Advancements in Energy Systems with Deep Learning

Energy Management through Heterogeneous Graph Neural Networks by Salah Ghamizi
The increasing complexity of power grids with new unpredicatble sources of energies (renewables) and storage facilities (household batteries, electric vehicles) is making the optimization and maintenance of the power grids more complex and challenging. On of the challenging problems is the Optimal PowerFlow (OPF) problem. It consists at finding the best way to distribute electricity through power grids, considering factors like demand, supply, storage, transmission limits, and costs, to ensure everything runs smoothly and efficiently without overloading the system. The objective of this optimization is to fullfill the demand while minimizing the cost of generation of various source of energies.

Traditional approaches often struggle with the complexity and diversity of modern grid components, offering solutions that may not adequately reflect the grid's operational constraints, leading to potentially suboptimal outcomes. Addressing these limitations, we introduce OPF-HGNN, a novel architecture leveraging heterogeneous Graph Neural Networks (GNNs) that incorporates grid constraints directly into the model's loss function through differentiable penalty regularization. This innovative approach significantly enhances the model's ability to generalize across varied grid topologies and manage real-world operational settings, achieving performance improvements by two orders of magnitude over existing solutions.

Deep Reinforcement Learning for Optimized Energy System Management by Jun CAO
The second segment explores the transformative potential of Deep Reinforcement Learning (DRL) in managing and optimizing complex energy systems. As renewable energy sources increasingly integrate into the grid, managing their variability and the dynamic demands of modern energy networks requires advanced, adaptive control strategies. DRL offers significant advantages by learning optimal policies for grid management, battery storage optimization, and real-time demand response, adapting to the energy system's complexities. This presentation will detail how DRL is applied to enhance grid stability, increase energy efficiency, and optimize operational costs, supported by real-world case studies that highlight the practical benefits and challenges of implementing DRL solutions in energy systems.