陈若冰,陈萱,董明媚,梁建峰,李雨森,王刚,刘晓光.基于机器学习的海洋浮标寿命及轨迹预测[J].海洋通报,2021,(3):
基于机器学习的海洋浮标寿命及轨迹预测
Lifetime and trajectory prediction of marine floats based on machine learning methods
投稿时间:2020-09-21  修订日期:2021-01-07
DOI:
中文关键词:  Argo 数据  寿命预测  轨迹预测  机器学习
英文关键词:Argo data  life prediction  trajectory prediction  machine learning
基金项目:自然资源部海洋信息技术创新中心 2018 年度开放基金
作者单位E-mail
陈若冰 南开大学 计算机学院天津 300350 chenrb@nbjl.nankai.edu.cn 
陈萱 南开大学 计算机学院天津 300350  
董明媚 国家海洋信息中心天津 300171 自然资源部海洋信息技术创新中心天津 300171  
梁建峰 国家海洋信息中心天津 300171 自然资源部海洋信息技术创新中心天津 300171  
李雨森 南开大学 计算机学院天津 300350 liyusen@nbjl.nankai.edu.cn 
王刚 南开大学 计算机学院天津 300350  
刘晓光 南开大学 计算机学院天津 300350  
摘要点击次数: 244
全文下载次数: 138
中文摘要:
      Argo 计划 (Array or Real-time Geostrophic Oceanography) 为海洋和大气研究提供了宝贵的资料,在短期天气预报和长期气候预测中起到了重要作用。为保证 Argo 观测阵列的正常运转,需要时刻关注浮标的运行状态,以保证研究区域内维持一定数量和密度的浮标。然而 Argo 浮标投放费用高昂,投放过早会导致资源浪费,投放过迟会导致信息资料的缺失。本文旨在使用机器学习的方法对 Argo 浮标在未来某个时间点的位置和状态 (仍在工作或已经损坏) 进行预测,以提前制定投放计划,保证在正确的位置和时间投放新的浮标,以减少资金投入。对于浮标寿命预测任务,除硬件特征之外添加额外的已存活时间作为动态属性,使用回归决策树、梯度提升回归树、随机森林和支持向量回归机等机器学习方法,对浮标剩余寿命 进行预测。对于浮标轨迹预测任务,使用基于 LSTM 的 Encoder-Decoder 模型对未来多个时间步后的浮标的经/纬度信息进行预测,有效地避免了传统的 LSTM 模型循环单步预测所带来的误差累积问题。实验证明本文提出的浮标剩余寿命和位置预测模型都能达到较高的预测准确率,对指导浮标投放有重要意义。
英文摘要:
      The Argo (Array or Real-time Geostrophic Oceanography) provides valuable data for ocean and atmospheric research, and plays an important role in short-term weather forecasting and long-term climate prediction. In order to ensure the normal operation of the Argo array, it is necessary to pay attention to the operation status of the floats at all times to ensure that a certain number and density of floats are maintained in the study area. The replacement cost of Argo floats is high. Too early release will result in waste of resources and too late release will result in the lack of information. This article aims to use machine learning methods to predict the position and status of Argo floats at a certain point in the future (still working or damaged), in order to formulate a delivery plan in advance to ensure that new floats are placed at the correct location and time and reduce capital investment. In the float life prediction, additional survival time is added as a dynamic attribute besides hardware features. Machine learning methods such as Decision Tree Regressor, Gradient Boosting Regression Tree, Random Forest and Support Vector Regression are used to predict the remaining life of the floats. For the float trajectory prediction, the Encoder-Decoder model based on LSTM is used to predict the latitude/longitude information of floats for multiple time steps in the future, which effectively avoids the error accumulation caused by the traditional LSTM model cyclic single-step prediction. The experiment proves that the float remaining life and position prediction model proposed in this paper can achieve high prediction accuracy, thus is of great significance to guide the float delivery.
查看全文    下载PDF阅读器
关闭