关注周期–趋势–波动的云平台任务数量预测方法

A Cloud Platform Tasks Number Prediction Method Focusing on Period-Trend-Fluctuation

摘要: 在云计算数据中心，准确预测未来一段时间内的用户任务达到数量并及时调整资源分配，对提升资源利用率和降低能耗十分重要。为应对云用户任务提交模式的高度动态性，以及长期时间序列预测过程中易产生误差累积的问题，提出了一种关注周期、趋势、波动特征的云平台任务数量预测方法。该方法首先从原始时间序列中提取单个可循环的周期模式，以便后续显式地利用时间序列中的稳定周期模式；然后引入周期注意力机制，通过设定周期长度对子序列进行聚合以提升周期模式的表达能力，并结合门控机制动态调节周期模式对预测结果的影响。最后，利用Transformer对去除周期分量的趋势–波动分量进行建模，实现了云任务数量的预测。利用Alibaba-cluster-gpu-v2020数据集，开展实验验证，实验结果表明该方法在长期预测精度和稳定性方面优于对比的基线方法，具有较好的实际应用价值。

Abstract: In cloud data centers, accurately forecasting the number of arriving user tasks over a future period and timely adjusting resource allocation are essential for improving resource utilization and decreasing energy consumption. To address the high dynamism of cloud users’ task submission patterns and the issue of error accumulation commonly encountered in long-term forecasting, this paper proposed a method for cloud task arrival prediction with Periodic, Trend, and Fluctuation features. The method first constructed a single recurrent periodic pattern from the original time series to explicitly capture stable periodic behaviors. Meanwhile, the paper proposed a Period-Attention mechanism that aggregates long-term subseries through the explicit period length, thereby enhancing the representation capability of periodic components. Additionally, a gating mechanism is incorporated to dynamically regulate the influence of periodic information on the prediction results. Finally, the Transformer architecture was utilized to method and forecasted the Trend-Fluctuation components after removing the periodic components, achieving the prediction of the number of cloud tasks. Experiments are conducted on the Alibaba Cluster Trace GPU v2020 dataset and the results demonstrate that the proposed method consistently surpasses several baseline models in both long-term forecasting accuracy and stability, showing strong potential for practical application in cloud task prediction scenarios.