Two hours of warning is worth more than a bigger chiller.
A demand model that runs ahead of the building, and a control layer with just enough authority to act on it before the peak arrives.
Three-dimensional model of the simulated campus. Drag to orbit, Ctrl or Command plus scroll to zoom, select a building to focus it. Arrow keys orbit, plus and minus zoom, R resets the view.
The load the model argues with
A campus does not draw power smoothly. A lecture block goes from nearly nothing to 120 kW in the twenty minutes either side of nine o'clock, and the plant that has to answer it only finds out once it has happened. Everything below is built on one simulated campus, described in full so the figures can be checked against it.
Modelled campus peak
1.46× the 176 kW daily mean
ModelledLecture block at 10:00
300 seats, full ventilation
ModelledDormitory overnight base
Flat from 01:00 to 06:00
ModelledLead time the plant needs
Chiller ramp plus store discharge
Design targetOne signal chain, five layers
Measure, predict, act, store, supply. The right-hand column is what each layer contributes to the peak reduction — two of the five contribute kilowatts, and the other three are what make those two possible. The column adds up, which is the only honest way to present a headline number.
Sensing mesh
312 nodes · 1 Hz · MQTT / TLS
Sub-metering at feeder level, CO₂ and occupancy per zone, flow and temperature difference across the chilled-water loop. Continuous rather than a 15-minute poll, because the control decisions below are only as good as the most recent sample and a poll interval is a delay you cannot recover.
Design targetDemand model
LSTM · 2 h horizon · 15 min steps · MAE 7 kW · MAPE 4.0%
One model per building rather than one for the campus, because a dormitory and a lecture block have almost nothing in common on a Tuesday morning. Backtested on simulated years of the campus model; the residual, not the accuracy headline, is what gets reported after every run.
ModelledLoad-shed authority
Lighting −5 kW · HVAC setpoint +1.5 °C −8 kW
Armed when CO₂ in occupied zones passes 600 ppm and the forecast shows the peak forming, not on a fixed clock. Both actions are reversible within one control interval, and both are bounded: the setpoint shift is capped at 1.5 °C so nobody has to negotiate with the building to get their room back.
Design targetPhase-change thermal store
SiC-doped paraffin · ~200 kJ/kg · k 0.8 W/m·K
Charged overnight when ambient is 8–10 °C cooler and the chiller runs at COP 3.3 instead of 3.0, then discharged into the chilled-water loop across the midday peak. Silicon carbide raises conductivity from roughly 0.2 to 0.8 W/m·K, which is what makes the discharge fast enough to matter over a two-hour window.
Design targetEnvelope and floor harvesting
Perovskite glazing 5% PCE at 70% VLT · piezo floor ≈5 mW/m² peak flow
Semi-transparent glazing at a 1.8 eV gap yields roughly 0.15 kWh/m²/day on a south façade — real, but small next to a 286 kW campus. The piezoelectric corridor tile is smaller still, about 5 mW/m² while a class changes and a fifth of a milliwatt averaged over a day. We keep both because that is enough to run the duty-cycled sensor nodes in the same corridor, not because they move the energy bill.
Design targetDrive the control layer
The building-management view, running on the same simulated campus the figures above are derived from. Arm the load-shed controls and the forecast peak flattens in front of you. Nothing here is metered from a real building, and the panel says so on its face.
Campus demand · control view
Loading forecast model, load-shed controls and site plan
What a year of this returns
Against the 1,546 MWh modelled baseline, over a full year of the simulated campus. Each figure carries how it was arrived at; none of them is a metered result, because there is nothing fielded yet to meter.
Consumption avoided per year
9% of a 1,546 MWh modelled baseline
ModelledCO₂e avoided per year
At 0.60 t CO₂e/MWh grid intensity
ModelledPeak demand removed
7.8% of the 257 kW modelled peak
ModelledForecast MAE at 2 h
4.0% MAPE against a 176 kW mean
ModelledChiller COP on stored cooling
From 3.0, by charging overnight
Design targetControl interval
Every action reversible within one step
Design targetWho sits in front of it
Three roles, three different views of the same data, and one decision each. If a role cannot point at the decision the system changes for them, it does not belong on this page.
Campus energy manager
01Forecast panel · peak window
Whether to arm load-shed for the coming window, or ride the peak and take the demand charge.
The decision is made from last month’s bill, which arrives long after the peak that set it. GEST moves it two hours before.
Building services engineer
02Zone load meters · chilled-water ΔT
Which zone is drifting away from its own baseline, and whether that is occupancy or a fault.
A failing valve reads as a slightly higher bill for months. Against a per-zone forecast it reads as a residual that stops coming back to zero.
Laboratory manager
03Feeder sub-metering · shed exemptions
Which circuits are exempt from setback, and proving that an experiment was never at risk.
Shedding is avoided entirely because nobody can guarantee what it will touch. An explicit exemption list is what makes it usable at all.
- University campuses with a demand charge and no sub-metering below the main incomer
- Research buildings where a single equipment start can add 30 kW without warning
- Estates teams running plant well past its design life, where new capacity is not on the table
What facility teams ask first
01How far ahead does GEST forecast demand, and how wrong is it?
Two hours, in 15-minute steps. In backtesting against our campus simulation the mean absolute error is 7 kW, or 4.0% of the 176 kW daily mean. That is a modelled result, not a metered one — we have no fielded deployment, and we would rather say so than quote an accuracy figure with nothing behind it.
02Where does the 20 kW of peak reduction actually come from?
Three places, and they add up: 5 kW from setting back non-critical lighting, 8 kW from raising the cooling setpoint by 1.5 °C, and 7 kW from discharging cooling banked overnight in the phase-change store. Against a 257 kW modelled peak that is just under 8%. Nothing here is a step change; it is three ordinary actions taken early enough to count.
03What happens if the forecast is wrong?
Every action is bounded and reversible within one 15-minute control interval, and the load-shed layer is armed rather than forced — it acts only when the forecast and the live CO₂ reading agree that a peak is forming. A bad forecast costs a setback that is reversed at the next step, and a logged residual that goes into the next training set.
04Is the instrument view on this page showing real data?
No. It runs on our campus simulation so the interface can be explored, and it is labelled as simulated wherever it appears. The model behind it is the same one the figures on this page are derived from, so the panel and the page cannot quietly disagree.
Run it against your own load curve
Send a week of half-hourly metering from a building you know well. We will fit the model to it and send back where the peak forms, what the forecast error looks like on your data, and how much of the peak this would actually reach.