Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Comparative Analysis of Deep Learning Models for Real-World ISP Network Traffic Forecasting

Josef Koumar, Timotej Smoleň, Kamil Jeřábek, Tomáš Čejka

TL;DR

The paper evaluates seven deep learning models for real-world ISP network traffic forecasting using the CESNET-TimeSeries24 dataset, spanning 40 weeks of multivariate time series across three aggregation levels. By modeling each metric univariately at hourly resolution and applying a standardized preprocessing and sliding-window protocol, it provides a reproducible benchmark and a multi-faceted comparison of accuracy and deployability. The study finds that GRU and LSTM variants achieve the strongest overall performance, with GRU-FCN delivering a favorable balance between accuracy and speed, while models like InceptionTime and ResNet underperform in this domain. It also demonstrates how granularity, forecast horizon, and metric type influence predictive success and highlights actionable insights for deploying forecasting systems in real-world ISP environments.

Abstract

Accurate network traffic forecasting is essential for Internet Service Providers (ISP) to optimize resources, enhance user experience, and mitigate anomalies. This study evaluates state-of-the-art deep learning models on CESNET-TimeSeries24, a recently published, comprehensive real-world network traffic dataset from the ISP network CESNET3 spanning multivariate time series over 40 weeks. Our findings highlight the balance between prediction accuracy and computational efficiency across different levels of network granularity. Additionally, this work establishes a reproducible methodology that facilitates direct comparison of existing approaches, explores their strengths and weaknesses, and provides a benchmark for future studies using this dataset.

Comparative Analysis of Deep Learning Models for Real-World ISP Network Traffic Forecasting

TL;DR

The paper evaluates seven deep learning models for real-world ISP network traffic forecasting using the CESNET-TimeSeries24 dataset, spanning 40 weeks of multivariate time series across three aggregation levels. By modeling each metric univariately at hourly resolution and applying a standardized preprocessing and sliding-window protocol, it provides a reproducible benchmark and a multi-faceted comparison of accuracy and deployability. The study finds that GRU and LSTM variants achieve the strongest overall performance, with GRU-FCN delivering a favorable balance between accuracy and speed, while models like InceptionTime and ResNet underperform in this domain. It also demonstrates how granularity, forecast horizon, and metric type influence predictive success and highlights actionable insights for deploying forecasting systems in real-world ISP environments.

Abstract

Accurate network traffic forecasting is essential for Internet Service Providers (ISP) to optimize resources, enhance user experience, and mitigate anomalies. This study evaluates state-of-the-art deep learning models on CESNET-TimeSeries24, a recently published, comprehensive real-world network traffic dataset from the ISP network CESNET3 spanning multivariate time series over 40 weeks. Our findings highlight the balance between prediction accuracy and computational efficiency across different levels of network granularity. Additionally, this work establishes a reproducible methodology that facilitates direct comparison of existing approaches, explores their strengths and weaknesses, and provides a benchmark for future studies using this dataset.

Paper Structure

This paper contains 14 sections, 3 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Dataset
  4. Deep Learning models for forecasting
  5. Methodology
  6. Missing values handling
  7. Time series splitting, normalization, and sliding window
  8. Models optimization
  9. Evaluation measures
  10. Experimental forecasting results
  11. Evaluation of models performance
  12. Deployability
  13. Discussion
  14. Conclusion

Figures (3)

  • Figure 1: Used methodology applied on each time series metric
  • Figure 2: Illustration of time series partitioning into training, validation, and test sets, along with the sliding window approach for generating overlapping training and prediction windows.
  • Figure 3: Average across all models of overall mean Harmonic score for each time series metric