Ultra-short Term Load Forecasting Using Multiple Models Extreme Learning Machine

With the increase of power quality and the development of technologies,the smart grid emerge as the times require. It's different from the traditional power grid with cleanliness, safety and interaction . To some extent, it changed the operation mode, the supply and the demand. As the requirements of real-time interaction, information is needed in a real-time, high-speed and two-way interaction way between the supply side and consumer side. So, it gives the new features to the smart grid.In this paper, it begins with the analysis of classification and features in load forecasting, and the development of present manners in ultra-short term load forecasting. At the same time, the meaning of ultra-short term load forecasting in smart grid is summarized, according to the requirement in real-time. Then the forecasting method based on Multiple Models Ensemble Extreme Learning Machine Using Suspending Criterion is proposed. It saves the training time and improves the computing speed because of the network parameter need not to be iterative during the training process. As the result of the random selection of the weight matrix in prime extreme learning machine, the output fluctuates. In order to reduce the fluctuation, the Multiple Models Ensemble Extreme Learning Machine Using Updating Criterion method is presented. The updating criterion based on output error is designed to classify each sub-model, and different models are processed by different ways. Finally, directed to the difference of forecasting precision for the load variation of two methods described above, the Weighted Multiple Models Extreme Learning Machine based on Switching Criterion is advanced to improve the accuracy furthermore. The experiments prove that great improvement both on speed and precision have achieved. When it is applied in the ultra-short term load forecasting of smart grid, the requirement can be satisfied.