中文摘要：

针对普通的递归神经网络训练过程较复杂，而且存在记忆渐消等问题，提出一种基于回声状态网络的船舶横摇运动预报方法．该网络将隐层设计成一个巨大的动态记忆库，具有记忆功能，隐层中的神经元在学习过程中不进行权值调整，而通过线性回归的方式训练网络，使网络记忆功能加强，学习速度加快．运用该网络对某型船舶在海浪遭遇角为90°海况下的横摇运动进行预报．结果表明：回声状态网络训练简单，加速了网络的训练速度，有更好的记忆性能，以预报60步为例，回声状态网络和对角递归神经网络预报的均方根误差分别为0．0039和0．0249，提高了近8倍，在相同的预报精度下，回声状态网络的预报时长明显增长，验证了该方法的可行性与有效性．

英文摘要：

Aimed at the complexity of training for recurrent networks and problem of existing memo- ries fading, a new method was put forward, which employed echo state networks （ESN） to forecast the ship rolling motion. The hidden layer of ESN was designed as a huge dynamic database with mem- ory function. The neurons in the hidden layer cannot adjust the weight value in the learning process, but training network in way of linear regression to strengthen the network memory function and make learning speed quickly. This method was applied to forecast the roll motion of a ship sailing in the beam sea condition. Simulation results show that training of ESN network is simple and training speed can be accelerated, and possesses better memory performance. Taking an example of forecasting 60 steps, the root mean square error of forecast of ESN and diagonal recurrent neural networks （DRNN） is respectively 0. 003 9 and 0. 024 9, so the accuracy is raised nearly eight times, and under the same forecast accuracy, the forecasting time of ESN is increased obviously. This method is hereby verified to be feasible and effective.