Feasibility-guaranteed unsupervised deep learning for real-time energy management in integrated electricity and gas systems

With the increasing integration of renewable energy sources and the shift toward a low-carbon economy, managing the complexities of integrated electric and gas systems (IEGS) has become more challenging. Traditional methods often struggle in real-time applications due to their computational intensity and inability to handle uncertainties effectively. This paper proposes a feasibility-guaranteed, unsupervised deep learning approach for the real-time operation of IEGS. Leveraging advancements in IEGS modeling and machine learning, the method addresses challenges related to pre-training datasets, policy gradient approximation, iterative energy flow solvers, and real-time feasibility. The approach integrates neural networks with non-learnable, physics-driven layers to predict control actions while reconstructing the remaining decisions based on energy flow equations, ensuring exact gradient computations. To enhance solution feasibility, an augmented Lagrangian method is adopted, circumventing the ill-conditioning of penalty methods. The model incorporates LSTM layers to handle dynamic constraints, and a convex safety layer ensures online feasibility. Simulation results on three IEGS systems validate the approach, demonstrating superior learning performance and computational efficiency compared to state-of-the-art methods, achieving over 97% time savings. Results position the approach as a promising tool for addressing the complexities of modern energy systems and energy market clearing and adapting to system changes.