基于混合深度学习的无分布假设光伏功率概率密度预测

Abstract Photovoltaic (PV) power generation exhibits high randomness, and probabilistic forecasting can effectively quantify its uncertainty. However, existing probabilistic forecasting models are constrained by prior distribution assumptions and incomplete representation, which weakens their ability to model the true data generation process, leading to suboptimal prediction results. To address this, this study proposes a distribution-free probabilistic density forecasting model for PV power based on B-spline-iTransformer-Multi-Head Cross Attention(BS-iMCFormer). The core of the model lies in: using B-splines to fit the probability density function (PDF), which is obtained through kernel density estimation (KDE), extracting control points that represent the characteristics of the PDF to construct the coefficient vector and transforming the PDF prediction into the prediction of the coefficient vector; utilizing iMCFormer to integrate meteorological, power, and coefficient vector cross-modal associations to achieve enhanced feature reconstruction; and employing B-spline interpolation to obtain the predicted PDF from the predicted coefficient vector, thus avoiding the constraints of parametric distribution assumptions and discrete representation. The study conducts experiments using data from a PV station in Hebei Province, China. The results show that the proposed model’s predicted PDF significantly outperforms all comparison models in terms of visual overlap and statistical consistency with the real data’s kernel density estimate. The core metric of distribution difference, KL divergence, is reduced by 54.82%, while MAE and RMSE decrease by 68.93% and 69.42%, respectively, and the correlation metric R2 improves by 6.76%.