在对船舶横摇预测研究的基础上,提出一种基于混沌和在隐层具有2个反馈权值的二阶对角递归神经网络的直接多步预测模型;给出了易于实现的动量梯度学习算法,并对其收敛性进行了验证.仿真结果表明,直接多步预测不依赖于单步预测的结果,对比单步预测模型能快速、准确地预测船舶横摇运动时间序列,具有更好的预测精度及较长的预测时间. Based on the research of ship roll prediction, a direct multi-step prediction model based on chaos and second-order diagonal recursive neural networks with two feedback weights in hidden layer is proposed. It is easy to realize the momentum gradient learning The simulation results show that the direct multi-step prediction does not depend on the result of single-step prediction, and the single-step prediction model can predict the time series of ship roll motion quickly and accurately, Prediction accuracy and long prediction time.