LGA-EL: a spatio-temporal adaptive ensemble method with local-global awareness for traffic prediction

doi:10.11947/j.AGCS.2026.20250340

Abstract

Abstract:

Traffic prediction is a core requirement for building intelligent transportation systems. In complex traffic scenarios, different prediction models exhibit significant performance variations across spatial regions and time periods, making it difficult for any single model to stably adapt to diverse prediction demands. Existing ensemble learning methods enhance prediction stability by leveraging the strengths of multiple models. However, they typically rely on globally fixed or locally optimal ensemble strategies, which overlook the synergistic constraints of global spatio-temporal correlations and spatio-temporal heterogeneity in the ensemble process, limiting the predictive performance and generalization ability. Therefore, this study proposes a spatiotemporal adaptive ensemble learning method with local-global awareness (LGA-EL) for traffic prediction tasks, which optimizes the performance of the base models under different traffic conditions by adaptively adjusting the ensemble parameters. The method first embeds road network topology and traffic state evolution characteristics to jointly capture spatio-temporal information of monitoring stations from both local and global perspectives, thereby collaboratively representing the spatio-temporal correlation and heterogeneity in the ensemble process. Based on the embedding vectors, the method adaptively solves the ensemble parameters for each spatio-temporal location and dynamically weights the output features of the base models to generate the final prediction. Experiments on short-term and long-term prediction tasks for traffic flow and speed demonstrate that the proposed method outperforms six mainstream ensemble prediction methods in terms of both prediction accuracy and computational efficiency. Further interpretability analysis shows that the method can accurately capture performance differences among models under different traffic states, leveraging the strengths of various models through adaptive ensemble weights, significantly enhancing the performance and robustness of ensemble learning in traffic prediction tasks.

Key words: traffic prediction, ensemble learning, spatio-temporal awareness, global spatio-temporal correlation, spatio-temporal heterogeneity

CLC Number:

P208

Lizeng WANG, Shifen CHENG, Yitao YANG, Peixiao WANG, Feng LU. LGA-EL: a spatio-temporal adaptive ensemble method with local-global awareness for traffic prediction[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(2): 206-221.

