Wind farm power output prediction based on machine learning recurrent neural networks

Scientists, investors and policy makers have become aware of the importance of providing near accurate prediction of renewable energy. This is why current studies show improvements in methodologies to provide more precise energy predictions. Wind energy is tied to variabilities of weather patterns, especially wind speeds, which are irregular in climates with erratic weather conditions. To predict wind power output, model technologies like autoregressive integrated moving average (ARIMA), variants of ARIMA, hybrid models involving ARIMA and artificial neural networks (ANN), Kalman filters and support vector regressions (SVR) have been applied for wind speed involving short, ultra-short, medium and long terms kind of predictions. ARIMA ensemble with ANN has shown better performance for short and ultra-short terms of two to three hours ahead. On the other hand, SVR, Kalman filters and ensemble of both has recorded good performance for medium-term kinds of wind speed predictions. Recently, neural networks in particular recurrent neural networks (RNN) have reported immense achievement in time series predictions particularly for medium and long-term. This is largely due to its retentive memory-mapping capabilities in fitting sequence in series. These capabilities are short-lived; when the sequence grows over time, the RNN tend to lose correlated information on back-propagation operations. This can lead to errors in the predicted potentials. Therefore, RNNs are exploited for enhanced wind-farm power output prediction. The main contribution of this research is the study of a model involving a combination of RNN regularisation methods using dropout and long short-term memory (LSTM) for wind-power output predictions. In this research, the regularisation method modifies and adapts to the stochastic nature of the wind and is optimised for the wind-farm power output (WFPO) prediction for up to 12-hours ahead – 72-timesteps. This algorithm implements a dropout method to suit the non-deterministic wind speed by applying LSTM to prevent RNN from overfitting. A demonstration for accuracy using the proposed method is performed on a 14-turbine wind farm with up to ten thousand wind samples for model training and five hundred for model validation and testing. The model out performs the ARIMA model with up to 90% accuracy and is expected to be applied to erratic weather condition, especially those observed in an off-shore wind farms.