Wind Turbine Condition Monitoring Through Artificial Neural Networks Using SCADA Data

Wind power has been an important part of the energy market for the last decade. Anyway, the main problems associated with wind farms are operation and maintenance costs, which represent up to one-third of the total cost of energy production. The purpose of this thesis is to analyze the behavior of the turbines located on a wind farm in southern Italy to detect in real-time any malfunctions or alarms generated by the control system. Several Feedforward Neural Networks (FNN) and Recurrent Neural Networks (RNN) models have been created starting from data coming from Supervisory Control And Data Acquisition systems (SCADA). This made it possible to build both the characteristic behavior of the critical components of wind turbines and a Statistical Process Control (SPC), useful to evaluate their anomalous behavior. This thesis is part of a collaboration between the Politecnico di Torino and the Turin-based company Sirius. The work was divided into several phases. The first of these concerned data acquisition. A preliminary analysis of the information available for a wind farm was carried out to analyze both the number of sensors available and the possible malfunctions that may fall within the cases of interest. An extraction and formatting of sensory data averaged over 10 minutes were then performed. The second phase regarded data preprocessing. Through various data filtering and cleaning operations, it has been possible to isolate the data in which the turbines do not show failures and have no operating limitations or anomalous behavior. The third phase concerned model processing. Different Multilayer Perceptrons and Long Short-Term Memory (LSTMs) models were created and trained to represent the normal behavior of the wind turbines which was used as a reference to identify unusual behaviors. The models with the best performance were selected for each output variable. The fourth and last phase regarded the analysis of the results. The output variable generated by each of the best models was compared with the actual value to evaluate the possible applicability of the system for condition monitoring. Real-time behavior was reproduced to achieve the monitoring and prevention goal. Unfiltered data from additional turbines of the same model and in the same wind farm were used to calculate the ideal behavior of the output of interest. Subsequently, an analysis of the deviation between the real value and the estimated one was carried out. The deviation is compared with the selected control charts and their limits, focusing on where the actual behavior differed significantly from the ideal. This solution tries to predict different types of output variables. With the aid of some known alarms, it is possible to demonstrate that the trained systems could detect failures and anticipate alarms signaled by the control system, in particular the prediction of a general state of emergency of the turbines and faults on the gearbox.