Abstract
Currently, deep-learning-based prediction of Significant Wave Height (SWH) is mostly performed for a single location in the ocean or simply relies on a single factor (SF). Such approaches have the disadvantage of lacking spatial correlations or dynamic complexity, leading to an inevitable growth of the prediction error with time. Here, attempting a solution, we develop a Multi-Factor (MF) data-driven 2D SWH prediction model for the Bohai, Yellow, and East China Seas (BYECS). Our model is developed based on a multi-channel PredRNN algorithm that is an improved deep-learning calculation of the ConvLSTM. In our model, the MF of historical SWH, 10 m surface winds, ocean surface currents, bathymetries, and open boundaries are used to predict 2D SWH in the next 1-72h. Our modeled SWHs show the correlation coefficients as 0.98, 0.90, and 0.87 for the next 6h, 24h, and 72h, respectively. According to the ablation experiments, winds are the dominant factor in the MF model and the memory-decoupling module is the key improvement of the PredRNN compared to the ConvLSTM. Furthermore, when the historical SWH is excluded from the input, the correlation coefficients remain around 0.95 in the 1-72h prediction due to the elimination of the error accumulation. It was worse than the MF-PredRNN with the historical SWH before 10h but better than it after 10h. Overall, for the prediction of SWH in the BYECS, our MF-PredRNN-based 2D SWH prediction model significantly improves the accuracy and extends the effective prediction time length.
Introduction
Predicting Significant Wave Height (SWH)—the average height of the highest one-third of waves—is a critical task for maritime safety, offshore engineering, and disaster prevention. Traditionally, this prediction relies on numerical "phase-averaging" models like SWAN or WaveWatch III. While these models are physically grounded and accurate, they are computationally intensive and often exhibit systematic biases.
Deep Learning (DL) has emerged as a promising, faster alternative. However, existing DL models for SWH often suffer from spatial isolation (predicting for single points) and factor simplicity (relying solely on historical wave data). These limitations lead to rapid error accumulation over time. This study, published in Frontiers in Marine Science, introduces a Multi-Factor (MF) 2D SWH prediction model specifically designed for the Bohai, Yellow, and East China Seas (BYECS) to address these challenges.
Methodology
The core of the proposed model is the PredRNN (Predictive Recurrent Neural Network), which evolves from the standard Convolutional LSTM (ConvLSTM). PredRNN incorporates two major innovations:
- Spatio-Temporal Memory Flow (STMF): This allows memory to flow both horizontally through time and vertically through network layers in a zigzag pattern, preserving spatial details that are often lost in deeper standard LSTMs.
- Memory Decoupling (MD): This module separates long-term motion patterns from short-term variations, enabling the network to capture complex wave dynamics more effectively.
The model adopts a Multi-Factor approach, integrating historical SWH, 10m surface winds, ocean surface currents, and bathymetry data.
The study utilizes the ECMWF ERA5 reanalysis dataset (2011–2019) for SWH and wind data, while ocean currents are derived from the Regional Ocean Modeling System (ROMS). All data were interpolated to a 0.5° × 0.5° spatial resolution.
Results and Discussion
The MF-PredRNN model demonstrates high predictive accuracy across various lead times. Correlation coefficients (CC) reached 0.98 for 6-hour forecasts, 0.90 for 24-hour forecasts, and 0.87 for 72-hour forecasts. Compared to standard ConvLSTM models, MF-PredRNN significantly reduced the Mean Absolute Error (MAE).
The Role of Physical Factors
Ablation experiments revealed that surface winds are the dominant driver in the model. Removing wind data caused the prediction error to grow five times faster. Additionally, the Memory Decoupling module was confirmed as the most critical algorithmic improvement over ConvLSTM for distinguishing between steady-state conditions and sudden changes.
The Historical SWH Paradox
An interesting finding was the impact of historical SWH on long-term forecasts. While including past SWH data improves short-term accuracy (under 10 hours), it contributes to error accumulation in longer-term forecasts. A model variant excluding historical SWH maintained a stable CC of approximately 0.95 for the 1-72 hour window, outperforming the full MF-PredRNN after the 10-hour mark by avoiding the feedback of its own previous errors.
Conclusion
The MF-PredRNN-based model represents a significant advancement in 2D SWH prediction for the BYECS region. By integrating multiple physical factors and utilizing advanced recurrent neural network architectures, the model achieves higher accuracy and extends the effective forecast window up to 72 hours. Future research will explore the integration of attention mechanisms to further enhance spatial dependency capture and refine predictions in complex coastal environments.
Summary updated