Noise-Resilient Hybrid Quantum-Classical Learning for Smart Grid Stability: A Systematic Study of Architecture, Embedding, and Robustness
SN Computer Science, cilt.7, sa.7, 2026 (Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 7 Sayı: 7
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s42979-026-05384-9
- Dergi Adı: SN Computer Science
- Derginin Tarandığı İndeksler: Scopus, Aerospace Database, Technology Collection (ProQuest)
- Anahtar Kelimeler: Data embedding, Hybrid neural networks, Noise resilience, Quantum machine learning, Smart grid stability, Variational quantum circuits
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Hybrid quantum-classical neural networks are an increasingly popular design pattern for near-term quantum machine learning, yet their behavior under systematic variation of architecture, data encoding, and noise remains underexplored in realistic application settings. This paper presents a comprehensive empirical study of a hybrid encoder-VQC-decoder architecture for smart grid stability assessment, extending our preliminary work with a fresh implementation and 1430 experimental runs on the full UCI Electrical Grid Stability dataset. Across 30 random seeds the hybrid model reaches 95.96% accuracy, statistically equivalent to a classical network in which the quantum layer is replaced by a bottleneck of the same width, and both significantly outperform SVM, XGBoost, random forest and gradient boosting baselines. A representational analysis of the trained networks attributes the equivalence to the classical encoder, whose output already supports a linear classifier as accurate as the full model. An ablation over qubit count, circuit depth and embedding, replicated with width-matched classical controls, identifies the bottleneck width as the factor governing trainability and yields a criterion for selecting the embedding dimension from a classical pilot experiment. The architecture tolerates depolarizing and amplitude damping channel noise up to strength 0.05 without measurable degradation, including when trained without noise and evaluated under it. Under input-level perturbations the hybrid and classical models degrade identically, indicating that input robustness is governed by the shared classical encoder. These results identify the conditions under which a compact variational circuit can be integrated into a classical pipeline and characterize the noise resilience of the resulting model.