ML-based Predictive Models for Power Consumption in Virtualised O-RANs
This paper proposes and evaluates machine learning models for predicting power consumption in virtualized O-RANs, demonstrating that a hybrid DNN-XGBoost approach achieves superior accuracy (under 0.5% error) compared to standard and regularized DNNs, thereby enabling more energy-efficient network orchestration.
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
Imagine the internet as a massive, bustling city where data is the traffic and cell towers are the traffic lights. For decades, these traffic lights have been getting smarter, but they've also been getting hungrier. As our phones demand faster speeds and more video streaming, the networks powering them are guzzling electricity like a marathon runner on a sugar rush. This isn't just bad for the planet's climate; it's also a huge bill for the companies running the networks. To fix this, engineers are trying to rebuild these networks from scratch, turning them into flexible, software-driven systems called "Open Radio Access Networks" (or O-RANs for short). Think of O-RAN as taking a rigid, pre-built house and turning it into a set of Lego blocks that can be rearranged instantly to fit the weather. But here's the tricky part: because these Lego blocks can change shape so quickly, it's incredibly hard to predict how much energy they will use at any given moment. Old ways of guessing energy use are like trying to predict the weather by looking at a static map; they just can't handle the sudden storms and sunny spells of a modern, digital network.
This is where the researchers from Trinity College Dublin step in with a new idea: let's teach a computer to learn the network's energy habits instead of trying to write a manual for it. They set up a real-life testbed—a miniature, working version of this flexible network—and measured exactly how much power it used under different conditions. They then trained three different types of "digital brains" (machine learning models) to look at the network's settings and guess the power bill. The first brain was a standard Deep Neural Network (DNN), a type of AI good at spotting patterns. The second was a "regularized" version of that same brain, tweaked to stop it from daydreaming too much (a problem called overfitting). The third was a hybrid team-up: the DNN acted as a scout to find the most important clues, and then passed those clues to a different, very sharp decision-maker called XGBoost to make the final guess.
The results were clear: the team-up was the star of the show. While the standard and regularized brains made some decent guesses, they stumbled when the network conditions got tricky, sometimes missing the mark by up to 5%. The hybrid model, however, was incredibly precise, consistently guessing the power usage with an error rate of less than 0.5%. In fact, in the most difficult scenarios, it was off by only about 0.1%. The paper suggests that this hybrid approach is the most reliable way to predict energy use in these complex, virtualized networks. By using this kind of smart prediction, network managers could potentially turn off unused parts of the system or adjust power levels in real-time, saving energy without slowing down our internet. The authors note that while this is a strong start, future work could involve teaching these models to understand how traffic changes over time, much like learning the daily commute patterns of a city, to make the energy savings even better.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.