On Model Selection for Scalable Time Series Forecasting in Transport Networks

Julien Monteil, Anton Dekusar, Claudio Gambella, Yassine Lassoued, Martin MevissenIEEE Trans. ITS 2022 · DOI

Abstract

Accurate traffic speed forecasting is a cornerstone of modern urban mobility, yet existing models often struggle with long-term horizons and network-wide scalability. This paper benchmarks various forecasting methods, from simple autoregressive models to advanced Graph Convolutional Neural Networks (GCNNs), using a large-scale dataset from Los Angeles. We demonstrate that while simple models excel at short-term predictions, localized GCNNs provide superior performance for horizons beyond two hours by capturing complex spatial dependencies. Crucially, the GCNN approach offers a scalable framework for city-wide deployment, reducing the operational burden of maintaining thousands of individual link-specific models.

Introduction: The Horizon and Scalability Gaps

Traffic prediction powers essential services like ETA estimation and ride-sharing dispatch. Despite its importance, research has historically focused on short-term windows (under 1 hour) and link-specific modeling. This study identifies two primary challenges in moving toward city-scale forecasting:

  1. The Horizon Gap: As the prediction window extends to 2–3 hours, the temporal "memory" of current conditions fades. This makes long-term forecasting significantly harder as local historical data becomes less predictive.
  2. The Scalability Gap: Traditional approaches that train one model per road segment do not scale to cities with millions of links. A more efficient, unified approach is required.
Autocorrelation plot showing the decay of temporal 'memory' over time, illustrating the difficulty of long-term forecasting.

The study utilizes 14 weeks of speed data from 1,098 road segments in Los Angeles, collected at 15-minute intervals, to evaluate how different models bridge these gaps.

Methodology

The researchers benchmarked several model categories:

  • Baselines & ML: Contextual averages, ARIMA, and Support Vector Regression (SVR). Gradient Boosting (XGBoost) was also included as a high-performing tabular baseline.
  • Link-based Deep Learning: Multi-Layer Perceptrons (MLP) and Long Short-Term Memory (LSTM) networks trained on individual segments.
  • Localized GCNN: The paper’s primary contribution. Unlike global spectral GCNNs, this model uses a localized approach, considering the $k$-nearest incoming and outgoing neighbors for every link. This allows the network to learn spatial "flow" dependencies across the graph efficiently.
Architecture of the localized GCNN, which incorporates spatial dependencies from k-nearest neighbors.

The architecture processes spatial-temporal input tensors through multiple convolutional layers, eventually concatenating them with contextual features for the final prediction.

Results and Key Insights

The evaluation revealed a distinct threshold for model effectiveness based on the prediction horizon:

  • Short-term (0–45 mins): Simple models like Gradient Boosting are dominant. For a 15-minute horizon, Gradient Boosting achieved an RMSE of 5.42 km/h.
  • Long-term (1–3 hours): Deep learning models, particularly the GCNN, show a significant "accuracy lift" as the horizon increases. For the 3-hour mark, the GCNN achieved the lowest RMSE of 7.87 km/h, outperforming the baseline (8.39 km/h) and standard AR models (9.64 km/h).
  • Spatial Importance: The GCNN’s success in long-term forecasting is attributed to its spatial awareness. At 3 hours, regional network conditions are more influential than the immediate history of a single link.

Discussion: Scalability and Future Directions

Beyond accuracy, the study highlights a critical operational trade-off. While simple models train quickly, the GCNN offers superior scalability. Maintaining 1,000+ individual models is an $O(N)$ problem, whereas a single GCNN represents the entire network ($O(1)$), making it far more manageable for large-scale urban deployment. Although the GCNN required over 5 hours to train initially, its unified nature simplifies long-term maintenance.

Future research will focus on testing these models in more volatile weather conditions and exploring hybrid architectures that combine the temporal strengths of LSTMs with the spatial capabilities of GCNNs.

Summary updated

BibTeX
@article{monteil2022model,
  author = {Julien Monteil and Anton Dekusar and Claudio Gambella and Yassine Lassoued and Martin Mevissen},
  title = {{On Model Selection for Scalable Time Series Forecasting in Transport Networks}},
  journal = {IEEE Trans. ITS},
  year = {2022},
  doi = {10.1109/tits.2021.3060959},
}
Figure