Energy-efficient quantum-spiking multi-agent reinforcement learning for adaptive energy management in microgrid networks

Interconnected micro-grids (MG) energy management requires coordinated decision-making under renewable uncertainty, distributed operation, and network-level trading constraints. Conventional centralized reinforcement learning (RL) approaches often suffer from scalability limitations, large communication overhead, and high computational cost for decentralized edge-level deployment. Existing multi-agent reinforcement learning (MARL) methods also commonly rely on dense artificial neural network (ANN) models and do not explicitly enforce global trading balance during decentralized execution. This paper presents a quantum-spiking-MARL (QS-MARL) framework for adaptive energy management in interconnected MG networks. The proposed method combines decentralized event-driven spiking neural network (SNN) policies for low-power local control with a centralized parameterized quantum circuit (PQC) critic that estimates the global value function under the centralized training and decentralized execution (CTDE) paradigm. Each MG employs a low-power, temporally-aware SNN policy, while a centralized PQC critic captures inter-agent trading dependencies during training under a CTDE paradigm. The PQC critic uses a qubit-per-agent representation with shallow nearest-neighbor entanglement, which maintains linear scaling with the number of MGs. Trading balance is enforced through a differentiable penalty mechanism without requiring centralized execution. Experiments on CityLearn urban districts and a synthetic Pecan+NREL mixed-use cluster show that QS-MARL reduces average daily operational cost by approximately 9% compared to ANN-based CTDE baselines and by approximately 14% compared to fully decentralized MARL, while remaining within 6%–7% of a centralized perfect-foresight oracle. QS-MARL also achieves significant emissions reductions (10%–25%), maintains the narrowest cost distribution under renewable and load uncertainty among all baselines, and outperforms multi-agent proximal policy optimization (MAPPO) and Q-value mixing network (QMIX) in both cost efficiency and constraint satisfaction. Ablation and sensitivity analyses confirm that both the SNN policies and the PQC critic are important to the observed performance gains. These results demonstrate that QS-MARL provides a computationally efficient and coordinated decentralized control method for distributed MG energy management.