FCG-ASpredictor: An Approach for the Prediction of Average Speed of Road Segments with Floating Car GPS Data

Difeng Zhu, Guojiang Shen, Duanyang Liu, Jingjing Chen, Yijiang Zhang2019 · DOI

TL;DR

This paper introduces FCG-ASpredictor, a novel approach for predicting the average speed of road segments using floating car GPS data. By combining historical speed data, weather, and date attributes with a Recursive Least Squares-Extended Kalman Filter (RLS-EKF), the model achieves high accuracy in traffic congestion prediction, outperforming existing methods.

What Problem Does This Paper Address?

Accurate prediction of average speed (AS) on road segments is crucial for effective intelligent transportation systems (ITS), particularly for traffic congestion estimation and signal control. Traditional methods often rely on fixed detectors, which can suffer from low accuracy and limited coverage. While floating car GPS data offers advantages in flexibility, real-time performance, and coverage, existing prediction models using this data face challenges. These include:

  • Data Sufficiency: Some models struggle when there's insufficient historical data or when unexpected events occur.
  • Training Quality Dependence: Machine learning approaches like Neural Networks (NNs), Support Vector Machines (SVMs), and Support Vector Regression (SVR) are highly dependent on the quality and quantity of training data.
  • Spatiotemporal Dynamics: Capturing the complex interplay between traffic environment and spatiotemporal information remains a challenge.
  • External Factors: Accurately incorporating external factors like weather and date attributes, which can significantly influence traffic conditions, is difficult.

This paper aims to address these limitations by proposing a model that leverages floating car GPS data while robustly handling data scarcity and incorporating external influences.

Methods

The FCG-ASpredictor model is built upon a hybrid approach combining spatiotemporal correlation analysis with a Recursive Least Squares-Extended Kalman Filter (RLS-EKF).

  1. Spatiotemporal Correlation Calculation:
    • Historical Data Analysis: The paper first analyzes the chronological nature of AS data, showing that current AS is correlated with previous timeslots and historical simultaneous timeslots.
  2. Pearson Correlation Coefficient: This statistical method is used to quantify the linear relationship between AS at adjacent timeslots and historical data.
  3. Impact of External Factors: The influence of weather and date attributes on AS is also analyzed and quantified.
  4. Recursive Least Squares-Extended Kalman Filter (RLS-EKF):
    • RLS for Parameter Identification: Recursive Least Squares (RLS) is employed to identify and update system parameters by fusing historical AS with external factors (weather, date attributes). This helps in real-time estimation and addresses issues with noisy data.
  5. EKF for Nonlinear Prediction: An Extended Kalman Filter (EKF) is used to handle the nonlinearities introduced by external factors and improve prediction accuracy. The EKF linearizes the nonlinear state-space model to make it compatible with Kalman filtering.
  6. Hybrid Integration: The RLS component identifies the historical AS and external factors, while the EKF integrates these identified parameters with measured and observed values to refine the prediction. The model establishes multiple regression equations to capture the relationships between AS and historical data, weather, and date attributes.

Datasets

The study utilizes floating car GPS data from Didi Chuxing's Chengdu branch.

  • Data Source: Chengdu, China.
  • Data Type: Floating car GPS data (order details).
  • Sampling Frequency: 3 seconds.
  • Data Size: 462 GB.
  • Record Fields: Driver ID, Order ID, Timestamp, Latitude, Longitude, Vehicle Status.
  • Study Area: Selected central urban districts of Chengdu (Wuhou, Jingjiang, Qingyang, Jinniu, and Chenghua).
  • Road Segments: Four specific road segments were chosen to represent different road types (main urban, general, ring, outer ring).
  • Time Span: Data from June to November 2016 was used, with specific subsets for training (June-October 2016) and forecasting (November 2016).

Key Insights and Ideas

  • Interconnectedness of AS: The AS of a road segment is not an isolated phenomenon; it's strongly influenced by its own historical AS, the AS of adjacent segments, and external factors like weather and time of day/week.
  • Limitations of Simple Prediction: Relying solely on the AS of the immediate preceding timeslot leads to low accuracy, especially during traffic rush hours. Incorporating a broader historical context is essential.
  • Value of External Factors: Weather and date attributes, though seemingly secondary, have a quantifiable impact on traffic speed and must be integrated for robust prediction.
  • RLS-EKF Synergy: The combination of RLS for parameter estimation and EKF for handling nonlinear dynamics provides a powerful framework for traffic speed prediction, outperforming individual methods.
  • Data Quality Matters: Removing vehicle status records indicating "empty" or "parking" is crucial to ensure the data reflects actual traffic flow and not just vehicle presence.

Key Results

The FCG-ASpredictor model demonstrated superior performance compared to two benchmark algorithms (ARIMA-KF and LSTM-RNN) across various prediction horizons (15 min, 30 min, 1 h) and evaluation metrics (RMSE, MAE, MAPE).

  • Overall Accuracy: The RLS-EKF approach consistently achieved lower RMSE, MAE, and MAPE values for all four road segments and all prediction horizons.
  • Prediction Horizon Impact: As the prediction horizon increased (from 15 min to 1 h), the error metrics generally increased, highlighting the challenge of long-term traffic forecasting.
  • Feature Importance: Analysis of missing factors indicated that historical AS (both simultaneous and forward) had the most significant impact on prediction accuracy, followed by weather conditions, and then date attributes.
  • Feasibility: The model's reliance on these identified factors (AS-hst, AS-ft, WC-ct, DA-ct) is deemed reasonable and feasible for improving prediction accuracy.

Limitations and Future Work

  • Long-Term Prediction: The paper acknowledges limitations in applying the model for long-term traffic prediction and suggests future work to improve its adaptation to spatiotemporal correlations for longer horizons.
  • Model Adaptation: While the model shows good results, further improvements in adapting it to complex spatiotemporal correlations are envisioned.
  • External Factors: Although weather and date attributes were included, other external factors (e.g., special events, road incidents) could further enhance prediction accuracy. The paper mentions that emergencies have greater randomness and unpredictability, suggesting a need for handling such events.

Summary updated

BibTeX
@misc{zhu2019fcg,
  author = {Difeng Zhu and Guojiang Shen and Duanyang Liu and Jingjing Chen and Yijiang Zhang},
  title = {{FCG-ASpredictor: An Approach for the Prediction of Average Speed of Road Segments with Floating Car GPS Data}},
  year = {2019},
  doi = {10.3390/s19224967},
}
Figure