Impact of target variable distribution type over the regression analysis in wind turbine data

摘要

The Wind sector has roughly 2200M euros of profit losses due to wind turbine failures and these failures do not contribute to the goal of reducing greenhouse gas emissions of many states. The 25-35% of the generation costs are operation and maintenance services. To lower this ratio, the wind turbine industry is backing on the Machine Learning techniques over SCADA data. Signal trending analysis supported on linear regression models presents the problem of how to carefully choose the right target variable, which reproduces as close as possible the behavior of a failure from a component. This document evaluates the impact of that choice by comparing as target different variables with discrete-non normal distribution, commonly selected by feature selection methods, versus variables that are continuous over time with a near normal distribution. Experimental results on real data show the use of continuous target variables selected by human expert on the field give better results than the use of targets obtained through feature selection algorithm.