Accurate prediction of wind power is significant for enhancing the efficiency and safety of power systems. Offshore wind energy is intermittency and randomness, making offshore wind power prediction challenging. A new model named Boruta-SSA-CKCformer, an enhanced version of the Crossformer model architecture, was proposed for optimizing offshore wind power prediction. The Boruta feature selection algorithm was integrated into the front end of the model to optimize input feature selection, thus improving prediction accuracy. The convolutional neural network was incorporated into the dimension-segment-wise embedding layer of the Crossformer model, forming the convolutional dimension-segment-wise embedding layer(CDSW). This effectively captured local features and cross-dimensional dependencies in multivariate time series data. Kolmogorov-arnold networks(KAN) connections were added to the two-stage attention layer, establishing the KAN-based two-stage attention layer (KTSA), which addressed deficiencies in non-Gaussian noise processing by the Crossformer model. Model parameters were optimized using the sparrow search algorithm(SSA) to further improve prediction accuracy. Empirical studies on data from offshore wind farms in Fujian were carried out. The Boruta-SSA-CKCformer model shows superior performance in offshore wind power prediction compared to other models.