Given the nonlinear and non-stationary data distribution characteristics of distributed photovoltaic system load, this paper proposes a suspended ensemble extreme learning machine (SEELM) method based on neural networks and a hanging criterion to implement power load prediction in distributed photovoltaic systems. First, multiple neural network models are built, and the initial input weights of each model are randomly assigned. Then the hanging criteria are designed to divide the models into two parts according to the numerical fluctuation ranges at different time spots. For large error models with larger fluctuation ranges, the online updates will be carried out in a probabilistic way to optimize the training error and input weights simultaneously. Finally, the outputs of all submodels are taken for the final output, which can reduce the error fluctuation impacts in the initial weight selection step. Based on an empirical simulation of the actual distributed photovoltaic system in a region, the advantages of the proposed method in terms of prediction accuracy and output stability under the scenarios of large fluctuation in photovoltaic load can be verified, and better capability and performance of load forecasting in the high-proportion photovoltaic systems can thus be achieved. This work is supported by the National Natural Science Foundation of China (No. 61803343).