As heat pump installations climb past one million across Britain, the body responsible for keeping the lights on is turning to artificial intelligence to predict exactly where, and when, the electricity grid is most likely to buckle. With Ofgem’s backing, the National Energy System Operator (NESO, formerly National Grid ESO) has begun a winter trial of a forecasting model dubbed the “Grid Co-Pilot” — software designed to anticipate localised demand spikes street by street and defer billions of pounds in physical network upgrades. The pitch is seductive: smarter software instead of dug-up roads. But energy analysts warn the technology could quietly redistribute the risk of blackouts toward the households least able to push back.
Why heat pumps are stressing local networks
The strain is not at the national level, where Britain typically has ample generation, but in the low-voltage “last mile” — the substations and cables serving a few hundred homes each. A gas boiler draws nothing from the grid; a cold-snap heat pump can pull 3 to 6 kilowatts continuously. When dozens of neighbours fire up heat pumps on the same frigid January evening, demand on a single substation can double in minutes.
“The distribution network was sized for an era of evening kettle peaks, not whole streets electrifying their heating simultaneously,” said Dr Priya Nandakumar, a power systems researcher at the University of Strathclyde. “The aggregate national picture looks comfortable. It’s the postcode-level coincidence of demand that triggers a transformer overload — and that’s historically been invisible to operators in real time.”
Replacing those substations and cables wholesale would, by NESO’s internal estimates, run into the tens of billions over the coming decade. The Co-Pilot is the cheaper hypothesis: if you can forecast the spike, you can shave it.
How the Grid Co-Pilot works
The model fuses weather forecasts, smart-meter data, half-hourly consumption patterns and a growing dataset on which homes have installed heat pumps. It then projects demand for each low-voltage feeder up to 48 hours ahead, flagging substations likely to breach their thermal limits.
Where a risk is identified, operators can intervene before the overload occurs — by gently delaying heat pump cycles, nudging electric vehicle charging to off-peak windows, or signalling flexibility providers to reduce draw. Crucially, much of this is intended to be automated and pre-emptive rather than reactive.
- Forecasting: neighbourhood-level demand predictions refreshed every half hour.
- Targeting: identification of at-risk substations days before a cold snap.
- Mitigation: coordinated, often invisible, dimming of flexible loads.
NESO describes the system as a decision-support tool. “We are not handing the grid over to an algorithm,” a NESO spokesperson said. “Human controllers remain in the loop. The Co-Pilot extends their visibility into parts of the network they have never been able to see clearly before.”
The equity question: who gets dimmed first?
Therein lies the controversy. To balance a stressed substation, the system must reduce demand somewhere — and the homes most exposed are those served by older, more constrained infrastructure, frequently in less affluent areas where networks have seen less investment.
“Load-balancing is, by definition, a decision about whose comfort gets traded away,” argued Tom Easterbrook, principal analyst at the consultancy Watt Logic Advisory. “If the optimisation simply protects the network at lowest cost, it will tend to lean on the weakest feeders. Those tend to serve the least-connected, least-affluent households — the very people who can least tolerate a cold home or a charging delay.”
Consumer groups have raised a second concern: transparency. If a heat pump quietly throttles back during a forecasted spike, will households know? And will they be compensated? Current flexibility schemes reward customers who volunteer to shift usage, but pre-emptive automated curtailment blurs the line between voluntary and imposed.
There is also a data-fairness dimension. The model’s accuracy depends on smart-meter coverage, which remains patchy. “Areas with poor smart-meter penetration are forecast with less precision,” Dr Nandakumar noted. “Paradoxically, the neighbourhoods with the oldest infrastructure are sometimes the ones the model understands least — which is precisely where errors are most costly.”
A test case for AI in critical infrastructure
Ofgem has framed the trial as a controlled experiment, with reporting requirements on both performance and distributional impact. The regulator has signalled that any rollout would need safeguards ensuring curtailment is shared fairly and that vulnerable customers are protected. Whether those safeguards survive contact with cost pressures remains the open question.
The stakes extend beyond heating. The same architecture could one day govern EV charging, home batteries and vehicle-to-grid exports — making this winter’s pilot an early test of how far Britain is willing to let predictive AI mediate access to essential services.
What this means
The Grid Co-Pilot represents a genuine bet: that intelligent forecasting can absorb the heat pump surge without the eye-watering cost of rebuilding the low-voltage network. If it works, consumers could be spared years of disruptive upgrades and higher bills. But the technology does not eliminate scarcity — it allocates it. The decisive issue this winter will not be whether the model can predict spikes, but whether the rules governing its interventions distribute the burden equitably, transparently and with the consent of those affected. Get that wrong, and an efficiency triumph could harden into a quiet new form of energy inequality.
Photo by Miguel Á. Padriñán on Pexels
