空天地海一体化海洋环境数据多步预测

辅助海洋管理决策的多步预测预警意义重大且极具挑战性。实时、稳定、高效的广域海洋环境数据获取是保障多步预测性能的前提，未来6G空天地海一体化网络部署将有效地提升海洋分布式态势感知能力，提供高质量的数据支撑。众所周知，多步长模式下数据间的时序依赖性被极大地弱化，对此，本文提出了基于多阶段特征学习的海洋环境数据多步预测模型。结构上，该模型主要包括卷积神经网络（Convolutional Neural Networks,CNN）、优化组合的长短期记忆网络（Long Short-Term Memory Network,LSTM）和全连接层。这里，CNN用于提取海洋环境数据的细粒度特征，而基于粒子群优化的多LSTM组合方法，可以有效地挖掘数据间的时序依赖关系（粗粒度特征）。实验结果表明该模型的预测性能明显优于CNN、LSTM以及门控制循环单元，并进行了统计验证。

Multi-step Prediction of Ocean Environmental Data Based on Space-Air-Ground-Sea Integration

? ?Multi-step prediction and early warning to assist ocean management and decision making were significant and challenging. Real-time， stable and efficient acquisition of wide-area ocean environment data was the premise to guarantee multi-step prediction performance. The air-space-ground-sea integrated 6G network deployment could effectively improve distributed perception capacity， and provide high-quality data support. As we all know， the temporal dependence between data in multi-step mode was greatly weakened. The paper proposed a multi-step prediction model of ocean environmental data based on multi-stage feature learning. In structure， it mainly included a convolutional neural networks （CNN）， optimally combined LSTMs （OLSTMs）， and a fully connected layer. Here， CNN was used to extract fine-grained features between data， while the multi-LSTM combination method based on particle swarm optimization could effectively capture the temporal dependencies between data （coarse-grained features）. Experimental results show that the prediction performance of our proposal is significantly better than CNN， LSTM， and gated recurrent unit （GRU）， and some statistical verifications are carried out.