基于模块化神经网络的船舶航迹航速预测

Ship Trajectory and Speed Prediction Based on Modular Neural Network

为提高船舶航迹航速预测精度,提出一种模块化神经网络(modular neural network,MNN)船舶航迹航速预测方法。首先,利用归一化互信息与专家知识确定预测目标的辅助变量从而分解任务;然后,将径向基函数(radial basis function,RBF)神经网络和Elman神经网络用于子网络搭建,使用减法聚类算法确定初始子网络结构,在此基础上提出误差反馈方法将RBF神经网络训练的最大误差所对应的样本作为隐含层新增神经元并通过粒子群算法(particle swarm optimization,PSO)优化RBF神经网络学习参数,运用性能函数动态调整Elman神经网络隐含层神经元数目以此构造模块化神经网络对目标进行预测;最后预测模型对比实验表明:所提方法具有更加简洁的网络结构与较高的预测精度。

In order to improve the accuracy of ship track speed prediction,a method of ship track speed prediction based on modular neural network(MNN)was proposed.Firstly,the task was decomposed by using normalized mutual information and expert know-ledge to determine auxiliary variables of the prediction target.Then,sub-networks were established by RBF neural network and Elman neural network,and the structure of initial sub-networks were determined by using subtractive clustering algorithm.On this basis,an error feedback method was proposed,and the sample corresponding to the maximum error trained by the RBF neural network was used as a new neuron in the hidden layer,and learning parameters of the RBF neural network were optimized through the particle swarm optimization(PSO).The number of hidden layer neurons in the Elman neural network was dynamically adjusted.The performance function was used to construct a modular neural network to predict the target.Finally,the accuracy and effectiveness of the proposed method were verified by the comparison experiments of prediction model.