태양광 발전량 예측 알고리즘 연구

Abstract

본 연구는 딥러닝 모델과 앙상블을 적용한 태양광 출력예측연구로 기존에 수행된 유사과제와의 차별화된 점은 첫째, 피어슨 상관관계 분석을 통해 기상예측 정보와 발전량 분석으로 가장 선형적인 기상예보를 선정하였고, 둘째, XGBoost, LSTM, TCN 등 딥러닝 모델 장점을 조합하여 랜덤포레스트 기반 앙상블 모델 알고리즘을 적용하였으며, 셋째, 친수성과 소수성 모듈코팅을 통해 환경변화에 의한 발전량 변화를 실증하였고, 이를 통해 예측정확도 향상을 모색하였고, 넷째, Markov Chain을 이용하여 통계기반의 기상예보 오차를보정하여 기상정보의 정확도를 향상시키는 방안을 제시하였고, 마지막으로 기상정보와 예측·발전량 패턴 분석을 통해 발전소의 O&M 방안을 도출하였다. 실증 대상 발전소는 한전의 태양광 전문 SPC인 캡코솔라에서 수행하였고, KPX 출력예측 실증사업에 참가하여 적격사업자로 선정되어, 평균정확도는 92∼93% 수준을 보이고 있다. 향후, 신재생에너지 전력시장 제도는 많이 개선되고 있어 본 연구사례가 태양광 발전 사업자의 이익을 최대로 확보할 수 있는 방안이 될 것이라고 사료된다.

I studied the power prediction model of solar power plant using deep learning model and ensemble. Many previous research models for power prediction of photo-voltaic power plants have been presented, but they have been focused on demonstration research in limited areas. However, as the power prediction project of new and renewable energy is in full swing recently, an application model that can be applied to various environments and regions is needed. This study is a "solar power prediction" study applying a deep learning model and ensemble and introduces the differences from the previous renewable energy power prediction study. First, through Pearson correlation analysis, the most linear weather forecast was selected by weather forecast information and power generation analysis. Second, the algorithm was selected as a random forest based ensemble model by combining the advantages of deep learning models such as XGBoost, LSTM and TCN. Third, through module coating of hydrophilicity and hydrophobicity, the change in power generation due to environmental changes was demonstrated and through this predictive accuracy was improved. Fourth, a plan to improve the accuracy of weather information was proposed by correcting statistical-based weather forecast errors using Markov Chain. Finally, the O&M plan of the power plant was derived through weather information and prediction and power generation pattern analysis. The plant to be demonstrated was carried out by KEPCO SOLAR, a solar power SPC of KEPCO, and was selected as a qualified business operator by participating in the KPX power prediction demonstration project, and its average accuracy is 92∼93%. In the future, the renewable energy power market system is improving a lot, so it is believed that this research case will be a way to maximize the profits of solar power generation operators.