Multi-modal multi-task artificial intelligence model for active distribution network scheduling with multi-agent reinforcement learning

This work proposes a multi-modal, multi-task artificial intelligence (AI) framework for intelligent real-time scheduling of active distribution networks (ADNs). The proposed approach leverages multi-agent reinforcement learning (MARL) for the coordinated operation of multiple controllable resources, including energy storage systems (ESS), flexible loads (FL), photovoltaic (PV) units, and static var compensation (SVC) devices. The framework is developed under the practical constraint that only two types of real-time data are available: electrical grid state measurements and solar panel image information. A hybrid convolutional neural network-vision transformer (CNN-ViT) model is employed to extract soiling loss features from solar panel images, providing crucial environmental insights that affect PV output. Subsequently, the visual features are integrated with electrical network variables to form a comprehensive multimodal state representation. Consequently, a decentralized MARL architecture is formulated where each agent — controlling an ESS, FL, PV unit, and SVC device — learns optimal policies to minimize operational costs, network losses, and carbon emissions. The proposed framework is validated on the large-scale 2289-bus Nizwa distribution network in Oman. Results demonstrate that the method achieves superior cooperative scheduling performance, effectively balances multiple operational objectives, enhances grid flexibility, and significantly outperforms traditional scheduling strategies.