A Deep Learning Framework for Forecasting Power in a Full-Scale Wind Farm

Accurate wind power forecasts are essential for efficient power trading and decreased penalty rates for wind power producers. Highly fluctuating wind speeds and other variables like atmospheric stability contribute to the high variability in the wind power produced. A deep learning framework is developed to forecast the future power predictions in a full-scale wind farm. For this purpose, Long Short-Term Memory (LSTM) neural networks which is known to be powerful in forecasting tasks using time series data is used. In this study, we use 10-minute average power data for 12 months obtained from a wind farm with 86 wind turbines and hourly-averaged meteorological (Met) data from the nearby Met towers to conduct the analysis. The LSTM model is trained to predict the total power generated from all 86 wind turbines. The objectives of the study are (a) to predict the future power generated by the wind farm for 0 - 4 hours ahead, (b) to investigate the effect of including previous one, five, ten and fifteen hours as input parameters for forecasting future power. The hyper parameters of the LSTM network are optimized using advanced design of experiments. This study also investigates the LSTM model trained to predict the average power produced by the wind farm. Root Mean Squared Error (RMSE) is used as a performance metric in this study.