Machine learning approaches for decentralized volt-VAr control in PV-rich distribution network

Presenter's Name(s)

Seide Saba Rafiei

Abstract

The increasing integration of renewable energy sources introduces chal- lenges, such as voltage fluctuations in power distribution networks. While optimal Volt/VAr control addresses these issues, it typically depends on computationally intensive centralized optimization and makes real-time implementation challenging. This study explores machine learning tech- niques, including linear regression and multi-layer perceptron (MLP), to approximate centralized optimization outcomes and enable efficient de- centralized reactive power control. Using simulated data from a 141-bus network generated via MATPOWER under various operating conditions, preliminary findings show that linear regression provides limited predic- tive accuracy, whereas MLP models significantly enhance predictive per- formance, improving accuracy by 60–92%.

Primary Faculty Mentor Name

Jihong Ma

Status

Undergraduate

Student College

College of Engineering and Mathematical Sciences

Program/Major

Electrical Engineering

Primary Research Category

Engineering and Math Science

Abstract only.

Share

COinS
 

Machine learning approaches for decentralized volt-VAr control in PV-rich distribution network

The increasing integration of renewable energy sources introduces chal- lenges, such as voltage fluctuations in power distribution networks. While optimal Volt/VAr control addresses these issues, it typically depends on computationally intensive centralized optimization and makes real-time implementation challenging. This study explores machine learning tech- niques, including linear regression and multi-layer perceptron (MLP), to approximate centralized optimization outcomes and enable efficient de- centralized reactive power control. Using simulated data from a 141-bus network generated via MATPOWER under various operating conditions, preliminary findings show that linear regression provides limited predic- tive accuracy, whereas MLP models significantly enhance predictive per- formance, improving accuracy by 60–92%.