Modelling the Temperature of a Data Centre Cooling System Using Machine Learning Methods

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Although the idea is interesting, the paper is missing more insights to the data centers research community. Otherwise, this study is a simple forecasting exercise that was able to be carried out using any other dataset. Moreover, the rationale behind using those sophisticated methods should be clarified and a careful correlation study along with a comparative study to naïve methods should be conducted.

1. The authors should clarify the reasons/ rationale behind selecting the 2 models (TiDE and TSMixer) among the multiple AI techniques being developed and explored in the filed. In addition, the authors should justify why they compare with random forest and XG-Boost. Why not other choices?
2. This paper is missing a correlation analysis between the input parameters and the outputs of the models. Without this study (which should include linear Pearson correlation and Spearman correlations), it is difficult to justify whether the selected inputs are likely to influence the outputs.
3. In Figure 3, the authors should include the actual (True) temperature for fair comparison.
4. The comparative study should consider other criteria in addition to the forecasters’ accuracy.
5. The authors should compare to naïve methods to justify the use of such sophisticated methods and they should quantify the improvement in terms of accuracy.

 

Author Response

Dear reviewer, thank you for your work. We believe that your comments help the readers better understand our work. Below are the answers to your comments.

1. The authors should clarify the reasons/ rationale behind selecting the 2 models (TiDE and TSMixer) among the multiple AI techniques being developed and explored in the filed. In addition, the authors should justify why they compare with random forest and XG-Boost. Why not other choices?

Answer:

Thank you for the comment. We added an explanation of why we selected such a model for our comparison:

To develop the prediction model, we compared several well-established machine learning algorithms such as Random Forest and XGBoost with state-of-the-art techniques, including Time-series Dense Encoder (TiDE) [21] and Time-Series Mixer (TSMixer)—two of the most recent approaches for time series forecasting. The TiDE model was selected due to its design for long-term forecasting tasks, which aligns with our objective of forecasting across a full cycle of daily seasonality while incorporating external features, such as the day of the week. In contrast, TSMixer is a general-purpose forecasting model that is significantly more lightweight than typical transformer-based architectures, enabling more efficient training while maintaining competitive performance. On the other hand, tree-based models are feature-scale independent and are popular for their high scalability and simplicity in parameter tuning

2. This paper is missing a correlation analysis between the input parameters and the outputs of the models. Without this study (which should include linear Pearson correlation and Spearman correlations), it is difficult to justify whether the selected inputs are likely to influence the outputs.

Answer:

Thank you for the comment. Such additional information is beneficial for the reader; therefore, an additional section 3.1.2. was added, including correlation analysis between input features and between input features and the output feature.

3. In Figure 3, the authors should include the actual (True) temperature for fair comparison.

Answer:

Thank you for the comment, but unfortunately, we cannot do this because these are figures obtained from the company from their internal monitoring system, and we don’t have full access to this data.

4. The comparative study should consider other criteria in addition to the forecasters’ accuracy.

Answer:

Thank you for the comment. We extended our analysis by adding an additional section concerning execution time, which we believe is very important since the model should be executed on a daily basis

5. The authors should compare to naïve methods to justify the use of such sophisticated methods and they should quantify the improvement in terms of accuracy.

Answer:

Thank you for the comment. We extended our analysis by adding AutoARIMA model. We used it as a reference to validate whether the relation is linear or requires a more complex nonlinear model. The results show that the AutoARIMA model is the weakest, justifying the use of more complex models.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript titled “Modelling the Temperature of Data Center Cooling System using Machine Learning Methods” addresses the critical issue of reducing energy consumption in data centers by forecasting temperature in the warm corridor using machine learning. The authors evaluate and compare four models—TiDE, TSMixer, XGBoost, and Random Forest—using real-world data collected from an office-integrated data center. The primary novelty lies in the application of two recently proposed neural network architectures (TiDE and TSMixer) for temperature forecasting in data centers, which has not been widely explored in this specific domain. While the models themselves are not novel, their evaluation in this use-case contributes meaningfully to the literature, particularly in demonstrating TiDE’s advantages in handling extreme values—an important factor for ensuring equipment safety.

