KAN-GRU: A Lightweight Temporal Kolmogorov-Arnold Network for Cloud Resource Utilization Forecasting

Accurate forecasting of cloud resource utilization is essential for proactive provisioning, autoscaling, and energy-aware scheduling in modern data centers. Temporal Kolmogorov-Arnold Networks (TKANs) provide a promising alternative to MLP-based models by combining Kolmogorov- Arnold representations with recurrent sequential modeling. However, prior TKAN formulation typically relies on LSTM-style recurrence, which increases parameter count and computational cost. This work introduces KAN–GRU, a lightweight TKAN-based architecture that substitutes the LSTM cell in the TKAN temporal component with a GRU, reducing parameters while preserving temporal modeling capacity. The evaluation benchmarks KAN–GRU for data-center resource-utilization forecasting under univariate and multivariate configurations and across single-step and multi-step prediction horizons. The results indicate that KAN-GRU improves computational efficiency while maintaining competitive predictive performance, making it suitable for scalable deployment in monitoring and control operations in cloud environments.