Abstract
This study develops a hybrid forecasting model combining Grey system theory with Radial Basis Function (RBF) neural networks to predict container sea-rail transport volumes at Ningbo Port. By leveraging the strengths of both methods, the Grey-RBF model addresses the challenges of small-sample, high-uncertainty data. Analysis of historical data from 2009 to 2014 demonstrates that the combined model achieves superior accuracy compared to standalone GM(1,1) and RBF models, providing a robust tool for port infrastructure planning and logistics optimization.
The Challenge of Sea-Rail Logistics
Ningbo Port stands as one of the world’s most critical maritime hubs. However, as of 2014, its sea-rail intermodal transport—the direct transfer of containers from ships to trains—accounted for less than 1% of total throughput. Expanding this link is essential for reducing logistics costs and supporting inland economic development under China's "One Belt, One Road" initiative.
Forecasting these volumes is inherently difficult due to "noisy" data influenced by complex factors like regional GDP, railway freight capacity, and international trade policies. Traditional linear models struggle with data randomness, while standard artificial neural networks often require larger datasets than are currently available for these emerging transport routes.
A Hybrid Forecasting Framework
To address these limitations, researchers proposed a Grey-RBF Neural Network Combined Forecasting Model. This approach integrates two complementary mathematical frameworks:
- Grey Model (GM 1,1): Ideal for "small sample" problems with high uncertainty. It uses an Accumulated Generating Operation (AGO) to transform random data sequences into more regular, exponential trends.
- RBF Neural Network: A three-layer feed-forward network known for strong nonlinear mapping and fast learning, excelling at fitting complex curves.
The innovation lies in using the Grey AGO as a pre-processing step to "smooth" raw data and reduce randomness before training the RBF network.
Data and Variables
The study utilized historical data from 2009 to 2014, sourced from the Statistical Yearbook of Ningbo City and official economic bulletins. The model considers five primary variables:
- Container sea-rail intermodal transport volume (Target)
- Total container throughput
- Regional GDP
- Total goods sent by rail
- Import and export volume
Performance and Predictive Insights
The hybrid model was benchmarked against standalone GM(1,1) and RBF models. The integration of Grey theory allowed the neural network to achieve high precision even with limited historical data.
Key results include:
- Superior Accuracy: The Grey-RBF model achieved an average relative error of 3.37%, outperforming the standalone RBF (3.85%) and GM(1,1) (6.93%).
- High Predictive Power: The model successfully captured the "low base, fast growth" characteristic of the data, forecasting a continued surge in volume to approximately 287,497 TEU by 2016.
- Stability: The combined model maintained lower maximum errors compared to the single Grey model, which saw errors exceed 10% in some years.
Conclusion and Future Work
While highly accurate, the model's precision could be further enhanced by incorporating additional variables such as fuel prices and government subsidies. Future research may also explore dynamic weighting to account for the changing impact of factors like GDP versus rail infrastructure over time. Optimizing the RBF network's hidden layer units remains a key area for increasing precision as more data becomes available.
Summary updated