Radio spectrum prediction is of great significance for dynamic spectrum management and alleviating spectrum congestion. Based on the real spectrum dataset, this paper constructs a multi-channel temporal-frequency fusion network (MTF 2 N) for spectrum prediction. The network consists of two parts: first, it uses CNN to extract the latent features of the occupancy state of multi-channel and multi-slot; then, it uses the latent features of the spectrum occupancy state to predict occupancy state through the memory property of LSTM. Experiments show that the network model designed in this paper can achieve comparable performance to the LSTM network in short-term spectrum prediction. In the long-term spectrum prediction, MTF 2 N is more accurate than LSTM, Seq2seq and GRU networks since MTF 2 N integrates more channel and time slot correlation features. The prediction accuracy of 200 prediction time slots for 40 channels in GSM1800DOWN service band reaches 92.26% and more robust performance is obtained.