With the rapid application of smart electric meters, a large amount of power consumption data is generated, which challenges the data storage and communication. Meanwhile, due to the extensive production properties and business modes, the power consumption states of factories are complicated, which makes data analysis difficult. In this paper, a new neural networks-based power consumption state recognition method is proposed. The main efforts include three aspects: the pre-processing of power data, the modeling of neural network to automatically recognize the power consumption states of factories and the performance testing of the proposed method. The experimental results demonstrate the rationality and effectiveness of our method. The correct recognition of power consumption states of factories is helpful to the power supply company to arrange the sensible power supply strategy and efficient power utilization plan.