Small-Sample Solar Power Interval Prediction Based on Instance-Based Transfer Learning

In the context of high photovoltaic (PV) penetration, high-quality solar power interval prediction is important for grid system operation. However, in some cases, sufficient amount of data are not available to train a reliable prediction model, especially for new installed PV stations. To tackle this problem, this article proposes a novel small-sample interval prediction model with improved TrAdaBoost (SIPTAB) method for solar power prediction. Sample weight is designed to select the important samples from other data sources. First, Extreme Learning Machine (ELM) with Direct Quantile Regression (DQR) is employed as the base predictor to construct interval boundaries. Second, an improved TrAdaBoost algorithm is proposed to iteratively construct a boosting ensemble interval predictor to enhance the prediction performance with limited amount of data. Third, a two-stage model training strategy is introduced in the architecture to optimize the boosting ensemble interval predictor and further improve prediction quality. Comprehensive experiments based on realistic solar power data are conducted to confirm the superiority of proposed model.