IEEE Transactions on Smart Grid
Quantum Privacy-Preserving Multi-Agent Deep Reinforcement Learning for Local Volt/Var Control
Qiong Liu, Changhong Zhao, Yuji Cao, Tongxin Li, Yue Chen
Abstract
High photovoltaic penetration poses challenges to the voltage regulation of active distribution networks. While multi-agent deep reinforcement learning (MADRL) provides a data-driven framework for local Volt/Var control (VVC), classical approaches face limitations in coordination under extreme partial observability, parameter efficiency, and privacy. To address these gaps, this paper proposes quantum MADRL with public information (QMADRL-PI), which formulates VVC as a public-information-enhanced Markov game in which shared exogenous signals enable communication-free coordination. A hybrid quantum-classical policy with data re-uploading variational quantum circuits (VQCs) is developed for edge deployment, repeatedly encoding classical features at each circuit layer to enhance approximation capability. A quantum differential privacy (DP) mechanism repurposes intrinsic VQC shot noise, reducing external noise injection. Simulations on IEEE 33-bus and 141-bus systems show that QMADRL-PI attains VVC performance comparable to traditional MADRL while using over 1000-fold fewer parameters per agent. Under the same output-level DP mechanism, the compact quantum actor achieves a better privacy-utility trade-off than an over-parameterized classical actor. Experiments on a real 180-qubit superconducting quantum processor further validate the trained policy’s hardware feasibility and control performance.