Feasibility study of solar power curtailment prediction using multivariate dynamic time warping - based machine learning: A case study of Jeju island

초록

Solar power generation is rapidly expanding as a key renewable energy source amidst the international trend toward energy security. However, variability in solar irradiance often leads to supply-demand imbalances, causing oversupply that exceeds power grid capacity. Consequently, curtailment measures are being implemented to enforce power generation limits. These measures, however, hinder the efficiency of renewable energy and cause power losses. Predictive models using statistics and machine learning have been proposed; nevertheless, these models overlook numerous curtailment events, limiting their broader effectiveness. Therefore, this study proposes a dual-layer framework based on multivariate dynamic time warping (MDTW) to reduce false positives while maximizing curtailment detection by threshold optimization of three prediction models. Using 41 months of time-series data from Jeju Island showed that the models increased the curtailment detection rate (recall) by 5.5 and 19.4 percentage points for XGBoost and random forest. Notably, logistic regression showed the highest gains, with a 25-percentage-point increase in the curtailment detection rate. This study verifies the feasibility of MDTW-based filtering that can be combined with various prediction models and suggests its potential as a practical decision-making tool for responding to power system risks and establishing strategies through effective curtailment prediction.

키워드