The contribution of renewable energy sources either in the power grid or in smart microgrids is becoming more and more prevalent, and a further significant future growth is expected due to various energy transition initiatives around the world, especially in the EU. Hence, innovative technologies shall be used to deliver economic, efficient, and sustainable solutions that conform with these initiatives. Data science and machine learning are revolutionary enabling technologies that can enhance the performance of renewable energy systems (Brunton and Kutz, 2019; Nelles, 2001). Among the many possible application areas, the focus of this work is on construction of reduced order models. The constructed models can deduce the internal states in the form of virtual measurements to support live condition monitoring, the synthesis of modern control laws and virtual sensing, among others (Bogoevska et al., 2017; Cross and Ma, 2014). These reduced models have a low computational cost compared to white-box physics-based models, but they are still able to accurately predict the behavior of quantities of interest. This work explores the performance of different neural network based techniques to build data-driven black-box models of wind turbine drivetrains. The data used for model synthesis can either be derived from high-fidelity numerical simulations and/or from field measurements. The objective of the reduced order models is to infer some of the internal dynamic states of interest of the drivetrain, such as internal transmitted forces and elastic deformations of gears and shafts, based on measurements provided by the typically available SCADA sensors (e.g., generator and rotor torques, and generator rotational speed). A main challenge of this problem is the reduction of the state space without forfeiting the quality of predictions. Here, this issue was solved using principal component analysis (PCA), which showed a high quality of predictions in the low frequency spectrum with a relatively low number of reduced states. The suitability of the mentioned approached is investigated with respect to the high frequency content of the predicted signals, such as tooth contact at different stages of the drivetrain for live condition monitoring applications and fatigue lifetime estimation. Good quality predictions in the frequencies of interest (low frequency spectrum) were obtained with Nonlinear Autoregression with Exogenous Input (NARX) and Hammerstein-Wiener (HW) models with a kernel based on a nonlinear feedforward neural network. Using long/short-term memory (LSTM) neural networks as kernel demonstrated an even higher quality of predictions, as LSTM cells can simultaneously infer both the high and low frequency content of the system’s response (Simpson, Dervilis and Chatzi, 2020) . Initial results show correlations exceeding 85% for the inferred states. The proposed approaches are validated in realistic IEC load cases, such as power production (with and without faults). These models have very high computational efficiency and are able to run faster than real time.

{"references": ["Bogoevska, S. et al. (2017) 'A Data-Driven Diagnostic Framework for Wind Turbine Structures: A Holistic Approach', Sensors, 17(4), p. 720. doi: 10.3390/s17040720", "Brunton, S.L. and Kutz, J.N. (2019) Data-driven science and engineering machine learning, dynamical systems, and control.", "Cross, P. and Ma, X. (2014) 'Nonlinear system identification for model-based condition monitoring of wind turbines', Renewable Energy, 71, pp. 166\u2013175. doi: 10.1016/j.renene.2014.05.035", "Nelles, O. (2001) Nonlinear System Identification. Berlin, Heidelberg: Springer Berlin Heidelberg. Available at: http://link.springer.com/10.1007/978-3-662-04323-3.", "Reference list Bogoevska, S. et al. (2017) 'A Data-Driven Diagnostic Framework for Wind Turbine Structures: A Holistic Approach', Sensors, 17(4), p. 720. doi: 10.3390/s17040720 Brunton, S.L. and Kutz, J.N. (2019) Data-driven science and engineering machine learning, dynamical systems, and control. Cross, P. and Ma, X. (2014) 'Nonlinear system identification for model-based condition monitoring of wind turbines', Renewable Energy, 71, pp. 166\u2013175. doi: 10.1016/j.renene.2014.05.035 Nelles, O. (2001) Nonlinear System Identification. Berlin, Heidelberg: Springer Berlin Heidelberg. Available at: http://link.springer.com/10.1007/978-3-662-04323-3 (Accessed: 30 March 2020). Simpson, T., Dervilis, N. and Chatzi, E. (2020) 'On the use of Nonlinear Normal Modes for Nonlinear Reduced Order Modelling'. Available at: http://arxiv.org/pdf/2007.00466v1."]}