Deep Reinforcement Learning-Based Time-Varying Parameter Identification Method for Grid-Connected Renewable Power Generation Systems

With the large-scale integration of renewables into the power grid, a number of parameters in power systems exhibit time-varying characteristics, posing new challenges for parameter identification. To address these challenges, this paper proposes a novel identification method for time-varying parameters based on deep reinforcement learning (DRL). First, a DRL environment based on a hierarchical soft-guided reward function is designed, which lays a foundation for parameter identification tasks. Second, to tackle the challenges of multi-parameter coupling, a new DRL algorithm called MMoE-TD3 is proposed. This algorithm applies the concept of the multi-gate mixture-of-experts (MMoE) network to the actor network of the twin delayed deep deterministic policy gradient (TD3) algorithm, aiming to improve its capability in parallel identification of multiple parameters. Last, the identification of varying parameters is formulated as a continual reinforcement learning task, and a training strategy based on knowledge acquisition and retention is proposed. The test results validate the convergence, accuracy, and robustness of the proposed parameter identification method.