Multi-Scale Temporal Modeling with Volatility-Aware Channel Attention for Multi-Step Cloud Workload Forecasting

Cloud platforms require forecasts of CPU, memory and network demand several steps ahead because capacity adjustments take time. Workload levels and variability drift, short bursts coexist with slower changes, and the usefulness of each resource channel varies across windows. This preprint describes reversible window normalization, attention conditioned on window mean and variance, and three paths for trend, convolution–recurrent dynamics and multi-scale residuals. The residual path uses kernels of sizes 3 and 7; sigmoid gates regulate its contribution and the trend correction. The supplied evaluation summary describes an Alibaba cluster trace, a 96-step history and horizons of 1, 5, 10 and 30 steps. It reports the lowest MAE among five methods at every horizon, including a 4.1% reduction relative to C-LSTM at one step and reductions of 7.3%–8.3% at longer horizons. It reports the lowest MSE from five steps onward, while two recurrent baselines perform slightly better at one step. Component removal, severe input corruption and post-spike transitions degrade performance. The manuscript formalizes one compatible architecture and an evaluation protocol. These are source-reported findings: absolute errors, uncertainty estimates and executable experiment artifacts were not supplied and the results have not been independently reproduced. The PDF and editable LaTeX source package are included.