LSTM 기반의 단계적 분해 방법을 이용한단기 전력 사용량 예측
Prediction of Short-term Power Consumption Using LSTM-Based Two-Step Disaggregation Methods
- 주제(키워드) Power Consumption , LSTM , Disaggregation Methods , Clustering Analysis , Power Consumption , LSTM , Disaggregation Methods , Clustering Analysis
- 주제(기타) 사회과학
- 주제(기타) 경영학
- 주제(기타) 경영학일반
- 주제(기타) 경영과학
- 설명문(URI) https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART002576691
- 등재 KCI등재
- 발행기관 한국경영과학회
- 발행년도 2020
- 총서유형 Journal
- URI http://www.dcollection.net/handler/ewha/000000169680
- 본문언어 한국어
초록/요약
The purpose of this paper is to improve the prediction accuracy of short-term power consumptions in day-ahead market. With aims to achieve a lower prediction error, this paper proposed two-step hierarchical methods that are based on the idea of risk pooling strategy. In the first step, LSTM (Long Short-Term Memory) with several inputs such as the demand history and temperature is used to predict daily power demand. We disaggregate the daily predictions into hourly demands in the next phase. The disaggregation methods determine the hourly power consumption rate within the day by conducting clustering analysis. We developed three disaggregation methods by employing three types of clustering approaches. With the power consumption data of an office building in Washington, U.S. from 2009 to 2011. a series of numerical analysis was performed to evaluate how much the proposed methods reduce the prediction errors. The numerical analysis revealed that the proposed two-step methods outperform a LSTM model that predicts hourly power consumptions directly from inputs. The prediction accuracy is highly related to the variation in power consumptions so that the performance improvements achieved by the proposed methods are sensitive to season, month of a year and time of a day.
more