Learning Volt-VAR Droop Curves to Optimally Coordinate Photovoltaic (PV) Smart Inverters

Abstract

Learning-based solutions for power systems operational tasks are earning more consideration as potential candidates to help overcome challenges brought upon by the aggressive integration of inverter-based resources (IBRs) in active distribution networks (ADNs). Despite achieving high evaluation accuracies, machine learning (ML) methods are not yet accepted at utility-scale primarily due to safety concerns and limited interpretability. This presents an opportunity for ML approaches which can satisfy both performance and regulatory requirements. In an effort to improve these shortcomings, this work proposes a robust Deep Reinforcement Learning (DRL) based model-free adaptive volt-VAR control (VVC) dispatch framework of solar photovoltaic (PV) smart inverters (SIs) for system-wide voltage regulation and loss reduction. The framework utilizes reward shaping with a barrier function (BF) filter to embed physical boundaries for Category B-type SIs specified by the IEEE 1547-2018 standard into the constrained Markov Decision Process (CMDP) formulation. Results carried out on the IEEE 123 bus test system show that the proposed method converges to a robust discrete policy offline, producing QV-droop curves compliant with IEEE 1547-2018, which outperform the baseline benchmark during overloaded conditions.