During ionospheric storm periods, the existing prediction methods for the critical frequency of ionospheric F2 layer (foF2) can not satisfy the requirements of the practical applications. In this paper, a new prediction method for the foF2 based on neural networks is proposed. This method is driven by the previous time series of the ap index and the monthly mean of the solar spot number, while the output is the estimations of the foF2 in the next 24 hours. The simulation results indicate that the new method outperforms the existing prediction methods (such as STORM model, or neural network method proposed by Cander).