[Paper Review] Data augmentation through multivariate scenario forecasting in Data Centers using Generative Adversarial Networks

Jaime Pérez, Patricia Arroba+1 more|arXiv (Cornell University)|Jan 12, 2022

Energy Load and Power Forecasting36 references26 citations

TL;DR

This paper proposes a GAN-based data augmentation framework for data center time-series data that enables multivariate scenario forecasting and on-demand anomaly generation. By leveraging conditional GANs and novel validation metrics (KL Divergence and MSE), the method generates realistic, heterogeneous synthetic data that improves model robustness and predictive performance without compromising system integrity, achieving near-baseline accuracy even when training on synthetic data alone.

ABSTRACT

The Cloud paradigm is at a critical point in which the existing energy-efficiency techniques are reaching a plateau, while the computing resources demand at Data Center facilities continues to increase exponentially. The main challenge in achieving a global energy efficiency strategy based on Artificial Intelligence is that we need massive amounts of data to feed the algorithms. This paper proposes a time-series data augmentation methodology based on synthetic scenario forecasting within the Data Center. For this purpose, we will implement a powerful generative algorithm: Generative Adversarial Networks (GANs). Specifically, our work combines the disciplines of GAN-based data augmentation and scenario forecasting, filling the gap in the generation of synthetic data in DCs. Furthermore, we propose a methodology to increase the variability and heterogeneity of the generated data by introducing on-demand anomalies without additional effort or expert knowledge. We also suggest the use of Kullback-Leibler Divergence and Mean Squared Error as new metrics in the validation of synthetic time series generation, as they provide a better overall comparison of multivariate data distributions. We validate our approach using real data collected in an operating Data Center, successfully generating synthetic data helpful for prediction and optimization models. Our research will help optimize the energy consumed in Data Centers, although the proposed methodology can be employed in any similar time-series-like problem.

Motivation & Objective

• To address the critical challenge of insufficient real-world data for training AI models in data centers, especially for energy efficiency optimization.
• To develop a scalable, secure, and realistic synthetic data generation method that preserves multivariate temporal dependencies and statistical properties.
• To enhance model robustness by introducing on-demand anomalies during data generation without requiring additional data collection or risking hardware integrity.
• To validate the realism and utility of generated data using novel metrics (KL Divergence and MSE) and benchmark against real data in predictive modeling tasks.
• To enable broader adoption of AI-driven optimization in data centers by overcoming data scarcity and privacy barriers.

Proposed method

• A conditional Generative Adversarial Network (cGAN) is trained on real multivariate time-series data from a production data center, including temperature, humidity, and other operational metrics.
• The model generates synthetic time-series scenarios conditioned on specific time or environmental contexts, enabling controlled, diverse scenario forecasting.
• On-demand anomalies are injected into the synthetic data by modifying latent space vectors during generation, without requiring labeled anomaly data or additional training.
• The realism of generated data is validated using Kullback-Leibler (KL) Divergence and Mean Squared Error (MSE) to compare multivariate distributions between real and synthetic data.
• A feedforward neural network is trained on synthetic, real, or combined data to evaluate predictive performance on unseen real data.
• The framework is designed to be adaptable to any time-series problem with mixed data types and complex dependencies.

Experimental results

Research questions

• RQ1 Can GAN-based data augmentation generate realistic multivariate time-series data that preserve the statistical and temporal characteristics of real data center operations?
• RQ2 How does injecting on-demand anomalies into synthetic data affect the robustness and generalization of downstream prediction models?
• RQ3 To what extent can models trained exclusively on synthetic data achieve performance comparable to those trained on real data?
• RQ4 Do KL Divergence and MSE serve as effective metrics for evaluating the realism of synthetic multivariate time-series data?
• RQ5 Can this approach be scaled to larger, more complex data center environments with higher-dimensional sensor data?

Key findings

• The proposed method successfully generated synthetic time-series data that closely matched the statistical distribution of real data, as confirmed by low KL Divergence and MSE values.
• Models trained solely on synthetic data with injected anomalies (TSTR) achieved a mean squared error (MSE) of 0.103 ± 0.008 for temperature and 0.0168 ± 0.003 for humidity, approaching the performance of models trained on real data (MSE: 0.0738 and 0.0094, respectively).
• The inclusion of synthetic anomalies improved model robustness, with TSTR results outperforming those from the same setup without anomalies, indicating enhanced generalization.
• Training on a combination of synthetic and real data (TSRTR) further improved performance, achieving an MSE of 0.0532 ± 0.004 for temperature and 0.0078 ± 0.004 for humidity, surpassing the performance of real-data-only training in some cases.
• The method demonstrated that synthetic data with controlled anomalies can be used to train models that generalize well to real-world scenarios without compromising predictive accuracy.
• The study confirms that KL Divergence and MSE are effective metrics for validating the realism of multivariate synthetic time-series data, offering a more comprehensive comparison than traditional metrics.