摘要
Abstract
In the context of the transformation of new power systems driven by the"double carbon"goal,the operation optimization of hydropower station units faces core challenges such as strong non-linearity,time-varying parameters and multiple constraints.To solve these problems,this paper proposes a multi-layer collaborative adaptive control algorithm.The algorithm innovatively introduces a dual-time scale adaptive mechanism.The local adaptive model predictive control(LAMPC)on the fast time scale(10 ms)is responsible for high-frequency power tracking and oscillation suppression,and the global adaptive efficiency optimization on the slow time scale(1 s).(GAEO)is based on a lightweight digital twin for online efficiency optimization.In order to overcome the limitations of traditional digital twins with large computing volume,this paper builds a reduced-order digital twin model coupled with hydraulic-electrical-mechanical three-domain.The state dimension is effectively compressed to 21 dimensions through MOC,Kron reduction and modal synthesis methods.Combined with deep kernel learning and residual Kalman filtering,online adaptation and error correction of model parameters are achieved.In addition,in order to ensure the safety of the system under model uncertainty or extreme disturbances,this paper designs an adaptive switching strategy under the constraint of Safe Reinforcement Learning(Safe-RL).By embedding the control barrier function(CBF)as a hard constraint into the Actor-Critic framework,the study realizes undisturbed switching between LAMPC and Safe-RL when the model confidence drops.Simulation results show that the proposed adaptive control algorithm is superior to traditional control methods in terms of power tracking performance,error convergence speed,control input smoothness and system frequency domain stability,effectively enhancing the robustness,efficiency,and safety of hydropower units in complex operating environments.关键词
自适应控制/双时间尺度/数字孪生/安全强化学习/模型预测控制Key words
adaptive control/dual time scale/digital twins/safety reinforcement learning/model predictive control分类
建筑与水利
