In response to the "dual carbon" strategy, a new type of power system with a high proportion of renewable energy access has become the next development goal. As one of the main forms of current energy generation, photovoltaic (PV) power generation has characteristics such as multi-source, heterogeneous, and high-dimensional data distribution, which makes the mechanisms and effects of different features relatively complex and subsequently increases the difficulty of predicting the output of distributed PV systems. To address this, multiple categories of data mining models are integrated to construct an K-I-ELM prediction method for short-term PV output prediction in complex data environments. First, a kernel principal component analysis (KPCA) model is constructed to extract principal components based on the effective information contained in different features in the feature space through a kernel function. An information entropy (IE) model is employed to measure the weighting coefficients based on the information load of each principal component and comprehensively solve the corresponding effect weights. Finally, based on the feature evaluation results, the network parameters of the extreme learning machine (ELM) are set specifically to reduce prediction uncertainty. A case study based on actual PV power generation data from a certain substation demonstrates the adaptability and high prediction accuracy of the proposed method in different data environments.