A Physics-Informed GAN Algorithm for Energy-Saving Optimization of Parallel Chiller Plants

Abstract

This study addresses energy saving in parallel chiller plants under wide fluctuations of air-conditioning demand. Built-in local controls of individual chillers, including partial-load valve regulation and variable-frequency control, are not sufficient when heterogeneous machines must be staged in parallel. The central algorithmic problem is to learn an accurate mapping from partial load to coefficient-of-performance (COP) so that online grouping and load allocation avoid efficiency cannibalization among mismatched chillers. To solve this problem, the study proposes a physics-informed generative adversarial network digital twin (PI-GAN-DT) that learns unit efficiency curves from sparse manufacturer and field data while enforcing cooling-energy consistency and thermodynamic monotonicity. The learned mapping is then embedded in a supervisory optimization model for optimal chiller sequencing and loading. Two industrial electronic-factory plants are retained only as benchmark cases for empirical verification. Using the archived manufacturer part-load tables and plant load trajectories, the benchmark/simulation study (i.e., results obtained on the archived part-load tables and on a contiguous 84-day campaign rather than on closed-loop deployment) reports daily chiller-energy reductions of 574.2 kWh/day and 673.5 kWh/day in Case A, 2,073.6 kWh/day in Case B at 7 ℃, and an additional 1,601.1 kWh/day when Case B is operated at 9 ℃. These benchmark/simulation results serve as evidence of the effectiveness of the proposed framework; the corresponding closed-loop field savings remain to be confirmed by the A/B field campaign described in the deployment-path discussion. The main contribution of the study is the PI-GAN-DT-based optimization algorithm rather than the benchmark cases themselves.