The paper focuses on multi-step prediction of non-stationary and non-Gaussian processes. First,hybrid deep decomposition (HDD) is proposed to deal with non-stationary and non-Gaussian processes combining wavelet packet decomposition (WPD), sample entropy, unit root test and variational mode decomposition (VMD). The complexity of measurement of wind speed and pressure is reduced and the predictability is improved. Second, according to Mercer′s theorem, the MH kernel function is constructed by the linear combination of Morlet kernel function and Hermite kernel function, which takes the advantages of local multi-resolution and global generalization. Particle swarm optimization (PSO) is used to optimize the parameters of MH kernel. Least square support vector machine (LSSVM) is used to build HDD-MH-LSSVM multi-step prediction model. The down-burst wind speed and the measured wind pressure on the surface of the large trans-membrane structure of the typhoon are used for large-step multi-step prediction verification. The results show that the HDD-MH-LSSVM prediction algorithm has higher prediction accuracy, good stability and strong versatility.