Large-scale model driven real-time economic generation control for integrated energy systems

As a result of the growing integration of renewable energy generation units into integrated energy systems (IESs), the coupling configurations of equipment within the IESs are constantly changing, and the fluctuations of renewable energy sources (RESs) are even more drastic. To mitigate frequency deviations and area control errors (ACEs) in IESs, this paper proposes a transformer-soft actor-critic (T-SAC) algorithm, which integrates the efficient feature extraction capability of the large-scale model transformer with the online learning capability of deep reinforcement learning, and enables the mining of rich feature information from frequency deviation and ACE signals to generate accurate control commands. Furthermore, this paper constructs the cyber-physical-social systems-centralized real-time economic intelligent generation control (CPSS-CREIGC) framework built upon the T-SAC algorithm, which employs virtual parallel systems to optimize the parameters of T-SAC and thereby enhances training efficiency. By issuing control commands every 4 s, the CPSS-CREIGC framework effectively mitigating the reverse regulation phenomenon. The T-SAC algorithm is simulated and compared with seven different comparison algorithms in two-area and four-area IESs under high RESs penetration. Compared to the comparison algorithms, the T-SAC algorithm reduces frequency deviations by at least 46.67 %. The numerical results confirm the effectiveness and feasibility of the CPSS-CREIGC framework