Figures/Tables 17

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Tab. 1

Tab. 2

Tab. 3

Tab. 4

Fig. 5

Tab. 5

Tab. 6

Fig. 6

Fig. 7

Fig. 8

Fig. 9

Fig. 10

Fig. 11

References 37

[1]	RAHMANI S, BAGHBANI A, BOUGUILA N, et al. Graph neural networks for intelligent transportation systems: a survey[J]. IEEE Transactions on Intelligent Transportation Systems, 2023, 24(8): 8846-8885.
[2]	王建龙. 基于图过程神经网络的城市道路拥堵时空预测[J]. 测绘学报, 2024, 53(8): 1657. DOI: . doi: 10.11947/j.AGCS.2024.20230167
	WANG Jianlong. Spatio-temporal prediction of urban road congestion based on graph process neural network[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(8): 1657. DOI: . doi: 10.11947/j.AGCS.2024.20230167
[3]	程涛, 张洋, James Haworth. 基于网络和图的时空智能：概念、方法和应用[J]. 测绘学报, 2022, 51(7): 1629-1639. DOI: . doi: 10.11947/j.AGCS.2022.20220236
	CHENG Tao, ZHANG Yang, HAWORTH James. Network and graph-based SpaceTimeAI: conception, method and applications[J]. Acta Geodaetica et Cartographica Sinica, 2022, 51(7): 1629-1639. DOI: . doi: 10.11947/j.AGCS.2022.20220236
[4]	CUI Haipeng, MENG Qiang, TENG T H, et al. Spatiotemporal correlation modelling for machine learning-based traffic state predictions: state-of-the-art and beyond[J]. Transport Reviews, 2023, 43(4): 780-804.
[5]	王培晓, 张恒才, 张岩, 等. 地理空间智能预测研究进展与发展趋势[J]. 地球信息科学学报, 2025, 27(1): 60-82.
	WANG Peixiao, ZHANG Hengcai, ZHANG Yan, et al. GeoAI-driven spatiotemporal prediction: progress and prospects[J]. Journal of Geo-information Science, 2025, 27(1): 60-82.
[6]	WOLPERT D H, MACREADY W G. No free lunch theorems for optimization[J]. IEEE Transactions on Evolutionary Computation, 1997, 1(1): 67-82.
[7]	DU Zhenhong, WU Sensen, KWAN M P, et al. A spatiotemporal regression-Kriging model for space-time interpolation: a case study of Chlorophyll-a prediction in the coastal areas of Zhejiang, China[J]. International Journal of Geographical Information Science, 2018, 32(10): 1927-1947.
[8]	邓敏, 彭翀, 谌恺祺. 顾及地理空间特性的空间面板数据可预测性评估理论与方法[J]. 测绘学报, 2025, 54(7): 1305-1317. DOI: . doi: 10.11947/j.AGCS.2025.20240448
	DENG Min, PENG Chong, CHEN Kaiqi. A predictability measurement methodology for spatial panel data considering geo-spatial effects[J]. Acta Geodaetica et Cartographica Sinica, 2025, 54(7): 1305-1317. DOI: . doi: 10.11947/j.AGCS.2025.20240448
[9]	郭浩, 董磊, 邬伦, 等. 空间异质性建模方法[J]. 地理学报, 2025, 80(3): 567-585.
	GUO Hao, DONG Lei, WU Lun, et al. Tackling spatial heterogeneity in geographical analysis: an overview[J]. Acta Geographica Sinica, 2025, 80(3): 567-585.
[10]	MALLICK T, MACFARLANE J, BALAPRAKASH P. Uncertainty quantification for traffic forecasting using deep-ensemble-based spatiotemporal graph neural networks[J]. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(8): 9141-9152.
[11]	LUO Yun, SU Shiliang. Spatiotemporal random forest and spatiotemporal stacking tree: a novel spatially explicit ensemble learning approach to modeling non-linearity in spatiotemporal non-stationarity[J]. International Journal of Applied Earth Observation and Geoinformation, 2025, 136: 104315.
[12]	SHAFIZADEH-MOGHADAM H, VALAVI R, SHAHABI H, et al. Novel forecasting approaches using combination of machine learning and statistical models for flood susceptibility mapping[J]. Journal of Environmental Management, 2018, 217: 1-11.
[13]	YAO Ronghan, ZHANG Wensong, ZHANG Lihui. Hybrid methods for short-term traffic flow prediction based on ARIMA-GARCH model and wavelet neural network[J]. Journal of Transportation Engineering, Part A: Systems, 2020, 146(8): 04020086.
[14]	GU Yuanli, LU Wenqi, XU Xinyue, et al. An improved Bayesian combination model for short-term traffic prediction with deep learning[J]. IEEE Transactions on Intelligent Transportation Systems, 2020, 21(3): 1332-1342.
[15]	CHEN Yuanyuan, CHEN Hongyu, YE Peijun, et al. Acting as a decision maker: traffic-condition-aware ensemble learning for traffic flow prediction[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(4): 3190-3200.
[16]	SHANG Pan, LIU Xinwei, YU Chengqing, et al. A new ensemble deep graph reinforcement learning network for spatio-temporal traffic volume forecasting in a freeway network[J]. Digital Signal Processing, 2022, 123: 103419.
[17]	MI Xiwei, YU Chengqing, LIU Xinwei, et al. A dynamic ensemble deep deterministic policy gradient recursive network for spatiotemporal traffic speed forecasting in an urban road network[J]. Digital Signal Processing, 2022, 129: 103643.
