Long-span bridges and high-rise buildings are widely distributed in southeastern coastal areas of China. However, this area is also affected by typhoons every year. Accurate prediction of typhoon wind speed is a very important means of increasing disaster prevention capabilities of engineering structures and auxiliary decision-making. In this paper, a comparative study in different signal decomposition methods used in multi-step forecasting of wind speed is carried out. First, the characteristics of eight typical signal decomposition methods are enumerated. Then, the least squares support vector machine (LSSVM) prediction model based on particle swarm optimization (PSO) optimization is established based on different signal decomposition methods. Finally, the multi-step ahead forecasting experiment is carried out using two measured wind speed, which are collected from the main tower of a long-span bridge and the roof of a high-rise building. The prediction results of the two groups of experiments show that the VMD-LSSVM-PSO model has the best prediction performance.