A farm-level wind power probabilistic forecasting method based on wind turbines clustering and heteroscedastic model

Abstract

Large grid-connected wind farms face challenges in predicting wind power output due to the uncertainty, volatility, and intermittency of wind. The heteroscedasticity of wind power prediction errors further complicates the reliability of forecasts. This study presents a novel approach, termed long short-term-memory-improved autoregressive conditional heteroskedasticity (LSTM-IARCH), which combines a long short-term-memory model with an improved autoregressive conditional heteroskedasticity model. We first propose a novel clustering technique to group wind turbines and develop deterministic wind power prediction models based on LSTM within each cluster. The prediction interval for wind energy is determined using the variance of the prediction error from the improved ARCH model. The performance of the approach is evaluated using real data from two wind farms and compared against various popular probabilistic prediction methods. The results of the comparison demonstrate the advantages of this method in probabilistic prediction at the wind farm level.