**1. Introduction**

If the air conditioning (AC) system experiences identical operation condition all of the time, the engineer can set up the optimal settings and maintain the system energy e fficiency easily. However, the fact is that the operation condition of AC systems involves many random variables over time, e.g., weather and occupancy. This makes real-time optimal control important. Each time the AC system experiences significant changes, the control system should respond and reset the previous optimal settings. Thus, the optimal control of the AC system should be repeated over the operation period.

Central AC systems contribute towards a large portion of a building's energy consumption [1,2]. Optimal control has been considered as a powerful measure to improve the operating e fficiency of AC systems [3–7], where an objective function is optimized through optimizing the control set-points or operation modes (e.g., chiller sequence). Although many advanced optimization algorithms

(e.g., reinforcement learning and model-predictive control [8]) and sophisticated modelling techniques (e.g., grey-box models or data-driven models [9] were developed, the mechanism to trigger the optimal control actions is still simple. In fact, an e fficient triggering of optimal control can enhance the energy efficiency with the same resource consumption [10].

In most current practices, the control actions are triggered periodically based on a time "clock". This mechanism is thus termed as a time-driven optimal control (TDOC) [11,12]. In the TDOC, the time interval between two neighboring optimizations is called "optimal control frequency". Typical optimal control frequency ranges from a few minutes to a few hours [13]. The TDOC scheme has been widely used due to its simplicity and e ffectiveness. For example, Kusiak, Li and Tang [14] performed an hourly optimization of the set-points of a supply air pressure and temperature, leading to a 7.66% energy saving. Huang, Zuo and Sohn [15] optimized the condenser water set-point for an existing chiller plant, and the hourly optimization o ffered a 9.67% energy saving.

However, since AC systems always experience stochastic changes in their operation conditions, such as weather and occupancy changes [10], using a periodic optimization mechanism, TDOC will have inherent drawbacks, e.g., stochastic (or aperiodic) changes cannot be captured promptly or correctly. Consequently, the TDOC may lead to delayed or unnecessary control actions, which would degrade the expected optimization performance. To deal with this issue, an event-driven optimal control (EDOC) was recently developed by Wang et al. for AC optimal control [11], where control actions are triggered only when pre-defined events occur. In the study of Wang et al. [11], two events were defined based on the part-load ratio (PLR), because PLR has grea<sup>t</sup> impact on the AC optimal control [16]. One event was defined as the significant change of the chiller plant PLR, and one was defined as the chiller sequence change, which is also triggered by the variation of PLR. Thus, this EDOC was titled as a PLR-based EDOC. Numerical studies showed that the PLR-based EDOC was able to achieve a better energy e fficiency (0.4–2.6% higher), and simultaneously reduce the computational cost by 60–84% when compared with a traditional TDOC approach.

Many studies use PLR as an indicator to optimize the AC operations, e.g., optimal chiller loading [17]. However, for triggering the optimal control, the real and direct trigger of optimal control is the system coe fficient of performance (SCOP) [18]. SCOP links the system cooling/heating output to the system power [19,20]. Only when the SCOP deviates from the optimal value, an optimal control action is needed. Since the relationship between the PLR and SCOP is nonlinear [12,21], the PLR cannot reflect the SCOP precisely. Using PLR variation to trigger would lead to non-optimal actions. For instance, when the PLR has a significant change (e.g., 10%) compared with the last optimization time instant, it is possible that the system operation e fficiency remains at its ideal level, but the PLR-based EDOC strategy will trigger optimization, leading to unnecessary control actions. When the PLR has a little change (e.g., 2%), it is possible that the system operation e fficiency deviates largely from its ideal level, but the optimal control is not triggered, leading to a degraded operation e fficiency.

For a more e fficient EDOC, events could be directly defined based on the control objective(s). For instance, Xu et al. [22] investigated a PMV-based event-triggered mechanism for building energy management. The control objective is to satisfy the thermal comfort of occupants, while minimizing the building electricity cost. Event 1 was defined as the equality between the predicted and actual occupied time, and Event 2 was defined by a thermal comfort range. The two events were directly linked to the control objective, and the simulation results show that the energy cost can be saved with the reduced power demand. Therefore, to optimize the operation e fficiency, events could be directly defined based on the SCOP. SCOP has been widely used to evaluate the performance of AC system operations [23,24]. A higher SCOP indicates that less power is required when the same cooling/heating capacity is provided. Thus, the SCOP of AC systems should be maintained at its ideal value to minimize the energy use. Accordingly, events can be defined such that the optimal control is triggered when the SCOP is lower than its ideal value.

However, using the SCOP to trigger the AC optimal control has not been systematically investigated in terms of its energy performance and robustness. This study aims to develop the SCOP-based EDOC for AC systems, and systematically compares SCOP-based and PLR-based EDOC, regarding their merits and demerits. Since the ideal SCOP varies with the working condition [22,23], an artificial neural network (ANN) is developed to predict the ideal SCOP under various conditions. The methodology of SCOP-based EDOC is firstly illustrated, including events definitions and a SCOP prediction model. Then, the EDOC approaches are evaluated by testing in a case AC system. The control performance and applicability of di fferent EDOC approaches are compared. Discussions and concluding remarks are given at the end.
