An offshore wind speed forecasting system based on feature enhancement, deep time series clustering, and extended LSTM

Offshore wind power, closely linked to marine conditions, exhibits stochastic and intermittent behavior, challenging power system stability. To address the complex characteristics of offshore wind speed data, this study proposes a novel wind speed prediction system integrating feature enhancement, deep temporal clustering, and extended long short-term memory (xLSTM). The system employs a three-stage optimization: First, antlion optimization autonomously adjusts variational mode decomposition parameters, while fast Fourier transform extracts long-term trends and fluctuations, constructing a feature enhancement strategy to suppress chaotic effects. Second, deep temporal clustering, using a convolutional neural network and bidirectional LSTM, dynamically groups wind speed sequences based on multi-modal similarity metrics. The TOPSIS-entropy weight method scores clustering models, ensuring precise test set matching. Finally, xLSTM independently models and predicts each cluster, adapting to varying conditions. Cluster-based modeling reduces computational burden and enhances efficiency. Experimental results show that the system performs well in the comparison models of three Chinese offshore wind farms, and the mean absolute error (MAE) is reduced by at least 36.9 % compared with the comparison models. Transfer learning verified the generalization ability of the system, and coefficient of determination (R2) reached more than 0.99 in eight of the nine target sites. This study provides a high-precision, regionally transferable solution for offshore wind speed prediction, supporting large-scale offshore wind integration.