There exist many uncertain factors in electric power system load. To address the low prediction of single model ,wavelet packet transformation is utilized to divide the original load curve into sub-sequences with different frequencies. On this basis, BP neural network, error feedback weighted time sequence and grey model are combined to form a new prediction modal,whose weight of every individual modal is optimized by crisscross optimization algorithm(CSO). Finally, continuous seven days' short-term load forecasting is implemented on certain district power grid in GuangDong. The results prove that the proposed model is apparently superior over other models like single prediction modal, equal weight prediction modal as well as the the inverse variance prediction model in terms of prediction accuracy.