Wind velocity is usually assumed to obey a stationary stochastic process in wind-resistant design, without taking into account the non-stationarity. The wavelet transform (WT) method was adopted to capture the time-varying properties of the low-frequency mean winds, and the associated turbulence features, including turbulent intensity, gust factor, probability density function, and power spectrum, were analyzed in depth. Furthermore, the three-dimensional evolutionary power spectral density (EPSD) of strong winds was estimated. Finally, this study presents a study on the modeling of the three-dimensional EPSD. The results show that the proposed EPSD models are in good agreement with the estimated EPSD. This study can be used for numerical simulation of non-stationary wind and analysis of wind-induced effects under turbulent atmospheric boundary layers with consideration of non-stationarity.