Abstract
A multitude of factors can influence road traffic accidents, often resulting in information overlap that can obscure inherent traffic characteristics and patterns. To enhance forecasting precision, this study introduces a hybrid model combining Principal Component Analysis (PCA) with Back-Propagation (BP) neural networks. Comparative analysis demonstrates that the PCA-BP model achieves superior accuracy over traditional forecasting methods.
Introduction: The Challenge of Information Overlap
Predicting traffic accidents is complex due to the intricate interplay of socioeconomic variables, including regional GDP, population density, and traffic enforcement levels. A significant technical barrier in modeling these events is "information overlap" or multicollinearity. Many indicators—such as economic growth, vehicle ownership, and infrastructure expansion—are highly correlated. Feeding these redundant variables directly into a standard neural network introduces noise that can degrade forecasting performance. This research addresses the need for a simplified input structure that retains critical information while eliminating statistical redundancy.
Methodology: A Two-Stage Hybrid Approach
The proposed forecasting pipeline consists of two primary stages:
- Dimensionality Reduction via PCA: Instead of utilizing 13 raw variables, PCA transforms the dataset into uncorrelated "principal components." The study found that just two principal components could capture over 91.6% of the total variance, effectively stripping away redundant data.
- Non-linear Prediction via BP Neural Network: A Back-Propagation (BP) neural network was employed for the final forecast. The researchers compared a standard model using all 13 raw indicators against the optimized PCA-BP model using only the 2 principal components. This reduction in the input layer simplifies the network architecture, enhancing efficiency and preventing the model from over-fitting to noise.
The number of hidden layer neurons was optimized using the formula $n_1 = \sqrt{n+m} + a$, with 12 neurons ultimately selected for the PCA-based model to provide the most stable and accurate results.
Datasets and Indicators
The study utilized a decade of statistical data (1995–2004) from "City A," covering 13 indicators across economic, demographic, and traffic-specific categories. These included GDP, urban road length, vehicle ownership, and the number of licensed drivers. Data from 1995 to 2002 was used for training, while 2003 and 2004 served as the evaluation period.
Key Results and Performance Analysis
The PCA-BP model significantly outperformed the standard BP model. Performance was measured using Relative Percentage Error (RPE). For the year 2003, the standard model had an error of 2.56%, while the PCA-BP model achieved a remarkably low error of 0.10%—predicting 1,173 accidents against an actual count of 1,172. In 2004, the PCA-BP model maintained its lead with an error of 1.33% compared to the standard model's 2.22%.
Conclusion and Future Directions
By collapsing 13 variables into 2 principal components, the researchers successfully removed the "statistical fog" of correlated data, leading to a more robust forecasting tool. While limited by a relatively small sample size and geographic focus on a single city, the results suggest that this hybrid approach is highly effective for complex forecasting problems involving multiple interdependent factors. Future research could explore the model's performance on larger datasets and its applicability to other domains like economic or environmental trend analysis.
Summary updated