A conventional BMS PID loop reacts to error: it measures a deviation from setpoint and corrects it after the fact. Predictive control inverts that relationship — it forecasts the load an hour or a day ahead and pre-positions plant to meet it before the error occurs. Australian commercial towers are now the proving ground for this shift, and the difference between a project that delivers real savings and one that quietly reverts to manual override sits almost entirely in the controls schematic, not the analytics licence.
PID Control Is Reactive by Design — That's the Problem Being Solved
A standard AHU discharge-air control loop compares a live sensor reading against a fixed or reset setpoint and drives a valve or damper proportionally to the error, with integral and derivative terms smoothing the response. This works well for steady-state conditions but is structurally blind to what's coming: a PID loop doesn't know a cool change is arriving in Sydney at 3pm, or that the building will be near-empty on the Tuesday after a public holiday. Model predictive control (MPC) adds a forecasting layer above the PID loop — it doesn't replace the low-level control, it feeds the low-level loop a better setpoint trajectory, calculated from a thermal model of the building plus a weather forecast feed and occupancy schedule.
What the Thermal Model Actually Needs
An MPC engine needs, at minimum, a resistance-capacitance (RC) thermal model of each zone or floor plate — effectively a simplified digital twin of how fast the space gains or loses heat — trained against 6-12 months of historical BMS trend data covering a full range of Australian seasonal conditions. This is the part most commonly under-scoped: a model trained only on autumn shoulder-season data will mis-forecast a 40°C Melbourne heatwave or a sub-5°C Canberra morning, because it has never seen the building behave under those conditions. Practically, this means predictive control projects should budget a shadow-mode observation period spanning at least one full cooling season and one full heating season before setpoint write-back is switched on.
- Weather forecast integration typically pulls from the Bureau of Meteorology API or a commercial forecast provider, refreshed hourly, feeding the MPC engine rather than the BMS controller directly.
- Zone-level RC models need at minimum: dry-bulb temperature, supply air temperature, valve/damper position and occupancy trend — most of which already exists in a compliant BMS trend database.
- Thermal mass pre-conditioning (pre-cooling overnight using off-peak tariffs, or pre-heating ahead of a Canberra cold snap) only pays off if the model can accurately predict the following day's peak load — a poor forecast wastes energy rather than saving it.
- Demand-response integration with AEMO wholesale price signals is increasingly bundled into the same predictive layer, shifting load away from price spikes rather than purely thermal optimisation.
Why the Controls Schematic Decides the Outcome
Layering MPC onto an existing BACnet/IP network doesn't usually require new field devices, but it does require a clear demarcation on the controls schematic between what the BMS controller still owns (safety interlocks, freeze protection, high-limit overrides) and what the MPC layer is permitted to adjust (setpoint trajectories, start/stop timing, economiser crossover points). Projects that skip this step and let the analytics vendor "just write to the points database" run into two recurring failure modes: the MPC layer fighting the BMS's own optimisation logic (both trying to reset the same setpoint on different schedules), and commissioning teams unable to tell, during a fault, whether an unexpected setpoint came from the BMS sequence or the ML layer.
Design takeaway: Specify a clear setpoint-authority hierarchy on the controls schematic before the analytics platform is selected — which points the MPC layer may write to, what override priority the BMS retains, and how a technician disables ML control from the front end during fault-finding.
Sequence-of-Operations Changes Worth Writing Down
Predictive control changes the actual sequence-of-operations narrative, not just the setpoint values, and Australian commissioning agents increasingly expect this documented separately from the standard BMS SOO. Typical additions include a pre-conditioning start-time algorithm (replacing a fixed optimal-start routine with one that varies by forecast), a demand-response override sequence, and an explicit fallback sequence describing exactly what the plant does if the weather feed or analytics platform becomes unavailable — this fallback is a genuine commissioning test item, not a theoretical edge case, given how often campus-wide internet outages affect Australian buildings during storm events.
Where Predictive Control Is Paying Off Fastest
The strongest, fastest-proven returns in Australian towers are in central plant sequencing — chiller staging, condenser water reset and static pressure reset — rather than zone-level comfort control, because central plant has fewer, higher-value control points and a more predictable thermal response than dozens of individually-occupied tenancies. Zone-level predictive comfort control is maturing but needs the sensor-density uplift discussed elsewhere in this series before the payback is reliable.
Frequently Asked Questions
Does predictive control need new field devices on an existing Australian BMS?
Rarely at the sensor level. Most predictive control layers reuse existing temperature, flow and valve-position points. What's usually missing is a weather-data input and enough historical trend depth — at least one full cooling and heating season — for the model to learn the building's actual thermal response.
How long does a predictive control model take to become reliable?
Most Australian deployments run in a shadow/advisory mode for a full seasonal cycle (typically 6-12 months) before setpoint write-back is enabled, so the model's forecasts can be validated against actual outcomes across a range of weather conditions before it's trusted with control authority.
Can predictive control replace the existing BMS controller logic?
No — and it shouldn't try to. Predictive control sits as a supervisory layer that adjusts setpoints and schedules within safe bands; the underlying PID and safety interlock logic in the BMS controller remains the layer that actually operates plant and protects equipment.