Load classification plays an important role in power grid to guide the power planning and guarantee the reliable operation of grid. However

traditional load classification is difficult to adapt to larger scale and higher dimensions of user load data. Firstly

according to the characteristics of mass load data

multiple types

high speed on the demand side

the back propagation neural network(BPNN)algorithm is improved to achieve the high efficient classification of mass load data on Spark platform. Then

the training samples are divided into blocks to reduce the learning time of each BPNN

the ensemble learning is adopted in order to compensate for the performance degradation of the single BPNN classifier due to the lack of learning samples. The final result is obtained by voting most of the sub classification results. In addition

this paper presents a training sample selection method based on K-means and K-medoids clustering algorithm. Analytical results show that the presented method can not only effectively classify load curves

but also greatly improve the processing efficiency of mass data.