The authors follow a well-structured methodology with a robust dataset, clear problem formulation, thorough hyperparameter tuning, and multiple evaluation scenarios. The use of normalized RMSE, MAE, and MAPE offers a solid basis for comparison. The TiDE model achieved the lowest error in most scenarios, and more importantly, demonstrated better performance in predicting high temperature extremes compared to XGBoost, which tends to underestimate in critical ranges. This observation supports the claim that TiDE is better suited for practical deployment where overheating risk is a concern. The paper builds on prior work in energy forecasting and data-driven HVAC optimization. While similar studies exist, few have evaluated newer architectures like TiDE and TSMixer in this context. The comparison with classical models (XGBoost and RF) is relevant and supports the motivation. However, the literature review could be strengthened with additional recent studies on data center thermal modeling for broader context.

The manuscript would benefit from thorough English copy editing. Several grammatical issues, unclear phrasing, and repeated use of “worm corridor” (likely a typo for “warm corridor” or “hot aisle”) affect clarity. Standardizing terminology and improving figure descriptions would also enhance readability. The paper presents a well-executed and relevant study with modest but valuable contributions. With improvements in writing quality and some minor clarifications in methodology and literature positioning, it has the potential to be a strong addition to the journal.

Comments on the Quality of English Language

The manuscript would benefit from thorough English copy editing. Several grammatical issues, unclear phrasing, and repeated use of “worm corridor” (likely a typo for “warm corridor” or “hot aisle”) affect clarity. Standardizing terminology and improving figure descriptions would also enhance readability

Author Response

Dear reviewer, thank you for your work. We believe that your comments help the readers better understand our work. Below are the answers to your comments.

The manuscript titled “Modelling the Temperature of Data Center Cooling System using Machine Learning Methods” addresses the critical issue of reducing energy consumption in data centers by forecasting temperature in the warm corridor using machine learning. The authors evaluate and compare four models—TiDE, TSMixer, XGBoost, and Random Forest—using real-world data collected from an office-integrated data center. The primary novelty lies in the application of two recently proposed neural network architectures (TiDE and TSMixer) for temperature forecasting in data centers, which has not been widely explored in this specific domain. While the models themselves are not novel, their evaluation in this use-case contributes meaningfully to the literature, particularly in demonstrating TiDE’s advantages in handling extreme values—an important factor for ensuring equipment safety.

The authors follow a well-structured methodology with a robust dataset, clear problem formulation, thorough hyperparameter tuning, and multiple evaluation scenarios. The use of normalized RMSE, MAE, and MAPE offers a solid basis for comparison. The TiDE model achieved the lowest error in most scenarios, and more importantly, demonstrated better performance in predicting high temperature extremes compared to XGBoost, which tends to underestimate in critical ranges. This observation supports the claim that TiDE is better suited for practical deployment where overheating risk is a concern. The paper builds on prior work in energy forecasting and data-driven HVAC optimization. While similar studies exist, few have evaluated newer architectures like TiDE and TSMixer in this context. The comparison with classical models (XGBoost and RF) is relevant and supports the motivation. However, the literature review could be strengthened with additional recent studies on data center thermal modeling for broader context.

Answer:

Thank you for the comment. We updated the Introduction section accordingly and extended it by adding information on building and data centre thermal modelling. The corrected paragraphs are:

In recent years, there has been significant progress in the field of energy modeling for buildings. Such modeling has been applied in various areas, including energy-saving initiatives [5], operational optimization [6], and building modernization projects [7]. The development of building modeling initially focused on the application of Computational Fluid Dynamics (CFD), which enables the simulation of airflow, temperature distribution, and heat transfer within buildings and data centers. Notable early approaches are presented in [26, 27], while extensions to more complex building geometries are discussed in [28]. CFD has also been effectively applied to the thermal modeling of data centers, as demonstrated in [29].

