Stability-Oriented Multiobjective Control Design for Power Converters Assisted by Deep Reinforcement Learning

Impedance characteristics of power converters are dependent on operating conditions, posing challenges to the stability-oriented design of control systems. This is because constant control parameters, designed according to a limited number of operating conditions, may cause instability in other conditions. In this letter, a deep reinforcement learning-assisted framework is proposed to achieve multiobjective optimization of multiple control parameters. With a focus on converter stability under weak/strong grids, adaptive control parameters are generated for different power setting points, in alignment with requirements on dynamic performance. The effectiveness of the proposed framework is validated with the deployment and real-time operation of the well-trained actor (a shallowneutral network) in a control hardware-in-the-loop converter system. With adaptive control gains, system stability can be guaranteed without compromising dynamic response, despite the variation of internal power setting point or external grid strength.

Automated summary

This letter proposes a deep reinforcement learning-assisted framework for multiobjective optimization of control parameters and validates its effectiveness in a control hardware-in-the-loop converter system.