environment. The results demonstrate a successful integration of Gerschgorin discs early in the synthesis process, followed by Deep Reinforcement Learning improvements, generating physically feasible pole values that significantly enhance controller performance.