基于模块化神经网络的船舶航迹航速预测

为提高船舶航迹航速预测精度，提出一种模块化神经网络MNN(modular neural network)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RBF(radial basis function)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PSO(particle swarm optimization)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，实验结果表明模块化神经网络预测精度与网络结构均优于传统BP与RBF神经网络，证明了所提方法的有效性。

Ship trajectory and speed prediction based on MNN

In order to improve the accuracy of ship track speed prediction, a method of ship track speed prediction based on Modular Neural Network (MNN) is proposed. Firstly, the task is decomposed by using normalized mutual information and expert knowledge to determine auxiliary variables of the prediction target. Then, sub-networks are established by using RBF neural network and Elman neural network, the structure of initial sub-networks are determined by using subtractive clustering algorithm, on this basis, an error feedback method is proposed, which sample corresponding to the maximum error trained by the RBF neural network is used as a new neuron in the hidden layer, and learning parameters of the RBF neural network are optimized through the particle swarm optimization PSO. The number of hidden layer neurons in the Elman neural network is dynamically adjusted using the performance function to construct a modular neural network to predict the target. Finally, the experimental results show that the prediction accuracy and network structure of the modular neural network are superior to that of the traditional BP and RBF neural network, which proves the effectiveness of the proposed method.