Short-term interval prediction of photovoltaic power based on decomposition optimization and adaptive bandwidth KDE

As the installed capacity of grid-connected photovoltaic systems continues to rise, the demand for highly accurate photovoltaic power prediction has intensified. However, current point prediction approaches provide only single-value forecasts and cannot quantify the inherent uncertainty in photovoltaic power, which may limit their effectiveness in supporting risk-aware dispatching strategies. Therefore, this paper proposes a photovoltaic power interval prediction model based on decomposition optimization and adaptive bandwidth kernel density estimation. Firstly, the improved complete ensemble empirical mode decomposition with adaptive noise is utilized to decompose the photovoltaic power series into a series of components with different frequencies, which are then grouped into high-frequency and low-frequency components based on time-series fuzzy entropy. The low-frequency components are predicted using gated recurrent unit networks while the high-frequency components are predicted by Transformer. The point prediction error is obtained after reconstructing the predictions obtained from each sub-model. Finally, an adaptive bandwidth kernel density estimation method based on density clustering and adaptive bandwidth fits the prediction error distribution to derive the prediction intervals so as to form the final interval prediction under certain confidence interval. Experimental results show that the proposed model can provide reliable photovoltaic power interval forecasts.