The invention discloses a spectrum prediction method in a cognitive radio system based on a convolutional neural network. The method can obtain the higher prediction accuracy through the shorter-timetraining, and belongs to the field of information and communication. The method comprises the steps: 1, performing the uninterrupted spectrum sensing of a frequency band F through a period T, recording S<t0>, S<t1>, S<t2>,... of the spectrum sensing from the moment t0 and I<t0>, I<t1>, I<t2>,... corresponding to S<t0>, S<t1>, S<t2>,..., marking t0+kT as tk, wherein k=0, 1, ..., wherein S<tk> represents the occupation conditions of all channels of the frequency band F at the moment tk, I<tk> represents the number of idle sensing periods of each channel of S<tk> from the moment tk+T; 2, inputting S<t0>, S<t1>, S<t2>,... and I<t0>, I<t1>, I<t2>,... into the convolutional neural network, and performing the training of the convolutional neural network, wherein I<t0>, I<t1>, I<t2>,... are used for making labels in the training; 3, continuously predicting the channel C with the maximum future idle probability according to S<t0>, S<t1>, S<t2>,... through the trained convolutional neural network.