[M-ZZ40] Renewable energy and earth science

[MZZ40-P03] Day-ahead PV power prediction based on tree-based machine learning methods

*Yuki Kanno¹, Daisuke Nohara¹ (1.Central Research Institute of Electic Power Industry)

Keywords:photovoltaic generation, machine learning

The installed amount of renewable energy, such as photovoltaic generation and wind power, has been increasing. The volatility nature of the renewable energy facilitates the use of prediction of the power generation by renewable energy sources for balancing energy demand and supply. We have developed a photovoltaic prediction based on a numerical weather prediction model, using a power conversion table empirically estimated from the relationship between area-averaged solar radiation and area-integrated photovoltaic generation (Nohara and Kanno, 2021). Here, to improve the accuracy of the prediction, we update the photovoltaic prediction by using tree-based machine learning methods, instead of the empirical power conversion table. As for the machine learning methods, we compared random forest (RF) and gradient boosting decision trees (GBDT). The updated prediction reduces root mean square error of the day-ahead prediction by 8.9% and biases in the morning by 75%. These machine learning methods particularly improve the prediction in cloudy days. RF outperforms GBDT in all lead time. In addition, we attempted to explain the constructed RF machine learning model by Shapley Additive exPlanations. We found that temperature and time of day partly explain the uncertainty in the conversion table.