A multi dataset validation model for hybrid feature selection in wind energy maximum power point tracking systems

Making Wind Turbines Smarter, Not Just Bigger

Modern wind farms are packed with sensors that monitor everything from wind speed and blade angle to temperatures deep inside the machinery. These data streams can run into hundreds of separate measurements per turbine, updated every few minutes. While this sounds like a gold mine for boosting energy output, it also overloads the computers that must react quickly to shifting winds. This study shows how carefully choosing a smaller, smarter set of measurements can make wind turbines respond faster and more accurately, potentially squeezing a few extra percent of electricity out of the same wind—enough to mean millions of dollars over the life of a large wind farm.

The Challenge of Too Much Information

Wind turbines use control systems known as Maximum Power Point Tracking to constantly adjust how they operate so they capture as much energy as possible from changing winds. In today’s large wind farms, each turbine can stream more than 400 different sensor readings, and control decisions must be made on the order of every 10 minutes or faster. Processing every signal all the time slows down the system and introduces noise from sensors that add little or no useful information. The key question is: which measurements really matter for predicting power output or rotor speed, and which can be safely ignored without harming performance? Finding that sweet spot is a balancing act between accuracy and the limited computing power available inside industrial controllers.

A Two-Step Way to Trim the Data

The authors propose a two-stage method that first narrows the field and then fine-tunes the choices. In the first step, a statistical filter scans all available measurements and scores how strongly each one is related to the quantity the operator cares about—either electrical power in full-size farms or rotor speed in the lab system. Only the top slice of these signals is kept, immediately shrinking the problem from hundreds of candidates to a more manageable group. In the second step, an optimization procedure inspired by musical improvisation explores different combinations within this reduced set. Instead of chasing a single “best” answer, it searches for a family of solutions that trade off prediction accuracy against how many sensors they require, producing a menu of options that operators can match to their hardware limits.

Testing Across Very Different Wind Setups

To check that the approach works in the real world and not just in simulations, the team tested it on three very different data sets. One covered five years of operation from a six-turbine farm in the United Kingdom, with 464 sensor channels capturing a temperate, maritime climate. A second came from a commercial site in tropical southern India, with 87 measurements reflecting highly variable monsoon winds. The third was a controlled laboratory turbine with only five signals but very fast sampling, used to study a power electronics controller in fine detail. Across these cases, the method cut the number of active features by roughly three quarters—down to as few as 58 out of 464 signals in the UK farm and 8 out of 87 in the Indian farm—while still predicting power or speed slightly better than when every sensor was used.

What the Gains Look Like in Practice

When the researchers used the streamlined feature sets to train machine-learning models that predict turbine power or rotor speed, the errors dropped by about 9–15% compared with models that used all available sensors. Against simpler selection techniques commonly used in data science, the improvement was even larger, up to roughly 30% lower error. Crucially, these gains came with big savings in computer effort: reducing 464 signals to 58 cut the processing burden by nearly 88%, making it feasible to run advanced prediction models on the modest hardware typically found in wind farm control rooms. The selected sensor sets also tend to favor physically meaningful quantities such as wind speed at the nacelle, rotor speed, generator torque, and derived measures of aerodynamic efficiency, helping engineers understand and trust what the models are doing.

Why This Matters for Clean Energy

Because even a small improvement in prediction can translate into better control decisions, the authors estimate that a 10% boost in forecasting accuracy can raise annual energy production by 2–3% for a utility-scale wind farm. Spread over many turbines and years of operation, this becomes a substantial financial and climate benefit, achieved without building a single new turbine—only by using data more wisely. The study’s two-step strategy offers a practical recipe: first, quickly filter down hundreds of possible measurements to those that actually relate to performance; then, systematically explore combinations to find compact sensor sets that fit within real-time computing limits. For grid operators, developers, and policymakers, it highlights that smarter data selection is a powerful and relatively low-cost lever for making renewable energy systems more efficient and reliable.

Citation: Duraisamy, S., Thangavelu, V. A multi dataset validation model for hybrid feature selection in wind energy maximum power point tracking systems. Sci Rep 16, 9747 (2026). https://doi.org/10.1038/s41598-026-41602-3

Keywords: wind energy, feature selection, maximum power point tracking, machine learning, renewable power forecasting