What AI will change in ventilation — and what conventional controls can do without it

Artificial intelligence is not needed to switch ventilation on to a schedule, maintain supply air temperature or increase outdoor airflow in response to CO₂. A more useful question is: which decision would improve if forecasts and a model of this particular building were added?

Where an additional challenge arises

Imagine an office with variable meeting room occupancy. A large meeting is scheduled for tomorrow morning, solar heating of the façade is expected in the afternoon, and cooling capacity is limited. Basic controls maintain the set parameters. A higher level of control can compare options in advance: when to prepare the rooms, which mode to select and how to avoid a coincident peak.

A weather forecast alone is not enough. An assessment is needed of how the particular room responds to outdoor conditions, occupants and setpoint changes. Model predictive control selects future actions while taking these constraints into account.

What a useful scenario looks like

Before the meeting starts, the system assesses room conditions and expected occupancy, selects an appropriate preparation time and then adjusts operation using actual data once people arrive.

A forecast cannot ‘store fresh air’ for the whole day and remove the need for mandatory ventilation. If the meeting is cancelled or the forecast is wrong, the controls must respond to reality. Otherwise, a complex algorithm becomes little more than a confident timer.

Which data are useful

The necessary data set is selected for the task: temperatures, flow rates, valve and damper positions, commands, modes, outdoor conditions and occupancy information. Assessing energy use requires consumption measurements or a justified calculation model.

It is important to know each signal’s source, units, sensor location and relationship to the equipment. Five verified parameters are more useful than a hundred tags with unclear names. A digital passport preserves this context, while a digital twin is needed only when the task genuinely requires a model of dynamics and interrelationships.

How to implement it without losing control

A sensible sequence is observation, recommendations, verification and only then authorised automatic actions. Limits are set in advance: permitted setpoints, their rate of change, minimum airflow rates, restrictions and priorities.

Standard protective and emergency functions remain independent of the optimisation model. Loss of communication must not leave equipment without basic control. Results are assessed by consumption, hours outside specified parameters, peak power demand or the number of manual interventions, not by the number of sensors or the appearance of a dashboard.

AI is not needed simply to label ventilation as intelligent. It is appropriate where it helps make a better engineering decision and verify the outcome.