基于Informer模型的海洋环境要素缺失数据时间序列预测

Based on the Informer Model Time Series Prediction of Missing Data in Marine Environmental Elements

针对数据采集设备及数据存储的局限性等多种原因导致海洋环境要素数据缺失的问题,提出了对海洋环境缺失要素时间序列预测的Informer模型。通过Informer模型的概率稀疏自注意力机制筛选海洋环境参数与特征变量之间的关系,采用卷积层和池化层对模型变量维度和网络参数进行自注意力蒸馏;由解码层生成式机制预测得到海洋环境要素的长序列数据。为提高模型预测结果的稳定性,将Informer模型编码器中概率系数自注意力机制的随机取样改为系统性取样;采用GeLu激活函数提高模型预测性能。将改进的Informer模型在ERA5数据集上进行实例验证,与LSTM、GRU、GRU-D模型相比,均方根误差降低了25.42%,平均绝对误差降低了26.53%,验证了在海洋环境要素长序列预测方面的有效性。

Data on elements of the marine environment missed for a variety of reasons,including limitations in data collection equipment and data storage.An Informer model was proposed for time series prediction of missing elements of the marine environment.The relationship between marine environmental parameters and feature variables was filtered by the probabilistic sparse self-attention mechanism of the Informer model.Convolutional and pooling layers were used to distil the model variable dimensions and network parameters with self-attention.The generative mechanism of the decoding layer could predict long sequence data of marine environmental elements.The random sampling of the probability coefficient self-attention mechanism in the Informer model encoder was changed to systematic sampling.The GeLu activation function was used to improve the prediction performance.The improved Informer model was validated on the ERA5 dataset.Compared with the LSTM,GRU,and GRU-D models,the root-mean-square error reduced 25.42%and the mean absolute error reduced 26.53%.