In this paper, large fluctuations in output instability problems with the model for extreme learning machine algorithm, the proposed learning algorithm using the limit switch model for ultra short term load forecasting method. Multiple neural network model of the algorithm by switching the model guidelines to be established to maintain the model into smaller error and error larger update the model in two parts. Keep the model without the need for online updates, thereby reducing volatility model output; update the model you need to take a random method for online updates, which makes the training error to a minimum, improve the generalization ability of the model. Finally, the power load forecast for an area of simulation, forecasting results show that the proposed method improves the prediction speed, save computing time, with better generalization and prediction accuracy.