An on-premise cloud-native AIOps infrastructure system with workload forecasting in Kubernetes

AIOps leverages monitoring and machine learning to manage complex IT systems, but commercial and third-party solutions often face high costs or limited customization for enterprise needs. Accurate cloud workload forecasting is critical for proactive resource allocation and service reliability. We present a fully on-premise AIOps system with two main contributions. First, at the system level, it integrates open-source tools within a Kubernetes-based deployment to collect real-time metrics from containerized workloads, perform online prediction, and enable incremental model updates, providing a practical foundation for enterprise deployment. Second, at the model level, we design a transformer-based forecasting model enhanced with feature-wise normalization and a trend-aware combined loss, enabling accurate capture of complex temporal patterns. Experiments across diverse workloads demonstrate that our approach consistently outperforms baseline models in prediction accuracy and temporal consistency, highlighting the effectiveness of both system-level and model-level innovations. This work enables automated, intelligent management of cloud-native environments while addressing enterprise-specific constraints.