Meta-Optimised TD3 Control for Photovoltaic qZSI Shunt Active Power Filters

Abstract

This paper presents a meta-optimized reinforcement learning control framework for a photovoltaic-fed quasi-Z source inverter shunt active power filter (qZSI-ShAPF) applied to grid-connected power-quality (PQ) enhancement. The proposed method combines a twin delayed deep deterministic policy gradient (TD3) controller with particle swarm optimization (PSO) of the actor-network width and a standards-orientated reward function. The control objective is to stabilize the DC-link voltage while reducing current harmonic distortion, correcting the power factor and suppressing current unbalance at the point of common coupling. A continuous-input-current voltage-source quasi-Z-source inverter is selected because it provides single-stage buck-boost conversion and smoother photovoltaic input current than the conventional Z-source inverter. The reward function normalizes DC-link, harmonic, power-factor and unbalance errors, increases the penalty severity under large constraint violations, penalizes excessive control effort and rewards sustained operation inside the acceptable performance region. Simulation results show that the particle-swarm-selected actor architecture of 42 neurons and 45 neurons reaches the stopping criterion in 38 episodes with an average reward of -21273, whereas fixed actor networks require more episodes or fail to stabilize within 500 episodes. Across five independent training seeds, the proposed adaptive reward reduces the mean source-current total harmonic distortion from 4.30 % to 2.04 % and limits the worst-case value to 2.4 %, compared with 6.1 % for the baseline quadratic reward. The results indicate that reward shaping and neural-architecture meta-optimization are not secondary tuning choices but essential design elements for reproducible reinforcement-learning-based active filtering.