基于可解释扩散模型的多变量时间序列生成方法

doi:10.19678/j.issn.1000-3428.0070676

摘要/Abstract

摘要： 准确地生成多变量时间序列(MTS)数据为解决数据规模不足问题提供了有效途径,对电力负荷预测和风光发电评估等下游任务至关重要。然而,现有方法难以同时捕捉长短期依赖与变量间关联关系,且缺乏可解释性,无法满足能源系统分析需求。为此,提出了一种基于可解释扩散模型的MTS生成方法IDMTS。首先,在扩散模型的去噪网络中引入了包含三重注意力(TA)的Transformer架构,以捕捉长短期依赖与变量特征关联关系;接着,结合多尺度趋势季节分解,利用双向长短期记忆(BiLSTM)网络和傅里叶注意力(FA)分别建模趋势项和季节项,提升生成数据的准确性和可解释性。同时,通过多尺度自适应最大均值差异(Ada-MMD)损失函数优化生成质量。实验结果表明,IDMTS在4个公开数据集上的生成准确性显著优于基线方法,其中Context-FID得分、相关性得分、判别得分和预测得分分别降低了51.5%~84.5%、4.1%~26.8%、24.0%~68.8%、0.3%~40.0%。同时,在可解释性实验、条件插补和预测实验中,IDMTS展现出良好的可解释性和泛化能力。

关键词: 时间序列生成, 扩散模型, 可解释性, Transformer, 多尺度分解

Abstract: The accurate generation of Multivariate Time Series (MTS) data provides an effective way to solve the problem of insufficient data scale and is crucial for downstream tasks such as power load forecasting and wind and solar power generation evaluation. However, existing methods cannot capture the long- and short-term dependence and correlation between variables easily, and they cannot meet the requirements of energy system analysis owing to a lack of interpretability. Therefore, an interpretable diffusion model for multivariate time series generation method IDMTS is proposed. First, a transformer architecture containing Triplet Attention (TA) is introduced into the denoising network of the diffusion model to capture long- and short-term dependencies and variable feature associations. Subsequently, combined with multiscale trend seasonal decomposition, the trend and season terms are modeled using Bidirectional Long Short-Term Memory (BiLSTM) network and Fourier Attention (FA), respectively. This improves the accuracy and interpretability of generated data. Simultaneously, the generation quality is optimized by multiscale Adaptive Maximum Mean Discrepancy (Ada-MMD) loss function. The experimental results show that the generation accuracy of IDMTS on four public datasets is significantly better than that of the baseline method, in which the Context-FID score, correlation score, discriminant score, and prediction score are reduced by 51.5% to 84.5%, 4.1% to 26.8%, 24.0% to 68.8%, and 0.3% to 40.0%, respectively. Additionally, IDMTS shows good interpretability and generalization ability in the interpretability, conditional interpolation, and prediction experiments.

Key words: time series generation, diffusion model, interpretability, Transformer, multiscale decomposition

中图分类号:

TP391

朱莉, 徐婉茹, 高靖凯, 朱春强, 邓凡. 基于可解释扩散模型的多变量时间序列生成方法[J]. 计算机工程, 2026, 52(10): 138-153.

ZHU Li, XU Wanru, GAO Jingkai, ZHU Chunqiang, DENG Fan. Multivariate Time Series Generation Method Based on Interpretable Diffusion Model[J]. Computer Engineering, 2026, 52(10): 138-153.

