ETC-Based Highway Traffic Flow Prediction: Methods Survey and Comparative Analysis

Forum topic · ✨步子哥 · 2025-11-03

Summary

This article presents an in-depth survey and comparison of highway traffic flow prediction methods based on ETC (Electronic Toll Collection) data, organized into three methodological families. First, models built purely on historical ETC time-series data are reviewed, ranging from classical statistical approaches (ARIMA, Historical Average) to deep learning models (LSTM, GRU, Transformer), with reported benchmarks such as LSTM outperforming ARIMA (MAPE 22.5% vs 24.2%). Second, dynamic models that fuse multi-source data—including speed, vehicle type, sensor, and floating-car data—are discussed, covering early/late/hybrid fusion architectures, CNN-based spatial modeling, spatio-temporal graph convolutional networks (STGCN), and heterogeneous graph neural networks (HetGNN). Third, comprehensive models that explicitly incorporate external factors—holidays, weather (rain can cut road capacity by 4–14%), accidents, and roadworks—are analyzed, using attention mechanisms, multi-task learning, and hybrid modeling. A multi-dimensional comparison table covers accuracy, computational complexity, real-time capability, data dependency, and interpretability, with recommendations for applicable scenarios. Future directions include data quality and fusion, model explainability and robustness, and prediction under atypical traffic conditions.

ETC-Based Highway Traffic Flow Prediction: An In-Depth Survey and Comparative Analysis

This article surveys and compares three major families of highway traffic flow prediction methods based on ETC (Electronic Toll Collection) data: (1) statistical/time-series models on historical ETC data, (2) multi-source dynamic prediction models, and (3) comprehensive models incorporating external factors.

Key points

1. Prediction from historical ETC data

Classical statistical models
ARIMA: Theories are mature, parameters (p, d, q) are interpretable, implementation is simple, and it works on small datasets. However, it is inherently linear, requires stationarity (via differencing), and handles poor multivariate/exogenous inputs. In one comparative study, ARIMA achieved a MAPE of 24.2% (RMSE 9.1) versus LSTM's 22.5% (RMSE 5.8). Best for stable, highly periodic, short-term forecasting (e.g., 15 min–1 hour ahead on commuter corridors) or as a baseline.
Historical Average (HA): Zero-cost and highly interpretable, but static—cannot capture dynamics, anomalies, or spatial influence. Used mainly as a baseline or auxiliary input.
Deep learning models
LSTM / GRU: Gated recurrent networks capture nonlinearity and long-term dependencies, automatically extracting features from raw time series. LSTM has been shown to outperform SAE, RBF, SVM, and ARIMA, especially for horizons beyond 15 minutes. A GRU variant (DiffGRU) reduced RMSE by 23–26% versus ARIMA/SVR (vs 53–55% for LSTM). GRU is simpler and faster; LSTM may be better for long-horizon tasks. Both need large training data and GPUs, and are less interpretable.
Transformer: Self-attention handles arbitrary long-range dependencies and parallel computation; combined with GNNs it forms strong spatio-temporal frameworks (e.g., STGAFormer for long-term forecasting and sudden traffic events). Drawbacks: huge parameter counts, high data/compute requirements, limited interpretability. Suited to network-level, long-horizon, multi-source scenarios.

2. Multi-source dynamic prediction models

Fusion architectures: Early (data-level), late (decision-level), and hybrid fusion. Examples include a multimodal deep fusion + heterogeneous graph neural network (HGNN) framework for anomalous event detection, and a two-stage approach using cross-modal reconstruction pretraining on mainline ETC data to predict ramp flows even where ramps lack detectors.
Preprocessing & feature engineering: Cleaning (missing values, outliers, anomaly-removal algorithms), temporal alignment/resampling to a unified granularity (e.g., 5 minutes), and feature extraction (flow, headway, heavy-vehicle ratio, weather variables).
Spatio-temporal modeling:
CNN grids the road network to extract local spatial patterns, but loses true (irregular) topology.
GNN / STGCN: Models gantries/stations/segments as graph nodes; STGCN combines graph convolutions (space) with temporal convolutions or RNNs (time). Attention-based variants (ASTGNN) dynamically compute spatial correlations and are state-of-the-art for ETC-based network-level prediction.
HetGNN (e.g., HetGAT): Handles heterogeneous node/edge types (gantries, toll stations, weather, incidents) for finer-grained modeling.
Trade-offs: Higher accuracy and robustness through information complementarity and finer management granularity—but significant data integration difficulty, higher model complexity, heterogeneity challenges, and reduced interpretability. Best for high-precision requirements, complex/variable traffic environments, and agencies with strong data infrastructure.

3. Models incorporating external factors

Factor taxonomy:
*Holidays/special events*: Generate tidal, tourism-driven surges (amplified by free-pass policies); holiday-specific models substantially outperform generic ones.
*Weather*: Reduces road capacity (light rain: 4–7%; heavy rain: up to 14%), suppresses demand, and low visibility increases risk; weather parameters improve capacity/demand estimation.
*Accidents & roadworks*: Sudden, stochastic capacity shocks requiring real-time signal injection into the model.
Modeling approaches: Attention mechanisms dynamically re-weight external factors (e.g., weather weight rises in storms); multi-task learning jointly predicts flow plus weather/event classification; hybrid models combine statistical and deep components (e.g., ARIMA for linear parts + LSTM for nonlinear residuals).
Trade-offs: Highest accuracy and robustness with some interpretability potential via attention weights—but the greatest data dependency, complexity, and development/maintenance cost. Targeted at national/regional traffic situation assessment, smart-city "traffic brains," and fine-grained highway operations management.

4. Comparative summary

5. Future directions

Data quality & fusion: Smarter cleaning/repair/fusion; integrating new sources such as mobile signaling, social media, and V2X data.
Explainability & robustness: Transparent decision-making for high-stakes settings; resilience to attacks, sensor failures, and extreme anomalies.
Atypical traffic scenarios: Dedicated models for holidays, major events, severe weather, and incidents, potentially using transfer learning and meta-learning for rapid adaptation.