💻 computer science

Parameter-performance scaling of hybrid quantum-classical policies for demand-charge-aware battery control

This study demonstrates that while a compact hybrid quantum-classical policy can marginally outperform a size-matched classical network in demand-charge-aware battery control, the advantage is narrow, task-specific, and does not constitute a general quantum advantage over faster classical baselines.

Original authors: Eunsung Oh

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eunsung Oh

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the commercial world, electricity bills are often more than just a charge for how much power a building uses. For many businesses, a significant portion of the cost comes from a "demand charge," a fee based on the single highest moment of power usage in a month. Imagine a building drawing a massive amount of electricity for just fifteen minutes; that brief spike can set the price for the entire month, regardless of how efficiently the building runs the rest of the time. To avoid this, building managers use large batteries to store energy and release it precisely when demand is about to peak, effectively shaving off that dangerous spike. However, controlling these batteries is a difficult puzzle. If a manager drains the battery too early to handle a small spike, the battery might be empty when the true, massive peak arrives later in the month, leaving the building exposed to the high fee. The goal is to find a control strategy that balances saving money on the bill with preserving the health of the battery, all while reacting to changing conditions in real time.

For years, scientists have explored using quantum computers to solve such complex control problems, hoping that the unique way quantum systems process information could lead to smarter, more compact decision-making tools. The idea is that a hybrid system, mixing classical computer logic with quantum circuits, might achieve the same high performance as a large traditional computer program but with far fewer adjustable settings, or "parameters." This would be a major advantage if the quantum part could do more with less. However, until now, this claim had rarely been tested rigorously against strong, traditional computer models across a wide range of sizes. A new study by Eunsung Oh at Gachon University puts this idea to the test in a realistic simulation of battery control for hundreds of office buildings, asking a simple but critical question: can a hybrid quantum-classical system really control a battery better than a standard computer program when both are limited to the same small size?

To answer this, the researchers created a massive, locked-down experiment involving 549 simulated office buildings in Los Angeles. They used a standardized electricity tariff that includes a heavy demand charge, mimicking the real-world pressure to avoid power spikes. The team trained three different types of control policies to manage the batteries: one purely classical neural network, and two hybrid models that used small quantum circuits. All three types of models were given the exact same information about the building's energy use, the time of day, and the battery's status, and they were all trained to mimic a perfect, all-knowing optimizer. The researchers then tested these models on a set of buildings they had never seen before, ensuring that the results were not just lucky guesses from the training data. They compared the models at four different sizes, ranging from very small to quite large, to see how performance changed as the number of parameters increased.

The results revealed a nuanced picture that challenges the broad idea of a universal quantum advantage. The study found that one specific hybrid design, which used a skip-connected quantum circuit, did indeed perform slightly better than a traditional computer model when both were restricted to a very small size of about 1,400 parameters. In this specific, compact configuration, the hybrid model managed to lower the building's electricity costs and reduce battery wear slightly more effectively than its classical counterpart. This suggests that for very small, resource-constrained applications, a hybrid approach might offer a marginal edge in efficiency. However, this advantage was narrow and did not hold up across the board. When the researchers looked at other sizes, the hybrid models did not outperform the classical ones; in fact, at larger sizes, the classical models were often superior or performed just as well. Furthermore, the study showed that the hybrid model's success was heavily dependent on a specific engineering design for how it made decisions, rather than being an inherent property of the quantum circuit itself.

Perhaps more importantly, the study highlighted a significant trade-off. While the hybrid model was compact in terms of the number of parameters it needed to store, it was not efficient in terms of time. Running the hybrid model on a standard computer processor took roughly thirty-five times longer to simulate than running the classical model. This means that while the hybrid model might be smaller in size, it is much slower to operate in a simulation, making it less practical for real-time control unless future quantum hardware can execute these circuits much faster. The researchers also noted that their results were based on perfect, noise-free simulations of quantum circuits. In a real quantum computer, which is currently prone to errors and noise, the performance might be even lower. Therefore, the study concludes that while there is a small, bounded benefit to using a hybrid quantum-classical approach for this specific task at a specific small size, it is not a general solution that outperforms classical methods in all cases. The classical neural network remains a strong, fast, and reliable baseline, and the promise of quantum advantage in this field remains a narrow, conditional possibility rather than a broad revolution.

Technical Summary: Parameter-Performance Scaling of Hybrid Quantum-Classical Policies for Demand-Charge-Aware Battery Control

Problem Statement
The paper addresses the control of behind-the-meter battery storage systems in commercial buildings, specifically targeting the minimization of monthly demand charges. This is formulated as a long-horizon, multi-objective control problem where the battery must balance bill reduction against battery degradation while avoiding premature depletion before the true monthly peak demand occurs. The study investigates the hypothesis that hybrid quantum-classical models can achieve comparable control performance to classical neural networks with fewer deployable actor parameters (parameter efficiency), rather than claiming a general superiority in absolute performance or computational speed.

