Beyond Model Complexity: DLinear's Superior Accuracy-Efficiency Trade-off in Real-World ISP Backbone Traffic Forecasting

Deihimfar, Hamid Reza; Shahzadi, Ali

doi:10.22075/mseee.2026.41063.1269

Beyond Model Complexity: DLinear's Superior Accuracy-Efficiency Trade-off in Real-World ISP Backbone Traffic Forecasting

Articles in Press

Document Type : Research Article

Authors

Hamid Reza Deihimfar

Ali Shahzadi

Faculty of Electrical and Computer Engineering, Semnan University, Semnan, Iran.

10.22075/mseee.2026.41063.1269

Abstract

Accurate short-term network traffic forecasting is essential for Internet Service Providers (ISPs) to support capacity planning, congestion mitigation, anomaly-aware monitoring, and timely operational decision-making. Cross-validation on both an ISP-specific dataset (CESNET-TimeSeries24) and a public benchmark (Electricity) confirms that model superiority is highly dependent on data characteristics. This study challenges the prevailing assumption that more complex deep learning architectures are inherently superior for time series forecasting. Through comprehensive experiments, we demonstrate that DLinear achieves the best accuracy-efficiency trade-off for ISP backbone traffic, while NHITS and PatchTST perform better on the Electricity dataset. These findings highlight the importance of aligning model inductive bias with domain-specific temporal patterns rather than relying solely on architectural complexity. To ensure a fair and operationally meaningful comparison, we design a unified experimental pipeline that includes consistent preprocessing, chronological data splitting, repeated runs, and evaluation across both predictive accuracy and computational efficiency. The analysis covers representative forecasting paradigms, including recurrent neural networks, decomposition-based linear models, block-based architectures, and recent Transformer-inspired approaches. The results reveal that increasing architectural complexity does not necessarily lead to superior forecasting performance in ISP traffic scenarios. In particular, DLinear emerges as the most effective and practical model on the CESNET-TimeSeries24 dataset, achieving the most favorable balance between forecasting accuracy, training cost, and inference latency. Its strong performance can be attributed to its ability to explicitly model trend and seasonal components, which aligns well with the temporal regularities commonly observed in backbone traffic. These findings challenge the assumption that more sophisticated architectures are always preferable and highlight the importance of matching model inductive bias to data characteristics and deployment constraints. Overall, this study provides practical guidance for network operators seeking robust, efficient, and deployable forecasting solutions in resource-constrained ISP environments.

Keywords

Network traffic forecasting

Deep learning

Transformer architectures

PatchTST

DLinear

TimesNet

iTransformer

N-BEATS

N-HiTS

ISP

Time series prediction

CESNET-TimeSeries24

Subjects

Computer

[1] Y. Liu et al., “iTransformer: Inverted transformers are effective for time series forecasting,” arXiv preprint arXiv:2310.06625, Mar. 2024. [Online]. Available: https://arxiv.org/abs/2310.06625

[2] C. Challu, K. G. Olivares, B. N. Oreshkin, F. Garza, M. Mergenthaler-Canseco, and A. Dubrawski, “N-HiTS: Neural hierarchical interpolation for time series forecasting,” arXiv preprint arXiv:2201.12886, Nov. 2022. [Online]. Available: https://arxiv.org/abs/2201.12886

[3] B. N. Oreshkin, D. Carpov, N. Chapados, and Y. Bengio, “N-BEATS: Neural basis expansion analysis for interpretable time series forecasting,” arXiv preprint arXiv:1905.10437, Feb. 2020. [Online]. Available: https://arxiv.org/abs/1905.10437

[4] A. Mahdi and A. al Saadi, “Deep learning-based smart traffic prediction for enhanced quality of service in software-defined networking,” International Journal of Engineering, Transactions B: Applications, vol. 39, no. 3, pp. 739–755, Mar. 2026, doi: 10.5829/ije.2026.39.03c.14.

[5] L. Chen and Q. Wu, “Spatiotemporal-decoupled interactive learning for traffic flow prediction,” Scientific Reports, vol. 16, no. 1, Art. no. 9050, Feb. 2026, doi: 10.1038/s41598-026-38244-w.

[6] O. B. Sezer, M. U. Gudelek, and A. M. Ozbayoglu, “Financial time series forecasting with deep learning: A systematic literature review: 2005–2019,” Applied Soft Computing, vol. 90, Art. no. 106181, May 2020, doi: 10.1016/j.asoc.2020.106181.

