Normal behaviour modeling of hawt fleets using scada-based feature engineering

Wind turbine O&M presents a challenge for large wind turbine fleets, where many turbines are scattered across different wind farms, exposed to non-homogeneous environmental conditions and varied operational histories. In this scenario, operators can leverage big data to address this challenge, with the objective of deriving O&M strategies common to subsets or communities of turbines, either to monitor their behavior collectively or to schedule custom maintenance. In this paper, we introduce an unsupervised data-driven framework to isolate communities of wind turbines based on their actual operational histories, as a precursor to asset-specific O&M strategies. The approach is based on a novel Meta Predictive Power Score (MPPS) to derive a behavior model for horizontal axis wind turbine fleets. The proposed framework relies on feature derivation, feature combination, and community detection algorithms. The feature derivation component involves the use of a multivariate feature selection algorithm based on the Combined Predictive Power Score (CPPS), in which a regression task quantifies the information content of combinations of variables for the prediction of one or more target parameters. A feature combination procedure defines the core of MPPS, where in a combination of CPPS scores and selected features for different regression tasks on diverse versions of the same turbine SCADA data results in a similarity matrix for the entire fleet. A community detection algorithm, based on Complex Network Analysis, is then used to identify groups of wind turbines within a fleet that exhibit similar behavior. The dataset used comprises 67 wind turbines, including diverse OEMs. The proposed algorithm employs a decision tree regressor considering lagged time series to compute the MPPS. The results demonstrate the flexibility of the multi-input multi-output formulation. The target of the regression tasks required to build the scores is a hypersurface that can describe the operational state of a wind turbine, defined by active power, blade pitch angle, and rotational speed.