Meta‐Optimized Twin Delayed DDPG Reinforcement Learning Agent for Multiobjective Control of Shunt Active Power Filters

This letter proposes a hybrid framework that synergistically combines reinforcement learning with particle swarm optimization to automate the architectural design and control optimization of shunt active power filters (SAPFs). By leveraging PSO to autonomously design the neural network architecture of a twin delayed deep deterministic policy gradient (TD3) RL agent, specifically meta‐optimizing the neuron count in its hidden layers of the actor network, this approach eliminates manual tuning and enhances adaptability and efficiency. A multiobjective composite reward function is designed to simultaneously minimize DC voltage error, reduce total harmonic distortion, penalize excessive control actions and enhance overall system stability, thereby addressing key performance challenges in three‐phase SAPF operation. The simulation results obtained by implementing the proposed approach demonstrated rapid convergence and superior performance, confirming its practical effectiveness in satisfying the dynamic operational requirements of modern power systems.