크리깅 보간법을 활용한 풍력 발전량 예측에 관한 연구

Abstract

When climate change accelerates due to continuous carbon emissions, the Earth's surface temperature rises 1.5°C compared to pre-industrial times, and the frequency of abnormal weather will be doubled. When 2°C rises, 20% to 30% of the world will be deserted and 54% of species will disappear. Accordingly, it is inevitable to increase the power generation ratio from renewable energy to decline carbon emission. In the matter of renewable power, however, the importance of predicting power generation is increasing because of the nature of renewable power generation that depends on uncontrollable weather environments.

In the case of wind power generation by new and renewable energy, it is the most important to predict wind speed and direction depending on topographical conditions. However, it is difficult to secure accurate weather information in an area where meteorological tower is not installed and operated that require expensive cost.

In this study, a wind power prediction model using machine learning techniques based on weather values of AWS and ASOS, which are ground weather observation equipment near the wind power generator, is developed and evaluated.

In order to select weather data for wind power prediction model, the Pearson correlation coefficient, between the wind power generation amount and the wind speed values by the observed region, was calculated. As a result of the calculation, the values were 0.67 for Yeonggwanggun ASOS, 0.72 for Yeomsan AWS, and 0.8 for prediction by Kriging interpolation. The result shows that the wind speed value predicted by Kriging interpolation was most related to the power generation. In order to improve the accuracy of the model’s learning, the meteorological values predicted by the kriging interpolation method were pre-processed. By using the values, four decision tree and ensemble-based learning models (decision tree, random forest, XGBoost, and LGBM) were learned and tested. As a result, the LGBM algorithm showed the highest correlation(R2_score) with the actual value and the lowest root mean square error(RMSE) compared to other algorithms. Additionally, hyperparameter optimization was performed to improve the predicted value of the model, allowing error value to reduce by 6.1%.

Finally, based on the measurement of Yeomsan AWS, Yeonggwanggun ASOS, and prediction value using kriging interpolation, the LGBM-based power generation prediction model was learned and tested, resulting in that the model using kriging had higher performance.

Based on the experience of this study, development and utilization or a wind power prediction model are intended, that can be used by collecting and utilizing local forecast model (LDAPS) information along with actual wind power generation, weather, and operation information.