配备碳捕集是实现燃煤机组低碳化改造的重要途径。燃煤电站-碳捕集系统存在复杂的电、热、碳耦合关系，需通过优化调度合理分配系统出力，以保障其经济、低碳、灵活运行。然而，电、碳负荷强随机波动、机组宽工况运行的非线性特征，使得常规基于模型的优化调度方法难以取得满意效果。为此，本文提出一种基于不确定负荷指令分解策略的双延迟深度确定性策略梯度（TD3）算法的数据驱动燃煤电站-碳捕集系统智能调度方法。首先构建了计及机组煤耗成本、维护成本和负荷偏差惩罚的优化调度目标，以及包含设备运行、能量平衡和日均碳捕集率在内的调度约束，并将其转化为马尔科夫决策过程。随后，通过智能体与仿真模型的离线交互训练，自适应学习不确定环境下的最优调度策略。在1000MW超临界热电联产机组耦合单乙醇胺吸收碳捕集系统中的仿真结果表明，所提方法相比基于线性模型的确定性数学规划算法和常规不确定TD3算法，可分别降低0.66%和0.52%的运行总成本。本文研究为煤电碳捕集系统的智能低碳运行优化提供了有效方法。

Equipping coal-fired power plants with carbon capture technology is an important means of achieving low-carbon retrofits. Coal-fired power plant integrated with carbon capture system contains complex electrical, thermal and carbon coupling relationships, thus requiring optimal scheduling to reasonably allocate system output to ensure economical, low-carbon and flexible operation. However, the strong random fluctuations in electricity and carbon loads, and the nonlinear characteristics of wide operating conditions for power units, make conventional model-based scheduling methods ineffective. To address this, this paper proposes a data-driven intelligent scheduling method for coal-fired power plants with carbon capture system based on twin delayed deep deterministic policy gradient (TD3) algorithm with an uncertainty load command decomposition strategy. Firstly, an optimized objective is established, considering coal consumption cost of power units, maintenance cost and load deviation penalty, along with scheduling constraints including equipment operation, energy balance and average daily carbon capture rate. Then, convert the aforementioned condition into a Markov decision process. Subsequently, through offline interaction training between the agent and the simulation model, the optimal scheduling strategy under uncertain conditions is adaptively learned. Simulation results from a 1000 MW supercritical combined heat and power unit coupled with a monoethanolamine absorption carbon capture system indicate that the proposed method reduces total operating cost by 0.66% and 0.52%, compared to deterministic mathematical programming algorithm based on linear models and conventional uncertain TD3 algorithm, respectively. This study provides an effective method for intelligent low-carbon operation optimization of coal-fired power plant with carbon capture system.

国家自然科学基金(52376002)；中国能源建设集团江苏省电力设计院有限公司科技项目(32-JK-2024-010)