Comparative study of automated deep learning techniques for wind time-series forecasting

Despite wind being clean and sustainable, the main reason for inconsistent power production from a wind farm is its uncertain nature. Due to its capability in addressing the energy demand, the researchers aimed toward the accurate forecasting of time-series wind data for more realistic results. Since deep learning techniques maneuver the intense nonlinearities, the authors in this study compared recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) for accurate forecasting of wind characteristics, wind direction and speed. However, the trial-and-error way of design of architecture and activation function makes the modeling of deep learning techniques as inefficient and tedious. Therefore, the authors propose a novel and generic automated machine learning strategy to design them optimally under the framework of multiobjective optimization solved by NSGA-II. The study in this work demonstrates the importance of forecasting and its impact in wind farm design and control.