E.ON tests quantum models for demand forecasting

E.ON and the Washington Institute for STEM, Entrepreneurship and Research have tested hybrid quantum-classical models for forecasting electricity demand across multiple customers.

Using anonymised smart-meter consumption data from 103 households, the project examined whether several related demand series could be processed together. The models were designed to identify correlations between customers whose electricity use may change in response to common weather, behavioural, or tariff conditions.

Two approaches were assessed during the work. Kernelised quantum reservoir computing with repeated measurement used recurrent quantum dynamics to model temporal behaviour and relationships within smaller customer groups, while a projected quantum-kernel Gaussian process was developed for multi-output forecasting across larger groups.

In a smaller benchmark, the projected quantum-kernel model reduced mean absolute error by 62.01% on a simulator and 40.37% on quantum hardware relative to a classical multi-output Gaussian-process baseline. The reservoir-computing approach reduced average simulation error by 36.92% against its selected comparator, although physical hardware produced less consistent results.

A further projected quantum-kernel experiment used more than 100 qubits, with 80% of customers falling into low or medium error categories. The trial demonstrated that the method could operate across a larger set of demand series while also exposing the continuing effect of hardware noise.

Forecasting becomes a distribution-system function

Demand forecasting has long supported generation scheduling and national balancing, but distribution systems now contain more heat pumps, electric vehicles, batteries, rooftop solar, flexible tariffs, and automated building controls. Their combined behaviour creates demand profiles that are less uniform and increasingly responsive to local conditions.

Forecasting each customer independently can overlook relationships between loads, while merging them into a single aggregate can conceal useful variation. Multi-output models seek to retain both elements by identifying shared patterns without discarding differences between individual time series.

Operational value will depend on the forecast horizon. Day-ahead predictions can inform market positions and flexibility procurement, whereas shorter forecasts may support local balancing, voltage management, or the operation of constrained feeders.

Errors also have different consequences depending on scale. A forecast deviation affecting one secondary substation may be manageable through local flexibility, while a similar percentage error across a distribution region can alter reserve requirements and wholesale-market exposure.

Network digitalisation is already bringing monitoring, forecasting, and control into closer alignment. The combination of Schneider Electric’s network systems with Kraken’s flexibility controls reflects the same movement towards more active coordination of distributed resources.

Quantum performance remains application-specific

The study does not establish a general quantum advantage over conventional forecasting. Results were measured against particular baselines on defined datasets, and performance changed when the models moved from simulation to physical hardware.

Quantum processors remain sensitive to errors introduced through gates, measurement, calibration, and environmental effects. Circuits that perform strongly in simulation can lose accuracy on hardware, particularly where repeated measurements or deeper processing increase exposure to device instability.

Classical forecasting methods are also advancing rapidly, with gradient-boosting models, neural networks, probabilistic systems, and conventional Gaussian processes operating on mature computing infrastructure. These methods already benefit from established tools for validation, cybersecurity, maintenance, and integration with utility platforms.

A quantum model would therefore need to deliver a sustained improvement in accuracy, processing demand, or scalability before deployment in an operational energy system. Hybrid approaches may offer a more practical route because they retain conventional processing around a tightly defined quantum component.

Further testing will need to cover larger and more varied datasets, different forecast horizons, seasonal changes, missing values, and operational disturbances. Engineers will also require clear measures of uncertainty so that abnormal results can be identified without surrendering control-room judgement to an opaque model.

The project provides a structured comparison using real energy data rather than a deployable utility product. Its next stage will depend on whether improved hardware and broader testing can preserve the simulated gains under live operating conditions.