Auto-tuned Deep Recurrent Neural Networks for Application in Wind Energy Conversion Systems

As the uncertain nature of wind energy is the main reason behind inconsistency in functioning of the wind energy conversion systems (WECS), the accurate forecasting of wind characteristics such as wind speed and wind direction is helpful for more realistic and consistent results. Due to the efficiency in addressing long term temporal features and extreme nonlinearities, the authors have utilized Long Short Term Memory Networks (LSTMs) to model the wind characteristics in this work. However, the heuristic way of hyper-parameter selection (architecture design and activation function choice) in LSTMs makes their modelling inefficient and tedious. A novel algorithm to design LSTMs optimally by balancing the trade-off between number of parameters and accuracy, is, therefore, proposed. These LSTMs are utilized for long term forecasting by training on 4 years of real wind data with an accuracy of $R^{2}=0.97$. Further, to demonstrate the significance of accurate forecasting, the predictions are used to generate wind frequency maps, which forms the basis for energy calculations in WECS. These maps are subsequently used for calculating energy produced by a windfarm layout containing 37 turbines, where a recently developed Gaussian wake model is utilized for accounting the turbulence due to the presence of multiple turbines. The results and comparative studies have shown that realistic energy production is possible with the use of original and accurately forecasted data obtained from optimal LSTMs.