기상 수치 모델 기반 실시간 풍력 발전량 예측 기술에 대한 최적화 연구 = (A) study on the optimization of real-time wind power prediction using numerical weather model

국문 초록 (Abstract)

본 연구는 풍력 에너지의 효율적 이용을 위한 현업용 실시간 풍력 발전량 예측 프로세스를 설계하고, 각 세부 프로세스의 최적화 연구를 통해 국내 풍력단지 운영에서의 적용 가능성을 도출한 것이다.
실시간 풍력 발전량 예측을 위한 기상 수치모델은 WRF를 기반으로 하며, 단기 예측을 위해 통계 기법인 ARIMA와 신경망 모형을 결합하여 적용하였다.
현재를 기준으로 48시간 이내의 실시간 풍속을 예측하여, 풍력 발전기의 성능곡선과 발전기별 위치에 따른 후류 효과 등을 적용하여 최종적인 풍력 발전량을 산정하였다.
연구 대상 단지는 제주도에 위치한 성산과 한경 풍력 단지이며, 2012년을 대상으로 정확도를 검증한 결과, 예측 선행 24시간 이내에서 오차율이 20% 이내로 나타났다.

다국어 초록 (Multilingual Abstract)

Wind energy technology is one of the most rapidly expanding areas among the renewable resources of energy instead of fossil fuels. Integrating wind power into the electric power system is the key process for making a dispatch plan, which is directly linked to the electricity security and reliability. Accurate prediction of wind power output is crucial to reduce the allocation of reserves in advance, particularly on a time-scale of several hours to days ahead of dispatch. Despite rapid growth of wind energy industry, actual electric power derived from wind generation is still very restricted mainly due to the large uncertainty of wind prediction.
From the meteorological perspective, wind is considered as the most challenging variable to accurately forecast due to its nature characterized by strong spatial and temporal variability. Furthermore, wind information for wind energy application should be provided at turbine height (typically around 80 m), rather than near surface like 10 m, a height at which traditional wind observations are routinely taken. Due to scarcity of measurement data, there have been few studies for validation of simulated wind as well as analysis of long-term climatology of wind at 80 m. Accordingly, a large uncertainty as well as lack of understanding of the wind characteristics make it difficult to derive reliable estimation of the potential wind resources.
The purpose of this study is to develope optimized real-time wind power prediction using a numerical weather model, WRF, for operating wind power in electrical grid, focusing on Jeju island, Korea. The prediction processes of wind power are composed of two sub-modules, in terms of prediction module of meteorological variables (e,g. wind direction and wind speed) and calculation module of wind power output. For the target of short-term period, prediction information is provided with 10 minutes interval within 30 minutes and with 30 minutes interval from 30 min up to 360 min. On the other hand, prediction information for mid-term is provided with 1 hour interval and updated every 1 hour.
For the prediction of wind speed and direction, the combined system of physically-based numerical model and hybrid-type statistical model is developed in this study. As for the numerical model to derive physically-based meteorological variables, a triple-nested system is applied with a focus on the Jeju island with a 1km resolution using WRF, which is the most popular mesoscale numerical weather model. Global Forecast System (GFS) data with 0.5° × 0.5° resolution is used as the initial and lateral boundary condition to drive the WRF.
To enhance the model performance, the optimizations for each module are carried out, WRF triple-nested system, hybrid statistical model, and wind power calculation. First, to take into account for uncertainties due to model physics imperfections and initial conditions, we conduct the sensitivity experiments with different combinations of the PBL (YSU & MYJ) and the LS (Thermal diffusion & Noah) schemes initialized at regularly lagged 6-hour intervals. The combination of the YSU PBL and thermal diffusion LS schemes shows slightly better performance in simulating the mean speed and direction of surface wind against the observations averaged over the four stations in Jeju. The results of this study provide the usefulness of a time-lagged ensemble method, reducing the RMSE compared to individual deterministic forecasts. It is considered that deterministic forecast with particular lead-time initialization may not be reasonable or sufficient to draw the best performance because there is no perfect initial condition and numerical model to represent the "true" state of the atmosphere.
Second, the hybrid statistical model (ARIMA+Neural network) is optimized for short-term prediction. The value of MAPE in ahead 48 hours is under 12% at both wind farms(Seongsan, Hangyeong). Also, the accuracy is improved by adjustment non-linear error using neural network method.
Lastly, the optimization of wind power calculation is carried out using three algorithms. Algorithm 1 is to calculate wind power using predicted wind speed from WRF. Algorithm 2 is to use wind speed applied correlation coefficient between predicted and measured wind speed. Algorithm 3 is to use wind speed applied wake loss ratio of each turbine with wind direction. Based on the error analysis, the NAME(%) of algorithm 3 presents the best performance, showing the lowest error. The range of NMAE(%) in ahead 48 hours is from 17.1% to 19.1% in case of Seongsan wind farm, from 18.0% to 20.4% in case of Hangyeong wind farm.
This study is the first attempt to estimate the capability of the real-time prediction system of potential wind power over Jeju island, based on the state-of-the-art numerical model and hybrid statistical model module for appropriate target lead time. Based on the validation, the real-time wind power prediction system developed in this study shows the reasonable performance in ahead 24, demonstrating the potential utilization for operating wind power in electricity grid. To derive a more robust statement, it is necessary to improve the accuracy of prediction algorithm for shutdown and wake losses.

