WGGAL: A Practical Time Series Forecasting Framework for Dynamic Cloud Environments

Abstract

In the era of widespread cloud service adoption, accurate forecasting of cloud resource demands and workloads is paramount. However, the dynamic nature of cloud environments leads to high temporal volatility and instability in time series data, challenging traditional forecasting methods in capturing complex features. Moreover, many existing methods prioritize accuracy without adequately considering the potential economic losses and erosion of user trust resulting from resource underestimation. To address these issues, we introduce the concept of the Underestimation Rate(UDR) and propose a prediction framework named WGGAL. This framework leverages the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), employing an ALinear model as the generator and a Multilayer Perceptron (MLP) as the discriminator. Additionally, we present the Adaptive Penalized Hybrid Loss Function (APH Loss) to constrain prediction values, effectively reducing resource underestimation while maintaining prediction accuracy and avoiding issues such as out-of-memory errors. Experimental results demonstrate that, compared to state-of-the-art forecasting methods, WGGAL reduces forecasting error by 30.5% and the UDR by 46.8% in one-step forecasting, and by 54.4% and 26.3%, respectively, in multi-step forecasting.