Short-Term Offshore Wind Power Forecasting Based on Dynamic Trend Clustering of Meteorological Factors

Accurate short-term power forecasting for offshore wind farms is crucial for efficient grid integration and consumption of large-scale wind energy. However, the dynamic coupling and trend abruptness characteristics of meteorological factors pose severe challenges to forecasting accuracy. In this article, a short-term power forecasting method combining dynamic meteorological trend clustering and temporal attention mechanism is proposed. First, a two-stage hybrid clustering framework is designed: the first stage optimizes the fuzzy C-means (FCMs) clustering using an improved ant colony optimization (IACO) algorithm, precisely identifying the coupled patterns of meteorological features by adaptively adjusting the clustering centers and membership functions; and the second stage performs fine-grained clustering on wind speed time-series trends using cosine similarity, constructing a training dataset with both meteorological state similarity and trend consistency. Subsequently, a forecasting model integrating a multilayer bidirectional LSTM (MBLSTM) network and temporal pattern attention (TPA) mechanism is developed, deeply mining the nonlinear spatiotemporal associations between meteorology and power through bidirectional temporal feature extraction and multiscale attention weighting. Experiments based on two years of SCADA data from an offshore wind farm show that compared with traditional LSTM, BiLSTM, and other benchmark models, the proposed method significantly reduces forecasting errors under extreme conditions such as typhoon passage and monsoon abrupt change. This demonstrates that the method effectively solves the collaborative modeling problem of long-term trends and short-term abrupt changes in meteorological factors, providing a new technical framework for the prediction of highly volatile offshore wind power.