Abstract: With the development of smart grids, communication network and sensor technology, the scale of power data is growing exponentially and complexly which gradually forms the big data. Up to now, traditional load forecasting methods have been unable to satisfy the analyzing requirements of massive load data. As a result, this paper put forward a short-term load forecasting method by incorporating multi-distributed BP neural networks combined with Hadoop Framework. First, on the basis of decomposing positive transferring process of input signals and back propagation of error signals in the BP neural network respectively, the load forecasting model of distributed BP neural network based on Hadoop framework was investigated and then implemented. Second, in order to weaken over-fitting problem, an clustering approach based on gray correlation and shortest distance method was utilized for the original number and the member set of multiple distributed BP neural networks after introducing the concept of the multiple, and then the valid index which determines the effectiveness of advantages and disadvantages of clustering results in order to determine the total number of models was also defined. Analytical results show that the prediction precision of multiple parallel BP neural network method is higher than that of the traditional BP neural network.