The random fluctuation of wind leads to the instability of wind power, particularly wind power intermittency is a critical event in case of grid connected power plants, leading to severe consequences like low system reliability, high reserve capacity and high operational costs. This paper proposes a novel wind speed prediction model based on wind speed ramp (WSR) and residual distribution. Firstly, the variational mode decomposition (VMD) is used to decompose original wind speed series to extract different fluctuation characteristics, and several sub-sequences are obtained. Then, we use 3 different neural networks to predict the main part of decomposition result and use auto-regressive moving average (ARMA) model to predict the rest fluctuation part of decomposition results. Next, WSR optimized by particle swarm optimization (PSO) is used to modify the prediction results of LSTM neural network to decrease the prediction errors caused by one-step lag, the kernel density estimation (KDE) is used to fit the distribution of VMD residuals and sample randomly from the distribution to get residual series. Finally, the final prediction results of V-PSOR-LSTM-KDE are obtained by adding prediction results of WSR modified LSTM, ARMA and random sampling based on KDE. This study decomposes wind speed into the main trend part, fluctuation part and residual part, analyzing each part and makes predictions with different models according to their characteristics, which provides a new thought for wind speed prediction and contributes to the construction of smart grid.