Japan Geoscience Union Meeting 2022

Presentation information

[J] Poster

M (Multidisciplinary and Interdisciplinary) » M-ZZ Others

[M-ZZ48] Renewable energy and earth science

Tue. May 31, 2022 11:00 AM - 1:00 PM Online Poster Zoom Room (34) (Ch.34)

convener:Hideaki Ohtake(National Institute of Advanced Industrial Science and Technology), convener:Daisuke Nohara(Central Research Institute of Electric Power Industry), convener:Teruhisa Shimada(Graduate School of Science and Technology, Hirosaki University), convener:Fumichika Uno(Nihon University, College of Humanities and Sciences), Chairperson:Hideaki Ohtake(National Institute of Advanced Industrial Science and Technology)

11:00 AM - 1:00 PM

[MZZ48-P06] A basic study on the prediction for wind power generation using gradient boosting decision trees and its interpretability

*Yuki Kanno1, Daisuke Nohara1, Masamichi Ohba1, Yu Fujimoto2 (1.Central Research Institute of Electic Power Industry, 2.Waseda University)

Keywords:wind power generation, gradient boosting decision trees

The amount of renewable energy sources, such as photovoltaic and wind power generation, has been increasing. Because the power generation by renewable energy sources highly depends on weather patterns, an accurate prediction of wind power generation using a numerical weather prediction is important for balancing the electricity supply and demand. In this study, we performed a prediction of wind power generation using gradient boosting decision trees (GBDT) at four windfarms (WFs) in Japan and discuss the interpretability of the GBDT model using SHapley Additive exPlanations (SHAP). The GBDT model is constructed by learning the relationship between the WF power generation data and the numerical weather prediction data for two hours before and after the target forecast time around the target WF for a year. The prediction using GBDT is compared with a prediction that uses wind speed data from a numerical weather prediction at the nearest grid point of target WF and catalogue power curve as a reference. An evaluation of 8 months prediction showed that our prediction substantially reduces prediction errors compared with reference one by reducing bias errors. We tried to interpret the obtained GBDT model using SHAP. The spatial distribution of the SHAP value, which is a metric for the importance for explanatory variables, around the target WF shows that the value is higher over land than ocean, which indicates that explanatory variables over land are more important that those over ocean.