Energy-Efficient Predictive Control and Reinforcement Learning Agent for Power Transformers Cooling Systems

The dissipation of the heat duty induced by energy conversion in power transformers relies on coolants and actuation devices (fans, pumps) which combination showcases various cooling system types. The operation of the cooling actuation devices comes with the cost of increased energy consumption, layered-up with operation and regulation-related constraints. To cope with these issues, this paper introduces a concurrent cooling control strategy where an optimum load-dependent Model Predictive Controller (MPC) competes with a model-free Reinforcement Learning (RL) agent to devise a cost-effective course of actuation plan for the coolers. Performance evaluations conducted on a triple rated mineral oil-filled transformer cooling system reveal the similarity of both control schemes with respect to the established operation reference being tracked, and their ability to induce energy saving. Among the conclusions are the feasibility and great potential of the use of an RL agent to circumvent the need to design an accurate and computational-intensive optimal cooling control model.