Access Official Technical PDF

Share Summary:LinkedInEmail

Machine learning control offers a way to manage thermal conditions in small edge data centers, where cooling demand, workloads and local environments can vary between deployments. As computing moves closer to users and network access points, these facilities need to operate with limited human oversight while maintaining suitable conditions for IT equipment. Data-driven models can help connect sensor measurements to control decisions without requiring every facility to be modelled from first principles.

Why edge data centers need adaptive control

Edge data centers support applications that benefit from computing and storage close to the point where data is generated or used. Their distributed locations can make continuous, hands-on operation difficult. At the same time, their configurations, available resources, environments and load profiles may differ, even when facilities use modular or standardized designs.

These differences matter because computing equipment converts electrical power into heat, and that heat must be removed to keep servers operating within appropriate thermal conditions. Cooling demand changes with computing activity and environmental conditions. An effective controller therefore needs to account for the relationship between workload, airflow, component temperatures and fan power—not just respond to a single temperature reading.

Autonomy also means preparing for disturbances such as intermittent power failures, component faults and overheating. Machine learning does not, by itself, address every aspect of autonomous operation: fault detection, predictive maintenance and power distribution are separate challenges. But data-driven models can support control of internal thermal processes and may help make efficient use of available cooling and computing resources.

Machine learning control for data center cooling

A useful starting point is a model that describes how server temperatures change over time. In one experimental approach, sensor data from a small server cluster was used to calibrate a data-driven thermal model. The model was structured to reflect heat-exchange behaviour, while its parameters were obtained through optimization. This creates a representation of the physical system that can be used to explore the effects of control actions.

The model was then used in a model predictive control (MPC) strategy for server cooling fans. MPC uses a model to evaluate how the system is expected to respond to potential actions over a future horizon, then selects a control action in line with its objective. In this application, the model supported real-time fan regulation. The work also compared the MPC approach in simulation with a self-tuning proportional-integral-derivative (PID) controller.

The distinction is useful for engineers assessing control options. PID controllers are widely used because they are relatively simple and transparent. MPC can incorporate a model of system dynamics into decisions, but depends on having a model that is sufficiently useful for the application. The experimental work demonstrates a way to construct and use a data-driven model; it does not establish that MPC is the best choice for every facility or that conventional control should be replaced.

Look beyond the server fan

Cooling performance depends on interactions between IT equipment and facility systems. Server fans affect airflow and power consumption, while facility cooling determines the air supplied to the servers. Treating these systems independently can miss opportunities to coordinate their operation.

Experimental analysis of air-based thermal management examined server fan strategies alongside facility airflow conditions. It considered operating temperatures, fan power, server power losses and the temperature difference across the server. The results were also applied theoretically to a direct fresh-air-cooled data center to examine how server and facility control choices relate to overall energy use.

One important implication is that an efficiency metric should not be interpreted in isolation. Power Usage Effectiveness (PUE) compares total facility energy with computing equipment energy. It describes infrastructure overhead, but a lower PUE does not necessarily mean lower total facility power consumption. The analysis reported that decreasing PUE can coincide with increased overall power use, depending on the operating conditions. Control objectives should therefore reflect the outcomes that matter to the facility, rather than optimize a single indicator without context.

Reusing heat and adapting models

Thermal control can also support the productive use of heat that would otherwise be exhausted. In a small-scale apple-drying study, an MPC controller previously used for server-cluster thermal management was repurposed to regulate exhaust-air temperature. The experiment investigated whether server heat could maintain conditions suitable for drying, including a case using 37 °C for eight hours. This illustrates a potential connection between data center operation and a heat-consuming process; it is a specific experimental setup, not a claim that every edge facility can supply useful process heat.

Deploying a learned model across multiple sites presents another practical challenge. If each modular data center requires a full training process from scratch, implementation effort may become prohibitive. Transfer learning addresses this problem by using knowledge from a model trained in one environment to support deployment in another. In the described work, a thermal-node recurrent neural network was trained on one small server cluster and deployed in a different server environment with minimal fine-tuning using a small set of observations.

This approach is relevant where facilities share some structural features but differ in configuration or surroundings. It does not eliminate the need for site data or validation. Instead, it provides a method to investigate whether an existing model can be adapted with less new training than would otherwise be required.

What engineers should consider

For system integrators and operations teams evaluating data-driven thermal control, the research points to several practical considerations:

  • Measure the system: Models depend on observations of the server and its thermal environment, so sensor data and operating records are central to model development.
  • Define the objective: Fan energy, equipment temperatures and facility energy are related, but optimizing one measure alone may not improve the overall result.
  • Account for the deployment: Local configuration and environmental context influence how a model applies to a particular edge site.
  • Validate adaptation: Transfer learning can reduce retraining requirements, but the adapted model still needs to be assessed in its target environment.
  • Keep scope clear: Thermal control is one part of autonomy; power management, fault response and maintenance require additional capabilities.

Conclusion

Machine learning control can help turn operational data into models for managing server cooling and other thermal processes in small data centers. Experimental work with server clusters demonstrates the use of learned thermal models in MPC, while studies of holistic cooling, heat reuse and transfer learning show how the control problem connects to broader facility decisions. For industrial automation teams, the practical opportunity is to evaluate these methods against site-specific measurements and clearly defined objectives. A model should be treated as one component of a dependable control strategy—not as a substitute for system validation, coordinated engineering or operational safeguards.

Access Official Technical PDF

About The Author