Machine Learning Approaches for Traffic Volume Forecasting: A Case Study of the Moroccan Highway Network

Abderrahim Khalifa, Younes Idsouguou, Loubna Benabbou, Mourad ZirariarXiv 2017 · arXiv

TL;DR

This paper explores the application of various machine learning techniques for forecasting traffic volume on the Moroccan highway network. By comparing traditional statistical methods with algorithms like Random Forest, Extra Trees, Multi-layer Perceptron, and LSTM, the authors demonstrate the potential of ML to improve prediction accuracy in a developing country context.

What Problem Does This Paper Address?

Accurate traffic volume forecasting is crucial for effective transportation planning, network upgrades, congestion management, and the development of Intelligent Transportation Systems (ITS). While traditional statistical models like ARIMA have been used, they often struggle with the complexities of real-world traffic data, such as high dimensionality, non-linear relationships, and spatiotemporal variations. This paper addresses the challenge of applying and evaluating modern machine learning approaches for traffic forecasting, specifically within the context of a developing country's highway network, where data quality and infrastructure might present unique hurdles.

Methods

The authors adopted a two-pronged approach:

  1. Statistical Analysis: Initially, they employed statistical methods for data exploration and wrangling. A beta model was used to understand traffic behavior.
  2. Machine Learning: The core of the study involved applying several machine learning algorithms:
  3. Random Forest: A supervised classification algorithm that builds an ensemble of decision trees.
  4. Extra Trees (Extremely Randomized Trees): Similar to Random Forest but with more randomness in split point selection.
  5. Multi-layer Perceptron (MLP): A type of artificial neural network consisting of interconnected neurons organized in layers.
  6. AdaBoost: An ensemble learning method that combines multiple weak learners to create a strong learner.
  7. Long Short-Term Memory (LSTM) Networks: A specialized type of Recurrent Neural Network (RNN) designed to capture long-term dependencies in sequential data, making them well-suited for time series forecasting.

The paper also details data preprocessing steps, including handling anomalies and creating a suitable learning matrix by expanding the date attribute into various explanatory variables and dummy variables for days of the week.

Datasets

The study utilized daily historical traffic volume data from the Moroccan highway network, spanning from 2008 to 2015, with predictions made for the subsequent two years. The data included three classes of vehicles: light and small vehicles (TC1), medium vehicles (TC2), and heavy vehicles (TC3). The application section focuses on predictions for the "light vehicles" class at a specific toll station.

Key Insights and Ideas

  • Limitations of Traditional Models: The paper highlights that traditional statistical models (like Holt-Winters and ARIMA) often provide poor performance on real-world highway traffic data due to their inability to capture complex, non-seasonal variations.
  • ML's Superiority: Machine learning algorithms, particularly those that can handle non-linearities and complex patterns, show promise for improving traffic volume forecasting accuracy.
  • Data Preprocessing is Key: Effective data cleaning, anomaly detection, and feature engineering (e.g., creating dummy variables for days of the week) are critical for enabling ML models to learn effectively.
  • LSTM's Potential for Time Series: LSTMs, with their inherent ability to remember long-term dependencies, are identified as a strong candidate for time series forecasting tasks like traffic volume prediction.
  • Ensemble Methods for Improvement: The authors suggest that combining different models using ensemble techniques (Bagging, Boosting, Stacking) could further enhance prediction performance.

Key Results

The paper presents performance summaries using the Symmetric Mean Absolute Percentage Error (SMAPE) metric. While specific absolute values are provided, the relative performance is more telling:

  • Extra Trees achieved the lowest SMAPE at 12.8%.
  • Random Forest followed closely with 13.4%.
  • AdaBoost achieved 15%.
  • MLP had a SMAPE of 17%.
  • LSTM had the highest SMAPE at 19%.

Despite LSTM's higher error in this specific comparison, the paper emphasizes its suitability for time series and suggests that further tuning or different configurations might improve its performance. Visualizations of forecasts from MLP, Extra Trees, Random Forest, and AdaBoost are provided, showing how well they track the measured traffic, with Extra Trees and Random Forest appearing to offer the closest fits.

Limitations and Future Work

Limitations: * The study focused on a single toll station and the "light vehicles" class for the application section, which might limit the generalizability of findings to the entire network or other vehicle types. * The LSTM model, while promising, did not achieve the best performance in this particular experiment, suggesting potential for further optimization.

Future Work: * The authors explicitly mention the potential for improving performance by using "Ensemble Methods" (Bagging, Boosting, Stacking) to combine the strengths of different models. * Further exploration and tuning of LSTM parameters and architecture could lead to better results. * Extending the analysis to other vehicle classes and more toll stations across the Moroccan highway network would provide a more comprehensive understanding.

Summary updated

BibTeX
@misc{khalifa2017machine,
  author = {Abderrahim Khalifa and Younes Idsouguou and Loubna Benabbou and Mourad Zirari},
  title = {{Machine Learning Approaches for Traffic Volume Forecasting: A Case Study of the Moroccan Highway Network}},
  year = {2017},
  eprint = {1711.06779},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1711.06779},
}
Figure