In order to further mine the evolution information of chaotic system, improve the prediction accuracy and reduce the training time, a wide area learning system based on maximum information mining, multiple kernel LS-SVM was proposed for short-term power load forecasting. Firstly, in order to effectively capture the nonlinear information of power load, an improved leaky integrator dynamic reservoir was introduced, which could not only obtain the current state information of the system, but also take into account the historical state information. Then, the nonlinear information could be mine by nonlinear random mapping. Then a multiple kernel LS-SVM prediction model was proposed, which effectively integrated the advantages of each kernel function. Through two power load forecasting cases, the prediction error of this method was compared with the traditional BP algorithm and SVM algorithm. The prediction results verify that the proposed chaotic time series prediction algorithm in this paper has high prediction accuracy and is suitable for short-term power load forecasting.