A probabilistic solar irradiance interval-valued prediction model with multi-objective optimization of reliability, sharpness and stability

Abstract

Improved interval-valued prediction models for solar power and irradiance forecasting allow enhanced planning and operation of solar power systems. Highly uncertain atmospheric and environmental factors are major challenges of solar irradiance forecasting. Existing upper and lower bound estimation methods mainly focus on narrowing the prediction intervals and minimizing forecasting errors. However, the sensitivity of the interval-valued prediction model is not considered. Sensitivity is described as the model's output fluctuations due to unseen samples. Models with high sensitivity may not perform well in real-life applications under uncertain environments. This paper presents a novel interval-valued prediction model, P_RSS, by simultaneously optimizing the reliability, sharpness, and stability (RSS) for probabilistic solar irradiance interval-valued prediction. With sensitivity regularization, P_RSS has reduced sensitivity to unseen samples with perturbations from training samples and enhanced robustness. An Extreme learning machine (ELM) model is constructed to directly output prediction intervals (PIs) of solar irradiance via a multi-objective optimization of the RSS. An evaluation framework is proposed to verify the RSS performance. Moreover, a new comprehensive evaluation indicator is proposed to evaluate the PIs. Case studies on three American solar irradiance datasets show that P RSS yields outstanding performance against state-of-the-art methods.