In response to the complex marine extreme environments, nonlinear aerodynamic effects, and strong stochastic fluctuations of offshore wind power, this paper proposes an offshore wind power forecasting method that integrates an improved wavelet convolutional long short-term memory network with the Kolmogorov-Arnold Network architecture (CKAN-IWTC-LSTM). First, an improved wavelet convolution (IWTC) feature extraction method incorporating a star aggregation and distribution module (STAD) is adopted to enhance the interactive representation of multidimensional meteorological features offshore. Second, an LSTM forecasting model for offshore wind power with enhanced output based on the KAN architecture is constructed to capture the temporal variation patterns in offshore wind power data. Finally, a multi-dimensional evaluation framework is established, and the Shapley additive exPlanations (SHAP) method is employed to quantify the contributions of temporal and environmental features to offshore wind power. Case study analysis demonstrates that the proposed method effectively achieves accurate forecasting of offshore wind power.
AFRASIABI M,MOHAMMADI M,RASTEGAR M,et al.Advanced deep learning approach for probabilistic wind speed forecasting[J].IEEE Transactions on Industrial Informatics,2021,17(1):720-727.
FINDER S E,AMOYAL R,TREISTER E,et al.Wavelet convolutions for large receptive fields[M].Switzerland,2024.
HOU Y,Zhang D.Graph neural network-enhanced multivariate time series forecasting with series-core fusion[C]//2025 8th International Conference on Advanced Algorithms and Control Engineering(ICAACE). Shanghai,China,2025:1850-1854.