The invention provides a photovoltaic power generation power seasonal multi-model interval prediction method based on an extreme learning machine and nuclear density estimation, and the method comprises the steps: firstly, analyzing the output power, power deviation, power change rate and other indexes of a photovoltaic power station, and indicating that the photovoltaic power output and fluctuation show obvious seasonal distribution characteristics through a result; establishing a deterministic prediction model of photovoltaic output in different seasons through the neural network of the extreme learning machine; secondly, fitting error distribution of deterministic prediction through a non-parameter kernel density estimation method, and then obtaining a photovoltaic power prediction interval meeting a certain confidence level. According to the method, possible fluctuation ranges of photovoltaic power under different confidence levels can be described, an approach for evaluating the reliability of a prediction interval is provided, and support is provided for risk evaluation and system reliability analysis of the photovoltaic power station.