https://www.ecice06.com/CN/Y2026/V52/I10/138

参考文献

[1] 许青, 张龄之, 梁琛, 等. 基于联合时序场景和改进TCN的高比例新能源电网负荷预测[J]. 广东电力, 2024, 37(1): 1-7. XU Q, ZHANG L Z, LIANG C, et al. Short-term load forecasting for power system with high proportion new energy based on joint sequential scenario and improved TCN[J]. Guangdong Electric Power, 2024, 37(1): 1-7. (in Chinese)
[2] 王保义, 邓晓平, 张桂青, 等. 基于高低频数据融合的非侵入式负荷监测系统设计与应用[J]. 建筑科学, 2024, 40(10): 62-69, 142. WANG B Y, DENG X P, ZHANG G Q, et al. Design and application of a non-intrusive load monitoring system based on the fusion of high-and low-frequency data[J]. Building Science, 2024, 40(10): 62-69, 142. (in Chinese)
[3] HO J, JAIN A, ABBEEL P. Denoising diffusion probabilistic models[EB/OL].[2024-11-01]. https://arxiv.org/abs/2006.11239.
[4] KONG Z F, PING W, HUANG J J, et al. DiffWave: a versatile diffusion model for audio synthesis[C]//Proceedings of International Conference on Learning Representations (ICLR). New York, USA: ACM Press, 2021: 101-108.
[5] LIM H, KIM M, PARK S, et al. Regular time-series generation using SGM[EB/OL]. [2024-11-01]. https://arxiv.org/abs/2301.08518.
[6] KOLLOVIEH M, ANSARI A F, BOHLKE-SCHNEIDER M, et al. Predict, refine, synthesize: self-guiding diffusion models for probabilistic time series forecasting[C]//Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS 2023). New Orleans, USA: [s.n], 2023:55-59.
[7] COLETTA A, GOPALAKRISHAN S, BORRAJO D, et al. On the constrained time-series generation problem[C]//Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS 2023). New Orleans, USA:[s.n], 2023: 93-96.
[8] FONS E, SZTRAJMAN A, EL-LAHAM Y, et al. HyperTime: implicit neural representation for time series[EB/OL]. [2024-11-01]. https://arxiv.org/abs/2208.05836.
[9] REN L, WANG H T, LAILI Y. Diff-MTS: temporal-augmented conditional diffusion-based AIGC for industrial time series towards the large model era[J]. IEEE Transactions on Cybernetics, 2024, 54(12): 7187-7197.
[10] GOODFELLOW I J, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial networks[EB/OL]. [2024-11-01]. https://arxiv.org/abs/1406.2661.
[11] MOGREN O. C-RNN-GAN: continuous recurrent neural networks with adversarial training[C]//Proceedings of Constructive Machine Learning Workshop. Barcelona, Spain: [s.n], 2016: 45-48.
[12] LOHACHOV M, KOREI R, OKI K, et al. RNN-based approach for broccoli harvest time forecast[J]. Agronomy, 2024, 14(2): 361.
[13] MASOOD Z, GANTASSI R, CHOI Y. Enhancing short-term electric load forecasting for households using quantile LSTM and clustering-based probabilistic approach[J]. IEEE Access, 2024, 12: 77257-77268.
[14] YOON J, JARRETT D, VAN DER SCHAAR M. Time-series generative adversarial networks[C] //Proceedings of Advances in Neural Information Processing Systems. Vancouver, Canada: Neural Information Processing Systems Foundation, Inc., 2019: 5508-5518.
[15] JARRETT D, BICA I, VAN DER SCHAAR M. Time-series generation by contrastive imitation[C]//Proceedings of Advances in Neural Information Processing Systems. [S.l.]: Neural Information Processing Systems Foundation, Inc., 2021: 28968-28982.
[16] DESAI A, FREEMAN C, WANG Z H, et al. TimeVAE: a variational auto-encoder for multivariate time series generation[EB/OL]. [2024-11-01]. https://arxiv.org/abs/2111.08095.
[17] KINGMA D P, WELLING M. Auto-encoding variational Bayes[C]//Proceedings of the 2nd International Conference on Learning Representations (ICLR) Workshop Track. Banff, Canada: International Machine Learning Society, 2014: 87-95.
[18] RASUL K, SEWARD C, SCHUSTER I, et al. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting[C]//Proceedings of the 38th International Conference on Machine Learning. New York, USA: International Machine Learning Society, 2021: 8857-8868.
[19] TASHIRO Y, SONG J M, SONG Y, et al. CSDI: conditional score-based diffusion models for probabilistic time series imputation[C]//Proceedings of Advances in Neural Information Processing Systems. New York, USA: Neural Information Processing Systems Foundation, Inc., 2021: 24804-24816.
[20] ALCARAZ J M L, STRODTHOFF N. Diffusion-based time series imputation and forecasting with structured state space models[EB/OL]. [2024-11-01]. https://arxiv.org/abs/2208.09399.
[21] LI Y X, CHEN W C, HU X Y, et al. Transformer- modulated diffusion models for probabilistic multivariate time series forecasting[C]//Proceedings the 12th International Conference on Learning Representations. Vienna, Austria: ICLR, 2024:1125-1229.
[22] LIN N, PALENSKY P, VERGARA P P. EnergyDiff: universal time-series energy data generation using diffusion models [J]. IEEE Transactions on Smart Grid, 2025, 16(5): 4252-4265.
[23] WEN Q S, ZHANG Z, LI Y, et al. Fast RobustSTL: efficient and robust seasonal-trend decomposition for time series with complex patterns[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York, USA: ACM Press, 2020: 2203-2213.
[24] DOKUMENTOV A, HYNDMAN R J. STR: seasonal-trend decomposition using regression[J]. INFORMS Journal on Data Science, 2022,1(1): 50-62.
[25] WU H X, XU J H, WANG J M, et al. Autoformer: decomposition transformers with auto-correlation for long-term series forecasting[C]//Proceedings of the 35th Conference on Neural Information Processing Systems. New York, USA: [s.n.], 2021,34:22419- 22430.
[26] ZHANG X Y, JIN X Y, GOPALSWAMY K, et al. First de-trend then attend:rethinking attention for time-series forecasting[C]//Proceedings of the 36th Conference on Neural Information Processing Systems. New York, USA: [s.n.], 2022: 89-93.
[27] MISRA D, NALAMADA T, ARASANIPALAI A U, et al. Rotate to attend: convolutional triplet attention module[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. Washington D.C., USA: IEEE Press, 2021: 3139-3148.
[28] GRETTON A, BORGWARDT K, RASCH M J, et al. A kernel method for the two-sample problem [J]. Journal of Machine Learning Research, 2012, 13(25): 723-773.
[29] PAUL J, MICHAEL B S, PEDRO M, et al. PSA-GAN: progressive self attention GANs for synthetic time series[EB/OL]. [2024-11-01]. https://arxiv.org/abs/2108.00981.
[30] LIAO S J, NI H, SZPRUCH L, et al. Conditional Sig-Wasserstein GANs for time series generation[J]. Mathematical Finance, 2022, 2: 622-670.
[31] YUE Z H, WANG Y J, DUAN J Y, et al. TS2Vec: towards universal representation of time series[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2022, 36(8): 8980-8987.
[32] XU T L, WENLIANG L K, MUNN M, et al. Cot-GAN: generating sequential data via causal optimal transport[EB/OL].[2024-11-01]. http://arxiv.org/pdf/2006.08571.
[33] 孟祥福, 石皓源. 基于Transformer模型的时序数据预测方法综述[J]. 计算机科学与探索, 2025, 19(1): 45-64. MENG X F, SHI H Y. Survey of Transformer-based model for time series forecasting[J]. Journal of Frontiers of Computer Science & Technology, 2025, 19(1): 45-64. (in Chinese)

选择文件类型/文献管理软件名称

选择包含的内容