Explainable and Energy-Aware Workload Forecasting for Proactive Auto-Scaling in Cloud Data Centers: A Hybrid Deep Learning Approach

Background: Proactive auto-scaling needs workload forecasts that are accurate, quantify their uncertainty and can be audited, yet most cloud workload predictors report point accuracy only. Methods: We propose XE-CLA, a hybrid convolutional, long short-term memory and temporal-attention network with joint point and quantile (τ = 0.8, 0.9, 0.95) outputs. Using 11,746 real PlanetLab virtual machine (VM) days of CPU traces with chronologically separated test days, we compared XE-CLA with 12 statistical, machine-learning, deep and ablated models over three seeds, explained it with Expected Gradients under a deletion-based faithfulness test, and evaluated its decisions in a SPECpower-calibrated auto-scaling simulator. Results: XE-CLA achieved the lowest mean absolute error at all horizons (5 min: 5.15 percentage points), 20.6% below persistence and 6.4% below ARIMA, although differences among learned models were below 2% and random forest had the lowest RMSE. Quantile coverage was 0.787, 0.889 and 0.938. Allocation at the 0.9 quantile used 5.7–10.5% less energy than point-forecast-plus-headroom policies at equal unmet demand (8.01%) and 77.5% less than static provisioning. Time of day accounted for 61.7% of attributions to the provisioning headroom versus 31.5% for the point forecast, whereas attention weights were near-uniform. Conclusion: For noisy VM-level workloads, converting forecasts into allocations through calibrated quantiles matters more than marginal gains in point accuracy.