Rapid warning of wind turbine blade icing based on MIV-tSNE-RNN

A fast early warning algorithm for wind turbine blade icing based on a RNN model is proposed. Through wind turbine blade history data and labels as model input, the evaluation of raw m-dimension data through mean impact value (MIV) indices eliminates data with an MIV index of less than one; the remaining n-dimension data is reduced to x-dimension by the tSNE method; dimensional data is inputted into the RNN, and the model output is the icing state of the wind turbine blade in a certain future period. Based on the SCADA data from a wind field, the model was verified by an example. Using a certain example case, if the model training data is 10 4 orders of magnitude, using the MIV-tSNE-RNN algorithm, the prediction accuracy can reach approximately 72 %; compared with the RNN model, the prediction accuracy is improved by approximately 150 % while reducing the algorithm running time by approximately 45 %. If the amount of data exceeds 10 4 orders of magnitude, using the MIV-tSNE-RNN algorithm, the prediction accuracy is improved by approximately 100 %. This algorithm can provide accurate and rapid prediction results for wind turbine blade icing according to actual needs.