The prediction of deck motion is an effective and potential means of improving the landing/ take-off safety of carrier-based aircraft using current and historical deck-motion measurements when deck motion in six degrees of freedom cannot be effectively controlled or restrained. The prediction models of deck motion should have excellent nonlinear fitting ability to cope with the deck-motion characteristics of randomness and nonlinearity caused by waves and wind; and should not use heavy computation to fulfill the requirement of real-time prediction for deck motion. It is generally believed that classical feed-forward neural networks, such as the back-propagation (BP) network, have excellent nonlinear fitting ability but suffer from slow training processes and reduced local optimum, thus failing to satisfy the requirements of real-time and high accuracy for deck-motion prediction. In addition, the extreme learning machine (ELM) is easy to train but it is difficult for ELM to determine the number of hidden layer nodes; an incorrect number of hidden layer nodes will introduce poor stability and generalization ability. To fulfill the requirements of deck motion prediction, a prediction method based on ELM, support vector machine, and particle swarm optimization [particle swam optimization kernel extreme learning machine (PSO-KELM)] is designed. In this method, the fundamental structure of the ELM is used and the kernel function from the support vector machine (SVM) is introduced to replace the hidden function in ELM. Further aiming at the acquisition of optimal parameters, including the penalty coefficient and kernel parameters for the kernel function, autoadaptive particle swarm optimization is adopted. Simulation results indicate that a prediction method based on PSO-KELM has the advantages of a simple structure, fast training speed, and powerful generalization ability, and thus can satisfy the requirements of real-time and high-accuracy deck-motion prediction. Compared with the prediction data from BP and the ELM, high-precision prediction data can be obtained with PSOKELM. PSO-KELM has a significantly reduced training time compared with BP.