Search

Ocean forecasting is critical for various applications and is essential for understanding air-sea interactions, which contribute to mitigating the impacts of extreme events. State-of-the-art ocean numerical forecasting systems can offer lead times of up to 10 days with a spatial resolution of 10 kilometers, although they are computationally expensive. While data-driven forecasting models have demonstrated considerable potential and speed, they often primarily focus on spatial variations while neglecting temporal dynamics. This paper presents TSformer, a novel non-autoregressive spatiotemporal transformer designed for medium-range ocean eddy-resolving forecasting, enabling forecasts of up to 30 days in advance. We introduce an innovative hierarchical U-Net encoder-decoder architecture based on 3D Swin Transformer blocks, which extends the scope of local attention computation from spatial to spatiotemporal contexts to reduce accumulation errors. TSformer is trained on 28 years of homogeneous, high-dimensional 3D ocean reanalysis datasets, supplemented by three 2D remote sensing datasets for surface forcing. Based on the near-real-time operational forecast results from 2023, comparative performance assessments against in situ profiles and satellite observation data indicate that, TSformer exhibits forecast performance comparable to leading numerical ocean forecasting models while being orders of magnitude faster. Unlike autoregressive models, TSformer maintains 3D consistency in physical motion, ensuring long-term coherence and stability in extended forecasts. Furthermore, the TSformer model, which incorporates surface auxiliary observational data, effectively simulates the vertical cooling and mixing effects induced by Super Typhoon Saola., Comment: 37 pages,12figures,1 table

Mesoscale eddies are of utmost importance in understanding ocean dynamics and the transport of heat, salt, and nutrients. Accurate representation of these eddies in ocean models is essential for improving model predictions. However, accurately representing these mesoscale features in numerical models is challenging due to their relatively small size. In this study, we propose a convolutional neural network (CNN) that combines data-driven techniques with physical principles to develop a robust and interpretable parameterization scheme for mesoscale eddies in ocean modeling. We first analyze a high-resolution reanalysis dataset to extract subgrid eddy momentum and use machine learning algorithms to identify patterns and correlations. To ensure physical consistency, we have introduced conservation of momentum constraints in our CNN parameterization scheme through soft and hard constraints. The interpretability analysis illustrate that the pre-trained CNN parameterization shows promising results in accurately solving the resolved mean velocity at the local scale and effectively capturing the representation of unresolved subgrid turbulence processes at the global scale. Furthermore, to validate the CNN parameterization scheme offline, we conduct simulations using the MITgcm ocean model. A series of experiments is conducted to compare the performance of the model with the CNN parameterization scheme and high-resolution simulations. The offline validation using MITgcm simulations demonstrates the effectiveness of the CNN parameterization scheme in improving the representation of mesoscale eddies in the ocean model. Incorporating the CNN parameterization scheme leads to better agreement with high-resolution simulations and a more accurate representation of the kinetic energy spectra.

The accumulated remote sensing data of altimeters and scatterometers have provided a new opportunity to forecast the ocean states and improve the knowledge in ocean/atmosphere exchanges. Few previous studies have focused on sea level anomaly (SLA) multi-step forecasting by multivariate deep learning for different modalities. For this paper, a novel multimodal fusion approach named MMFnet is used for SLA multi-step forecasting in South China Sea (SCS). First, a grid forecasting network is trained by an improved Convolutional Long Short-Term Memory (ConvLSTM) network on daily multiple remote sensing data from 1993 to 2016. Then, an in-situ forecasting network is trained by an improved LSTM network, which is decomposed by the ensemble empirical mode decomposition (EEMD-LSTM), on real-time, in-situ and remote sensing data. Finally, the two single-modal networks are fused by an ocean data assimilation scheme. During the test period from 2017 to 2019, the average RMSE of the MMFnet (single-modal ConvLSTM) is 4.03 cm (4.51 cm), the 15th-day anomaly correlation coefficient is 0.78 (0.67), the performance of MMFnet is much higher than those of current state-of-the-art dynamical (HYCOM) and statistical (ConvLSTM, Persistence and daily Climatology) forecasting systems. Sensitivity experiments analysis indicates that, the MMFnet, which added CCMP SCAT products and OISST for SLA forecasting, has improved the forecast range over a week and can effectively produce 15-day SLA forecasting with reasonable accuracies.In an extension of the validation over the North Pacific Ocean, MMFnet can calculate the forecasting results in a few minutes, and we find good agreement in amplitude and distribution of SLA variability between MMFnet and other classical operational model products.