Posters

Abstract

We presenta novelmachine learningmethodologyenablingtheidentification of windturbines characterized by similar energy profiles. The methodology leverages on contextual factors such as wind speed and wind direction to cluster turbines in a context-aware fashion. Turbines in the same cluster are subsumed by a single prototypical model which can be used as a model for energy estimation for that cluster. Such a model is dynamically associated with a turbine to cope with the time-variant behavior of the turbine.This methodologysupports the control and monitoringof the farm. First, context-aware clustering canrevealthe existence of topologically-dependent connections among turbines and potentially reveal anomalies. Second, computationally time-consuming modelsused forfleetsimulation(to guarantee power network stability)can be replaced by a limited number of prototypical models that can run fast in parallel. Third, the estimation of the compensation associated with an energy loss (e.g. due to turbine shut down) can be based on the expected energy production of the prototypical model. To validate the methodology the real-world data coming from a fleet of 40+offshorewind turbines was used. As a first case, we applied context-aware clustering on the fleet and then visualized turbine positions in the farm. This highlighted that the identified turbine clusters are topologically meaningful (e.g. they reflect the existence of the wake effect). As a second case, we extracted the prototypical models (from the identified clusters) for energy forecast. By describing the entire fleet usingsuchmodels we achieveda good forecast accuracy.