基于多浮标空间多特征融合的海水溶解氧浓度预测  

Prediction of seawater dissolved oxygen concentration based on multi-buoy spatial multi-feature fusion

作　　者：朱奇光[1] 申震 李享 魏祯 乔文静 张淋淞 陈颖[2] Zhu Qiguang;Shen Zhen;Li Xiang;Wei Zhen;Qiao Wenjing;Zhang Linsong;Chen Ying(Key Laboratory for Special Fiber and Fiber Sensor of Hebei Province,School of Information Science and Engineering,Yanshan University,Qinhuangdao 066004,China;Key Laboratory of Measurement Technology and Instrument of Hebei Province,School of Electrical Engineering,Yanshan University,Qinhuangdao 066004,China)

机构地区：[1]燕山大学信息科学与工程学院河北省特种光纤与光纤传感器重点实验室,河北秦皇岛066004 [2]燕山大学电气工程学院河北省测试计量技术及仪器重点实验室,河北秦皇岛066004

出　　处：《海洋学报》2025年第1期104-116,共13页

基　　金：国家自然科学基金(62275228);河北省重点研发计划(19273901D,20373301D);河北省自然科学基金(D2024203002)。

摘　　要：溶解氧浓度是衡量海水水质的重要指标之一。为了及时掌握海水水质变化情况,降低海水污染风险及其带来的损失,建立海洋水质参数预测机制至关重要。为此,本文提出了一种基于浮标网络时空信息融合和改进生成对抗网络(Generative Adversarial Networks,GAN)的海水溶解氧浓度预测模型,旨在整合监测区域内浮标网络的拓扑信息并实现浮标传感器的多特征融合。该模型利用图注意力网络(Graph Attention Mechanism,GAT)挖掘不同近邻点对目标节点的影响,计算邻接节点的权重,从而捕获浮标数据的时空特征;通过双头注意力机制与双时间尺度更新规则(Two Time-Scale Update Rule,TTUR)优化GAN预测网络及网络训练过程,改善生成对抗网络的训练速度平衡问题,提高生成器网络的拟合效果。以均方误差、均方根误差、平均绝对误差与决定系数为评价指标进行模型预测性能对比,结果表明,所提出模型的各项评价指标均优于其他模型,能够有效挖掘多浮标的空间信息,克服了传统方法在海水溶解氧浓度预测中存在的精度低、无法灵活利用历史空间数据、训练稳定性差和速度慢等不足,可为海洋水质监测及预测提供重要的技术支撑。Dissolved oxygen concentration is one of the important indexes to measure seawater quality.In order to grasp the change of seawater quality in time and reduce the risk and loss of seawater pollution,it is very important to establish the prediction mechanism of marine water quality parameters.Therefore,this paper proposes a prediction model of dissolved oxygen concentration in seawater based on temporal and spatial information fusion of buoy Networks and Generative Adversarial Networks(GAN),which aims to integrate topological information of buoy networks in the monitoring area and realize multi-feature fusion of buoy sensors.The model uses the Graph Attention Mechanism(GAT)to mine the influence of different nearest neighbor points on the target node and calculate the weights of the adjacent nodes,so as to capture the spatio-temporal characteristics of the buoy data.The twohead attention mechanism and the two-time-scale Update Rule(TTUR)were used to optimize the GAN prediction network and the network training process,improve the training speed balance of the generated adversarial network,and improve the fitting effect of the generator network.The mean squared error,root mean squared error,mean absolute error and R-Square are used as evaluation indexes to compare the model prediction performance.The results show that the evaluation indexes of the proposed model are superior to other models,and can effectively mine the spatial information of multiple buoys.It overcomes the shortcomings of traditional methods in the prediction of dissolved oxygen concentration in seawater,such as low accuracy,inability to flexibly use historical spatial data,poor training stability and slow speed,and can provide important technical support for marine water quality monitoring and prediction.

关 键 词：溶解氧浓度预测 空间多特征融合 GAT GAN TTUR 

分 类 号：X834[环境科学与工程—环境工程]