In general, advancements in data aggregation and visualization technologies have brought substantial benefits to research on building energy efficiency [8]. However, a major challenge remains in the integration of these complex systems [9]. A promising solution lies in the adoption of artificial intelligence techniques, which enable self-learning, autonomous decision-making, and continuous model adaptation [10].

 

The manuscript would benefit from thorough English copy editing. Several grammatical issues, unclear phrasing, and repeated use of “worm corridor” (likely a typo for “warm corridor” or “hot aisle”) affect clarity. Standardizing terminology and improving figure descriptions would also enhance readability. The paper presents a well-executed and relevant study with modest but valuable contributions. With improvements in writing quality and some minor clarifications in methodology and literature positioning, it has the potential to be a strong addition to the journal.

Answer:

Thank you for the comments. We evaluated grammar and corrected the text in several places. All typos were removed. We also standardized the type of English, selecting UK English as the base.

Reviewer 3 Report

Comments and Suggestions for Authors

Strengths:

Relevance and Novelty: The paper addresses a timely and relevant issue, energy efficiency in data centers, by applying modern machine learning models like TiDE and TSMixer, which are not commonly used in similar work. 

Comprehensive Methodology: The authors provide a detailed explanation of dataset preprocessing, feature engineering, and hyperparameter optimization, making the work reproducible and transparent.

Weaknesses:

Limited Discussion on Real-World Application: While the models are evaluated extensively, the paper lacks a clear discussion on how these predictions can be integrated into real-time energy management systems in practice.

Underdeveloped Error Analysis: Although model performance is compared with standard metrics (MAE, MAPE, RMSE), there is limited insight into why certain models underperform in specific conditions, especially regarding temporal patterns or feature importance.

Comments on the Quality of English Language

The overall quality of the English language is good and the paper is understandable. However, there is room for improvement in terms of clarity and flow in some parts.

Author Response

Dear reviewer, thank you for your work. We believe that your comments help the readers better understand our work. Below are the answers to your comments.

Strengths:

Relevance and Novelty: The paper addresses a timely and relevant issue, energy efficiency in data centers, by applying modern machine learning models like TiDE and TSMixer, which are not commonly used in similar work. 

Comprehensive Methodology: The authors provide a detailed explanation of dataset preprocessing, feature engineering, and hyperparameter optimization, making the work reproducible and transparent.

Answer:

Thank you for the comments.

Weaknesses:

Limited Discussion on Real-World Application: While the models are evaluated extensively, the paper lacks a clear discussion on how these predictions can be integrated into real-time energy management systems in practice.

Answer:

Thank you for the comment. We extended the description in the introduction. In this part of the study, we focus on finding the most suitable prediction model, and in the following step, we plan to incorporate the model into the data centre – as that is the subject of the PhD thesis of one of the authors.

The updated paragraph is:

(…) Energy efficiency can be improved by increasing the temperature setpoints of the air conditioning system, thereby lowering energy consumption. However, this approach introduces the risk of exceeding the thermal safety thresholds of computing units, which is unacceptable in an operational data centre environment. To address this challenge, a reliable temperature forecasting model for the warm aisle is required. Such a model can support the safe adjustment of air conditioning parameters, ensuring that temperature limits are not breached, and the integrity of the compute infrastructure is preserved.

To achieve the goal of energy efficiency the initial step is the development of the temperature forecasting model. In our research, we evaluate and identify the most suitable model for estimating temperature conditions. In this study, we compared five different machine learning models, including two modern neural network architectures (…)

Underdeveloped Error Analysis: Although model performance is compared with standard metrics (MAE, MAPE, RMSE), there is limited insight into why certain models underperform in specific conditions, especially regarding temporal patterns or feature importance.

Answer:

Thank you for the comment. We added a section describing feature importance separately for Saturdays and Mondays, representing weekends and working days.

Comments on the Quality of English Language

The overall quality of the English language is good and the paper is understandable. However, there is room for improvement in terms of clarity and flow in some parts.

Answer:

Thank you for the comment. We evaluated the text, trying to make it sound better.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I recommend acceptance of the present version.