Data-based frequency control of power systems

Abstract

Driven by climate change and global warming, modern power systems are undergoing an essential transition, the heart of which is the replacement of fossil fuel generations with renewables. These will cause significant changes to power system dynamics and introduce new frequency stability and control issues, which challenge the conventional frequency control paradigm. For one thing, the high variability over time and limited predictability of renewables introduce uncertainties into generations, bringing new issues in maintaining the instantaneous power balance. For another thing, most renewable generators are converter-interfaced, providing no inertia to the system. The loss of synchronous generators will reduce system inertia and make it more variable, which makes the system more sensitive to contingencies and hard to control. This thesis focuses on the frequency regulation problem from a data-based control perspective with the goal to allow for the complexities arising in future power systems including those arising from increasing renewables. First, to explore more effective control strategies under increasing uncertainties, the frequency regulation problem is investigated in a centralized data-based control framework. Based on behavioral system theory, a centralized data-enabled predictive control (CDeePC) method is developed. The CDeePC utilizes a finite set of persistently excited historical data to describe the system without offline training. An optimization problem is then formulated using trajectory data to calculate the control signals. To avoid the burden of online computation in the CDeePC, a reinforcement learning (RL) algorithm with CDeePC-guided policy search is further proposed. In the offline process, the CDeePC is used to guide the training of neural networks in RL. Then, well-trained RL agents conduct a real-time mapping between observations and control signals in the online execution. Second, in view of the heavy communication and scalability issues that may be caused by centralized control, a new distributed attention-enabled proximal policy optimization (DAPPO) algorithm is proposed to learn optimal frequency control strategies in a distributed way. To achieve this goal, two distributed information sharing mechanisms are designed in the DAPPO. In addition, the importance of observations from different neighbors is adjusted in an adaptive way to help agents focus on the essential information. Third, the time-varying system inertia caused by the high penetration of renewables is considered. By combining behavioral system theory, distributed control, and switching control, a novel distributed switching data-enabled predictive control (DSDeePC) approach for frequency regulation with time-varying inertia is proposed. A supervisory framework is developed to organize the switch between alternative model-free controllers when the power system evolves from one mode (e.g., a certain inertia) to another. In addition, a distributed communication scheme is introduced to implement the DSDeePC method in a distributed way. To avoid frequent switching, a data-driven adaptive predictive control (DAPC) scheme is proposed. In the DAPC, a moving horizon estimation based algorithm is developed to update the data-based system representation adaptively to allow for time-varying inertia. Based on the estimated data-based system representation, the optimal control signals are calculated within a data-driven predictive control framework.