Optimization of Distinct Time-Series Neural Architectures for Cloud-Edge Workload Prediction

Optimizing resource utilization in cloud-edge networks requires accurate workload prediction under strict computational constraints. Traditional methods like Recurrent Neural Networks (RNNs) struggle to balance accuracy and efficiency in these dynamic environments. In this paper, we evaluate distinct state-of-the-art time-series architectures for containerized workload prediction, including Graph Neural Networks (GNNs), temporal Convolutional Neural Networks (CNNs), Transformers, and sparse Multi-Layer Perceptrons (MLPs). Through empirical analysis using Bayesian optimization, we identify optimal model configurations for cloud-edge workload prediction. Notably, the MLP model achieves comparable memory prediction accuracy to the best CNN model, while using approximately 7,000× less parameters (i.e., 35 vs. 247,091) and 5× faster inference (1.84µs vs. 7.15µs). Our findings provide concrete guidance for model selection across different cloud-edge deployment scenarios, from resource-constrained edge devices to high-capacity cloud servers.