Improving Digital Twin Using AI and SCADA

Renewable energy transition is undergoing at a fast pace. To ensure an economic viability, thus ensuring its continuous development, an accurate estimation, and means of improvement regarding the asset production capability and expected operational lifetime is demanded. For a wind turbine, its controller is a key aspect governing its lifetime performance, with a direct impact on the development of fatigue [1]. In this work, we seek to improve wind farm digital twins by identifying the controller through a black box approach leveraging on-site SCADA data and deep learning. Because current controller's implementations mostly rely on virtual controller with hand-tweaked parameters [2], we aim to develop this method to automatically find these laws. To effectively leverage deep learning techniques, SCADA data are preliminary pre-processed (filling missing values, removing abnormal operation, differentiating, etc.). A random forest-based surrogate model is then trained, to predict the pitch and torque time evolution, based on a selected time window. The surrogate model is integrated in close loop within the digital twin, providing accurate prediction in region 2 & 3 (see appendix [a] for regions details), but the model struggles to understand the behaviour when reaching region 2.5 where the wind turbine highly undergoes strong cycling loading and experience fatigue. Our preliminary results indicate that the deep learning model shows promising improvements in predictive accuracy over the previous method. We expect that further fine-tuning and work on SCADA will enhance the model's understanding of region 2.5. Preliminary simulations suggest a better reproduction of the controller behaviour in this critical region, resulting in a more reliable representation of the wind turbine's behaviour. We then propose a comparison with ROSCO, an open-source reference controller for wind turbines.