Aiming at the problems of the Informer model, where information is easily lost during the distillation process and training convergence efficiency is low, and the KDE algorithm, where prediction intervals are inaccurate due to fixed bandwidth and the tendency to ignore the temporal correlation of adjacent days’ output, this paper proposes a photovoltaic power probability interval prediction model based on PInformer-HKDE. Firstly, sparse entropy attention is introduced to improve the sparse attention mechanism in the distillation process of the Informer model, solving the issue of important information loss. Secondly, a novel decay weight loss function considering time series patterns is adopted to better capture the dependency relationships between sequences in the Informer model and improve convergence efficiency. Finally, an improved probability density function is employed in KDE, and a dynamic time window mechanism is added to accurately quantify the uncertainty prediction intervals. Experimental results show that the uncertainty prediction intervals generated by this method at different confidence levels exhibit high reliability and good sharpness.