【目的】 以空调负荷为主体的柔性负荷虚拟电厂（virtual power plant， VPP）在实际运行中易受控制时延、模型与量测误差等不确定性的影响，使削峰需求响应（demand response， DR）的效果偏离预期。导致该现象的重要原因为现有的DR策略往往依据静态的目标负荷曲线进行负荷跟踪，难以适应时变的运行环境。【方法】 以园区大规模分体式变频空调为例，提出柔性负荷VPP参与削峰DR的自适应控制方法，在需求响应邀约允许范围内依据当前的运行环境来调整后续DR时段的目标负荷曲线，有效提升了削峰DR的经济性和鲁棒性。在所提闭环控制模型中，被控过程被解耦为小规模且线性的进度偏差模型和削峰电量修正模型，二者分别置于控制器和反馈环节。其中，进度偏差模型将计划削峰电量分摊至空调，其解满足功率约束和用户舒适度约束；削峰电量修正模型结合削峰DR的响应实况来自适应调整后续控制时刻的目标负荷曲线，用于抵消不确定性对控制效果的负面影响。【结果】 算例以4类变频空调集群为研究对象，在市场模式和邀约模式下，探讨了不同削峰策略、模型与量测误差和控制时延对VPP参与削峰DR的影响，侧重验证所提方法的经济性和鲁棒性。【结论】 结果表明，所提基于动态目标负荷曲线的削峰DR自适应控制方法，能够根据实际响应情况自适应地调整目标负荷曲线，在控制精度、经济收益和鲁棒性方面表现优越。

[Objective] Virtual power plants （VPPs） centered on air-conditioning loads are susceptible to uncertainties， such as control delays and discrepancies between models and measurements， leading to deviations in the efficacy of demand response （DR） strategies from anticipated outcomes. A key contributor to this phenomenon is the reliance of existing DR strategies on static target load profiles， hindering their adaptability to dynamic operational environments.[Methods] To address this issue， this study introduced an adaptive control methodology for flexible-load VPPs participating in peak-shaving DR， utilizing a large-scale split-type inverter air conditioner on campuses as a case study. This approach allowed the adjustment of target load profiles for subsequent DR periods within the permissible range of the DR invitation based on the current operational environment， thereby enhancing the economic and robust nature of peak-shaving DR. In the proposed closed-loop control model， the controlled process was decoupled into a small-scale linear progress deviation model and a peak-shaving electricity correction model， each placed within the controller and feedback loop. The progress deviation model allocated planned peak shaving electricity to air conditioners， ensuring compliance with power constraints and user comfort levels. The peak-shaving electricity correction model， with the actual response to the peak-shaving DR， adaptively adjusted the target load profile for subsequent control moments to mitigate the adverse effects of uncertainties on control effectiveness.[Results] The case study focused on four types of inverter air conditioner clusters and examined the impact of different peak-shaving strategies， models， measurement errors， and control delays on the participation of the VPP in peak-shaving DR under market and invitation modes. This study verified the proposed method’s economic efficiency and robustness.[Conclusions] The results show that the proposed adaptive control method for peak-shaving DR based on a dynamic target load curve can autonomously adjust the target load curve based on the actual response conditions， demonstrating superior performance in terms of control accuracy， economic benefits， and robustness.