PPTNet: A Hybrid Periodic Pattern-Transformer Architecture for Traffic Flow Prediction and Congestion Identification

Introduction

Accurate prediction of traffic flow parameters and real-time identification of congestion states are essential for the efficient operation of intelligent transportation systems. This paper proposes a Periodic Pattern-Transformer Network (PPTNet) for traffic flow prediction, integrating periodic pattern extraction with the Transformer architecture, coupled with a fuzzy inference method for real-time congestion identification. Firstly, a high-precision traffic flow dataset (Traffic Flow Dataset for China’s Congested Highways & Expressways, TF4CHE) suitable for congested highway scenarios in China is constructed based on drone aerial imagery data. Subsequently, the proposed PPTNet employs Fast Fourier Transform to capture multi-scale periodic patterns and utilizes two-dimensional Inception convolutions to efficiently extract intra and inter periodic features. A Transformer decoder dynamically models temporal dependencies, enabling accurate predictions of traffic density and speed. Finally, congestion probabilities are calculated in real-time using the predicted outcomes via a Mamdani fuzzy inference-based congestion identification module. Experimental results demonstrate that the proposed PPTNet significantly outperforms mainstream traffic prediction methods in prediction accuracy, and the congestion identification module effectively identifies real-time road congestion states, verifying the superiority and practicality of the proposed method in real-world traffic scenarios.

Environments

python 3.9, pytorch 2.5.1, CUDA 12.4

git clone https://github.com/ADSafetyJointLab/PPTNet.git
conda create -n pptnet python=3.9.20
conda activate pptnet
pip install -r requirements.txt

Dataset

The TF4CHE (Traffic Flow Dataset for China’s Congested Highways & Expressways) is derived from the UAV-based AD4CHE dataset, calibrated to approximately 5 cm accuracy at 100 m altitude. TF4CHE is pre-processed into a consolidated time-series format suitable for traffic flow prediction and congestion identification.

Coverage
- 11 road segments (Road_segment_1.csv … Road_segment_11.csv), each corresponding to a distinct expressway section in five Chinese cities (originally 68 AD4CHE segments, consolidated by route).
- Lane layout: Four travel lanes + emergency lane in each direction; world coordinate origin at top-left of UAV frame, X increasing in travel direction, Y downward.
File format
Each Road_segment_{i}.csv contains one row per second, with the following columns:

| Name | Description | Unit | |--------------------|-------------------------------------------------------------------------------------------|-----------| | Month/Year | Month and year of video recording (virtual date information) | – | | Weekday | Completion date of video recording (virtual date information) | – | | TimeCode | Specific start time of video recording (virtual time information) | – | | second | Video time sequence in seconds | s | | drivingDirection | Traveling direction of the recorded segment | – | | car | Number of cars in frame | veh | | bus | Number of buses in frame | veh | | truck | Number of trucks in frame | veh | | G(t) | Equivalent vehicle count, with cars as the reference, (conversion coefficients: $\alpha_{\text{bus}} = 2,\ \alpha_{\text{truck}} = 2.5$) | veh | | k(t) | Average density | veh/m | | q(t) | Average flow | veh/s | | xVelocity(t) | Mean speed along X-axis | m/s | | yVelocity(t) | Mean speed along Y-axis | m/s | | xAcceleration(t) | Mean acceleration along X-axis | m/s² | | yAcceleration(t) | Mean acceleration along Y-axis | m/s² | | OccupancyRatio | Lane space occupancy | – | | File_ID | Original AD4CHE segment index | – |

Note: TF4CHE converts per-frame trajectory data (xx_tracks.csv) and metadata (xx_recordingMeta.csv, xx_tracksMeta.csv) into uniformly spaced time-series, applying conversion formulas from rail transit theory to compute densities, flows, and occupancy, thus streamlining downstream forecasting tasks.

