基于原型校正时空网络的多步海表面高度时空预测

MULTI-STEP SEA SURFACE HEIGHT SPATIOTEMPORAL PREDICTION BASED ON PROTOTYPE CORRECTED SPATIOTEMPORAL NETWORK

海表面高度作为一项重要的海洋观测指标,对海洋生态系统和气候变化研究具有重要影响。传统基于循环神经网络(recurrent neural network,RNN)的时空序列预测模型在更新记忆状态时,存在一个关键问题:旧的记忆状态会被立即刷新,模型无法有效地保留时间序列的长期依赖关系和变化趋势。这导致在多步海洋预测中,模型无法充分挖掘时间域的重要特征,从而使预测误差随着预测步长的增加而严重累积。为解决这个问题,设计一种原型校正时空网络(prototype corrected spatiotemporal network,PCST-Net)来实现准确的端到端多步海表面高度时空预测。PCST-Net采用基于RNN的网络结构,并设计记忆状态更新(memory state update,MSU)单元作为模型的核心构建单元。MSU单元利用原型校正模块(prototype correction module,PCM)来学习海表面高度样本的原型特征,从而提取时间域中的关键信息,并校正当前时间步的海表面高度高维特征,有效缓解多步预测中的严重误差累积问题。此外,提出一种多步信息输入策略,使模型能够从更广泛的时间步长中获得更全面准确的上下文信息,进而提高预测性能。通过对热带太平洋日平均海表面高度异常(sea surface height anomaly,SSHA)数据的多步时空预测验证了所提出的模型。结果表明:PCST-Net模型对未来5 d多步SSHA时空预测的平均均方根误差、平均绝对误差和皮尔逊相关系数分别为0.634 cm、0.488 cm和0.995。研究表明,PCST-Net模型可以准确地预测SSHA的时空变化趋势,这为多步海表面高度时空预测模型提供了一种可行性方法。

Sea surface height,as an important ocean observation indicator,has an important impact on marine ecosystem and climate change research.When updating the memory states of the traditional spatiotemporal sequence prediction models based on recurrent neural network(RNN),a key problem would occur:the old memory states will be refreshed immediately,the long-term dependence and change trend of the time series cannot be effectively saved,and the model cannot fully exploit important features in the temporal domain in multi-step marine prediction,resulting in serious accumulation of prediction errors as the prediction time step increases.To solve this problem,a prototype corrected spatiotemporal network(PCST-Net)was designed to achieve accurate end-to-end multi-step spatiotemporal prediction of sea surface height.The PCST-Net adopts an RNN-based network structure and designs a memory state update(MSU)cell as the core cell of the model.The MSU cell utilizes the prototype correction module(PCM)to learn the prototype features of the sea surface height samples,thereby extracting key information in the time domain and correcting the high-dimensional features of the sea surface height at the current time step,alleviating effectively the serious error accumulation problem in multi-step spatiotemporal prediction.In addition,a multi-step information input strategy was proposed to enable the model to obtain more comprehensive and accurate contextual information from a wider range of time steps,thereby improving prediction performance.The proposed model was validated through multi-step spatiotemporal predictions of daily mean sea surface height anomaly(SSHA)data in the tropical Pacific.The results show that the average ERMS(root mean square error),EMA(mean absolute error),and R(Pearson correlation coefficient)of the PCST-Net model for multi-step SSHA spatiotemporal prediction in the next 5 days are 0.634 cm,0.488 cm and 0.995,respectively.This study indicated that the PCST-Net model could accurately predict the spatiotemporal change trend of SSHA,and provided a feasible method for the multi-step sea surface height spatiotemporal prediction model.