Come meet the poster presenters to ask them questions and discuss their work
Check the programme for our poster viewing moments. For more details on each poster, click on the poster titles to read the abstract.
PO025: Large-scale turbine performance monitoring using adaptive Machine Learning
Bruno Pinto, Chief Technology Officer, Sereema
When we speak about applying machine learning algorithms to turbine data and more specifically to power curve data, everybody has most likely seen (or done) studies and examples where one specific Machine Learning (ML) algorithm has been applied to one standalone dataset in order to, for example, filter external restrictions, identify underperformance, etc. It is widely known that ML algorithms can provide good results to large operational datasets. However, when it comes to developing an industrial application using ML algorithms that are valid for a wide range of turbine types and different datasets, things start to get more complicated. In this presentation, we will be showcasing a method developed in-house to apply ML algorithms to wind turbine data for performance monitoring. Clustering techniques such as Gaussian Mixture Models (GMM) & Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithms have been combined and applied to a large dataset of wind speed, turbine power and pitch angle data from different turbines and turbine models in order to obtain a calibrated, multi-layered model that can be used for different applications. In order to handle the high variability of power curves from the different turbine models, the uncertainty of the input data (imprecise wind speed data for example) and how the external conditions can change the power curve, the ML algorithms and their filtering parameters are automatically adapted to the dataset, this creates an adaptive ML model that will filter data differently depending on each turbine normal power curve behaviour. The inputs of the clustering algorithms are a set of pitch and wind speed data that have previously been binned by turbine power. The model calibration was processed from a training dataset composed of multiple years of historical data on over 50 wind turbines of 10 different turbine models. The filtering of the external restrictions from the power curve is then obtained by applying the multiple layers of filtering and Machine learning algorithms described above.The large-scale applicability of the model was tested and validated by applying it to over 10.000 monthly power curves from over 1.000 operating wind turbines. The presentation will include cases where different turbine operational conditions have been identified and how they correlate with external conditions such as wind sector or time of day. In addition, results will be presented to detail how the different components of the models help improve certain aspects of the filtering, for example: how adding pitch angle data to the ML model significantly improves the accuracy to identify, for example, turbine restrictions such as sound modes. To conclude, the overall accuracy of the model will be presented and potential model improvements will be detailed. To summarise, this presentation will make emphasis on what are the key aspects to take into consideration when applying Machine Learning models to the development of an industrial and automated approach to monitor wind turbine performance.