Long Short-Term Memory (LSTM) is widely used in time series prediction, however, few people have noticed its predictions always lag behind corresponding measurements, which greatly reduces the predictive and warning capabilities of LSTM. The present paper proposes a new hybrid model that couples empirical/variational mode decomposition (E/VMD) and LSTM to solve the time lag problem in LSTM forecasting. Field significant wave height and wind speed data from three typical stations around the Shandong Peninsula: Cheng Dao (CD), Zhi-Fu Dao (ZFD) and Xiao-Mai Dao (XMD), were used to test the new model. All sites proved that the proposed model mitigated the time lag problem (for 12-h ahead prediction of CD, the lag is reduced from 3 to 1; for 12-h ahead prediction of XMD, the lag is reduced from 14 to 1) of LSTM while also increasing the accuracy (for 12-h ahead prediction of CD, the R2 is improved from 0.544 to 0.906; for 12-h ahead prediction of XMD, the R2 is improved from 0.291 to 0.781). The improvement is because VMD separates the high-frequency noise from the slowly varying low-frequency components, enhancing the stationarity of the original sequence. This method can be extended to other time series predictions.