A modular multi-step forecasting method for offshore wind power clusters

Offshore wind farm clusters, driven by economies of scale, are emerging as a prevalent trend. However, the intermittency and volatility of offshore wind power due to wind resource uncertainties pose significant challenges for forecasting. Existing research on offshore wind farm cluster power forecasting remains limited. This paper addresses this gap by proposing a modular and decoupled multi-step forecasting method for offshore wind farm clusters. The modular design enables adaptability to various forecasting scenarios, particularly with and without Numerical Weather Prediction (NWP) data, providing a flexible framework for future research and applications. The method leverages the spatiotemporal information of all wind farms within the cluster by first preprocessing the historical power output series using signal processing techniques, including Fast Fourier Transform (FFT) and Singular Value Decomposition (SVD), to decompose and denoise the data. Spatiotemporal feature extraction is then achieved through an Enhanced Long Short-Term Memory (LSTM) network, combining two-dimensional convolution and LSTM layers. Subsequently, incorporating NWP data, a Light Gradient Boosting Machine (LightGBM) model is employed for final forecasting. Finally, the proposed method is validated on a wind farm cluster in the eastern coastal region of China, demonstrating its effectiveness, accuracy, and generalizability at both individual wind farm and cluster levels.