Real-Time Control of Sustainable Data Centers: A Two-Layer Model Predictive Control Framework with Workload Flexibility and Heat Recovery
This paper proposes a two-layer model predictive control framework that integrates stochastic optimization and adaptive tube-based MPC to enable real-time, sustainable, and grid-supportive operation of data centers equipped with renewable energy, storage, and heat recovery systems.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The modern world runs on a vast, invisible network of digital services. From streaming movies to training artificial intelligence, every click and query is processed by massive data centers. These facilities are the engines of our digital age, but they come with a heavy price tag: they consume enormous amounts of electricity and generate significant heat. As the demand for computing power grows, these buildings are becoming some of the most energy-intensive structures on the planet. At the same time, the electrical grids that power them are changing. They are increasingly relying on renewable sources like solar and wind, which are unpredictable; the sun does not always shine, and the wind does not always blow. This creates a delicate balancing act. Data centers need a steady, reliable supply of power to keep their servers running, yet the grid they connect to is becoming more volatile. Furthermore, the carbon footprint of these facilities is a major environmental concern, with the industry projected to contribute significantly to global emissions in the coming decade. The challenge for engineers is to find a way to run these data centers that is not only cost-effective but also carbon-conscious and flexible enough to help stabilize the grid rather than strain it.
Researchers at the École polytechnique fédérale de Lausanne have developed a new control system designed to solve this complex problem. They propose a two-layer framework that acts as a smart brain for a sustainable data center, coordinating its internal operations with the outside world. Imagine the data center not just as a consumer of electricity, but as a flexible hub that can generate its own power through solar panels, store energy in large batteries, and even capture the waste heat from its servers to warm nearby buildings or generate extra electricity. The researchers' system manages all these moving parts simultaneously. It must decide when to run heavy computing tasks, when to charge or discharge batteries, and how much heat to recover, all while reacting to fluctuating electricity prices and the carbon intensity of the grid. The goal is to minimize costs and carbon emissions while ensuring that the data center never fails to deliver the computing services its users expect.
The core of this solution is a hierarchical control strategy that operates on two different time scales. The upper layer of the system acts like a strategic planner. It looks ahead at the day, using forecasts for weather, electricity prices, and computing demand to make broad decisions. Every fifteen minutes, this layer calculates the best course of action for the coming hours. It decides how much electricity to buy or sell in the intraday market, sets targets for how much work the servers should process, and determines the optimal state for the battery and the heat recovery systems. Crucially, this planner does not guess blindly; it uses a method that considers many possible future scenarios. It asks, "What if the clouds cover the solar panels? What if the price of electricity spikes?" By weighing these possibilities, it creates a robust plan that is resilient to uncertainty, aiming to keep costs low and carbon emissions down.
However, a plan made fifteen minutes in advance cannot account for the sudden, rapid changes that happen in real time. Solar output can drop in seconds due to a passing cloud, and computing workloads can surge unpredictably. To handle these fast fluctuations, the system employs a second, lower layer that operates every minute. This layer acts as a precise navigator, constantly monitoring the actual performance of the data center and comparing it to the plan set by the upper layer. If the solar panels produce less power than expected, or if the servers get hotter than anticipated, this lower layer instantly adjusts the controls. It tweaks the cooling systems, shifts the battery charging, or slightly modifies the workload execution to correct any errors. This rapid response ensures that the data center stays on track, preventing costly imbalances and maintaining stability even when the environment is chaotic.
The researchers tested this framework using a highly detailed simulation that mimics a real-world data center equipped with solar panels, a battery storage system, and a mechanism to recover waste heat. They simulated various operating conditions, including clear sunny days and overcast skies, to see how the system would perform under stress. The results showed that the two-layer approach was highly effective. While simpler control strategies struggled to keep up with rapid changes, leading to deviations from the plan and higher costs, the new framework maintained tight control. It successfully tracked the dispatch targets despite fast fluctuations in solar power and computing demand. By using the lower layer to correct short-term errors, the system significantly reduced the financial penalties associated with being out of sync with the grid.
Beyond just keeping costs down, the system demonstrated a natural ability to adapt to environmental signals. It responded to changes in the carbon intensity of the grid, shifting operations to times when the electricity was cleaner. It also managed the complex trade-offs between generating power from waste heat and selling it to a district heating network. The study confirms that by combining long-term strategic planning with minute-by-minute tactical adjustments, it is possible to run a data center that is economically efficient, environmentally sustainable, and supportive of the electrical grid. This approach offers a practical path forward for the future of digital infrastructure, proving that even the most energy-intensive facilities can be managed with intelligence and flexibility.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.