ICE CLASSIFICATION ON WIND TURBINE BLADE USING RECURRENT NEURAL NETWORKS WITH ATTENTION LAYER

Sammanfattning

The wind industry is expanding and areas in cold and remote locations can have good potential for location of wind turbines. Due to the cold climate ice can occur on sensors, wind turbine and the blades. The annual production losses can according to IEA be up to 20% in some areas located in the north. Methods to prevent ice on the blades can be using anti or de-icing systems. The wind turbines can also be shut down to prevent damage on the equipment due to ice conditions or for safety reasons such as risk for ice throw. Ice detection is an important part considering these conditions and for estimating the annual estimated production for sites in cold climate. A method to detect and classify ice from SCADA data is to use deep learning. Deep learning is continuously improving and developed. To increase the performance of deep learning models they can be combined to use the properties for different models together. To increase the performance of a RNN (recurrent neural networks) model the proposed models in this paper is RNN with added attention layer. Three RNN models is used with and without attention layer in order to compare the results. The models is LSTM (Long term short term memory), GRU (Gated recurrent unit) and SimpleRNN (Simple recurrent neural network). Open-source SCADA data provided by IEA is used for training and validating. The result shows increase in accuracy for each model and increase in F1 scores, both with respect to classification of true positive and true negative. This means that the models are more balanced with higher precision and recall. The highest increase for F1 score (true positive) was 0,86% for LSTM with attention layer and the highest increase for F1 score (true negative) was 1,19% for SimpleRNN model. The highest accuracy increase was also for SimpleRNN with attention layer by 0,86%. The overall highest performance had LSTM with attention with an accuracy of 0,922 and F1 score (true positive) 0,925. When compared to other study using the same open- source data the RNN models used in this study had high performance. The performance was high both with and without attention layer with an accuracy >0,907 and F1 score over >0,909 for each model.