Sliding window-based LightGBM model for electric load forecasting using anomaly repair

Sungwoo Park, Seungmin Jung, Seungwon Jung, Seungmin Rho, Eenjun Hwang

Research output: Contribution to journalArticlepeer-review

1 Citation (Scopus)

Abstract

Smart grids have attracted much attention recently for their potential to reduce power system operating and management costs. Smart grid core components include energy storage, renewable energy source(s), and smart meters. Smart meters collect diverse data regarding smart grid operation, which can lead to inefficient operation if the meter data are damaged or tampered with during collection or transmission. Therefore, it is important to identify abnormalities in smart grid data and process them accordingly. Various anomaly detection models have been proposed using statistical methods, but they cannot detect some anomaly patterns accurately, and the models generally did not consider repair strategies for the detected anomalies. Anomaly repair should be included with model training to improve forecasting performance. This paper proposes a robust sliding window-based LightGBM model for short-term load forecasting using anomaly detection and repair. We first show how to detect anomalies using a variational autoencoder and then how they can be repaired using a random forest method. Finally, we verify that the proposed sliding window-based LightGBM achieves superior forecasting performance in combination with anomaly detection and repair.

Original languageEnglish
JournalJournal of Supercomputing
DOIs
Publication statusAccepted/In press - 2021

Keywords

  • Anomaly detection
  • Data repair
  • Electric load forecasting
  • LightGBM
  • Random forest
  • Sliding window
  • Variational autoencoder

ASJC Scopus subject areas

  • Software
  • Theoretical Computer Science
  • Information Systems
  • Hardware and Architecture

Fingerprint

Dive into the research topics of 'Sliding window-based LightGBM model for electric load forecasting using anomaly repair'. Together they form a unique fingerprint.

Cite this