Multivariate Resource Usage Forecasting and Temporal Accuracy in Private Cloud Systems

Background: Accurate resource forecasting is crucial for optimizing performance,cost, and energy in private clouds. However, predicting dynamic workloads like CPUusage and quantifying forecast uncertainty remain challenging.Objectives: This thesis compares diverse forecasting models (classical, ensemble,deep learning) for private cloud CPU prediction. It assesses the value of multivariateinputs (CPU and memory data), analyzes accuracy degradation over forecast horizons, and evaluates non-Bayesian uncertainty quantification (UQ) methods (QuantileRegression - QR, Conformal Prediction - CP, Conformalized QR - CQR).Methods: Using historical private cloud hypervisor data and engineered temporal features, a naïve persistence baseline (predicting the next value equals the lastobserved value) and six advanced models (ARIMA, LR, RF, XGB, MLP, Transformer) were evaluated via chronological splits. Experiments compared univariatevs. multivariate inputs, examined performance across horizons (1-20 steps) usingdirect forecasting, and assessed 80% nominal prediction intervals from QR, CP, andCQR based on empirical coverage and width.Results: All advanced forecasting models significantly outperformed the naïve persistence baseline, demonstrating their utility. CPU history was the dominant predictor; tree ensembles (RF, XGB) performed slightly better than other ML models.Adding memory data offered marginal accuracy gains but slightly reduced underprediction risk (MUE/MCU). Point accuracy remained relatively stable across horizons with direct forecasting. Standard CP yielded conservative intervals (often >90%coverage), while QR/CQR achieved empirical coverage closer to the 80% target withcompetitive widths.Conclusions: For private cloud CPU forecasting, resource history is key, and treeensembles are effective, offering substantial improvements over simple baselines. Direct forecasting manages medium-term horizon errors well. Reliable uncertaintyquantification is vital; QR/CQR offer a practical alternative to potentially overconservative standard CP, presenting a trade-off between theoretical guarantees andempirical interval tightness