基于安全强化学习和多能源惯性协调的综合能源系统优化调度

The integrated energy system (IES), which synergistically couples electricity, heat, and natural gas, is pivotal for achieving high-quality energy development and carbon neutrality goals. However, the divergent inertial characteristics and multi-time-scale dynamics of these energy subsystems pose significant challenges to secure and economically optimal operation. The declining inertia in power grids, a consequence of high penetration of power-electronic-interfaced renewables, undermines system frequency stability and anti-disturbance capability. Concurrently, the disparate response times-with electricity operating on a second/minute scale and heat/gas on a minute/hour scale-complicate coordinated scheduling. Existing studies often have limitations: some overlook the full potential of multi-energy coupling equipment in providing grid inertia support, others fail to adequately model the impact of device participation on source-network energy balance, and many cannot effectively resolve the security-economic trade-off in high-dimensional decision-making across multiple time scales. To holistically address these intertwined challenges, this paper proposes a novel multi-time-scale coordinated optimization scheduling method for IES based on event-triggered safe reinforcement learning (SRL), designed to coordinate multi-energy inertia for enhanced security and economy. The proposed methodology is structured around three key innovations. First, a comprehensive "equipment-inertia-energy" model is established to characterize how multi-energy coupling devices participate in grid inertia support. Unlike existing models that simplify devices as fixed inertia sources or focus solely on energy balance, this model meticulously analyzes the participation degree of devices like CHP units, power-togas/hydrogen, fuel cells, and thermal storage, and maps their operational states to equivalent grid inertia constants, considering the energy buffering provided by thermal and gas networks. Second, a multi-time-scale energy coordination strategy is developed that leverages the inherent tolerance of thermal and gas systems to short-term power fluctuations. This strategy explicitly treats the power grid's second/minute-level inertia regulation demands as an energy buffer that can be absorbed and managed within the longer minute/hour-level dispatch cycles of the thermal and gas networks, eliminating the need for highly precise short-term predictions. Third, a bi-level SRL optimization framework with an event-triggering mechanism is constructed to solve the model efficiently. The upper-level SRL optimizes short-term grid economic operation and inertia security, while the lower-level SRL minimizes long-term thermal and gas subsystem operational costs. Crucially, the event-triggering mechanism, based on a Lyapunov safety function, activates a computationally intensive safety recovery strategy only when the system state approaches or violates security boundaries; otherwise, a more efficient evolutionary policy is executed, thus balancing computational load and safety assurance. Simulation studies were conducted on a modified IES testbed comprising an IEEE 30-bus power system, a 20-node gas system, and a 14-node thermal system. Five different schemes were compared to validate the proposed method's effectiveness. The results demonstrate that the proposed method (Scheme 5) effectively maintains grid inertia within a high-inertia security zone across various time periods, significantly reducing the probability and associated cost of frequency and rate-of-change-of-frequency (RoCoF) limit violations. Specifically, the frequency/RoCoF violation cost was reduced to nearly zero, and the thermal/gas network violation cost was cut by over 80% compared to the sub-optimal scheme. Furthermore, the multi-time-scale coordination reduced total multi-energy operational costs, with the comprehensive cost in Scheme 5 being at least 12.7% lower than schemes without full coordination. A comparative analysis with other reinforcement learning algorithms, including PPO, SAC, and CPO, showed that the proposed event-triggered bi-level SRL achieved higher and more stable reward values and a significantly lower cost function, indicating superior security constraint handling. Sensitivity analysis on the conversion gain coefficient confirmed its critical role in frequency stability, while computational efficiency tests proved that the event-triggering mechanism reduced the average cycle computation time by a factor of 3.5, making it highly suitable for the online dispatch of complex IES.