A hybrid forecasting framework based on support vector regression with a modified genetic algorithm and a random forest for traffic flow prediction

Lizong Zhang, Nawaf Alharbe, Guangchun Luo, Zhiyuan Yao, Ying LiTsinghua 2018 · DOI

Abstract

The ability to perform short-term traffic flow forecasting is a crucial component of intelligent transportation systems. However, accurate and reliable traffic flow forecasting is still a significant issue due to the complexity and variability of real traffic systems. To improve the accuracy of short-term traffic flow forecasting, this paper presents a novel hybrid prediction framework based on Support Vector Regression (SVR) that uses a Random Forest (RF) to select the most informative feature subset and an enhanced Genetic Algorithm (GA) with chaotic characteristics to identify the optimal forecasting model parameters. The framework is evaluated with real-world traffic data collected from eight sensors located near the I-605 interstate highway in California. Results show that the proposed RF-CGASVR model achieves better performance than other methods.

Introduction

Short-term traffic flow forecasting—predicting traffic conditions within the next 30 minutes—serves as the backbone of modern Intelligent Transportation Systems (ITS). Accurate forecasts enable proactive route guidance, optimized traffic light control, and effective congestion relief. However, traffic modeling is inherently difficult because flow is highly non-linear, variable, and influenced by complex spatiotemporal relationships between road segments.

While Support Vector Regression (SVR) is a powerful tool for handling non-linear data, it faces two primary challenges: feature overload and parameter sensitivity. Using excessive input features introduces noise and computational lag, while SVR performance depends heavily on three hyper-parameters ($C, \sigma, \text{ and } \epsilon$). Finding the optimal values for these parameters is often computationally expensive and prone to settling in sub-optimal "local optima."

Methodology: The RF-CGASVR Framework

To address these challenges, the researchers developed RF-CGASVR, a three-stage hybrid framework designed to prune redundant data and fine-tune model parameters.

Flowchart of the proposed hybrid forecasting framework.
  1. Feature Selection via Random Forest (RF): Instead of utilizing all available sensor data, the framework uses RF to rank feature importance. It iteratively removes the least significant features while monitoring the impact on model accuracy, ensuring only the most informative data points remain.
  2. Parameter Optimization via Chaotic Genetic Algorithm (CGA): Standard Genetic Algorithms (GAs) often suffer from "pre-maturation." This framework introduces Tent Mapping and Chaotic Mutation to inject mathematical chaos into the search process. This ensures the algorithm explores the entire solution space thoroughly, avoiding local optima.
  3. Prediction via SVR: The final model employs the optimized feature subset and the "fittest" parameters to predict traffic flow for the subsequent 5-minute interval.

Experimental Setup and Datasets

The framework was validated using real-world data from the Caltrans Performance Measurement System (PeMS), focusing on eight sensors along the I-605 interstate highway in California. The researchers evaluated two distinct spatial layouts to test the model's robustness.

Sensor layout for the crossroad scenario on the I-605 highway.

The study examined a "Straight Road" scenario (predicting flow based on upstream sites) and a "Crossroad" scenario (predicting flow at an intersection center based on four surrounding directions). Data was collected over 10 weekdays in March 2017, specifically targeting morning and evening rush hours to capture peak complexity.

Results and Discussion

The RF-CGASVR model was benchmarked against standard models including ARIMA, Back-Propagation Neural Networks (BPNN), and standard SVR variants.

Key Insights

  • The Power of Chaos: The use of chaotic characteristics (Tent Mapping) maintained "ergodicity" in the search space, preventing the model from settling for mediocre parameter sets.
  • Temporal Complexity: The model revealed that evening rush hour traffic is more complex than morning traffic. For the straight road scenario, the model required 8 features for the evening but only 4 for the morning, suggesting commuter behavior is more predictable at the start of the day.
  • Efficiency through Pruning: Reducing the feature set (e.g., from 20 to 8 in the crossroad case) actually increased accuracy by eliminating noise, proving that "more data" is not always better.

Performance Metrics

RF-CGASVR consistently achieved the lowest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). For instance, in the "Straight-M" (morning) dataset, RF-CGASVR achieved an RMSE of 49.81, significantly outperforming ARIMA (73.74) and BPNN (62.34). Qualitative analysis also showed that the model's residuals remained stable even during erratic traffic spikes.

Conclusion

The RF-CGASVR framework demonstrates superior accuracy and stability for short-term traffic forecasting. While the model excels at spatial relationships, future work will focus on integrating long-term patterns (e.g., weekly cycles) and external factors such as weather conditions and road accidents to further improve robustness in extreme conditions.

Summary updated

Abstract

The ability to perform short-term traffic flow forecasting is a crucial component of intelligent transportation systems. However, accurate and reliable traffic flow forecasting is still a significant issue due to the complexity and variability of real traffic systems. To improve the accuracy of short-term traffic flow forecasting, this paper presents a novel hybrid prediction framework based on Support Vector Regression (SVR) that uses a Random Forest (RF) to select the most informative feature subset and an enhanced Genetic Algorithm (GA) with chaotic characteristics to identify the optimal forecasting model parameters. The framework is evaluated with real-world traffic data collected from eight sensors located near the I-605 interstate highway in California. Results show that the proposed RF-CGASVR model achieves better performance than other methods.
BibTeX
@inproceedings{zhang2018hybrid,
  author = {Lizong Zhang and Nawaf Alharbe and Guangchun Luo and Zhiyuan Yao and Ying Li},
  title = {{A hybrid forecasting framework based on support vector regression with a modified genetic algorithm and a random forest for traffic flow prediction}},
  booktitle = {Tsinghua},
  year = {2018},
  doi = {10.26599/tst.2018.9010045},
}
Figure