Optimizing multi-energy fow scheduling of hydrogen-inclusive virtual power plants based on deep reinforcement learning under dual-carbon targets

Virtual power plants, as a comprehensive energy network with multi-energy flow interconnection, have become an important player in China's accelerated pursuit of its dual carbon goals. However, it is difficult to coordinate internal resources with low-carbon emission when facing challenges such as tight coupling of multi-energy flows, subjectivity of traditional carbon trading model parameters and difficulty of online optimization with high-dimensional dynamic parameters. To address these issues, this paper proposes a virtual power plant multi-energy low-carbon dispatching method that integrates the attention mechanism (AM) and soft actor-critic (SAC) algorithm. Firstly, based on the random carbon flow characteristics of virtual power plants, an improved stepped carbon trading mechanism based on Bayesian optimization is established for dynamic parameters. Next, an economic benefit and carbon emission-based objective function is constructed for the decoupling model of multi-energy flows in virtual power plants. Considering the high-dimensional nonlinearity and real-time updating of weight parameters in this model, the improved SAC deep reinforcement learning algorithm with integrated attention mechanism is used to solve it in a continuous action space. Finally, simulation analysis and comparative experiments are conducted to verify the feasibility of the proposed method and its efficiency compared with SAC algorithm.