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

This article is an English summary of a Chinese forum post reviewing methods for highway traffic volume prediction based on ETC (Electronic Toll Collection) data.

Key points

1. Statistical and time-series models on historical ETC data

• ARIMA: Classic linear model; requires stationarity (ADF test, differencing) and uses ACF/PACF to select orders (p, d, q). Weak on nonlinear dynamics and sudden disruptions. In a Kigali (Muhima intersection) study, ARIMA achieved MAPE 24.2% (RMSE 9.1) vs LSTM's 22.5% (RMSE 5.8).
• Historical Average (HA): Simplest baseline; predicts future flow as the average of the same time slots historically. Cheap and interpretable, but fails under atypical events.
• LSTM: Gated RNN capturing long-term dependencies; beats SAEs, RBF networks, SVM, and ARIMA, especially for horizons beyond 15 minutes. High computational cost and many hyperparameters.
• GRU: Simplified LSTM with a single update gate; fewer parameters, faster, often comparable accuracy; a DiffGRU comparison showed RMSE reductions of 23%/26% vs ARIMA/SVR (LSTM: 53%/55%).
• Transformer: Self-attention handles long-range dependencies and parallel computation; variants like STGAFormer improve long-horizon prediction and handle sudden traffic events; often combined with GNNs for spatial modeling.
• Selection guidance: ARIMA suits stable, periodic, small-data, interpretable, real-time baseline settings; LSTM/GRU suit nonlinear, long-term-dependency scenarios with ample data (GRU preferred when compute-constrained); Transformer suits network-scale, long-horizon, multi-source scenarios with abundant compute.

2. Multi-source dynamic prediction models

• Fusion architectures: early fusion (feature-level concatenation), late fusion (decision-level weighted/voting combination), and hybrid fusion. Examples include a heterogeneous graph neural network (HGNN) framework fusing image/point-cloud/sound/temperature data for anomalous event detection, and a two-stage cross-modal reconstruction pretraining approach enabling ramp flow prediction from mainline ETC data only.
• Preprocessing: cleaning missing/anomalous values, temporal alignment to a uniform granularity (e.g., 5 minutes), and feature engineering (flow, headway, heavy-vehicle ratio, rainfall, visibility).
• Spatial-temporal modeling:
• CNN: grid-based spatial feature extraction; efficient but loses irregular road topology.
• STGCN / GNN: models road networks as graphs (nodes = gantries/stations/segments); spatio-temporal blocks combining graph convolution with temporal convolution or attention (e.g., ASTGNN) are now state-of-the-art for network-level ETC prediction.
• HetGNN: handles heterogeneous entities (gantries, toll stations, weather, incidents), e.g., HetGAT for highway speed prediction.
• Trade-offs: higher accuracy, robustness, and richer traffic-state information vs. difficult cross-system data integration, higher complexity, heterogeneity challenges, and reduced interpretability. Best for high-precision, complex environments and agencies with strong data infrastructure.

3. Models incorporating external factors

• External factors: holidays/special events (tidal, tourism-driven surges, e.g., under toll-free holiday policies), weather (light rain reduces road capacity 4-7%; heavy rain up to 14%; fog lowers speeds and increases speed variance), and incidents/roadworks that create sudden bottlenecks.
• Modeling techniques:
• Attention mechanisms: dynamically weight weather/event features (e.g., ASTGNN).
• Multi-task learning: jointly predict flow plus weather classification or event detection via shared representations.
• Hybrid models: e.g., a classifier flagging holidays/bad weather feeding an LSTM/Transformer; or ARIMA for the linear component plus LSTM on nonlinear residuals.
• These models deliver the highest accuracy and robustness but demand the most data, compute, and cross-disciplinary maintenance. Suitable for national/regional traffic situation assessment, smart-city "traffic brain" engines, and refined highway operations.

4. Comparative summary and outlook

| Dimension | Statistical/time-series | Multi-source dynamic | External-factor comprehensive | |---|---|---|---| | Accuracy | Medium | High | Highest | | Complexity | Low-medium | High | Highest | | Real-time | Good | Medium | Poor | | Data dependency | Medium (historical ETC) | High (synchronized multi-source) | Highest (real-time external data) | | Interpretability | Good (ARIMA), poor (DL) | Poor | Medium (via attention) | | Use cases | Stable patterns, baselines | Critical segments, complex areas | Macro assessment, traffic brains |

Future directions: smarter data cleaning and fusion (including mobile signaling, social media, V2X), improved explainability and robustness of deep models, and better prediction of abnormal scenarios (holidays, extreme weather, incidents) via transfer/meta-learning.