A Data-Driven Approach for Modeling a Digital Twin of a Wind Turbine under Ideal Conditions

Wind turbine generators (WTGs) are one of the most widely used sources of renewable energy currently available. To accurately predict their production and quickly notice any anomalies, it is important to analyze the data produced by these turbines to understand their behavior and patterns. The purpose of this thesis is to create a data-driven digital twin of a wind turbine generator capable of simulating its ideal behavior. To carry out this task, the model receives input data of environmental variables, including wind speed and ambient temperature, and produces output values of parameters that a turbine should have under ideal conditions, including produced power, rotor speed, and more. The digital twin serves as a reference model that can be used as a comparison metric for the real turbines to evaluate their real-time performance and verify that the turbine is working properly by comparing the parameters of the internal components. The work was carried out in collaboration with the Turin-based company Sirius s.r.l. and exploits data provided by the company itself and collected at some wind farms in southern Italy. The data is acquired through Supervisory Control And Data Acquisition (SCADA) systems installed on the turbines with the aim of monitoring and collecting data both on the environment and on the internal components of the turbine. The work is structured in several parts: the initial part is characterized by data extraction and dataset creation. The data is in the form of a ten-minute average and is taken from turbines of the same model belonging to the same wind farm. Subsequently, a significant work was done on filtering the data with the goal of keeping only the data related to moments in which the behavior of the turbine can be defined as ideal. In this way, the algorithms can be trained only with ideal data and can adequately learn its trends without being misled by other non-ideal data. To obtain this result, multiple filters have been used considering both environmental and turbine variables, also using algorithms such as the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for outlier removal. Finally, several models have been created with the filtered data using different algorithms from both machine learning and deep learning, trying many combinations of inputs and outputs. Specifically, the Feedforward Neural Network (FNN), Support Vector Regression (SVR), and Gated Recurrent Unit (GRU) were tested. The tests can be divided into two parts: the first part is characterized by the fact that the algorithms are trained on multiple turbines and tested on others never previously seen by the algorithm. Instead, the second part concerns some variables representing the temperature of the internal components of the turbine, whose behavior is highly variable from turbine to turbine, even if they belong to the same model. In these cases, further studies were needed, leading to alternative solutions. In both cases, satisfactory results were achieved.