A deep reinforcement learning method based on Mamba model with adaptive cross-attention for multi-energy microgrid energy management

The multi-energy microgrid (MEMG) can achieve dynamic mutual support of electricity, heat, and gas through energy cascade utilization, thereby promoting the economic operation benefits of the MEMG. However, the complex multi-energy coupling operation process leads to the problem that the optimal operation of MEMG faces high-dimensional decision variables and difficulty in obtaining global optimal solutions. Therefore, this paper proposes a deep reinforcement learning (DRL) method based on the Mamba model to solve the above problems. Among these components, the Mamba model, by virtue of its sequence processing architecture with linear time complexity and its ability to efficiently capture long-range dependencies, can identify effective variables in high-dimensional decision spaces. It lays a solid foundation for accurate state representation in the subsequent DRL method. In addition, the method combines actor-critic networks and weighted double Q-learning to achieve optimal management of MEMG. Subsequently, an adaptive cross-attention mechanism is proposed to accurately identify key state variables affecting MEMG optimization by dynamically adjusting the weights of different state variables, thereby improving operational efficiency. Simulation results indicate that compared to common methods, the proposed method can enhance economic efficiency of MEMG and absorption of renewable energy.