A Study on Real-Time OLTC Control in PV-Integrated Distribution Systems Using Machine Learning

Abstract

PV(Photovoltaic) systems increase voltage fluctuation in distribution networks due to their intermittent characteristics. Additionally, since PV generation occurs only during the day when solar irradiance is present and ceases at night, ESS (Energy Storage Systems) are commonly integrated to enhance profitability. However, the integration of ESS can further increase voltage fluctuation in distribution networks, making voltage profiles more challenging to predict. As a result, extreme voltage fluctuations in distribution networks lead to frequent OLTC (On-Load Tap Changer) adjustments to meet voltage regulation standards, complicating control operations. This increases transformer maintenance costs and may cause issues such as reduced power quality. These challenges can ultimately result in productivity losses in both residential and industrial environments. To address these issues, this paper proposes a machine learning-based OLTC tap adjustment method to improve voltage stability in distribution networks with distributed energy resources such as PV and ESS. The proposed method estimates the voltage distribution of feeder buses, which cannot be directly measured from the substation, and uses these estimated voltages to control the OLTC. By doing so, it aims to enhance voltage stability in distribution networks and improve the quality of power delivered to consumers. The test system used in this study is the IEEE 13-bus network, assuming three-phase balance, with integrated PV and ESS operations. PV generation and demand data are based on actual measurements, and the ESS schedule is optimized using the GUROBI Optimizer. Training data for the machine learning model, designed to optimize OLTC tap adjustments, are generated using OpenDSS. The trained model predicts the voltage profiles at each bus in the network and suggests the optimal OLTC tap position. The applicability of the trained machine learning model is validated through real-time simulations using RTDS, demonstrating its potential for implementation in real systems. © 2025 IEEE.