Abstract
Wind power is one of the fastest-growing renewable energy sectors and is considered instrumental in the ongoing decarbonization process. However, wind turbines (WTs) present high operation and maintenance costs caused by inefficiencies and failures, leading to ever-increasing attention to effective Condition Monitoring (CM) strategies. Nowadays, modern WTs are integrated with sensor networks as part of the Supervisory Control and Data Acquisition (SCADA) system for supervision purposes. CM of wind farms through predictive models based on routinely collected SCADA data is envisaged as a viable mean of improving producibility by spotting operational inefficiencies. However, given the large number of variables monitored by SCADA systems, selecting those that contribute the most to the modelling of wind turbine health conditions is an open challenge. In this paper, we propose an unsupervised feature selection algorithm based on a novel multivariate Predictive Power Score (PPS). Unlike other approaches in literature that only consider relationships between pairs of variables, here we propose a Combined PPS (CPPS), where the information content of combinations of variables is considered for the prediction of one or more key parameters. The algorithm has been tested on 9 turbines belonging to the same wind farm located in the Italian territory. The results show that the proposed approach is more flexible and outperforms standard PPS.
