Sea Surface Temperature Error Correction Method Based on Transformer

In light of the low accuracy of the numerical model system in predicting sea surface temperature (SST) on an hourly timescale, this paper thoroughly examines the spatiotemporal relationships inherent in the SST error data and delves deeply into the realm of machine learning, with a particular focus on the application of the Transformer deep learning model for this challenge. By engaging in profound learning of the error data from the past three days, the Transformer model offers an in-depth perception and understanding of the evolutionary patterns within the SST error data. Utilizing this advanced machine learning approach, the model successfully predicts the SST errors for the subsequent three days, significantly enhancing the prediction capability of the numerical model at the hourly scale. Experimental results further corroborate that the Transformer model demonstrates exceptional correction performance in machine learning across different time intervals such as 3 hours, 6 hours, and 24 hours.