The classification prediction of the typical mode of user consumption is an important part of electric power load forecasting. The single-core fuzzy C-means algorithm cannot balance the prediction accuracy and the universal performance in electric power big data mining, so this paper presents an improved unsupervised learning multi-core fuzzy C-clustering algorithm in the short-term power load scenario. A power data load forecasting model of the double-layer neural network is established to compare the effects of the improved algorithm. User data is accelerated by MapReduce parallelization. The numerical experiments show that the improved algorithm has wide applicability and effectiveness in the actual power user data set, and can significantly improve the accuracy of short-term load forecasting.