Marine weather prediction method and system based on space-time hybrid convolution attention

The invention discloses a multivariate space-time prediction method and system for marine weather forecast. The method comprises the following steps: S1, making a sea surface weather data set; s2, constructing an StHCFFormer model, training the StHCFFormer model by using the sea surface weather data set as a sample, and carrying out training and optimization; and S3, based on the trained StHCFFormer model, inputting sea surface weather data to carry out sea surface weather prediction. According to the method, global representation and local features of a space are deeply explored by using a convolution-space self-attention mixing mechanism, a time dependency relationship of time sequence data is captured by using time sequence self-attention, and channel-level scaling is performed through a local feature extraction module to capture dynamic correlation of multiple variables; and meanwhile, uncertainty related loss is introduced to dynamically adjust the multi-task weight, so that the globally optimal solution