The holistic optimization of district cooling systems is a computationally intensive undertaking, owing to the sheer number of conflicting decision variables and nonconvex nature of the problem. This is the primary reason which inhibits the real-time deployment of optimization algorithms for the operations of district cooling systems. To overcome this challenge, we adopt a model-based, decomposed approach involving the concurrent use of reinforcement learning and mixed integer linear program to holistically optimize the thermal and physical interactions while still capturing the tight coupling between the components of the system. Resolution speed and solution accuracy are paramount for a realtime optimization algorithm thus, the critical advantage of the proposed approach is two-fold – the mixed integer linear program drastically reduces the action space of the reinforcement learning problem, promoting accuracy and when trained, the agent neural network can then rapidly determine the optimal values of the remaining actions, improving resolution speed. The current work makes the two ensuing vital contributions: (1) we introduced a decomposed optimization approach with resolution speeds which are compatible with real-time deployment, (2) through the application on a real test-case, we compare both the resolution time and solution quality against an approach used in our previous work, which deployed the genetic algorithm instead of a reinforcement learner. Results indicate that the impact on solution quality is below 7.52%, thereby, validating the feasibility of the proposed approach.
Keywords reinforcement learning, district cooling, energy efficiency, cyber-physical systems, holistic optimization, mixed integer linear program