WISER and E.ON Test Quantum Machine Learning for Energy Forecasts
WISER and the energy company E.ON have tested whether quantum machine learning can forecast household electricity demand, and posted the results as an arXiv preprint. In plain terms, the task is to predict how much power a hundred households will use over the next few hours, given that their habits move together. The team ran it on both simulators and IBM quantum computers with more than 100 qubits.
How the quantum forecasting models were tested
According to the July 24 report, the study, a collaboration between WISER and E.ON’s quantum team, built two hybrid quantum-classical models and pointed them at an anonymized smart-meter dataset of 103 residential customers. The hard part is forecasting several correlated customers at once, rather than one at a time, and doing that with quantum machine learning has been awkward on real hardware. As WISER’s Vardaan Sahgal said:
“We can now run these complex, multi-output time-series forecasts on real quantum computers with over 100 qubits. While the ‘perfect quantum advantage’ is still waiting for the hardware to get a bit quieter and more reliable, we come very close to it.”
The framing there is candid, and it fits a run of industrial quantum tests on real data, like IQM and Deutsche Bahn’s rail scheduling experiment.
What the results show
The headline results are strong on paper. The QGP model cut average forecasting error, measured as mean absolute error, by 62% on a simulator and 40% on hardware against a classical Gaussian Process baseline. The second model, KQRC-RM, cut it 37% on a simulator, though its hardware version was more sensitive to noise, with no clean hardware figure given.
Three qualifiers travel with those numbers. The researchers measure them against selected classical baselines in a chosen comparison and not the strongest classical method available. The simulator beat the hardware by a wide margin, which is where the noise still bites. And it’s a preprint from the utility’s own quantum scientists, without independent benchmarking, so it sits alongside other hedged industrial trials. A recent hybrid portfolio-optimization test and the ongoing energy-to-solution debate are cut from the same cloth.
On the 100-qubit run, 80% of customers landed in low or medium error, which shows both the reach and the ceiling. That scale of experiment echoes other IBM-hardware evaluations, like Singapore’s quantum mission-planning test. The genuine read is a promising, carefully hedged benchmark on a real problem, as opposed to proof that quantum beats classical at load forecasting.