Abstract
The prediction of car ownership is fundamental for sustainable urban traffic development. This paper utilizes Principal Component Analysis (PCA) to identify key influencing factors in Wuhan City. By combining Genetic Algorithms (GA) with Neural Networks (NN), the study optimizes initial network weights to improve training speed and generalization while avoiding local minima. A specific example validates the model's predictive effectiveness.
Introduction
Accurate forecasting of car ownership is a cornerstone of sustainable urban planning. Underestimating growth leads to congestion and parking shortages, while overestimation results in wasted infrastructure resources. The complexity of car ownership stems from its non-linear relationship with various socio-economic factors such as GDP, fuel prices, and road infrastructure. While standard Back-Propagation (BP) Neural Networks can model these complexities, they often suffer from slow convergence and the tendency to settle in local minima due to random weight initialization.
Methodology
This research proposes a three-stage hybrid pipeline to overcome traditional modeling hurdles:
- Dimensionality Reduction (PCA): Eleven socio-economic indicators—including population, GDP, and road length—were compressed into three principal components. These components retained 98.844% of the original variance, significantly reducing input complexity and improving computational efficiency.
- Global Optimization (Genetic Algorithm): A GA was employed to search the solution space for optimal initial weights and thresholds for the neural network. By treating the weights as "chromosomes," the GA finds a globally strong starting position, preventing the network from getting stuck in sub-optimal solutions.
- Neural Network Training: A 3-layer BP Neural Network (3-15-1 architecture) was initialized with the GA-optimized weights and trained using an adaptive learning rate method.
Results and Analysis
The model was trained on Wuhan City data from 1995 to 2009 and tested on data from 2010 to 2012. The training process demonstrated high efficiency, reaching the performance goal of $10^{-3}$ in just 234 epochs, taking approximately 19.9 seconds.
The predictive performance was highly accurate when compared against actual historical data. For the test period of 2010–2012, the model achieved a Mean Absolute Error (MAE) of 1.49%. Notably, the absolute error for the year 2012 was as low as 0.35%, indicating that the model becomes increasingly robust as it processes more historical context.
Conclusion
The integration of PCA and GA significantly enhances the performance of neural networks in predicting urban car ownership. By simplifying input data and providing a globally optimized starting point for training, the hybrid model achieves superior accuracy and speed compared to traditional methods. Future work should consider larger datasets and the impact of external policy shocks, such as vehicle registration quotas, on ownership trends.
Summary updated