Background. Private cloud systems face significant challenges in resource manage-ment due to dynamic workloads and the need for sustainable operations. Traditionalmonitoring tools often lead to over-provisioning, increasing energy consumption andoperational costs. Machine learning (ML) offers potential solutions, but existingresearch focuses primarily on public clouds, leaving private cloud forecasting under-explored.

Objectives. This thesis aims to develop ML models for traffic and load forecast-ing in private cloud environments. It evaluates the impact of historical data featureson prediction accuracy, assesses achievable forecasting horizons, and investigates howdomain expertise can enhance model robustness.

Methods. A mixed methods approach combines a snowballing literature review,quantitative case study, and expert surveys. Five models, XGBoost, LSTM, Trans-former, ARIMA, and Linear Regression—were tested on 24-hour telemetry data fromEricsson’s private cloud, including CPU, memory, network, and transaction metrics.Statistical validation (Friedman/Nemenyi tests) and domain-informed feature engi-neering were applied.

Results. XGBoost demonstrated superior performance for CPU and networkmetrics , while LSTM achieved better memory error reduction when enhanced withexpert insights. Key findings include: quadratic error growth patterns across most ofthe metrics, critical dependencies between transactional features, excluding TPS_4,(Policy signaling rate), increased TPS_3, (Charging speed), dramatically, and Trans-formers unexpected proficiency in short-term TPS forecasting.

Conclusions. Effective private cloud forecasting requires model specialization:XGBoost for resource metrics, domain-tuned LSTM for memory optimization, andTransformers for transactional patterns. The study establishes that human expertiseamplifies model capabilities when applied architecturally rather than through sim-ple feature engineering. These insights enable improvements in resource allocationaccuracy, directly supporting energy-efficient cloud management.