Neural Networks for Offshore Wind Turbine Converter Failure Prognosis

As offshore wind continues to expand in size and to more challenging environments, costs need to decrease to ensure it remains financially viable. Operations and maintenance costs, which are estimated at 30% of lifetime costs offer an opportunity for cost savings. This could be achieved using a condition-based maintenance strategy to minimise unnecessary maintenance actions while eliminating failures in operation. To achieve this, models that can predict turbine failures must be developed. The analysis of an operational wind farm’s data reveals that converter failures account for a significant portion of maintenance downtime and revenue loss due to their sudden and unpredictable failures in operation. To tackle the problem of predicting their failures an artificial neural network has been trained using SCADA data. The prognosis task is defined as a binary classification problem where the target variable is a time to failure interval. It is shown that it is difficult to train a model which generalises across multiple failures; however, there is indication that deep learning can be used to make failure predictions. It is further demonstrated that performance of models must be considered in the context of maintenance decision making. 10 decision making scenarios, based on varying prediction certainty thresholds are considered. When model performance is considered in the context of a correct maintenance decision, performance more than doubles. Producing models with generalizable predictive power and combining them with optimal maintenance scheduling methods will generate extra wind farm revenues, reduce the number of maintenance actions taken and facilitate the work of operators.