Methodology
The study employs a rigorous, locked-test experimental design to evaluate parameter-performance scaling across different model families.

Dataset and Cohort: The benchmark utilizes 549 simulated office buildings from the ComStock 2025 Release 2 (AMY2018), filtered for Los Angeles County (ASHRAE climate zone 3B) and built in 1980 or later. The dataset is split at the building level into 329 training, 110 validation, and 110 locked-test buildings.
Control Task: The task involves 15-minute interval control over a full summer month (June–September) under a standardized PG&E B-10 demand-charge tariff. The objective is a weighted sum of electricity costs and degradation costs, conditioned on a preference weight $w$ .
Model Families: Three policy families were compared, all sharing the same 16-dimensional observation vector, preference vector, and inference-time isotonic wrapper:
1. MLP-Cap: A classical Multi-Layer Perceptron.
2. VQC-Cap: A skip-connected hybrid Variational Quantum Circuit (VQC) where only the observation vector is encoded into a 4-qubit circuit, while the preference vector bypasses the quantum block via a classical skip connection.
3. VQC-RegimeCap: A variant with a smooth preference-regime-gated demand-cap head.
Experimental Protocol: Models were trained via behavioral cloning from a perfect-foresight oracle. The study evaluated four parameter budgets (~500, ~1,400, ~4,000, and ~12,000 parameters). Crucially, all architecture choices, hyperparameters, and performance thresholds were fixed using only the validation split before the locked test was executed.
Simulation Environment: All quantum circuits were evaluated using exact analytic expectation simulations (state-vector backend) on a classical CPU to isolate parameter efficiency from hardware noise and finite-shot sampling effects.
Metrics: Performance was measured using regime-balanced normalized regret, bill savings, and degradation costs. Statistical significance was assessed via building-level paired bootstrapping and Holm–Bonferroni-adjusted permutation tests.

Key Contributions

Frozen Benchmark: Established a 549-building demand-charge benchmark with a pre-registered, locked train/validation/test split and SHA256-anchored model manifests to prevent overfitting and selection bias.
Parameter-Performance Frontier: Conducted a same-action-projection comparison of hybrid VQC families against strong classical MLP baselines across multiple parameter budgets, isolating the architectural contribution of the hybrid block.
Validation-Derived Thresholds: Introduced a "threshold ladder" derived solely from validation data to define comparable performance regimes, replacing single tuned thresholds with a robust sensitivity analysis.
Resource Audit: Distinguished between deployable parameter compactness and computational efficiency, explicitly reporting inference latency and circuit evaluation costs.

Results

Compact Hybrid Advantage: The 1,393-parameter VQC-Cap model achieved the lowest macro-regret (0.581) among all tested configurations, modestly outperforming the size-matched 1,393-parameter MLP-Cap (0.596). The paired difference was statistically significant at the boundary ( $p=0.050$ after Holm adjustment).
Non-Monotonic Scaling: The performance advantage was not monotonic. The 1.4k VQC-Cap was the best point, but the 0.5k and 4k VQC-Cap configurations performed significantly worse than their matched MLP counterparts. The 12k MLP-Cap also underperformed smaller MLP variants, indicating a non-monotone frontier for both families under the fixed training dose.
Conditional Advantage: The hybrid advantage was found to be dependent on the specific "demand-cap" action head architecture. Preliminary ablation suggests that removing this head diminishes the hybrid's relative performance, indicating the benefit arises from the combination of the skip-connected VQC and the specific control head rather than the quantum circuit alone.
Computational Cost: Despite having fewer parameters, the hybrid models were computationally slower. The 1.4k VQC-Cap required approximately 42 hours of replay time per seed on a 32-core CPU, compared to ~1.2 hours for the matched MLP-Cap.
Tail Risk: All retained models satisfied the high-weight tail-failure rate constraint (saturated at 0.20), though the dense preference-grid analysis showed that regret increases sharply for all learned policies at high degradation-avoidance weights ( $w=0.9$ ).

Significance and Claims
The paper claims a narrow, bounded parameter-efficiency benefit rather than a general quantum advantage. Specifically:

The tested skip-connected demand-cap hybrid VQC shifts the compact deployable-parameter frontier for this specific benchmark, achieving comparable performance to a classical network with fewer parameters at the 1.4k budget.
This result does not imply that hybrid quantum-classical policies are generally superior, faster, or computationally more efficient; in fact, the inference latency is significantly higher.
The advantage is conditional on the specific task (demand-charge-aware control), the specific architecture (skip-connected VQC with a demand-cap head), and the simulation backend (noise-free).
The study concludes that while a compact hybrid design can be effective, the classical network remains a strong, faster baseline. Any claim of hardware-level quantum advantage would require a separate validation pass accounting for device noise and finite-shot sampling, which was not performed in this study.

More like this