Wireless network control is a favorable factor to promote the intelligence of high-speed trains. As a typical time series, wireless network delay has strong randomness, large volatility and other problems leading to difficult prediction. In view of these problems, a wireless network delay prediction model with singular spectrum analysis-improved particle swarm optimization and LSSVM is proposed. The length of the window was first determined by the Cao method, the delay sequences were analyzed by singular spectral analysis to obtain a series of subsequences. Each subsequence was predicted using the LSSVM model optimized for the chaotic particle swarm. Finally, all the subsequence predicted values were superimposed to obtain the final prediction results, the simulation results show that the average absolute percentage error (MAPE), mean squared error (MSE) and average absolute error (MAE) are 2.8%, 1.055 and 0.44 lower respectively compared with the wavelet decomposition model. Compared with the EMD decomposition model, 7.4%, 3.377 and 1.118 decreased, respectively. Compared with the CEEMD decomposition model, it was reduced by 6.2%, 2.568, and 0.974, respectively. The accuracy was significantly higher than that in the other models.