QCNEWS
The German utility E.ON and a US research lab used a quantum forecasting method built to survive hardware noise to forecast 103 anonymized households' power use on actual IBM quantum hardware.
A 100-qubit quantum machine learning model beat a classical benchmark at forecasting electricity demand for 103 European households at once. The result, posted this month on the preprint server arXiv by researchers at the Washington Institute for STEM, Entrepreneurship and Research (WISER) and the German utility E.ON, is one of the clearest published tests of "utility-scale" quantum machine learning on a real industrial problem rather than a synthetic dataset.
The team ran two hardware-aware quantum algorithms on IBM Quantum hardware, focused on multi-output time-series forecasting across 103 anonymized smart meters — the kind of data an electricity retailer uses to predict how much power a household block will pull in the next hour. A qubit is the basic unit of information in a quantum computer, and 100 qubits is roughly the boundary of what counts as "noisy intermediate-scale quantum" (NISQ) hardware: machines large enough to be interesting and small enough that environmental noise still degrades results.
The two algorithms were the story. The first, Kernelized Quantum Reservoir Computing with Repeated Measurement (KQRC-RM), combines coupled quantum reservoirs with a kernel ridge regression readout. It modeled temporal patterns within smaller household subsets. The second, Projected Quantum Kernel Gaussian Process (QGP), replaced the global fidelity-based quantum kernels that dominate the quantum ML literature with kernels built from local reduced-density state observables. Instead of asking the chip to compute an expensive global property of the whole 100-qubit wavefunction, the team asked it to estimate smaller, local features that are less vulnerable to noise.
That architectural choice, not raw qubit count, is what made the 100-qubit run possible. QGP cut mean absolute error (MAE) by 62.01% against a classical multi-output Gaussian Process baseline in simulation, and by 40.37% on the actual IBM hardware. KQRC-RM, run in simulation, posted a 36.92% MAE reduction against an Echo State Network baseline; on real hardware it was "more sensitive to environmental device noise," and the aggregator summary does not give a comparable hardware-side percentage.
The 22-point gap between QGP's simulated and real-hardware improvements is what the authors call the honest cost of NISQ noise — the degradation that comes from a quantum chip interacting with its environment faster than the computation finishes. It is also the gap that most quantum ML press coverage flattens by quoting the simulation figure alone.
A separate 100-qubit, topology-aware QGP execution forecasted 100 multi-output customer time series in a single pass. According to the paper, 80% of the forecasted outputs fell into low- or medium-error buckets, though the source excerpt truncates the exact threshold definitions.
The result carries three honest limits. It is a preprint, not a peer-reviewed paper. The 103-meter dataset is anonymized and small by industrial standards. And the projected-kernel advantage is local to this benchmark until independent teams reproduce it.
By reframing the 100-qubit boundary as an architectural question rather than a headcount, the WISER-E.ON benchmark gives industrial quantum ML a sharper test: which kernel designs survive real device noise, and which collapse. The team asked the chip for local observables rather than global properties, and the model held. On the WISER project page, the work sits inside a broader quantum-learning program; the Quantum Computing Report podcast with Corey O'Meara, E.ON's chief quantum scientist, is the natural place to hear how the utility side is reading the result.