Abstract: The ultra short term power prediction of photovoltaic power generation usually relies on support vector machines to obtain prediction results, and the prediction accuracy is low in dynamic and complex scenarios. Therefore, a prediction method for ultra short term power of centralized photovoltaic power generation considering fluctuations in optical power is proposed. Apply Hi1bert transform algorithm to preprocess historical power generation data and analyze the time-frequency entropy characteristics of power fluctuations. Based on multiple time-sharing short-term memory neural networks, an intelligent prediction model is established to accurately derive ultra short term photovoltaic power generation prediction values by capturing long-term dependencies between power fluctuation characteristics. The experimental results show that the root mean square error of the power prediction value provided by this method is always below 1 MW, which improves the accuracy of photovoltaic power generation prediction.