Downloads
We provide download link from Google Drive and Baidu Yunpan to facilate users from all over the world.
- Baidu Yunpan Extraction code : xrvr
- Google Drive
Preparation
You should refer to the preprocess.py and change your input_dir、utput_dir path. And divide the original TF4CHE dataset into 'train', 'val', and 'test'.

python preprocess.py

Train

You should refer to the config.py and change your raw_data_path、 processed_data_path and pred_dir path. And start predicting network training

python train.py

Results

Prediction

Performance comparison between PPTNet and mainstream prediction models at different prediction horizons

Bold denotes the best result, underline denotes the best result of previous methods, ↑ denotes performance improvement, and ↓ denotes performance degradation.

| Model | MAE @15 | MSE @15 | RMSE @15 | MAE @30 | MSE @30 | RMSE @30 | MAE @45 | MSE @45 | RMSE @45 | |-------|-------------|-------------|--------------|-------------|-------------|--------------|-------------|-------------|--------------| | LSTM | 0.1719 | 0.0477 | 0.2184 | 0.2517 | 0.0954 | 0.3088 | 0.2911 | 0.1179 | 0.3433 | | RNN | 0.1528 | 0.0316 | 0.1776 | 0.2747 | 0.0997 | 0.3157 | 0.2993 | 0.1201 | 0.3465 | | ConvLSTM | 0.1362 | 0.0236 | 0.1535 | 0.2372 | 0.0772 | 0.2779 | 0.2747 | 0.1010 | 0.3179 | | Bi‑LSTM | 0.1703 | 0.0410 | 0.2024 | 0.1908 | 0.0544 | 0.2331 | 0.2425 | 0.0929 | 0.3084 | | GRU | 0.1106 | 0.0184 | 0.1356 | 0.1816 | 0.0442 | 0.2102 | 0.2060 | 0.0606 | 0.2462 | | CNN | 0.1117 | 0.0166 | 0.1287 | 0.1623 | 0.0345 | 0.1857 | 0.1666 | 0.0358 | 0.1891 | | TCN | 0.1036 | 0.0174 | 0.1319 | 0.0934 | 0.0151 | 0.1230 | 0.1547 | 0.0374 | 0.1934 | | ConvGRU | 0.0883 | 0.0119 | 0.1091 | 0.1224 | 0.0202 | 0.1420 | 0.1061 | 0.0167 | 0.1291 | | LSSL | 0.0859 | 0.0095 | 0.0973 | 0.0763 | 0.0078 | 0.0882 | 0.1333 | 0.0237 | 0.1539 | | Transformer | 0.0946 | 0.0125 | 0.1117 | 0.0677 | 0.0069 | 0.0828 | 0.1379 | 0.0296 | 0.1721 | | Reformer | 0.0955 | 0.0110 | 0.1049 | 0.0558 | 0.0044 | 0.0664 | 0.0711 | 0.0073 | 0.0854 | | FEDformer | 0.0973 | 0.0087 | 0.0934 | 0.0566 | 0.0048 | 0.0693 | 0.0670 | 0.0065 | 0.0809 | | LSTNet | 0.0856 | 0.0094 | 0.0972 | <ins>0.0523</ins> | <ins>0.0040</ins> | <ins>0.0633</ins> | 0.0645 | 0.0059 | 0.0768 | | TimesNet | <ins>0.0778 | <ins>0.0085</ins> | <ins>0.0925</ins> | 0.0573 | 0.0048 | 0.0690 | <ins>0.0615</ins> | <ins>0.0056</ins> | <ins>0.0746</ins> | | PPTNet (Ours) | 0.0821 | 0.0074 | 0.0861 | 0.0512 | 0.0033 | 0.0574 | 0.0660 | 0.0050 | 0.0709 | | Improvement (%) | ↓ 5.53 | ↑ 12.94 | ↑ 6.92 | ↑ 2.10 | ↑ 17.50 | ↑ 9.32 | ↓ 7.32 | ↑ 10.71 | ↑ 4.96 |

Congestion Identification

Acknowledgment

This work was supported in part by the Science and Technology Development Program of Jilin Province under Grant 20240302052GX and in part by the National Natural Science Foundation of China under Grant 52075213.
This library is inspired by AD4CHE, Time-Series-Library and many other related works, we thank them for sharing the code and datasets.

PPTNet

Install / Use

README