In order to improve the prediction accuracy of the ship’s motion attitude during the launch of the carrier-based target drone， this paper uses a ship attitude prediction method based on the long short-term memory （LSTM） network. In view of the error accumulation problem caused by long-term prediction， this paper proposes an improved window sliding method， which eliminates the prediction result oscillation caused by the cumulative error by filtering each prediction result with variational mode decomposition （VMD）. The wave tank test was carried out through finite element simulation and a self-designed ship model experimental platform to collect time series data of key attitude parameters such as roll， pitch， and heave. The experimental setting covers typical sea conditions from level 1 to level 5. The experiment shows that the maximum mean squared error（MSE） reduction of the model in the prediction of heave displacement， roll angle and pitch angle is 99.4%， and the MAPE is reduced to 2.11%， which verifies the effectiveness of its engineering application. The research results can provide high-precision ship motion situation prediction for the launch guidance system of the carrier-based target drone， which has important engineering value for improving landing safety.