Accelerating self-optimization control of refrigerant cycles with Bayesian optimization and adaptive moment estimation

This paper presents a model-free self-optimization control algorithm for modulating multiple inputs simultaneously to minimize the power consumption of a vapor compression system (VCS). We propose the use of Bayesian Optimization (BO) to warm-start a state-of-the-art extremum seeking control (ESC) algorithm and then accelerate the ESC on-line with Adam, a well-studied adaptive moment-based optimization method used to solve high-dimensional non-convex optimization problems such as training deep neural networks. BO initializes the ESC at conditions favorable for rapid convergence while concurrently learning a surrogate map of VCS power consumption as a function of the inputs. In addition, the warm-start increases the likelihood of attaining a global optimum for locally convex, but globally non-convex, objective functions by identifying regions where the global optimum most likely resides. The proposed algorithm is evaluated using a Modelica model of an air conditioning system with variable compressor speed, an electronic expansion valve and two variable speed fans. We demonstrate the acceleration of this algorithm in simulations of an occupied space with a realistic heat pump model with realistic ambient temperature profiles, variation in heat-load, and different actuation rates. We also show that, in spite of the presence of unknown exogenous disturbances, the proposed algorithm computes better set-points, faster than the ESC. We also observe that the proposed method improves transient performance compared to the state-of-the-art. •Learn calibration cost functions instead of dynamical models for setpoint optimization.•Energy model learned analytic model of underlying system dynamics.•Accelerate convergence of extremum-seeking algorithms via Bayesian optimization.•Optimize many setpoints at various operating modes without re-estimating gradient.•Case study with high-fidelity Modelica heat pump model in realistic environment.