⚡ electrical engineering

Hybrid VMD-2TCDNet Model for Accurate Significant Wave Height Forecasting Using Meteorological Predictors

This paper proposes a novel hybrid VMD-2TCDNet model that integrates Variational Mode Decomposition with a Two-dimensional Temporal Convolutional Dilated Network to significantly improve the accuracy of significant wave height forecasting across major Indian ports compared to state-of-the-art deep learning baselines.

Original authors: Rituparna Saud, Mihirkumar Patel

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rituparna Saud, Mihirkumar Patel

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The ocean is a restless, powerful force, and for the millions of people who live along its edges, understanding its moods is a matter of safety and survival. In places like India, where the coastline stretches for thousands of kilometers, the rhythm of the waves dictates the daily lives of fishermen, the safety of shipping routes, and the stability of coastal towns. For decades, scientists have tried to predict how high the waves will rise using complex computer models based on the laws of physics. These models are powerful, but they are also heavy and slow, requiring vast amounts of data and computing power to run, which often makes them too sluggish for real-time decisions. In recent years, a different approach has emerged: using artificial intelligence to learn from the data itself, spotting patterns that human-made equations might miss. However, ocean waves are notoriously difficult to predict because their behavior is chaotic and changes constantly, making it hard for even advanced computer programs to look far into the future without losing accuracy.

A team of researchers from Gauhati University and the National Institute of Technology Silchar has developed a new method to tackle this challenge, aiming to give forecasters a sharper, more reliable tool for predicting significant wave height, which is the average height of the highest third of the waves. Their work, published in a research article, introduces a hybrid system that combines two distinct techniques to handle the messy, unpredictable nature of ocean data. First, they use a signal processing method called Variational Mode Decomposition. Think of this as a sophisticated way of untangling a knotted rope; it takes the complex, jumbled history of wave data and breaks it down into simpler, smoother strands, each representing a different underlying rhythm or pattern. By separating the noise from the signal, the researchers make the data much easier for a computer to understand. Once the data is cleaned and organized, they feed it into a specialized deep learning network they designed, which they call a two-dimensional temporal convolutional dilated network. This network is built to look at the data in a unique way, scanning both the timeline of the waves and the relationships between different weather factors, such as wind speed and air pressure, all at once.

The researchers tested their new system against data collected from six major ports along the Indian coast, including Mumbai, Mangaluru, and Chennai. They compared their results with several other leading artificial intelligence models that are currently used for weather forecasting, such as long short-term memory networks and transformer models. The results showed that their combined approach was consistently more accurate. For example, when predicting wave heights 24 hours in advance, their model produced errors as small as 0.0060 meters in Cochin and 0.0087 meters in Mumbai, outperforming the other models in most locations. The system proved particularly good at looking further ahead, maintaining high accuracy even when forecasting 72 hours into the future, a task where many other models begin to struggle and lose precision. The researchers found that by breaking the data down first and then using their specialized network to learn from the simplified pieces, they could capture both the quick, short-term shifts in the waves and the slower, long-term trends that drive them.

This success is not just a theoretical improvement; it has direct implications for the people who depend on the sea. The study highlights that accurate, long-range forecasts can help fishermen choose safer times to go out, potentially saving lives and protecting livelihoods. It also aids in managing coastal infrastructure and planning for storms. The researchers noted that while their model is highly effective, it is not a perfect crystal ball; its performance relies heavily on the quality of the data it receives, and it still faces challenges when predicting extremely rare or unusual weather events that have never been seen before. They also pointed out that the inner workings of such advanced artificial intelligence can be difficult to interpret, meaning it is sometimes hard to explain exactly why the computer made a specific prediction. Despite these limitations, the study demonstrates that combining signal decomposition with modern deep learning creates a robust framework for understanding the ocean. By turning chaotic, non-stationary data into clear, learnable patterns, this new method offers a significant step forward in making coastal regions safer and more resilient against the unpredictable power of the sea.

Technical Summary: Hybrid VMD-2TCDNet Model for Accurate Significant Wave Height Forecasting

Problem Statement
Accurate forecasting of Significant Wave Height (SWH) is critical for maritime navigation, naval operations, and coastal management, particularly for nations with extensive coastlines like India, where the fishing industry and coastal economy rely heavily on sea conditions. Traditional physics-based numerical models (e.g., SWAN, WAM, Wave Watch III) offer high precision but are computationally intensive, requiring extensive environmental input data and high-resolution simulations that hinder real-time application. Conversely, conventional Machine Learning (ML) and early Deep Learning (DL) approaches, such as ARIMA, LSTM, and standard CNNs, struggle with the strong nonlinearity, non-stationarity, and complex spatial-temporal interdependencies inherent in ocean wave dynamics. These models often fail to capture long-term dependencies effectively or degrade in accuracy over extended forecasting horizons, limiting their utility for operational decision-making.

