• A transformer-enhanced reinforcement learning method with diffusion-driven data augmentation is proposed. • A decentralized partially observable Markov game-based formulation is developed for Volt-VAR control problems. • A conditional diffusion model and an adaptive mixing mechanism are designed. • A cloud-edge collaborative framework with centralized-training-decentralized-execution paradigm is implemented. • The proposed approach effectively optimizes system power loss and mitigates voltage violations. Volt-VAR control to address the challenges of limited historical data and high environmental uncertainty in distribution networks (DNs), and to meet the requirement of Volt-VAR control for real-time scheduling, this paper proposes a transformer-enhanced multi-agent reinforcement learning method integrated with diffusion-driven data augmentation. This method integrates a conditional diffusion model to generate synthetic training samples to expand the replay buffer, thereby alleviating the problem of data scarcity. Meanwhile, a multi-agent twin delayed deep deterministic policy gradient (MATD3) architecture based on transformer is adopted, where the self-attention mechanism of the transformer serves as the feature encoder to capture the complex spatiotemporal coordination relationships among distributed photovoltaic (PV) inverters, and its output is sent to the actor-critic network for policy learning. The coordinated real-time Volt-VAR control of PV inverters using only local information is realized via the framework of centralized offline training and decentralized online execution. The proposed strategy exhibits strong adaptability to DNs with constrained communication resources, while achieving computationally efficient control and high operational economy. Case studies on the modified IEEE 33-bus system and 141-bus system demonstrate superior performance of the proposed method.
Quan et al. (Sun,) studied this question.