바람자원평가 및 경제성 분석을 위한 기계학습 적용 방안

Abstract

In order to apply machine learning method to wind resource assessment and economic viability analysis, an investigation was carried out in terms of wind data quality check, met tower shadow correction, Measure-Correlate-Predict (MCP) method and system marginal price (SMP). K-Nearest Neighbor (KNN), Light Gradient Boosting Machine (LGBM) and Support Vector Machine (SVM) were used as algorithm for supervised learning. The accuracy was evaluated using root mean square error (RMSE) and the coefficient of determination (R2) for numerical data, while using the accuracy ratio of the number of data points classified correctly to the total number of data points for categorical data. As for wind data quality check, machine learning was performed using Sangmyeong onshore met mast data and LGBM was found to be the best among the three algorithms for data quality check. The LGBM was applied to Daejeong offshore met mast wind data for the quality check, which led to the accuracy ratio of 93.51%. The wind flow distortion due to tower shadow effect was detected using statistical method for the offshore wind data. Then the normal data points within an azimuth angle of no-shadow effect were used as train data points, which resulted that the best algorithm was SVM among the three algorithms. The SVM algorithm was applied to the data points distorted by tower shadow at 82.5 m height. When the data points without shadow effect were a reference, the RMSE decreased from 1.009 m/s to 0.2611 m/s and R2 increased from 0.8910 to 0.9926. That is, the tower shadow effect could be corrected using machine learning technique. For long-term wind data correction, MCP method has been utilized, which was adopted in WindPRO software. The reanalysis wind data at four positions around the offshore met mast were collected to use them as reference data for MCP application. The above three algorithms at Machine learning method were applied to predict the long-term wind data from Oct. 1st 2013 to Jan. 31st 2015, after being applied to train using the wind data from Jul. 1st 2015 to Jun. 30th 2016. It was found that the LGBM had the highest accuracy among the three algorithms, while the neural network did the highest accuracy among the three algorithms of WindPRO software. The predictions from the LGBM were compared with those from the other three algorithms of WindPRO that were the regression, the matrix and the neural network. The result was that the LGBM had the RMSE of 0.12 m/s and the R2 of 0.99, while the RMSE and the R2 were 0.20 m/s and 0.98, respectively, at the neural network algorithm. In other words, the machine learning could improve MCP prediction accuracy more than the traditional MCP algorithms at WindPRO. For the purpose of estimating long-term SMP from 2020 to 2030 in the mainland of South Korea, the machine learning technique were also applied in this study. Liquefied Natural Gas (LNG), West Texas Intermediate Crude Oil (WTI) and FOB Kalimantan price data were collected from public data portal sites since those were considered main factors deciding the mainland SMP. As a result of correlation analysis, LNG and WTI were determined as the factors being used for the machine learning. The most accurate algorithm was LGBM, which was employed for forecasting the long-term SMPs. Using Japanese LNG price estimates and world crude oil price estimates, the long-term SMPs were estimated, which had a decreasing trend from 2020 to 2022 and then maintain 72 KRW/kWh until 2030.