A bi-level optimization strategy for flexible and economic operation of the CHP units based on reinforcement learning and multi-objective MPC

Enhancing the comprehensive performance of the combined heat and power (CHP) units is crucial for accommodating renewable energy and achieving energy conservation. To this end, a bi-level optimization strategy based on reinforcement learning (RL) and multi-objective model predictive control (MOMPC) is proposed to enhance the CHP units flexibility and economic performance. Firstly, a CHP unit model is constructed, and its various parameters are incorporated into the rolling optimization of the MOMPC, serving as the lower-level follower to solve the fundamental control. Secondly, a bi-level optimization strategy integrating the twin delayed deep deterministic policy gradient (TD3) algorithm with MOMPC (TD3-MOMPC) is proposed. The TD3 agent is designated as the upper-level leader. By decomposing the complex flexibility requirements and the optimization control sequence of the CHP unit, tasks are assigned to both the upper-level leader and the lower-level follower for bi-level interactive optimization. Thirdly, with power flexibility, heating quality, and operational economy serving as leader guidance, a multi-criterion optimization reward function is designed for the upper-level. Then, the actions of the upper-level TD3 agent are designed as dynamic weights and time-varying prediction horizons for the rolling optimization of MOMPC, serving as a bridge to connect and guide the bi-level optimization. Finally, to verify the effectiveness of the bi-level optimization strategy, extensive tests on load variation and disturbance rejection were conducted on a 300 MW CHP unit. The results show that the proposed strategy enhances the unit's load flexibility, heating quality, and operational economy.