Learning-based energy management in multi-carrier energy systems: architectures, algorithms, and open issues

Multi-energy systems interconnect electricity, heating and cooling, gas, and emerging chemical carriers through tightly constrained conversion and storage pathways, which complicates scalable and reliable energy management under renewable variability, nonlinear device behavior, and multi-actor operation. This paper examines learning-enabled energy management in multi-carrier energy hubs and their interconnections, with emphasis on how learning is functionally integrated into hub architectures and operational decision processes rather than on algorithmic taxonomy alone. Existing studies show that learning-based components are increasingly used to augment predictive control through data-driven forecasting and surrogate models, to enable adaptive operational policies via reinforcement learning, and to support decentralized coordination among coupled hubs. At the same time, the literature reveals a growing but uneven treatment of uncertainty, risk, feasibility enforcement, and resilience to operational and cyber-related disruptions. While reported results consistently indicate improved adaptability to time-varying demand and partial observability compared to purely deterministic or scenario-driven optimization, performance is often contingent on data coverage, constraint-handling strategies, and robustness across operating regimes. By synthesizing comparative evaluation practices across heterogeneous studies, this paper exposes recurring trade-offs between operational flexibility and feasibility and identifies research gaps related to benchmarking, communication-aware coordination, and scalable hybrid learning optimization frameworks for next-generation multi-energy systems.