A Multi-Resolution Multi-Domain Pre-Training Framework for Universal Traffic Forecasting

Zhouyang Liu; Jindong Han; Hao Wang; Xinyue Liu; Hui Gao; Dongsheng Li; Hao Liu

cs.LGSep 12, 2026

A Multi-Resolution Multi-Domain Pre-Training Framework for Universal Traffic Forecasting

Authors: Zhouyang Liu, Jindong Han, Hao Wang, Xinyue Liu, Hui Gao, Dongsheng Li, Hao Liu

Organizations: PDL Lab, National University of Defense Technology, Hunan, China · Shandong University, Shandong, China · HKUST (GZ) · Didichuxing Co. Ltd, Beijing, China ·

Abstract

Spatio-temporal traffic data are central to intelligent transportation systems, yet their heterogeneity poses significant challenges for large-scale modeling. Existing pre-trained models often rely on a homogeneous modeling paradigm to handle highly heterogeneous traffic data. This fundamental mismatch not only limits model generalization but also leads to computationally expensive and parameter-inefficient designs. To this end, we propose FlexST, a novel pre-training framework that introduces modularity and adaptivity for traffic modeling. Specifically, we first propose a multi-resolution spatio-temporal diffusion module that captures both short-term fluctuations and long-range trends, effectively reconciling inputs with divergent temporal and spatial resolutions. After that, we construct a domain-adaptive mixture-of-experts that dynamically routes data to specialized sub-networks, enabling selective knowledge transfer while preventing negative interference across diverse domains. Moreover, we devise a unified periodic encoding strategy that injects resolution- and domain-aware inductive biases to harmonize periodic inconsistencies across datasets. Extensive experiments on 23 real-world traffic datasets demonstrate that FlexST significantly outperforms state-of-the-art baselines in zero- and few-shot settings, showcasing superior generalization, adaptability and efficiency. This work offers a new direction for building general-purpose pre-trained models capable of handling the complexity and variability of urban traffic systems.

Figures & tables

Panel of Figure 1 — Fig. 1 : Weekly average traffic time series across regions, highlighting heterogeneity across cities (top: Hangzhou vs. Jinan), temporal resolutions (middle: Beijing Subway 10 min vs. 15 min), and variables (bottom: PEMS08 flow vs. speed). Z-score normalization handles large value differences.

Fig. 2 : The model pipeline.

Fig. 3 : Aligning different temporal resolutions by time span ensures that patches of the same absolute duration are compared consistently, regardless of their length or resolution.

Dataset	PEMS04		PEMS08		Didi-SZ		Beijing Taxi				HZMetro				CHI-Bike				Beijing Subway
Dataset	PEMS04		PEMS08		Didi-SZ		Inflow		Outflow		Inflow		Outflow		Inflow		Outflow		Inflow		Outflow
	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
HA	38.67	57.49	31.93	47.72	3.19	4.81	56.39	94.83	56.47	94.93	130.30	235.99	130.00	219.45	6.02	11.18	5.99	11.53	238.66	425.64	232.28	454.64
TimesFM	36.84	56.62	30.43	47.10	3.09	4.70	48.84	85.42	48.93	85.53	113.21	216.89	114.88	204.20	5.69	10.57	5.63	10.83	201.74	385.23	197.52	415.69
Timer	125.88	152.52	117.77	141.59	7.83	10.58	90.80	127.08	90.84	127.18	150.13	258.28	150.24	245.23	6.24	12.91	6.24	13.23	224.25	425.30	219.84	463.39
Sundial	36.56	55.35	30.20	46.14	3.18	4.84	50.90	88.25	51.01	88.38	123.18	221.96	124.59	221.96	5.85	10.99	5.82	11.35	224.25	425.30	219.84	463.39

TABLE I : Zero-shot performance on short-term prediction. We bold the best and underline the second models.

Dataset	PEMS04		PEMS08		Didi-SZ		HZMetro
Dataset	PEMS04		PEMS08		Didi-SZ		Inflow		Outflow
	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
HA	123.73	160.18	104.67	137.75	5.04	6.86	106.77	193.77	106.63	183.58
TimesFM	95.63	132.00	72.03	105.05	4.28	5.96	87.759	218.089	78.832	182.676
Timer	111.52	132.88	106.06	128.96	6.39	8.82	142.714	242.061	145.531	236.142
Sundial	101.46	136.81	67.46	97.97	4.82	6.50	159.44	313.21	157.96	304.78

TABLE II : Zero-shot performance on long-term prediction.

