Key points are not available for this paper at this time.
Highly accurate solar power forecasting using machine learning has been developed for efficient power grid operations. However, from the viewpoint of power transmission system operation, it is important to reduce not only the average accuracy of forecasts, but also the risk of serious errors (especially overestimation). Despite its importance, this risk prevention has not been focused on in the forecast model design. We aim to address this subject by developing low risk models with good average accuracy through the following approach: adoption of a machine learning ensemble, and optimization of the metamodel using quadratic programming with downsampled validation data. In the Tokyo Electric Power Company area solar irradiation forecast, it was confirmed that the proposed integration method enabled forecasts to maintain or improve the average accuracy and suppress the 99.87th percentile error by more than 17% compared with the benchmark forecast. This result is equivalent to maintaining the average accuracy and reducing the risk of overestimation by more than 4.8% when compared to the metamodel constructed by simple averaging. In addition, the performance of the optimized metamodel is discussed, with a trade-off analysis and visualization of the contribution to serious forecast errors. • Day-ahead solar radiation forecast model proposed. • Model improves average accuracy and reduces risk of serious over-estimation. • Prediction data, ML ensembles, domain optimization, down-sampling used. • Over-estimations reduced by > 17% while maintaining average accuracy. • Optimized integrated model using negative coefficients is optimal.
Takamatsu et al. (Sat,) studied this question.