In terms of the optimization and distribution problem of power metering, a neural networks based power metering scheduling model is developed to improve the prediction accuracy. The production scheduling process is influenced by multi-factors including demand, purchasing, and calibration plans, etc. In this paper, by mining the historical data, Holt-Winters model, BP neural network model and RBF neural network model are used to predict, compare, and analyze the demand of electricity meters. In addition, Shapley method is adopted to obtain a combined model based on two models with smallest performance measures, where weights of the corresponding models are calculated to obtain the optimal production scheduling scheme. Numerical simulation indicate that the prediction accuracy of RBF neural network model is higher than those of BP neural network and Holt-Winters model. Moreover, in contrast to the single-model based method, Shapley combined model is more effective and more practical, which can be used for grid companies to establish efficient and scientific production scheduling plan.