Scientific Journal of Control Engineering June 2013, Volume 3, Issue 3, PP.106-110
Prediction of Ship Roll Based on Second Diagonal Recurrent Neural Network Liang Xu 1, Zhanying Li 2, Yuzhi Song 3, Yanping Wang 2 1. Hull Workshop, COSCO (Dalian) ship yard Co., Ltd, Dalian 116113, China 2. School of Electronic Engineering and Automation, City Institute, Dalian University of Technology, Dalian 116600, China 3. Det Norske Veritas (China) Co., Ltd. Dalian Branch 116011, China #
Email: l_zy1979@126.com
Abstract An optimized second diagonal recurrent neural network is proposed to develop a model of prediction of ship rolling motion. This approach is based on an algorithm of optimization second diagonal recurrent neural networks (OSDRNN). The stochastic gradient descent algorithm is used to optimize parameters of this network. Using this model to predict the situation of one certain type of ship sailing in the beam sea condition, simulation results show that the optimization of this network improves network performance and the generalization performance of the network, and it has higher prediction on the accuracy and forecast rate. The presented network model used in contrast to SDRNN model can quickly and accurately predict the time series of ship rolling. Keywords: Second Diagonal Recurrent Neural Network (SDRNN); Stochastic Gradient Descent Algorithm; Time Series Prediction; Ship Rolling Motion
1 INTRODUCTION Recently, there have been increasing researches interests of artificial neural networks and many efforts have been made on applications of neural networks of various fields [1-4]. Neural network was extensively used to study on nonlinear control system by many scholars. Most researchers used feed-forward neural network, combined with tapped delays and the back propagation training algorithm (BP) to identify dynamical systems. Then recurrent neural networks and fully connected recurrent neural networks were developed, but these methods needed long time to converge at weights. Diagonal Recurrent Neural Network (DRNN) is proposed by Ku Chao-chee and Lee K wang in 1995[5], DRNN has simple structure and easy realization training method. The self-feedback of the hidden neurons ensures that the outputs of DRNN contain the whole past information of the system even if the inputs of DRNN are only the present states and inputs of the system. The structure of DRNN may be simple than that of recurrent neural networks, so it attracted attention by many scholars. However, the recurrent weights are updated only using the previous state and can’t use other states directly. Many scholars started to make the improvement on the basis of it [612] . In the literature [12], second diagonal recurrent neural network that contains two recurrent weights for every hidden neuron was proposed by Ali Kazemy in 2007, more historical states of neurons can be incorporated directly into the training algorithm in this network. It also raises concern in time series prediction [13-16]. Due to increasing one feedback weight, the structures and parameters of network become complex, thereby it is not easy to select the parameters for network during actual system’s prediction. This article tries to optimize the parameters of secondorder diagonal recurrent neural network on the basis of literature [16], so that improve the network performance. It is used for the ship rolling motion prediction, and by contrast of experiment of non-optimized second-order diagonal recurrent neural network. The accuracy, time and other characters of prediction after optimization of second-order diagonal recurrent neural network can be evaluated. This paper is organized as follows. In section II, one SDRNN model is developed. In Section III, the optimized parameter was considered as to train various weights based on stochastic gradient descent algorithm. In Section IV, a simulation prediction of ship rolling motion will be given. Finally, concludes the paper. - 106 http://www.sj-ce.org/