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

This article focuses on a novel farm-level probabilistic wind power forecasting method called long short-term-memory-improved autoregressive conditional heteroskedasticity (LSTM-IARCH). The approach integrates a new wind turbine clustering algorithm based on wind speed and direction similarities, a deterministic forecasting model using a seven-layer long short-term memory (LSTM) neural network for each cluster, and an improved autoregressive conditional heteroskedasticity (IARCH) model to estimate the heteroscedastic variance of prediction errors. Evaluated on real datasets from two wind farms in Pennsylvania and Shanghai, the LSTM-IARCH method demonstrated superior overall forecasting performance compared to classical clustering, non-clustering, and other probabilistic forecasting methods, particularly in terms of mean skill score (MSC) and continuous ranked probability score (CRPS). The study also highlights the importance of accounting for heteroscedasticity in prediction errors and suggests future improvements including dynamic clustering updates and incorporation of additional exogenous variables.