Abstract: For the problem that geosynchronous earth orbit(GEO) satellite spectrum changes unsmoothly and nonlinearly when low earth orbit(LEO) satellite accesses GEO satellite spectrum, an improved particle swarm optimization(IPSO) based on improved particle swarm optimization, is proposed. IPSO optimizes the interstar-spectrum prediction model of graph convolutional networks and long short term memory (GCN-LSTM). The model uses GCN-LSTM network to learn the time-frequency domain characteristics of spectral data and adjust the weight allocation of key information by combining the attention mechanism. The particle swarm algorithm is improved by using the nonlinear adjustment of inertia weight strategy and Cauchy’s variance strategy seeks to optimize the number of the first layer of LSTM units, the number of the second layer of LSTM units, the learning rate, dropout and batch size, which in turn improves the prediction accuracy of the model. Using the collected high-orbit satellite spectrum dataset, experimental comparisons are completed for three spectrum prediction scenarios of 1 second, 30 seconds and 1 minute, and the results show that compared with the ConvLSTM baseline model, the mean absolute error(MAE) of the proposed model has been reduced by 27.00%, 17.63%, 17.68%, respectively, and it has better spectrum prediction capability.