insider brief
- WISER and E.ON have completed a joint research effort exploring a hybrid quantum machine learning approach for electricity demand forecasting.
- In this project, we evaluated our quantum model using an anonymized dataset of 103 residential customers, both in a simulator and on real quantum hardware.
- The researchers found that the projected quantum kernel model improves predictive accuracy against a selected classical baseline under short-term quantum computing conditions.
Press Release – Washington Institute for STEM Entrepreneurship Research (WISER) announces successful completion of research collaboration with E.ON Energy demand forecasting using quantum machine learning. The joint project is arXivWe consider two hybrid quantum-classical approaches for predicting correlated power consumption time series: Kernelized Quantum Reservoir Computing with Repeated Measurements (KQRC-RM) and Projected Quantum Kernel Gaussian Process (QGP).
Industrial use case problem statement
Imagine trying to predict how much electricity 100 different families/households will use in the next few hours. Some families use more electricity when it’s cold, others use more electricity when watching TV, and their habits often mirror each other.
This research addresses a central challenge in energy system planning: how to accurately forecast demand across multiple correlated customers while operating under realistic short-term quantum hardware constraints. The team evaluated both simulator and real quantum hardware models using an anonymized smart meter dataset of 103 residential customers, showing that the quantum-enhanced approach can already be studied experimentally in practical prediction tasks at a meaningful scale.
This is important for the reliable operation of modern energy systems, including load balancing and renewable energy integration.
“It’s possible!… We can now perform these complex multi-output time series predictions on real quantum computers with more than 100 qubits. “Full quantum supremacy” is still waiting for hardware to become a little quieter and more reliable, but we are very close to it. And in the process, we establish that hybrid quantum models can outperform certain classical baselines in structured energy prediction tasks under NISQ constraints. This represents an important step towards commercialization. Quantum-assisted prediction. ” Vardhan Sagar, said the sage.
From quantum methods to energy prediction
This project combined two complementary approaches. KQRC-RM used iterative quantum dynamics and repeated measurements to model temporal structure and cross-stream correlation in a small subset of customers. QGP, on the other hand, supported multi-output prediction for larger groups of customers using a more hardware-efficient predictive quantum kernel.
In a small benchmark study, the QGP model reduced the average MAE by 62.01% in simulator and 40.37% in hardware compared to a classical multi-output Gaussian process baseline. In the same study, KQRC-RM reduced the average MAE by 36.92% on a simulator compared to an echo state network using kernel ridge regression, but the hardware implementation remained sensitive to noise.
“Quantum machine learning models that can predict multiple time-series values have been somewhat elusive in this field, but classically they exist everywhere in industry. We were excited to push the boundaries of hybrid quantum algorithm development and bring it to life in a real-world use case, benchmarking it with more than 100 qubits on an IBM quantum computer,” said Corey O’Meara, Principal Quantum Scientist at E.ON Digital Technology GmbH. says Dr.
Practical relevance to energy systems
Forecasting electricity demand is difficult due to nonlinear dynamics, multiscale seasonality, and strong dependence between correlated series. Classic statistical models often struggle with these nonlinearities, while flexible machine learning approaches typically require large amounts of data and computational resources.
This study shows that quantum machine learning can be useful for structured multi-output forecasting, where relationships between time series matter, especially as quantum hardware improves.
This paper also demonstrates the importance of hardware-aware design. The QGP model was extended to a 100-qubit utility-scale experiment, with 80% of customers falling into the low or medium error category, highlighting both the promise of this approach and the impact of device noise on performance.
The findings are particularly relevant for utilities and energy providers dealing with volatile load patterns, distributed consumption, and increasing demand for predictive intelligence.
