Real-Time Implementation of Deep Reinforcement Learning-Driven Active Power Filters for Nonlinear Load Compensation in Hybrid Wind-PV Microgrids

The increasing presence of nonlinear loads and the fluctuation of the renewable energy sources in hybrid wind-photovoltaic (PV) microgrids have increased the demand for advanced power quality (PQ) control solutions.In this paper, a Deep Reinforcement Learning (DRL)-based real-time control method for the Shunt Active Power Filters (SAPFs) to reduce current harmonics and its free-flow bidirectional reactive power distortion in the dynamically changing microgrid environment is proposed.Based on the Deep Deterministic Policy Gradient (DDPG) algorithm, a DRL agent is trained to perform real-time interactions with the grid environment for continuously optimized SAPF control signal in order to achieve the least harmonic distortion and enhanced power factor.The hybrid microgrid consists of wind and PV sources with time-varying profiles, nonlinear loads, and a Voltage Source Converter (VSC)-based SAPF.The DRL control was trained in a high-fidelity MATLAB/Simulink simulator and verified for real-time operation in a TMS320F28379D Digital Signal Processor (DSP).Comparative simulation with conventional PI + SRF controllers shows the enhanced converter performance, with the THD always less than 2.5% for all the operating conditions.In addition, the DRL controller provided a fast transient response to disturbances, dynamically operated under various load profiles, and maintained the stability of the system with respect to model uncertainties.The real-time experimental results verify the feasibility of DRL-based power filters in guaranteeing the IEEE 519 standard in the emerging microgrids.This paper sets the stage for the implementation of intelligent, self-learning control in distributed renewable energy, contributing to the resilience, adaptability and quality delivery of power of the smart grid of the next generation.