TCformer: Fine-Grained Cloud Workload Prediction via Multi-Scale Convolution and Channel-Aware Temporal Attention

Abstract

Accurate prediction of cloud workloads is vital for optimizing resource allocation and reducing energy consumption in data centers. However, cloud workloads exhibit non-stationarity, burstiness, and weak periodicity, which make forecasting highly challenging. Existing time-series models often fail to capture fine-grained short-term fluctuations, tend to oversmooth long-term trends, and struggle to capture dynamic inter-channel dependencies. These limitations hinder their ability to track abrupt spikes and cross-domain signals that are critical in real-world scenarios. To overcome these issues, we propose TCformer, a Transformer-based prediction model that integrates two key components: a multi-scale convolutional embedding module to capture global and local feature variations across temporal cycles adaptively, and a temporal channel attention mechanism to dynamically adjust channel importance for modelling cross-channel dependencies over time. Extensive experiments on three real-world cloud workload datasets covering 27 workload types demonstrate that TCformer consistently outperforms state-of-the-art baselines, achieving superior accuracy in both short-term and long-term forecasting.