Abstract: Objectives The motion of ships in real sea are random and uncertain, which cause great influence on offshore operations. Predicting the stochastic motion state of ship in advance can better support offshore operations. Methods This study uses the Variational Mode Decomposition (VMD) method to decomposes data features. Based on the Long Short-Term Memory (LSTM) network, an extreme short-term prediction neural network model for ship motion is built, taking ship motion data as input. Extreme short-term prediction of ship motion is achieved through multi-input and multi-output predictions using simulation data and real ship data. Results Through analysis, it is found that when the optimal prediction duration is approximately one motion period, this neural network’s prediction accuracy can reach 75%-90% or above. During simulated real-time prediction, this neural network demonstrates superior performance with strong real-time capability, with each prediction step taking less than 0.05s. Conclusions Compared with complex theoretical models, it significantly enhances computational efficiency and provides technical support for practical real-time ship motion prediction applications.