Methodology
The authors propose a novel hybrid architecture, VMD-2TCDNet, which integrates Variational Mode Decomposition (VMD) with a Two-dimensional Temporal Convolutional Dilated Network (2TCDNet) to address the challenges of long-term, multivariate time series forecasting.

Data Preparation: The study utilizes hourly ERA5 reanalysis data (2021–2024) covering six major Indian ports. The input features include SWH (target), Wind Speed (WS), and Sea Level Pressure (SLP). WS is derived from zonal and meridional wind components.
Variational Mode Decomposition (VMD): As a preprocessing step, VMD decomposes the non-stationary original time series into a set of Intrinsic Mode Functions (IMFs). This process transforms the complex signal into simpler, more stationary sub-series, effectively reducing noise and isolating distinct frequency components to facilitate better feature extraction.
2TCDNet Architecture: The core forecasting engine is the 2TCDNet, designed specifically for multivariate inputs.
- 2D Convolutional Layers: Unlike 1D CNNs, this model employs 2D convolutional kernels that process both the temporal dimension and the feature dimension simultaneously. This allows the model to capture interdependencies between meteorological variables (SWH, WS, SLP) and temporal patterns.
- Causal and Dilated Convolutions: The network utilizes causal convolutions to prevent future information leakage and dilated convolutions with varying dilation rates (1, 2, 4, 8) to expand the receptive field. This enables the capture of both short-term fluctuations and long-term dependencies without increasing computational complexity or requiring recurrent layers.
- Temporal Dense Blocks: Stacked blocks with dense connectivity concatenate outputs from previous layers, promoting gradient flow and feature reuse.
- Direct Multi-Step (DMS) Forecasting: The model employs a DMS strategy, directly mapping historical data to future time steps (24, 48, and 72 hours ahead) to mitigate error accumulation common in iterated multi-step approaches.
Optimization: Hyperparameters were tuned using Bayesian optimization, with GELU activation, a learning rate of ~4.61e-4, and a batch size of 32. The model is trained using Mean Squared Error (MSE) loss.

Key Contributions

Novel Hybrid Framework: The integration of VMD for signal decomposition with 2TCDNet for forecasting creates a robust pipeline for handling non-stationary, multivariate climate data.
2D Temporal Convolutional Dilated Network: The proposal of 2TCDNet introduces a 2D convolutional approach for multivariate time series, utilizing causal and dilated convolutions to better model the spatial-temporal relationships between variables compared to traditional univariate or 1D approaches.
Long-Horizon Performance: The study focuses on hourly long-term predictions (up to 72 steps), addressing a gap in existing literature that often prioritizes short-term or daily forecasts.
Comprehensive Evaluation: The model is rigorously tested across six distinct geographical locations along the Indian coast, demonstrating robustness across diverse oceanic conditions.

Results
Extensive evaluations demonstrate that VMD-2TCDNet outperforms state-of-the-art baselines, including LSTM, GRU, CNN-LSTM, TCN, and Transformer models.

Accuracy: The proposed model consistently achieved the lowest Root Mean Square Error (RMSE) across most locations for 24-step forecasts. Notable RMSE values include 0.0087 (Mumbai), 0.0069 (Mangaluru), and 0.0060 (Cochin).
Long-Term Stability: While prediction errors naturally increased with longer horizons (48 and 72 steps), the model maintained high accuracy (average ~90% $R^2$ ) and low bias, indicating strong capability in capturing long-range temporal dependencies.
Ablation Study: Removing key components (VMD, dense connectivity, or dilation) resulted in performance degradation, confirming that the decomposition step and the specific architectural choices (2D kernels, dilation depth) are critical for the model's success.
Comparative Performance: At Paradip, the model was competitive with TCN and significantly outperformed recurrent and Transformer-based models. In other locations, it consistently surpassed all baselines.

Significance and Claims
The paper claims that the VMD-2TCDNet framework represents a significant advancement in wave forecasting techniques by effectively combining signal decomposition with advanced deep learning. The authors assert that this approach enhances forecasting accuracy and robustness, offering a reliable tool for improving maritime safety and supporting sustainable fishing practices in coastal regions.

The study highlights that the model's ability to handle non-stationary data and capture complex interdependencies makes it suitable for broader applications in environmental forecasting, potentially extending to variables like sea surface temperature and air pressure. However, the authors modestly acknowledge limitations, noting that the model's performance is dependent on data quality and that its "black box" nature presents challenges for physical interpretability. They suggest future work should focus on integrating physical dynamics (e.g., wind forcing principles) and expanding data diversity to improve generalization for rare weather events.

More like this