Experimental Prediction Intervals for Monitoring Wind Turbines: an Ensemble Approach

This paper proposes a data-driven approach for monitoring the operative conditions of wind turbines in wind farms. An ensemble of polynomial regression models with l1-norm regularization is derived via pairs bootstrap, to predict the active power produced and to provide a reliable prediction interval. Further, to improve the prediction performance, more complex ensemble multi-models have been derived by sampling the data from different input data clusters reflecting the turbine operative conditions. These models are then combined by using a specific fuzzy membership function fitted from the experimental data. For a given turbine the models are built using as inputs the operational and non-operational (i.e., environmental) variables of the turbine. Then, in order to improve the prediction accuracy, other correlated wind-field information, provided by the environmental variables measured by the nearby turbines have been added as further input regressors as these came from a distributed sensors network. The proposed methods have been designed and validated on experimental data from the Supervisory Control And Data Acquisition (SCADA) system of five turbines in a wind farm in Italy.