[7] S.-Y. Shih, F.-K. Sun, and H. Lee, “Temporal pattern attention for multivariate time series forecasting,” Machine Learning, vol. 108, nos. 8–9, pp. 1421–1441, Sep. 2019, doi: 10.1007/s10994-019-05815-0.

[8] G. Bontempi, S. Ben Taieb, and Y.-A. Le Borgne, “Machine learning strategies for time series forecasting,” in Business Intelligence, 2013, pp. 62–77, doi: 10.1007/978-3-642-36318-4_3.

[9] G. P. Zhang, “Time series forecasting using a hybrid ARIMA and neural network model,” Neurocomputing, vol. 50, pp. 159–175, Jan. 2003, doi: 10.1016/S0925-2312(01)00702-0.

[10] G. T. Wilson, “Time Series Analysis: Forecasting and Control, 5th Edition, by George E. P. Box, Gwilym M. Jenkins, Gregory C. Reinsel and Greta M. Ljung, 2015. Published by John Wiley and Sons Inc., Hoboken, New Jersey, pp. 712. ISBN: 978-1-118-67502-1,” Journal of Time Series Analysis, vol. 37, no. 5, pp. 709–711, Sep. 2016, doi: 10.1111/jtsa.12194.

[11] K. Cho et al., “Learning phrase representations using RNN encoder–decoder for statistical machine translation,” in Proc. 2014 Conf. Empirical Methods in Natural Language Processing (EMNLP), 2014, pp. 1724–1734, doi: 10.3115/v1/D14-1179.

[12] J. Zhang, Y. Zheng, and D. Qi, “Deep spatio-temporal residual networks for citywide crowd flows prediction,” in Proc. AAAI Conf. Artificial Intelligence, vol. 31, no. 1, Feb. 2017, doi: 10.1609/aaai.v31i1.10735.

[13] S. Guo, Y. Lin, N. Feng, C. Song, and H. Wan, “Attention based spatial-temporal graph convolutional networks for traffic flow forecasting,” in Proc. AAAI Conf. Artificial Intelligence, vol. 33, no. 1, pp. 922–929, Jul. 2019, doi: 10.1609/aaai.v33i01.3301922.

[14] H. Nguyen Nguyen et al., “ConEm: A novel framework for integrating external factors with inner and outer correlations in time series forecasting,” Knowledge-Based Systems, vol. 329, Art. no. 114312, Nov. 2025, doi: 10.1016/j.knosys.2025.114312.

[15] K. Bandara, C. Bergmeir, and S. Smyl, “Forecasting across time series databases using recurrent neural networks on groups of similar series: A clustering approach,” Expert Systems with Applications, vol. 140, Art. no. 112896, Feb. 2020, doi: 10.1016/j.eswa.2019.112896.

[16] Q. Wen et al., “Time series data augmentation for deep learning: A survey,” in Proc. 30th Int. Joint Conf. Artificial Intelligence (IJCAI), Aug. 2021, pp. 4653–4660, doi: 10.24963/ijcai.2021/631.

[17] B. Lim, S. Ö. Arık, N. Loeff, and T. Pfister, “Temporal fusion transformers for interpretable multi-horizon time series forecasting,” International Journal of Forecasting, vol. 37, no. 4, pp. 1748–1764, Oct. 2021, doi: 10.1016/j.ijforecast.2021.03.012.

[18] D. Salinas, V. Flunkert, J. Gasthaus, and T. Januschowski, “DeepAR: Probabilistic forecasting with autoregressive recurrent networks,” International Journal of Forecasting, vol. 36, no. 3, pp. 1181–1191, Jul. 2020, doi: 10.1016/j.ijforecast.2019.07.001.

[19] G. Lai, W.-C. Chang, Y. Yang, and H. Liu, “Modeling long- and short-term temporal patterns with deep neural networks,” in Proc. 41st Int. ACM SIGIR Conf. Research and Development in Information Retrieval, Jun. 2018, pp. 95–104, doi: 10.1145/3209978.3210006.

[20] A. V. Gulyaev, “Hybrid deep learning model for time series forecasting,” Neurocomputers, pp. 62–72, Mar. 2025, doi: 10.18127/j19998554-202503-08.

[21] Z. Li, Y. Zhang, Z. Zhang, X. Wang, and L. Zhu, “Adaptive spatial-temporal convolution network for traffic forecasting,” 2022, pp. 287–299, doi: 10.1007/978-3-031-10986-7_23.

