Predicting Ocean Temperature in High-Frequency Internal Wave Area with Physics-Guided Deep Learning: A Case Study from the South China Sea

Song Wu, Xiaojiang Zhang, Senliang Bao, Wei Dong, Senzhang Wang, Xiaoyong LiJournal of Marine Science and Engineering 2023 · DOI

TL;DR

This paper introduces a "physics-guided" deep learning model (PGSTCN) that leverages oceanic stable stratification to improve the accuracy and physical consistency of ocean temperature predictions, particularly in areas prone to internal waves. By integrating physical laws into the training process, the model outperforms traditional deep learning approaches like LSTM and ConvLSTM, especially when data is scarce.

What Problem Does This Paper Address?

Accurate prediction of ocean temperature is crucial for various applications, from climate forecasting to marine activities. Traditional physics-based numerical models, while grounded in physical laws, face significant challenges with parameterization, initial conditions, and computational cost. On the other hand, purely data-driven machine learning models often act as "black boxes," lacking interpretability and sometimes violating fundamental physical principles. This can lead to unreliable predictions, especially in dynamic environments like the South China Sea, which experiences frequent internal waves that cause localized temperature shifts. The authors aim to bridge this gap by developing a model that combines the strengths of both approaches: the predictive power of deep learning and the physical grounding of established oceanographic principles.

Methods

The core of the proposed solution is a Physics-Guided Spatio-Temporal Convolutional Neural Network (PGSTCN). This model builds upon the ConvLSTM architecture, which is adept at handling spatio-temporal data. The key innovations lie in:

  • Physics-Guided Loss Function: The model incorporates a "physics-guided" loss term that penalizes predictions violating the principle of oceanic stable stratification. This physical constraint is derived from the relationship between density and depth, and its negative values are used to guide the model's learning process.
  • Two Training Schemes:
    • Vector Output Scheme: This scheme trains the model layer by layer, focusing on time-domain characteristics.
  • Multiple-Parallel-Input and Multi-Step-Output (PIMO) Scheme: This approach processes data from multiple depth layers simultaneously as input features, allowing it to better capture inter-depth relationships and significantly reducing training time.
  • Pretraining: To address data scarcity, the model utilizes a pretraining step where weights from a trained model are used to initialize the new model, accelerating convergence and improving performance with limited data.
  • Interpolation: Missing data points (NaN values) are handled through mean imputation or discarding data layers, ensuring data quality for training.

Datasets

The study utilizes temperature data collected from the northeastern South China Sea (SCS) mooring array system. Specifically, data from the single-point dataset EW6 (117°26′, 21°16′) was used, spanning from June 5, 2014, to June 9, 2015. This dataset provides temperature measurements at various depths (85m-475m) with a 3-minute sampling interval. The data exhibits missing values at different depth layers, necessitating preprocessing.

Key Insights and Ideas

  • Physical Constraints Enhance Predictions: The most significant insight is that incorporating physical knowledge (like stable stratification) into the deep learning framework directly improves prediction accuracy and, crucially, the physical consistency of the results. This moves beyond "black box" models.
  • Spatio-Temporal Modeling is Key: The problem inherently involves both spatial (depth) and temporal (time) dependencies. The ConvLSTM backbone, combined with the PIMO scheme, effectively captures these complex relationships.
  • Trade-offs in Training Schemes: The vector output scheme offers higher accuracy but is computationally more intensive. The PIMO scheme provides a significant speedup in training time while maintaining competitive accuracy, making it more practical for real-world applications.
  • Data Scarcity is a Real Challenge: The analysis on reduced datasets highlights the vulnerability of purely data-driven models. Physics-guided approaches show greater resilience and better performance even with less data.

Key Results

  • Improved Accuracy: The PGSTCN model consistently outperformed traditional LSTM and ConvLSTM models in terms of Root Mean Square Error (RMSE) and accuracy across various historical time steps and prediction horizons.
  • Reduced Physical Inconsistency: The physics-guided loss function significantly reduced the proportion of physically inconsistent predictions, indicating that the model's outputs better adhere to oceanographic principles.
  • High Accuracy Achieved: In one-step predictions using the vector output scheme, PGSTCN achieved a maximum accuracy of 0.997. Even in multi-step predictions with the PIMO scheme, accuracy remained high, reaching up to 0.98.
  • Robustness to Data Volume: When data volume was reduced to 30%, PGSTCN's accuracy was still superior to LSTM with 100% data, demonstrating its effectiveness in data-scarce scenarios.
  • Training Efficiency: The PIMO scheme significantly reduced training time compared to the vector output scheme, making it more scalable.

Limitations and Future Work

  • Limited Data Scope: The study was limited to a single point (EW6) in the South China Sea and focused on intra-year prediction. The authors acknowledge this limitation and plan to:
  • Obtain more interannual data to enable interannual trend analysis.
  • Extend the data dimension to explore global sea temperature prediction.
  • PIMO Scheme Accuracy: While faster, the PIMO scheme's maximum accuracy (0.98) was slightly lower than the vector output scheme's (0.994). This trade-off between speed and peak accuracy is noted.
  • Exploring Other Physical Information: The authors suggest that integrating other relevant physical information beyond stable stratification could further enhance the model's performance and validity. They plan to explore and compare the effectiveness of different physical constraints.

Summary updated

Abstract

Higher-accuracy long-term ocean temperature prediction plays a critical role in ocean-related research fields and climate forecasting (e.g., oceanic internal waves and mesoscale eddies). The essential component of traditional physics-based numerical models for ocean temperature prediction is solving partial differential equations (PDEs), which has immense challenges in terms of parameterization, initial values, and boundary conditions setting. Moreover, the existing machine learning models for ocean temperature prediction have “black box” problems, and the influence of external dynamic factors is not considered. Moreover, it is hard to judge whether the model satisfies certain physical laws. In this paper, we propose a physics-guided spatio-temporal data analysis model based on the widely used ConvLSTM model to achieve long-term ocean temperature prediction and adopt two schemes to train the model in vector output and multiple parallel input and multi-step output. Meanwhile, considering the spatio-temporal correlation, physical information such as oceanic stable stratification is introduced to guide the model training. We evaluate our proposed approach on several popular deep learning models in different timesteps and data volumes in the northern coast of the South China Sea, where the frequent occurrence of internal waves leads to an intensity trend of a local transformation of sea temperature. The results show higher prediction accuracy compared with the traditional LSTM, and ConvLSTM models, and the introduction of physical laws can improve data utilization while enhancing the physical consistency of the model.
BibTeX
@article{wu2023predicting,
  author = {Song Wu and Xiaojiang Zhang and Senliang Bao and Wei Dong and Senzhang Wang and Xiaoyong Li},
  title = {{Predicting Ocean Temperature in High-Frequency Internal Wave Area with Physics-Guided Deep Learning: A Case Study from the South China Sea}},
  journal = {Journal of Marine Science and Engineering},
  year = {2023},
  doi = {10.3390/jmse11091728},
}
Figure