Dataset	PEMS08		Didi-SZ		TrafficZZ		NYC-Taxi				CHI-Taxi				NYC-Bike				CHI-Bike
Dataset	PEMS08		Didi-SZ		TrafficZZ		Inflow		Outflow		Inflow		Outflow		Inflow		Outflow		Inflow		Outflow
	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
HA	32.19	48.07	3.17	4.79	1.46	2.09	16.99	38.73	25.15	51.74	4.42	11.53	4.80	12.46	14.47	10.48	14.30	24.16	6.16	11.54	5.99	11.54
Informer	69.10	94.23	3.05	4.70	1.09	1.79	18.54	41.33	30.07	57.14	4.13	13.15	4.62	14.81	13.45	24.82	13.42	24.50	4.37	8.89	4.27	8.76
PatchTST	22.26	35.58	2.82	4.34	1.16	1.80	11.37	26.91	17.25	38.02	3.30	8.42	3.60	9.40	10.87	19.77	10.86	19.34	5.00	9.68	4.88	9.55
STGCN	28.99	42.40	3.39	5.09	1.07	1.69	17.71	38.17	28.68	52.78	3.66	11.17	4.10	13.04	11.11	21.87	11.10	21.66	3.53	7.24	3.52	7.18

TABLE III : Few-shot performance on short-term prediction using training data of 7 days.

Dataset	PEMS08		Didi-SZ		TrafficZZ		NYC-Taxi				NYC-Bike
Dataset	PEMS08		Didi-SZ		TrafficZZ		Inflow		Outflow		Inflow		Outflow
	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
HA	105.80	139.29	4.95	6.75	1.18	1.66	13.87	31.39	20.99	42.47	11.86	20.41	11.66	20.04
Informer	87.83	118.34	3.69	5.38	1.09	1.77	19.56	43.31	31.40	59.84	13.45	24.50	13.37	24.13
PatchTST	46.88	73.00	3.35	5.07	0.59	0.95	7.31	18.71	11.45	27.58	7.44	14.28	7.68	14.55
STGCN	51.81	74.70	4.11	5.82	1.09	1.72	22.84	45.49	34.75	60.09	10.73	21.24	10.66	20.81

TABLE IV : Few-shot performance on long-term prediction using training data of 7 days.

Fig. 4 : Comparative performance evaluation of few-shot prediction with varying training data volumes (1, 3, and 7 days).

Panel of Figure 9 — Fig. 5 : Peak memory, inference throughput and model parameter counts.

Fig. 6 : Ablation study.

Notation	Description
$G=(\mathcal{V},\mathbf{A})$	Graph with spatial units $\mathcal{V}$ and adjacency matrix $\mathbf{A}$ representing spatial connections
$\mathbf{X}\in\mathbb{R}^{N\times T}$	Single-domain traffic data with $T$ observations per spatial unit $n\in N$ , characterized by task type $Q$ , temporal resolution $R$ , spatial resolution $D$ and spatial graph $G$
$S$	Time span of time series, determining its absolute duration
$\mathbf{ts}$	timestamps associated with observations
$D,\mathbf{D}_{i}$	Spatial resolution and spatial receptive field at resolution $i$
$\mathcal{X}$	Multi-domain traffic dataset

TABLE V : Summary of notations used in this paper.

Panel of Figure 12 — Fig. 7 : Spatio-temporal dynamics observed across different forecasting tasks (left) and temporal resolutions (right).

Dataset	Traffic Flow	Taxi Demand	Traffic Speed	Metro Flow	Congestion Level	Total
# of Regions	1,957	263	2,676	564	524	3,584
# of Records	71,834,110	9,240,768	28,015,352	5,621,616	9,054,720	123,766,566
Sampling Rate	5T	30T	5T, 30T	10T, 15T	10T	5T, 10T, 15T, 30T
Source	[ 3 , 12 ]	[ 36 ]	[ 8 , 38 ]	[ 39 , 40 ]	[ 41 ]	-

TABLE VI : Statistical information of the pre-training datasets, which are aggregated based on the measured variables. For sampling rate: T=minute, H=hour. We provide detailed description of each dataset in supplementary material.

