Predictive & Model-Based Control

The overwhelming majority of building control is reactive feedback. A PID controller compares a measurement to a setpoint and drives the error toward zero: room too warm, open the cooling valve; too cold, close it. It is robust, well understood and entirely appropriate for tight inner loops - holding a supply-air temperature, a coil valve, a fan speed. But feedback has a structural blind spot: it only responds to what has already happened. By the time the room is warm, the heat is already in the building, and you are paying to remove it at whatever the electricity costs right now.

Buildings, though, are slow and predictable in ways feedback ignores. A concrete floor takes hours to warm; a crowd's heat and CO2 arrive on a schedule; tomorrow's weather is forecast with useful accuracy; the price of energy is often known in advance. Feedforward control begins to use this - nudging output from a disturbance you can see coming, like boosting cooling as outdoor temperature climbs. MPC is feedforward taken to its logical, optimised conclusion: not a single nudge but a whole plan over a horizon, chosen to minimise cost while respecting comfort. Where reactive control asks what is wrong now?, predictive control asks what is about to happen, and what should I do now to be ready?

MPC runs a loop of three ingredients. First, a model of the building's thermal behaviour - how zone temperatures respond to plant, weather and internal gains. It can be physics-based (a reduced-order thermal-network model), data-driven (learned from the building's own history, and a natural job for the digital twin), or a grey-box blend. Second, forecasts over a horizon of typically a few hours to a day: weather, occupancy, solar gain, and energy price or grid-carbon signals. Third, an optimiser that searches over possible control sequences and picks the one that minimises a cost - energy spend, peak demand, carbon - subject to constraints, the most important being that comfort stays inside an agreed band.

The defining trick is the receding horizon. MPC computes an optimal plan across, say, the next twelve hours, but only executes the first step - then, at the next interval, it takes fresh measurements, re-forecasts and re-optimises, sliding the horizon forward. This is what keeps it honest: every cycle corrects for whatever the model and forecast got wrong last time, folding feedback back into a fundamentally feedforward scheme. It sits as supervisory control above the existing BMS - MPC decides setpoints and sequences, and the trusty PID loops underneath still do the fast, safe work of holding them.

The cleanest illustration is pre-cooling. An office in a hot climate faces a punishing afternoon: outdoor temperature peaks, everyone is in, and electricity is dearest exactly then - often with a demand charge on the single highest half-hour of the month. A reactive system simply runs the chiller flat-out through the peak, buying the most expensive kilowatt-hours at the worst moment and setting a costly demand peak.

MPC, seeing the forecast the night before, does something cleverer. In the cool early morning, when power is cheap and the chiller is efficient, it deliberately over-cools - pushing the building slightly colder than needed and charging the thermal mass of the slabs and structure with coolth. When the afternoon peak hits, it lets the space coast: the stored cool keeps occupants comfortable while the chiller throttles right back, shaving both the energy bill and the demand charge, and easing load on the grid at its most stressed. Crucially, comfort never leaves the agreed band - the optimiser treats it as a hard constraint, not a variable to sacrifice. The same logic drives load-shifting for batteries and thermal storage, and demand response where a building is paid to trim load on signal. In every case the win comes from one thing reactive control cannot do: acting on the future before it arrives.

MPC's hardest ingredient is the model, and there are three broad ways to get one. A white-box (physics-based) model derives thermal behaviour from first principles - envelope, mass, plant characteristics - and is interpretable but laborious to build and calibrate. A black-box (data-driven) model learns the input-output relationship purely from the building's own logged history using regression or machine learning; it is quicker to fit where good data exists but only as trustworthy as that data and blind outside the range it has seen. A grey-box blend - a reduced-order physical structure with parameters fitted from data - is the common pragmatic middle, capturing the essential physics while staying calibrated to the real building.

This is exactly where the digital twin earns its place in the control story. A twin that already carries the building's structure and is continuously fed live data is a natural home for the predictive model MPC needs: it can be identified from operational history, kept calibrated as the building changes, and used to simulate candidate control plans before any of them touch the real plant. That simulate-before-you-act ability is both an optimisation engine and a safety net - bad plans can be caught in the model. It also clarifies the architecture: MPC is not a monolith bolted onto the BMS but a supervisory brain that leans on the twin to think and on the existing control loops to act. The better and better-maintained the twin, the more the controller can safely plan - and the more fragile or stale the model, the more conservative the guardrails around it must be.

The advantage of MPC is structural, not incremental. Because it optimises over a horizon against a real objective, it naturally does things reactive control cannot: it trades a little comfort headroom now for big savings later, coordinates many pieces of plant toward one goal, shifts load in time, and respects constraints explicitly. Field studies of MPC on commercial HVAC commonly report meaningful energy and peak-demand savings versus conventional control - the exact figure depends heavily on the building, climate and tariff, so treat any single headline number sceptically and insist on measured results for your building.

The honesty clause matters, because MPC is genuinely hard. It lives or dies by its model and its forecasts - a model that misjudges the building's thermal response, or a bad weather or occupancy forecast, produces confidently wrong plans; the receding horizon limits the damage but does not erase it. Models are costly to build and must be maintained as the building changes. And autonomy raises the stakes: an optimiser told only to minimise energy will, unless constrained, cheerfully drift comfort, over-discharge a battery, or fight another system. So MPC needs hard guardrails - comfort bands, equipment limits, rate limits - and it must never touch life-safety functions. A sensible deployment therefore always keeps a fallback: if the MPC layer loses confidence, times out, or is switched off, the building drops cleanly back to conventional reactive control rather than losing control entirely, and the optimiser is best commissioned in an advisory shadow-mode first, so that its proposed plans can be checked carefully against measured reality before it is ever trusted to act on the plant. Done well it is one of the highest-value things a digital twin enables; done carelessly it is an expensive way to make confident mistakes.