[18]	LIU Yang, LIU Zhiyuan, VU H L, et al. A spatio-temporal ensemble method for large-scale traffic state prediction[J]. Computer-Aided Civil and Infrastructure Engineering, 2020, 35(1): 26-44.
[19]	WANG Lizeng, CHENG Shifen, LU Feng. Decomposing spatio-temporal heterogeneity: matrix-informed ensemble learning for interpretable prediction[J]. Knowledge-Based Systems, 2025, 309: 112906.
[20]	程诗奋, 彭澎, 张恒才, 等. 异质稀疏分布时空数据插值、重构与预测方法探讨[J]. 武汉大学学报(信息科学版), 2020, 45(12): 1919-1929.
	CHENG Shifen, PENG Peng, ZHANG Hengcai, et al. Review of interpolation, reconstruction and prediction methods for heterogeneous and sparsely distributed geospatial data[J]. Geomatics and Information Science of Wuhan University, 2020, 45(12): 1919-1929.
[21]	CHENG Shifen, WANG Lizeng, WANG Peixiao, et al. An ensemble spatial prediction method considering geospatial heterogeneity[J]. International Journal of Geographical Information Science, 2024, 38(9): 1856-1880.
[22]	LI Longxiang, BLOMBERG A J, STERN R A, et al. Predicting monthly community-level domestic radon concentrations in the greater Boston area with an ensemble learning model[J]. Environmental Science & Technology, 2021, 55(10): 7157-7166.
[23]	GUO Fangce, POLAK J W, KRISHNAN R. Predictor fusion for short-term traffic forecasting[J]. Transportation Research Part C: Emerging Technologies, 2018, 92: 90-100.
[24]	朱阿兴, 闾国年, 周成虎, 等. 地理相似性：地理学的第三定律?[J]. 地球信息科学学报, 2020, 22(4): 673-679.
	ZHU Axing, LÜ Guonian, ZHOU Chenghu, et al. Geographic similarity: third law of geography?[J]. Journal of Geo-information Science, 2020, 22(4): 673-679.
[25]	LI Longxiang, BLOMBERG A J, LAWRENCE J, et al. A spatiotemporal ensemble model to predict gross beta particulate radioactivity across the contiguous United States[J]. Environment International, 2021, 156: 106643.
[26]	WANG Jingyuan, JIANG Jiawei, JIANG Wenjun, et al. LibCity: an open library for traffic prediction[C]//Proceedings of the 29th International Conference on Advances in Geographic Information Systems. New York: ACM Press, 2021: 145-148.
[27]	YU Bing, YIN Haoteng, ZHU Zhanxing. Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting[C]//Proceedings of the 27th International Joint Conference on Artificial Intelligence. Stockholm: [s.n.], 2018: 3634-3640.
[28]	LI Yaguang, YU R, SHAHABI C, et al. Diffusion convolutional recurrent neural network: data-driven traffic forecasting[EB/OL]. [2025-12-18]. https://arxiv.org/abs/1707.0192.
[29]	ZHENG Chuanpan, FAN Xiaoliang, WANG Cheng, et al. GMAN: a graph multi-attention network for traffic prediction[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(1): 1234-1241.
[30]	BELKIN M, NIYOGI P. Laplacian eigenmaps and spectral techniques for embedding and clustering[C]//Proceedings of 2001 Neural Information Processing Systems Conference. Cambridge: The MIT Press, 2001: 585-591.
[31]	HUBER P J. Robust estimation of a location parameter[M]//Breakthroughs in statistics: methodology and distribution. New York: Springer, 1992: 492-518.
[32]	XU Yi, HAN Liangzhe, ZHU Tongyu, et al. Generic dynamic graph convolutional network for traffic flow forecasting[J]. Information Fusion, 2023, 100: 101946.
[33]	SHEN Jinxing, LIU Qinxin, FENG Xuejun. Hourly PM_2.5 concentration prediction for dry bulk port clusters considering spatiotemporal correlation: a novel deep learning blending ensemble model[J]. Journal of Environmental Management, 2024, 370: 122703.
[34]	WU Zonghan, PAN Shirui, LONG Guodong, et al. Connecting the dots: multivariate time series forecasting with graph neural networks[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York: ACM Press, 2020: 753-763.
[35]	GUO Shengnan, LIN Youfang, FENG Ning, et al. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting[C]//Proceedings of the 33rd AAAI Conference on Artificial Intelligence and 31st Innovative Applications of Artificial Intelligence Conference and 9th AAAI Symposium on Educational Advances in Artificial Intelligence. New York: ACM Press, 2019: 922-929.
[36]	BAI Lei, YAO Lina, LI Can, et al. Adaptive graph convolutional recurrent network for traffic forecasting[C]//Proceedings of the 34th International Conference on Neural Information Processing Systems. New York: ACM Press, 2020: 17804-17815.
[37]	FANG Zhice, WANG Yi, PENG Ling, et al. A comparative study of heterogeneous ensemble-learning techniques for landslide susceptibility mapping[J]. International Journal of Geographical Information Science, 2021, 35(2): 321-347.