목차 (Table of Contents)

• 제 1 장 연구개요 1
• 1.1. 연구 배경 및 필요성 1
• 1.1.1. 연구 배경 1
• 1.1.2. 연구 필요성 3
• 1.2 연구 목적 4
• 제 2 장 기술 현황 및 기반 기술 6
• 2.1 기술 개요 및 현황 6
• 2.1.1 풍력 발전량 예측 기술 개요 6
• 2.1.2 기존 연구 및 기술 수준 7
• 2.1.3 풍력 예측 정보의 경제적 효과 13
• 2.2 기반 기술 15
• 2.2.1 기상 수치 모델 개요 15
• 2.2.2 풍속 예측 기술 16
• 2.2.3 풍력 발전량 산정 알고리즘 19
• 2.3 풍력 에너지 예측 정보의 활용 프로세스 20
• 2.3.1 전력수급 및 예비력 20
• 2.3.2 전력시장 23
• 제 3 장 기초 자료 분석 25
• 3.1 연구 대상 지역 및 발전기 25
• 3.2 기상 현황 분석 27
• 3.2.1 성산 27
• 3.2.2 고산 30
• 3.3 풍력단지 분석 33
• 3.3.1 분석 개요 및 방법 33
• 3.3.2 분석 결과 34
• 제 4 장 실시간 풍력 발전량 예측 기술 개발 38
• 4.1 실시간 발전량 예측 기술 개요 38
• 4.1.1 프로세스 설계도 38
• 4.1.2 발전량 예측 정보 생성 시간 40
• 4.1.3 실시간 프로세스 42
• 4.2 기상 수치 모델 구축 44
• 4.2.1 상세 지형자료 구축 및 예측 영역 44
• 4.2.2 초기 경계조건 45
• 4.2.3 단지 내 실측 자료 동화 46
• 4.3 기상 수치 모델 최적화 47
• 4.3.1 최적화 개요 47
• 4.3.2 분석 방법 설계 48
• 4.3.3 분석 결과 51
• 4.4 단기 예측 모델 개발 63
• 4.4.1 단기 예측 모델 개요 63
• 4.4.2 ARIMA 모형 66
• 4.4.3 ARIMA 모형 검증 71
• 4.4.4 신경망 모형 86
• 4.4.5 신경망 모형 검증 96
• 4.5 발전기별 풍속 산정 알고리즘 107
• 4.5.1 알고리즘 개요 107
• 4.5.2 상관식을 이용한 후류 손실 109
• 4.5.3 풍향별 후류 손실 113
• 제 5 장 예측 정확도 평가 117
• 5.1 평가 방법 및 기준 117
• 5.2 풍속 예측 결과 120
• 5.2.1 성산 풍력단지 120
• 5.2.2 한경 풍력단지 126
• 5.3 풍력 발전량 예측 결과 132
• 5.3.1 성산 풍력단지 132
• 5.3.2 한경 풍력단지 142
• 제 6 장 요약 및 결론 152
• 참 고 문 헌 155
• Abstract 162