[22] W. Li and K. L. E. Law, “Deep learning models for time series forecasting: A review,” IEEE Access, vol. 12, pp. 92306–92327, 2024, doi: 10.1109/ACCESS.2024.3422528.

[23] T. Wu, F. Chen, and Y. Wan, “Graph attention LSTM network: A new model for traffic flow forecasting,” in Proc. 5th Int. Conf. Information Science and Control Engineering (ICISCE), Jul. 2018, pp. 241–245, doi: 10.1109/ICISCE.2018.00058.

[24] Y. Zhao, Y. Xu, X. He, and D. Zhang, “Spatiotemporal graph attention networks for urban traffic flow prediction,” in Proc. IEEE 33rd Annual Int. Symp. Personal, Indoor and Mobile Radio Communications (PIMRC), Sep. 2022, pp. 340–345, doi: 10.1109/PIMRC54779.2022.9977794.

[25] Y. Okamoto, “Transformers for time-series forecasting: A comprehensive survey through 2024,” Nonlinear Theory and Its Applications, IEICE, vol. 17, no. 1, pp. 260–278, 2026, doi: 10.1587/nolta.17.260.

[26] D. Ge, B. Liu, Y. Cheng, and Z. Zhu, “Rethinking decomposition for time series forecasting: Learning from temporal patterns,” Expert Systems with Applications, vol. 287, Art. no. 128044, Aug. 2025, doi: 10.1016/j.eswa.2025.128044.

[27] Y. Cheng, G. Li, X. Zhou, and S. Ye, “Research on time series forecasting models based on hybrid attention mechanism and graph neural networks,” Informatica, vol. 49, no. 21, May 2025, doi: 10.31449/inf.v49i21.7580.

[28] N. Maragkos and I. Refanidis, “A comparative evaluation of time-series forecasting models for energy datasets,” Computers, vol. 14, no. 7, Art. no. 246, Jun. 2025, doi: 10.3390/computers14070246.

[29] I. Alawe, A. Ksentini, Y. Hadjadj-Aoul, and P. Bertin, “Improving traffic forecasting for 5G core network scalability: A machine learning approach,” IEEE Network, vol. 32, no. 6, pp. 42–49, Nov. 2018, doi: 10.1109/MNET.2018.1800104.

[30] A. Oluwasanmi, M. Aftab, Z. Qin, M. Sarfraz, Y. Yu, and H. Rauf, “Multi-head spatiotemporal attention graph convolutional network for traffic prediction,” Sensors, vol. 23, no. 8, Art. no. 3836, Apr. 2023, doi: 10.3390/s23083836.

[31] K. Papagiannaki, N. Taft, Z.-L. Zhang, and C. Diot, “Long-term forecasting of Internet backbone traffic: Observations and initial models,” in Proc. IEEE INFOCOM 2003, vol. 2, 2003, pp. 1178–1188, doi: 10.1109/INFCOM.2003.1208954.

[32] C. Wu, H. Ding, Z. Fu, and N. Sun, “Multi-scale spatio-temporal attention networks for network-scale traffic learning and forecasting,” Sensors, vol. 24, no. 17, Art. no. 5543, Aug. 2024, doi: 10.3390/s24175543.

[33] S. Troia, R. Alvizu, Y. Zhou, G. Maier, and A. Pattavina, “Deep learning-based traffic prediction for network optimization,” in Proc. 20th Int. Conf. Transparent Optical Networks (ICTON), Jul. 2018, pp. 1–4, doi: 10.1109/ICTON.2018.8473978.

[34] H. Mezaache, H. Bouzgou, C. Raymond, and N. Zemouri, “A novel approach for accurate wind speed time series forecasting using ICEEMDAN decomposition and sample entropy through integration of deep learning models,” International Journal of Engineering, Transactions B: Applications, vol. 39, no. 2, pp. 309–320, Feb. 2026, doi: 10.5829/ije.2026.39.02b.03.

[35] I. Boukrouh, S. Idiri, F. Tayalati, A. Azmani, and L. Bouhsaien, “A hybrid approach for sales forecasting: Combining deep learning and time series analysis,” International Journal of Engineering, Transactions A: Basics, vol. 38, no. 4, pp. 859–870, Apr. 2025, doi: 10.5829/ije.2025.38.04a.16.