A Hybrid VMD–TCN–GRU Model with Attention for Cloud Workload Prediction

Accurate workload forecasting is essential for proactive resource allocation and scheduling in cloud systems. However, real-world workload traces are often nonlinear and nonstationary, which makes robust forecasting challenging. This paper proposes a hybrid approach that integrates variational mode decomposition (VMD) with a temporal convolutional network (TCN)–gated recurrent unit (GRU) predictor and attention module. VMD decomposes the workload sequence into a set of frequency-aware components, which helps suppress noise and irregular fluctuations. Each component is modeled separately: the TCN captures long-range temporal patterns, the GRU models short-term dynamics, and the attention module highlights informative time steps. The final forecast is obtained by summing the component-level predictions, consistent with VMD reconstruction. Experiments on Alibaba Cloud traces show that the proposed method achieves lower prediction errors than several representative baselines.