Data-driven condition monitoring reduces downtime of wind turbines and increases reliability. Wind turbine operation and maintenance (O\&M) cost is a significant factor that calls for automated fault detection systems in wind turbines. In this manuscript, the anomaly detection problem for wind turbine gearbox is formulated based on adaptive threshold and twin support vector machine (TWSVM). In this work, SCADA data from wind farms located in the UK is considered with samples from thirteen months before failure. Gearbox oil and bearing temperatures are used as two univariate time-series for analyzing adaptive threshold. The effectiveness of the proposed method is compared with standard classifiers like support vector machines (SVM), k-nearest neighbors (KNN), multi-layer perceptron neural network (MLPNN), and decision tree (DT). Anomaly detection of wind turbine gearbox using TWSVM and adaptive threshold results in an accurate performance, thus increasing the reliability. The missed failure and false positive rate that indicate the proposed methodology's ability is also investigated to discriminate between false alarms, and comparison with previous studies shows superior performance.

The intelligent condition monitoring of wind turbines reduces their downtime and increases reliability. In this manuscript, a feature selection-based methodology that essentially works on regression models is used for identifying faulty scenarios. Supervisory control and data acquisition (SCADA) data with 1009 samples from one year and one month before failure are considered. Gearbox oil and bearing temperatures are treated as target variables with all the other variables used for the prediction model. Neighborhood component analysis (NCA) as a feature selection technique is employed to select the best features and prediction performance for several machine learning regression models is assessed. The results reveal that twin support vector regression (99.91%) and decision trees (98.74%) yield the highest accuracy for gearbox oil and bearing temperatures respectively. It is observed that NCA increases the accuracy and thus reliability of the condition monitoring system. Furthermore, the residuals from the class of support vector regression (SVR) models are tested from a statistical point of view. Diebold–Mariano and Durbin–Watson tests are carried out to establish the robustness of the tested models.

Fluctuation in wind speed over the period of time can impact operation of a utility grid. Sudden large scale variation in turbine power, often termed as ramp event, can threaten the security of power system. In this work, we predict ramp events through a hybrid method based on discrete wavelet transform (DWT) and learning algorithms such as Twin Support vector regression (TSVR), random forest regression (RFR), and Convolutional neural networks (CNN) for onshore, offshore and hilly sites. Wavelet transform‐based signal processing helps extract features from wind speed. Results suggest that SVR based prediction models are the best in forecasting among the available models. Besides, CNN predicts ramp events closer to the TSVR model and gives a better prediction performance for larger training datasets. The proposed hybrid version of TSVR, RFR, and CNN models are compared with existing SVM, ANN, and ELM models and indicate significant improvement in predicting ramp events. Compared to SVM, TSVR and RFR are 17.88% and 4.87% efficient in terms of RMSE, respectively. Further, randomness in ramp event signal for all the wind farm sites considered is evaluated using log‐energy entropy. Results reveal that compared to wavelet transform, empirical mode decomposition yields lower randomness in predicted ramp event signals.