MMCAformer — Traffic Speed Prediction with CV-Based Microscopic Driving Behavior

MMCAformer (Macro–Micro Cross-Attention Transformer) is a freeway traffic-speed forecasting framework that integrates macroscopic traffic flow states with fine-grained microscopic driving behavior extracted from connected-vehicle (CV) trajectories.

Rather than treating all inputs as homogeneous traffic variables, MMCAformer separately represents macro features (segment speed and CV volume) and micro features (speed volatility and hard/medium/light acceleration and braking), and then learns their spatial and temporal interactions through macro–micro cross-attention. A Student-t negative log-likelihood objective enables both point-wise speed prediction and uncertainty estimation.

This repository provides the processed modeling datasets, scripts, and pre-trained weights used to reproduce the main experiments in our paper: https://arxiv.org/abs/2602.16730

---

Paper Overview

1. Why microscopic driving behavior matters

In freeway systems, sudden hard-braking events can introduce localized deceleration disturbances that propagate through car-following interactions and contribute to subsequent traffic-speed deterioration. Hard acceleration and speed volatility may also reflect unstable traffic dynamics and provide informative signals of traffic-flow evolution. However, traditional aggregated traffic-flow data cannot capture such fine-grained driving behavior information, which may limit further improvements in traffic speed prediction.

Fig. 1 — A hard-braking event in the current segment slows traffic upstream and downstream; future speeds depend on both corridor context and local driving behavior.

2. Model Overview

CV trajectories are first aggregated into two feature groups:

• Macro traffic states: segment speed and CV volume.
• Micro driving behaviors: CV speed volatility and frequencies of hard, medium, and light acceleration and braking.

MMCAformer independently embeds the two feature groups and applies:

• Spatial macro self-attention to capture dependencies among road segments;
• Spatial macro–micro cross-attention to integrate behavior information across segments;
• Temporal macro self-attention to model historical traffic evolution; and
• Temporal macro–micro cross-attention to capture time-dependent interactions between traffic states and driving behaviors.

The final prediction head directly generates multi-step future speeds together with the parameters of a Student-t distribution, allowing the model to quantify predictive uncertainty while accommodating heavy-tailed forecasting errors.

Fig. 2 — MMCAformer: embedding → spatial CA → temporal CA → speed & uncertainty heads.

3. Speed prediction with calibrated uncertainty

Under midday congestion, MMCAformer tracks ground-truth speed and provides 90% prediction intervals. Panel (a) uses macro features only (without micro); panel (b) uses the full macro–micro model. Adding micro features yields tighter intervals (lower MPIW) while maintaining coverage. Panel (c) shows co-occurring spikes in volatility and hard braking/acceleration during the same congested periods.

Fig. 3 — Case study on I-4 Hillsborough: (a) w/o micro features, (b) full MMCAformer with micro features, (c) micro dynamics during congestion.

4. Cross-attention focuses on congested segments

When segment speed drops below 40 mph, cross-attention scores between macro and micro features increase significantly—indicating the model relies on micro dynamics most during congestion.

Fig. 4 — Macro/micro inputs, cross-attention (CA) score matrix, and mean CA scores across speed bins.

5. Validation against MVDS loop detectors

CV-based speed predictions from MMCAformer are compared with MVDS (inductive loop detector) measurements—an independent ground-truth source. Predictions closely follow MVDS on both weekday and weekend traffic patterns.

Fig. 5 — Segment-level CV speed vs. MVDS speed (May 2024); weekday and weekend zoom-ins.

---

Requirements

• Python 3.8+
• PyTorch
• NumPy
• scikit-learn
• SciPy
• iTransformer package (for the iTransformer baseline)
• einops (for the TLAST baseline)

How to Use

Clone the repository (Git LFS is required for the large .npz files):

git lfs install
git clone https://github.com/UCFLeiHan/MMCAformer.git
cd MMCAformer

Supported datasets: I4_Hills, I75_Hills, I275_Hills, I4_Orange (or all).

# Evaluate the main MMCAformer model
python run_MMCAformer.py --dataset I4_Hills --mode test

# Train MMCAformer
python run_MMCAformer.py --dataset I4_Hills --mode train

# Ablation studies (Micro / Cross / S / T)
python run_MMCAformer_ablations.py --ablation Micro --dataset I4_Hills --mode test

# Uncertainty evaluation (Gaussian or Student-t)
python run_MMCAformer_uncertainty.py --dataset I4_Hills --distribution student

# Baseline models (e.g., LSTM, STGCN, TLAST, ...)
python run_baselines.py --model LSTM --dataset I4_Hills --mode test

Pre-trained weights are loaded from checkpoints/ by default. Use --mode both to train and then evaluate.

Data Format

Input data

Each processed .npz file contains:

• sequences: input tensor with shape (S, T, N, C)
• targets: future speed tensor with shape (S, T, N)

where:

• S: number of samples
• T: number of time steps
• N: number of freeway segments
• C: number of input features

The model uses the past 12 time steps (1 hour at 5-min intervals) to predict the next 12 time steps (1 hour).

Input features

Each processed .npz file stores 14 raw channels in this order:

Index	Field	Role
0	Road_direction	Segment direction ID (not used by MMCAformer)
1	mean_of_speed_mean	Macro: segment-level average CV speed
2	n_journeys	Macro: CV volume (unique journeys per 5-min interval)
3	mean_of_speed_std	Micro: CV speed volatility within the segment
4	total_stop	Stop-event count (not used by MMCAformer)
5	total_hard_acc	Micro: hard acceleration event count
6	total_median_acc	Micro: medium acceleration event count
7	total_light_acc	Micro: light acceleration event count
8	total_hard_dec	Micro: hard braking event count
9	total_median_dec	Micro: medium braking event count
10	total_light_dec	Micro: light braking event count
11	total_LC	Lane-change event count (not used by MMCAformer)
12	hour	Time-of-day index (used as temporal embedding)
13	weekday	Day-of-week index (used as temporal embedding)

In the main MMCAformer scripts, channels 0, 4, and 11 are dropped. The remaining continuous inputs are grouped as:

Macro features (2)

• Segment speed (mean_of_speed_mean)
• CV volume (n_journeys)

Micro features (7)

• Speed volatility (mean_of_speed_std)
• Hard / medium / light acceleration frequency (total_hard_acc, total_median_acc, total_light_acc): number of hard/medium/light acceleration events within each 5-min interval
• Hard / medium / light braking frequency (total_hard_dec, total_median_dec, total_light_dec): number of hard/medium/light braking events within each 5-min interval

Temporal embeddings (2)

• hour and weekday are kept as discrete embeddings (not as continuous macro/micro channels)

Overall, each road segment is represented by 2 macro features + 7 micro features at every 5-min time step, plus hour/weekday embeddings.

Prediction target

The prediction target is the future segment-level average traffic speed in mph.

Adjacency matrix

Graph-based baselines use an adjacency matrix stored as a PyTorch tensor:

• shape: (N, N)
• required by DCRNN, STGCN, GraphWaveNet, and ASTGCN

Experimental Notes

• The training/test split is performed by date to avoid temporal information leakage.
• Only observations from 06:00–22:00 are used because nighttime CV penetration is substantially lower.
• All scripts use early stopping and retain the model with the lowest validation MAE.
• Speed is normalized during training and inverse-transformed before evaluation.
• MMCAformer and its ablation variants use mixed-precision training (AMP + GradScaler).
• Evaluation is reported at three levels:
  • Overall performance
  • Per-step performance across 5–60 min prediction horizons
  • Per-bin performance under different observed speed ranges (0–20, 20–40, 40–60, >60 mph)