Energy demand-based optimal start control for multi-chiller plants empowered by physics-guided AI

In hot climates, central cooling systems in buildings often need to be activated in advance for precooling. However, in practice, chiller start control typically relies on load-based or expert rule-based control methods and often fails to reach optimum due to the inherent uncertainty of the transient precooling process. The core challenge lies in accurate prediction of precooling requirements. Prior studies have predominantly relied on load-based predictions to estimate optimal precooling duration and cooling capacity provision, but this approach is unreliable, as the cooling load in precooling period is strongly influenced by the actual cooling capacity provision. This study proposes an energy demand-based optimal chiller start control strategy for systems with multiple chillers to minimize the cooling energy consumption while ensuring thermal comfort targets are met. A novel concept, “precooling energy demand”, is proposed to quantify the total cooling demand, which is independent of actual cooling capacity provision according to the precooling mechanism. This approach eliminates the impact of cooling load measurement uncertainty on precooling demand prediction. A Light Gradient Boosting Machine (LightGBM) model, enhanced with a Tree-Structured Parzen Estimator (TPE) for hyperparameter optimization, is developed to predict the precooling energy demand. Field implementation in a real central cooling system shows that the strategy improved chiller plant COP by 5 %. Simulation tests conducted during a typical summer month show that the strategy could shorten the precooling time by 25 min and reduce precooling energy use by up to 28.2 % compared with conventional strategies.