Posters

Abstract

The theoretical power production of a turbine is the amount of energy that the turbine can produce at a certain set of operating conditions. The original equipment manufacturers (OEM) provide this value by the supplying any buyer of their turbine with the contractual power curve which they guarantee their turbine can produce. It represents a lower limit of energy production capabilities of a turbine. Turbines, however, can significantly exceed the theoretical values provided by the OEM. Contrary to any contractual or legal work, this is important when detecting under performance or estimating lost energy due to stand still or under performance. In many cases, the contractual power curve is used for these calculations, neglecting the over performance of the turbine. Machine learning (ML) regression models are a widely accepted way of estimating the theoretical power production of a wind turbine. The estimated values are no longer based on the power curve provided by the manufacturer, but instead on the history of the turbine. To create a ML model that estimates the theoretical power of the wind turbine accurately, the ML model is usually trained on normal operation data only. This requires a manual filtering process, separating underperforming data points from normally operating data points. This presentation introduces a new approach to the problem, that the normal operation patterns of a turbine must be known a priori to train the ML model. Instead of a manual filter process, it is proposed that the machine learning model itself separates normal behavior from abnormal behavior. It can then base its theoretical power prediction only on operation data points and leave out the abnormal operation points. This is done by enriching the feature base of the model. It is shown that the new approach reaches at least the same accuracy when predicting theoretical turbine power. This approach eliminates the need for manually curating the training data set and, thus can easily be automated. This makes this innovative approach suitable for use in automated underperformance detection algorithms.