ANALYSIS AND DETECTION OF WIND TURBINE BLADE PERFORMANCE USING MACHINE LEARNING ALGORITHM

Abstract

Wind energy is one of the fastest growing renewable energy enterprises in the world. The structural integrity of wind turbine blades is of prime importance for the efficient, safe and economic operation of wind farms. Slow erosion and corrosion of turbine blades especially from high velocity impact from rain, hail, sand and UV radiation cause large aerodynamic performance losses and unpredicted maintenance costs. Tra ditional condition monitoring approaches mainly rely on human periodic inspections using deterministic threshold approaches, which are not adaptive to the dynamic non linear operational environment of modern wind turbines. This thesis proposes a pre dictive maintenance framework for wind turbine blades based on machine learning. The method is meant to classify the structural health state of blades into four opera tionally relevant classes: Good, Minor degradation, Advanced degradation and Major degradation. The data used in the present study was obtained from the National Re newable Energy Laboratory (NREL) Wind Resource Database for the Mumbai coast line region and consisted of 8,762 hourly records of atmospheric parameters at multi ple altitudes (wind speed, temperature, air pressure and wind direction). But there is no explicitly available blade health indictors on the NREL data set. Thus a physics informed stress index was developed to indicate the state of blade health categories. The stress index has taken the main environmental variables wind speed at 100 and 120 meters weighted by experimental coefficients and ambient temperature in order to indicate an appropriate blade health degradation proxy. Four supervised classifica tion techniques,Logistic Regression,Decision Tree,Random Forest and XGBoost are well-trained and well-tested. Training is carried out on 80% of the processed data and testing is carried out on the rest 20% of the data in terms of Accuracy, precision, recall, F1-score and confusion matrix evaluation. It is also concluded that XGboost achieved the best classification accuracy of 99.32% which is followed by Logistic Regression accuracy of 99.20%, Random Forest accuracy of 99.09% and Decision Tree accuracy of 98.57% . Analysis of feature importance of tree based models identified wind speed at height of 100m and 120m height as the key factors to predict blade degradation. The framework also proves the feasibility of data driven predictive maintenance system for wind turbine blade and acts as a foundation for real time monitoring and scheduling of the wind turbines under operation.