数据集名称	预测变量	研究区域	时间间隔/min	监测站点数量	时间步数量	时间跨度
TR-Wuhan	交通流量/辆	中国武汉市	5	175	8064	2021-02-01—2021-02-28
PEMS07（M）	交通速度/（km/h）	美国加利福尼亚州	5	228	12 672	2012-05-01—2012-06-30

模型结构	参数名称	最优设置
局部-全局时空感知模块	局部嵌入向量维度e_l	24/24
	全局嵌入向量维度e_g	32/48
	多尺度时空融合层结构	（64）/（64）
时空自适应加权模块	时空自适应加权网络结构	（128，32）/（256，64）
训练策略	损失函数超参数δ	0.1/10
	优化器	Adam
	随机失活率	0.2/0.1
	批处理大小	64
	学习率	5×10^－5/1×10^－4

类型	方法	短期预测任务（15 min）			长期预测任务（60 min）
类型	方法	MAE/辆	RMSE/辆	MAPE/（%）	MAE/辆	RMSE/辆	MAPE/（%）
基模型	STGCN	4.079	7.445	16.869	7.756	15.259	28.210
	DCRNN	4.353	8.041	17.065	7.155	14.348	27.062
	GMAN	4.862	8.926	18.887	7.558	14.831	28.201
其他单一模型	MTGNN	4.014	7.551	16.812	7.070	14.133	26.402
	ASTGCN	3.992	7.313	16.755	6.986	13.799	25.931
	AGCRN	3.963	7.271	16.193	6.891	13.151	25.314
集成方法	Avg-EL	4.063	7.438	15.866	6.731	13.296	24.749
	AW-EL	4.053	7.420	15.835	6.717	13.259	24.747
	LinReg-EL	4.010	7.379	15.856	6.689	13.316	24.792
	GTWR-EL	3.969	7.280	15.653	6.651	13.185	24.660
	DDPG-EL	3.962	7.292	15.711	6.647	13.232	24.772
	MI-EL	3.941	7.228	15.406	6.606	13.082	24.563
	LGA-EL	3.915	7.166	15.283	6.495	12.785	23.615

类型	方法	短期预测任务（15 min）			长期预测任务（60 min）
类型	方法	MAE/（km/h）	RMSE/（km/h）	MAPE/（%）	MAE/（km/h）	RMSE/（km/h）	MAPE/（%）
基模型	STGCN	1.799	3.286	4.133	2.984	5.825	7.315
	DCRNN	1.780	3.386	4.023	2.993	6.094	7.364
	GMAN	1.826	3.264	4.206	2.837	5.257	7.028
其他单一模型	MTGNN	1.716	3.144	4.170	2.758	5.291	6.723
	ASTGCN	1.765	3.221	4.288	3.190	5.693	7.957
	AGCRN	1.683	3.160	3.958	2.696	5.371	6.708
集成方法	Avg-EL	1.718	3.164	3.930	2.813	5.273	6.790
	AW-EL	1.720	3.169	3.927	2.802	5.224	6.768
	LinReg-EL	1.723	3.158	3.927	2.797	5.204	6.635
	GTWR-EL	1.703	3.147	3.901	2.740	5.196	6.601
	DDPG-EL	1.695	3.122	3.859	2.771	5.193	6.624
	MI-EL	1.681	3.114	3.856	2.718	5.122	6.599
	LGA-EL	1.664	3.091	3.776	2.658	5.038	6.480

数据集	方法	短期预测任务（15 min）			长期预测任务（60 min）
数据集	方法	MAE	RMSE	MAPE	MAE	RMSE	MAPE
TR-Wuhan	Avg-EL	3.645	3.654	3.675	3.499	3.848	4.583
	AW-EL	3.395	3.414	3.486	3.302	3.577	4.575
	LinReg-EL	2.368	2.885	3.613	2.889	3.989	4.749
	GTWR-EL	1.355	1.567	2.363	2.338	3.035	4.238
	DDPG-EL	1.181	1.728	2.724	2.280	3.379	4.671
	MI-EL	0.654	0.853	0.798	1.532	1.679	3.473
	平均提升	2.100	2.350	2.776	2.640	3.251	4.382
PEMS07（M）	Avg-EL	3.135	2.306	3.939	5.504	4.471	4.567
	AW-EL	3.218	2.456	3.865	5.134	3.569	4.252
	LinReg-EL	3.414	2.140	3.846	4.975	3.200	2.330
	GTWR-EL	2.261	1.787	3.213	2.987	3.046	1.830
	DDPG-EL	1.800	1.001	2.160	4.072	2.990	2.171
	MI-EL	0.982	0.746	2.083	2.202	1.646	1.801
	平均提升	2.468	1.740	3.184	4.146	3.154	2.825