Quantum machine learning can boost energy demand forecasts

The joint project was aimed to overcome the challenge in energy systems planning of how to forecast demand accurately across multiple correlated customers.

Or put another way, imagine trying to predict how much electricity 100 different households would use in the coming hours, some using more power when it's cold, others when they watch TV, and their habits often mirroring each other.

The project drew on two complementary hybrid quantum classical approaches to model a dataset comprised of anonymised real smart meter consumption data for 103 household consumers.

Also of interest
UK quantum computing firm gets $4.5m for hydrogen project
Quantum AI framework to reduce data centre energy consumption

The first model, kernelised quantum reservoir computing with repeated measurement (KQRC-RM), used recurrent quantum dynamics and repeated measurements to model temporal structure and cross-stream correlations in smaller customer subsets.

The second, a projected quantum kernel gaussian process (QGP), used a more hardware-efficient projected quantum kernel to support multi-output forecasting on larger groups of customers.

In the smaller scale benchmarking study, the QGP model reduced the average error (mean absolute error, MAE) relative to a classical multi-output Gaussian process baseline by 62.01% on the simulator and 40.37% on the hardware.

In the same study, KQRC-RM reduced the average MAE relative to an echo state network with kernel ridge regression by 36.92% on the simulator, while the hardware implementation remained more sensitive to noise.

“Quantum machine learning models that can forecast multiple time series values have been somewhat elusive in the field, yet classically exist everywhere in industry,” commented Dr Corey O’Meara, Chief Quantum Scientist at E.ON Digital Technology.

“We were happy to push the boundaries of hybrid quantum algorithm development to make that happen for a real world use-case and run benchmarks using 100+ qubits on IBM quantum computers.”

Multi-output time series forecasting in energy systems is challenging because of non-linear dynamics, multi-scale seasonality and strong dependencies across the correlated series. Classical statistical models often struggle with these non-linearities, while flexible machine learning approaches typically require substantial data and computational resources.

Forecasting at scale

The study has demonstrated that quantum-enhanced approaches can already be studied experimentally on practical time series forecasting tasks at meaningful scale.

It also has demonstrated the importance of hardware-aware design. The QGP model scaled to a 100-qubit utility-scale experiment, where 80% of customers fell into low or medium error categories, underscoring both the promise of the approach and the influence of device noise on performance.

This is critical for the reliable operation of modern energy systems, including load balancing and renewable integration, especially as quantum hardware improves.

Vardaan Sahgal, Quantum Solutions Head at WISER, said the project represents a significant step toward practical quantum assisted forecasting.

"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. And in the process establish that hybrid quantum models can outperform specific classical baselines in structured energy forecasting tasks under [quantum computation] constraints.”

In conclusion, the study indicates that comparing the KQRC-RM and QGP models, KQRC-RM should be used when the training dataset is large but the number of time series to predict is small, whereas QGP is potentially the better choice when there is a small amount of training data and a large number of time series to simultaneously predict.

E.ON has been a pioneer in the application of quantum computing to current energy sector challenges, such as peer-to-peer trading in energy communities and scheduling EV charging and discharging to support the grid.

WISER is a not-for-profit organisation driving applied R&D across quantum, AI, machine learning, and computational science for commercial, government and academic partners worldwide.