Abstract
Reliable and accurate prediction of time series plays a crucial role in the maritime industry, impacting economic investment, transportation planning, and port design. However, maritime time series often exhibit complex, nonlinear, and non-stationary properties. To address these challenges, we propose a three-step prediction framework that utilizes Empirical Mode Decomposition (EMD) and Ensemble EMD (EEMD) to separate signals into high- and low-frequency components. While low-frequency components are predicted using traditional neural networks (NN), high-frequency components—which lack strong mathematical regularity—are handled through a novel similarity grouping strategy. By using Dynamic Time Warping (DTW) to identify and group historical segments with high geometric similarity, we create more effective training datasets for the NN. Experiments on port cargo throughput and vessel traffic flow demonstrate that this regrouping strategy significantly enhances prediction accuracy and robustness.
Introduction
Predicting maritime trends, such as port cargo throughput (PCT) and vessel traffic flow (VTF), is essential for operational efficiency and strategic planning. Despite their importance, these time series are notoriously difficult to forecast due to their non-stationary nature. Traditional neural network approaches often struggle with the "noise" or high-frequency volatility inherent in maritime data. Standard models typically treat all historical data with equal weight or rely solely on chronological proximity, failing to exploit the inherent self-similarities hidden within chaotic signal fluctuations. This paper introduces a framework that leverages frequency decomposition and shape-aware similarity grouping to overcome these limitations.
Methodology: The Three-Step Framework
The proposed methodology follows a structured three-step process designed to isolate and specifically target the different behavioral patterns within a maritime signal.
1. Decomposition
The original time series is first decomposed using Empirical Mode Decomposition (EMD) or Ensemble EMD (EEMD). This process breaks the complex signal into several Intrinsic Mode Functions (IMFs). These IMFs represent different frequency scales, ranging from stable, slow-moving trends (low-frequency) to rapid, volatile fluctuations (high-frequency).
2. Similarity Grouping-Guided Prediction
The framework treats components differently based on their frequency:
- Low-Frequency Components: These exhibit strong regularity and are predicted directly using standard NN architectures like Back-Propagation Neural Networks (BPNN) or Generalized Regression Neural Networks (GRNN).
- High-Frequency Components: These are the primary challenge. The signal is divided into overlapping segments, and Dynamic Time Warping (DTW) is employed to find historical segments that share a similar geometric shape with the current reference segment.
Unlike standard Euclidean distance, which compares points at fixed time intervals, DTW is "shape-aware." It can align patterns that are similar but slightly shifted or stretched in time, as shown in the comparison below.
By training the NN on these "geometrically similar" historical patterns rather than just the most recent chronological data, the model learns the underlying recurring dynamics of the high-frequency noise.
3. Combination
In the final step, the individual forecasts for each high- and low-frequency component are integrated (summed) to produce the final time series prediction.
Experimental Evaluation
The framework was validated using two distinct maritime datasets:
- Port Cargo Throughput (PCT): Monthly data from 18 major Chinese ports (1996–2016).
- Vessel Traffic Flow (VTF): Real-time AIS data from the Wuhan Yangtze Great Bridge, sampled at 3-hour intervals.
The EEMD+DTW+NN approach consistently outperformed standalone NNs and standard EMD+NN combinations across various architectures (WNN, ENN, FNN, BPNN, and GRNN). For vessel traffic flow, the proposed method achieved a Relative Error (RE) as low as 0.02 to 0.03, a significant improvement over the 0.13 to 0.15 range seen in standalone models. Furthermore, the similarity grouping strategy improved robustness, showing a much lower standard deviation in error across multiple runs, which mitigates the risks associated with random neural network initialization.
Conclusion
By combining frequency decomposition with shape-based similarity grouping, the proposed framework effectively addresses the non-linear and non-stationary challenges of maritime time series. The use of DTW to curate training datasets for high-frequency components represents a significant shift from chronological modeling to pattern-based modeling. While the computational cost is higher than simpler models, the gains in accuracy and reliability make it a powerful tool for maritime logistics and beyond.
Summary updated