Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Distribution System Voltage Prediction from Smart Inverters using Decentralized Regression

Zachary R. Atkins, Christopher J. Vogl, Achintya Madduri, Nan Duan, Agnieszka K. Miedlar, Daniel Merl

TL;DR

Using simulated data and a constellation of 15 inverters in a ring communication topology, the Cola algorithm is shown to accomplish the learning task required for voltage magnitude prediction with far less communication overhead than fully connected P2P learning protocols.

Abstract

As photovoltaic (PV) penetration continues to rise and smart inverter functionality continues to expand, smart inverters and other distributed energy resources (DERs) will play increasingly important roles in distribution system power management and security. In this paper, it is demonstrated that a constellation of smart inverters in a simulated distribution circuit can enable precise voltage predictions using an asynchronous and decentralized prediction algorithm. Using simulated data and a constellation of 15 inverters in a ring communication topology, the COLA algorithm is shown to accomplish the learning task required for voltage magnitude prediction with far less communication overhead than fully connected P2P learning protocols. Additionally, a dynamic stopping criterion is proposed that does not require a regularizer like the original COLA stopping criterion.

Distribution System Voltage Prediction from Smart Inverters using Decentralized Regression

TL;DR

Using simulated data and a constellation of 15 inverters in a ring communication topology, the Cola algorithm is shown to accomplish the learning task required for voltage magnitude prediction with far less communication overhead than fully connected P2P learning protocols.

Abstract

As photovoltaic (PV) penetration continues to rise and smart inverter functionality continues to expand, smart inverters and other distributed energy resources (DERs) will play increasingly important roles in distribution system power management and security. In this paper, it is demonstrated that a constellation of smart inverters in a simulated distribution circuit can enable precise voltage predictions using an asynchronous and decentralized prediction algorithm. Using simulated data and a constellation of 15 inverters in a ring communication topology, the COLA algorithm is shown to accomplish the learning task required for voltage magnitude prediction with far less communication overhead than fully connected P2P learning protocols. Additionally, a dynamic stopping criterion is proposed that does not require a regularizer like the original COLA stopping criterion.

Paper Structure

This paper contains 7 sections, 8 equations, 6 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Decentralized Learning Method
  3. The CoLa Algorithm
  4. Stopping Criterion
  5. Inverter Regression
  6. Results
  7. Conclusion

Figures (6)

  • Figure 1: Network topology of the power distribution network used for the experiments. This distribution network has a peak demand of approximately $25$MW. Solar panels for a uniform $80\%$ grid penetration are indicated as yellow parallelograms with sizes represented relative to their peak capacity. The red arrow indicates the node where the voltage is being predicted by the regression-based algorithm.
  • Figure 2: Linear least-square regression testing statistics, where each day is individually used to train a model that is testing on the remaining $364$ days (nominal voltage is 7200V).
  • Figure 3: Local subproblems and global objective function values for the inverter voltage dataset (left: original data, right: preprocessed data).
  • Figure 4: Regression comparison after $500$ iterations of the CoLa algorithm applied to the elastic net problem with the inverter voltage data.
  • Figure 5: Local updates (dashed) vs loss function (solid) for the inverter voltage dataset (top: original data, bottom: preprocessed data)
  • ...and 1 more figures