Abstract: The traditional research on the PUE in data centers mainly focuses on hyperscale data centers. However, hyperscale data centers differ significantly from small - scale data centers in terms of scalability, workload volatility, network infrastructure and connectivity, and refined management capabilities. As a result, there is limited research on PUE prediction for small - scale air - cooled data centers at present. This paper proposes a PUE prediction model for data centers based on the LSTM-Transformer model. Built on the LSTM and Transformer architectures, this model can effectively integrate multi - source and heterogeneous energy data from data centers, analyze the influence of different factors on the PUE of data centers, and make predictions. The experimental results on a real - world data center dataset show that the mean squared error of this model for PUE prediction in small - scale air - cooled data centers is as low as 0.0007, the mean absolute error is 0.0214, and the coefficient of determination reaches 0.9583. The experiments demonstrate the superiority of this model over other models and provide valuable support for PUE prediction in small - scale air - cooled data centers.

Key words: small air-cooled data centers, Power Usage Effectiveness prediction, LSTM-Transformer, PUE prediction, sensitivity analysis

摘要： 传统数据中心电源使用效率（Power Usage Effectiveness, PUE）预测工作主要以超大规模数据中心为研究对象。但是，超大规模数据中心在扩展能力、工作负载波动性、网络基础设施与网络连接、精细化管理能力等方面，同小型数据中心具有较大的差别。因此，当前对小型风冷数据中心PUE预测的工作不多。本文提出了一种基于LSTM-Transformer模型的小型风冷数据中心PUE预测模型。该模型基于LSTM和Transformer架构，能有效整合数据中心多源异构能源数据，分析不同因素对数据中心PUE的影响，并进行预测。在真实数据中心数据集上的实验结果显示，该模型在小型风冷数据中心PUE预测的均方误差低至0.0007，平均绝对误差为0.0214，决定系数达到0.9583。实验证明了该模型相对于其他模型的优越性，为小型风冷数据中心PUE预测提供了有价值的支持。

关键词: 小型风冷数据中心, 电源使用效率预测, LSTM-Transformer, PUE预测, 敏感度分析