What AI can change in VRF, DX and chiller–fan coil systems — without promises of universal savings

Schedules, group control, setpoint limits and some energy-saving functions may already be available in the standard controls without separate AI. Such capabilities are described, for example, in Daikin controller documentation.

So the right question is not ‘how do we make an air conditioner smart?’, but what data and decisions does the whole site lack to make the installed equipment work together?

Different systems need different solutions

For VRF, we would start with the zones: which spaces each unit serves, when loads change, what the sensors measure and which commands are actually available.

For ducted DX, the relationship between the refrigeration unit and the air distribution network matters. There is little point optimising start-up time before establishing whether the required airflow reaches the rooms.

For chiller–fan coil systems, the key is the interaction between the cooling source, water circuits and terminal units. In each case, the data required is determined by the task, not by a desire to collect as many tags as possible.

An example of useful forecasting

Suppose an office opens at 9:00. A basic schedule starts the air conditioning at 7:00 regardless of the weather or the building's condition.

A model can estimate when preconditioning needs to start on a particular day. But reduced operating time is not the only thing to check: were comfortable conditions reached by opening time, was a higher peak demand created, and did energy consumption increase later?

Accounting for building dynamics, forecasts and constraints is what distinguishes predictive control from simply changing a timer. Field studies of MPC also highlight the importance of data quality and engineering work during implementation. ²

Where AI is not yet needed

In a server room with three air conditioners, rotating the duty pair and automatically bringing the standby unit online can be implemented using conventional control logic. We would not call rotation itself artificial intelligence.

The next task is to analyse changes in how temperature zones respond under comparable conditions. This may help identify which unit needs attention. But without additional measurements, such observations should not be presented as precise measurements of cooling capacity or COP.

Simple logic is not a weak solution. Adding complexity without a clear benefit is.

What should come before automatic commands

Our approach is to validate the data and observe first, then develop recommendations, and only move to permitted automatic adjustments after approval.

Each project needs defined setpoint limits, rates of change, local control priority and behaviour on loss of communication. External analytics must not bypass factory protection functions or create a single point of failure for critical cooling.

In this model, AII Systems is the technological foundation of AI Engineering, not a one-size-fits-all solution for every building.

How to demonstrate value

First, we define the metric: energy, cost, hours outside agreed temperature limits, peak power demand or the number of manual interventions.

A calculated scenario demonstrates the mechanism and expected effect. Operational verification shows how it works on site. When shifting loads to a cheaper tariff, cost and kWh must be reported separately: they are different outcomes. LBNL research also emphasises that the feasibility and implementation costs of MPC depend on the scale and characteristics of the system. ³

The value of AI is not its presence on a diagram. It lies in a better decision for specific equipment, with an explainable rationale and a verifiable result.