Dataset

PEMS04

PEMS08

GBA

NYC-Taxi

CHI-Taxi

Beijing Taxi

NYC-Bike

CHI-Bike

TrafficZZ

Beijing Subway

HZMetro

Didi-SZ

Task

Traffic Flow

Taxi Demand

Bicycle Sharing

Traffic Speed

Metro Flow

Congestion Level

# of Regions

307

170

2,352

263

1,024

540

270

676

276

627

# of Records

5,216,544

3,035,520

124,637,184

46,102,848

2,698,080

43,745,280

113,633,280

2,384,640

948,428

993,600

292,000

10,834,560

Sampling Rate

30T

15T

10T

Source

[ 3 ]

[ 12 ]

[ 36 ]

[ 8 ]

[ 36 ]

[ 37 ]

[ 8 ]

[ 38 ]

[ 40 ]

[ 41 ]

TABLE VII : Statistical information of the downstream evaluation datasets. Notice that, the evaluation datasets differ from the pre-training datasets in terms of spatial regions, sampling rates, tasks, or years .

Traffic Forecasting

Traffic forecasting predicts future flows, speeds, demand, travel times, and congestion from spatial, temporal, and contextual data.

See all 87 papers

UniST-Pred: A Robust Unified Framework for Spatio-Temporal Traffic Forecasting in Transportation Networks Under Disruptions
Feb 15, 2026Yue Wang, Djellel Difallah, Areg Karapetyan +1Intelligent Transportation Systems Spatiotemporal Forecasting
New York University Abu Dhabi
SpecLST: Spectral-Enhanced Long-Short Term Spatio-Temporal Multimodal Modeling for Traffic Forecasting
Oct 26, 2026Kenan Kang, Matthew M. Y. Kuo, Weihua LiSpatiotemporal Forecasting Traffic Forecasting
Department of Computer Science and Software Engineering, Auckland University of Technology, Auckland 1010, New Zealand · Department of Data Science and Artificial Intelligence, Auckland University of Technology, Auckland 1010, New Zealand
Continuous Cross-Domain Traffic State Prediction via Memory-Augmented Graph Liquid Time-Constant Networks
Jun 14, 2026Jinrong Xiang, Ming XuCross-Domain Transfer Learning Spatiotemporal Forecasting
Software College, Liaoning Technical University, Huludao, Liaoning 125100, China

Feb 15, 2026J/K move · Enter open · S save

UniST-Pred: A Robust Unified Framework for Spatio-Temporal Traffic Forecasting in Transportation Networks Under Disruptions

Yue Wang, Djellel Difallah, Areg Karapetyan, Samer Madanat

New York University Abu Dhabi

Intelligent Transportation Systems Spatiotemporal Forecasting Traffic Forecasting

Open paper AskPDF arXiv

Spatio-temporal traffic forecasting is a core component of intelligent transportation systems, supporting various downstream tasks such as signal control and network-level traffic management. In real-world deployments, forecasting models must operate under structural and observational uncertainties, conditions that are rarely considered in model design. Recent approaches achieve strong short-term predictive performance by tightly coupling spatial and temporal modeling, often at the cost of increased complexity and limited modularity. In contrast, efficient time-series models capture long-range temporal dependencies without relying on explicit network structure. We propose UniST-Pred, a unified spatio-temporal forecasting framework that first decouples temporal modeling from spatial representation learning, then integrates both through adaptive representation-level fusion. To assess robustness of the proposed approach, we construct a dataset based on an agent-based, microscopic traffic simulator (MATSim) and evaluate UniST-Pred under severe network disconnection scenarios. Additionally, we benchmark UniST-Pred on standard traffic prediction datasets, demonstrating its competitive performance against existing well-established models despite a lightweight design. The results illustrate that UniST-Pred maintains strong predictive performance across both real-world and simulated datasets, while also yielding interpretable spatio-temporal representations under infrastructure disruptions. The source code and the generated dataset are available at https://anonymous.4open.science/r/UniST-Pred-EF27

Abstract

Traffic Forecasting

Explore similar work

UniST-Pred: A Robust Unified Framework for Spatio-Temporal Traffic Forecasting in Transportation Networks Under Disruptions

SpecLST: Spectral-Enhanced Long-Short Term Spatio-Temporal Multimodal Modeling for Traffic Forecasting

Continuous Cross-Domain Traffic State Prediction via Memory-Augmented Graph Liquid Time-Constant Networks