전력수요예측을 위한 딥러닝 모델의 성능 비교 연구

다국어 초록 (Multilingual Abstract)

In recent years, power demand has soared every year due to continued economic growth and development of new industries, and since domestic energy environment is highly dependent on foreign situations and very influenced by oil price fluctuations, it becomes important to find the ways for stable energy control. Responding to growing power demand, it is necessary to make accurate estimations of power demand for investment in the facilities used in power systems, or for efficient operation of the existing facilities and stable energy supply. Since electric energy is consumed at the same time as it produced, once started the production, it has the property to be difficult in storing and being wasted immediately, so that the overestimated demand of power will increase the production costs, while the underestimated demand will cause the instability in the supply. Therefore, accurate demand forecast of power should be made for optimal operation of power system. However, responding to fast-growing power demand, power grid is gradually expanding on a large scale to be more complex, and the patterns of power consumption are diversifying, making it difficult to secure adequate reserve margin or to forecast power demand.
This paper aims to find the most suitable forecasting method by applying to power demand forecasting and evaluating the performance of machine learning and deep learning techniques, which have recently shown excellent performances in various fields. It implemented a forecasting model through machine learning techniques, such as logistic regressions, decision tree, and support vector machine algorithm as well as deep neural network algorithm as a deep learning technique and compared it. In addition, it applied the results from the learning of machine learning models to deep neural networks for the implementation of complex forecasting models, and as the results from comparison of the performances, a complex forecasting model with decision tree and deep neural networks was found the most outstanding, and also as the results from comparison of the data forecasted by the KSLF (KPX Short-Term Load Forecaster) program, which is currently used by Korea Power Exchange, for 100 days from January 1, 2019 to April 10, 2019, an improvement by 0.87% in the performance was confirmed. Through this study, it is expected that a foundation can be provided for the ways to reduce the waste of electric energy, as well as distribute and utilize the energy efficiently by close connection with the smart grid technology.

목차 (Table of Contents)

• 제1장 서론 1
• 제1절 연구배경 1
• 제2절 연구방법 3
• 제2장 관련 연구 6
• 제1절 전력수요예측 6
• 1. 전력수요예측 방법 6
• 2. 전력수요의 특성 7
• 3. 기존 전력수요예측 9
• 제2절 전력수요예측을 위한 머신러닝 기법 11
• 1. 머신러닝의 개념 11
• 2. 머신러닝 알고리즘 12
• 가. 로지스틱 회귀 12
• 나. 의사결정나무 14
• 다. 서포트 벡터 머신 18
• 제3절 전력수요예측을 위한 딥러닝 기법 21
• 1. 인공신경망 21
• 2. 심층 신경망 23
• 가. 다층 퍼셉트론 23
• 나. 심층 신경망의 구조 24
• 다. 경사 하강법 26
• 라. 활성 함수 29
• 제3장 전력수요예측 기법의 구현 및 실험 33
• 제1절 실험 데이터 33
• 1. 데이터 수집 33
• 2. 데이터 분석 36
• 3. 데이터 전처리 42
• 제2절 단일 예측모델 46
• 1. 로지스틱 회귀 예측모델 46
• 2. 의사결정나무 예측모델 50
• 3. 서포트 벡터 머신 예측모델 54
• 4. 심층 신경망 예측모델 58
• 제3절 복합 예측모델 63
• 1. 복합 예측모델의 연구방법 63
• 2. 로지스틱 회귀와 심층 신경망 복합 예측모델 63
• 3. 의사결정나무와 심층 신경망 복합 예측모델 68
• 4. 서포트 벡터 머신과 심층 신경망 복합 예측모델 72
• 제4장 실험 결과 및 성능 비교 76
• 제1절 실험 결과 76
• 제2절 성능 비교 77
• 제5장 결론 및 향후 과제 81

참고문헌 (Reference)

1 오일석, "“기계학습 MACHINE LEARNING”", 서울: 한빛아카데미, , p.620, 2017

2 이건명, "“인공지능 튜링테스트에서 딥러닝까지”", 서울: 생능출판, , pp.170~171, 2018

3 이시카와 아키히코, "“인공지능을 위한 수학”, 신상재, 이진희", 서울: ㈜프리렉, , pp.284~286, 2018

4 류성호, "실시간 수요예측 개발 및 온라인 수요예측 방안", 전력거래소 , pp. 18~20, 2015

5 정응수, "“실시간 관제운영 분석 시스템 구축방안 연구”", : , , pp.61~62, 2017

6 조종만, "단기 전력수요예측 기법 및 적용방안에 관한 연구", pp.172~177, 2011

7 구본길, "“기온 데이터를 이용한 하계 단기전력수요예측”", The Korean Institute of Electrical Engineers, vol64:8, 2015

8 김광호, 최황규, 장병훈, "“원-핫 인코딩을 이용한 딥러닝 단기 전력수요 예측모 델”", Institute of Korean Electrical and Electronics Engineers, vol23:3, 2019

9 노윤지, "“LSTM 기반의 Top-down 분해방식을 이용한 단기 전력수요예측”", 이화여자대학교,석 사학위청구논문 , pp.1~3, 2020