何柏岩,杜金辉,杨绍琼^{, ,} ,张润锋,牛文栋^{, ,} ,兰世泉
AuthorsHTML:何柏岩¹，杜金辉¹，杨绍琼^{1, 2, 3}，张润锋¹，牛文栋^{1, 2, 3}，兰世泉¹
AuthorsListE:He Baiyan,Du Jinhu,Yang Shaoqiong,Zhang Runfeng, Niu Wendong,Lan Shiquan
AuthorsHTMLE:He Baiyan¹，Du Jinhuⁱ¹，Yang Shaoqiong^{1, 2, 3}，Zhang Runfeng ¹， Niu Wendong^{1, 2, 3}，Lan Shiquan¹
Unit:1. 天津大学机械工程学院，天津 300350；
2. 天津大学青岛海洋技术研究院，青岛 266237；
3. 青岛海洋科学与技术试点国家实验室，海洋观测与探测联合实验室，青岛 266237

Unit_EngLish:1. School of Mechanical Engineering，Tianjin University，Tianjin 300350，China；
2. Qingdao Institute for Ocean Engineering of Tianjin University，Qingdao 266237，China；
3. The Joint Laboratory of Ocean Observing and Detection，Pilot National Laboratory for Marine Science and Technology(Qingdao)，Qingdao 266237，China

Abstract_Chinese:水下滑翔机是海洋立体观测网络的重要组成部分．本文以天津大学“海燕-L”水下滑翔机为平台，针对其 航行过程中易受到海流等因素影响发生航向偏离的问题，提出基于变分模态分解(variational mode decomposition， VMD)和最小二乘支持向量机(least squares support vector machine，LSSVM)方法的深平均流(depth-averaged current velocity，DACV)预测模型，从而指导水下滑翔机进行有效航向修正或局部路径规划．首先，根据水下滑翔机运动 原理，针对水下滑翔机剖面运动理论出水位置与实际出水位置的偏离现象，结合岸基操作与控制条件以及水下滑翔 机的通信方式，建立了短数据精简深平均流计算模型；其次，对得到的深平均流数据按照时间顺序进行序列构造， 得到了实时性多剖面深平均流历史序列，并利用 VMD 将其分解为若干分量序列；最后，基于 LSSVM 对深平均流 数据进行预测，得到了未来深平均流预测数据．海试结果表明，在以 4 个剖面数据作为输入的情况下，VMD\u0002LSSVM 的预测结果在经度、纬度方向的流速均方根误差分别为 0.025 7、0.020 5，相关系数分别为 0.948 4、 0.973 7，优于 LSSVM、经验模态分解(empirical mode decomposition，EMD)的 LSSVM(EMD-LSSVM)、长短期记 忆神经网络(long short-term memory，LSTM)、EMD-LSTM 及 VMD-LSTM 预测性能，更好地预测深平均流的数值 并追踪其变化趋势，证明所提方法的有效性．
Abstract_English:Underwater glider is an important part of the ocean stereo observation network. This study considered the “Petrel-L” underwater glider of Tianjin University as the research platform，aiming at the problem of course deviation caused by ocean currents and other factors during its navigation process，and proposed a depth-averaged cur-rent velocity(DACV)prediction model based on the least squares support vector machine(LSSVM)method and variational mode decomposition(VMD)to guide the effective course correction or local path planning for underwater gliders. First，according to the motion principle of an underwater glider，in view of the deviation of the theoretical water outlet and the water exit positions of the underwater glider profile motion，combined with shore-based operation and control conditions，as well as the communication mode of the underwater gliders，a short-data streamlined DACV calculation model was established. Next，the obtained DACV data was structured in chronological order，the real\u0002time historical sequence of multiprofile DACV was then obtained，and VMD was used to decompose the historical sequence into several components. Finally，the future DACV forecast data was predicted based on the LSSVM. The sea trial result shows that in the case of four-profile data as input，the root mean square error of the prediction result of VMD-LSSVM is 0.025 7 in the longitudinal direction and 0.020 5 in the latitudinal direction；the correlation coefficient is 0.948 4 in the longitudinal direction and 0.973 7 in the latitudinal direction，superior to the prediction performance of LSSVM，LSSVM using EMD(EMD-LSSVM)，long short-term memory neural network(LSTM)， EMD-LSTM，and VMD-LSTM，proving that VMD-LSSVM can better predict the value of the DACV and track its changing trend，demonstrating the effectiveness of the proposed method.
Keyword_Chinese:水下滑翔机；深平均流；变分模态分解；最小二乘支持向量机；预测方法
Keywords_English:underwater glider；depth-averaged current velocity(DACV)；variational mode decomposition(VMD)； least squares support vector machine(LSSVM)；prediction method

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6616