Abstract: To make power and load matching because of the changing load operation of nuclear power plant, it needs to know the precise value of reactor power to make accurate and timely adjustment. But it is hard to acquire the precise value of reactor power through measurement of reactor power based on physical or empirical model for the reason that there is uncertainty in measurement instrument. Reactor power prediction was made in several cases through developing BP neural network and preprocessing of input data of network. The results indicate that not only the disadvantage of uncertainty in measurement instrument is overcome, but also the reactor power can be properly predicted in some cases of partial data lack.