Three-dimensional variational data assimilation method based on automatic encoder

The invention provides a three-dimensional variational data assimilation method based on an automatic encoder, and relates to the field of marine science and machine learning. The method is mainly divided into two main steps: feature compression based on an automatic encoder and three-dimensional variational data assimilation result prediction based on a minimum loss function. In the first step, a matrix involved in a data assimilation process is usually too large and cannot be easily operated or stored in a memory, so that the data must be compressed, and the data assimilation data is compressed through an auto-encoder, so that the state dimension is remarkably reduced, and calculation is facilitated; after data features are compressed, a data assimilation result used for analysis is calculated in a loss function minimization mode. 本发明提供了一种基于自动编码器的三维变分数据同化的方法，涉及海洋科学、机器学习领域。本发明将方法主要分为两个主要步骤：基于自动编码器的特征压缩和基于最小化损失函数的三维变分数据同化结果预测。在第一步中，由于数据同化过程中所涉及到的矩阵通常太大，无法轻易操作或者存储到内存中，因此必须在对数据进行压缩，为此本发明提出通过自编码器对数据同化数据进行压缩，以此显著