As offshore wind farms expand in our country, their share in the power system is increasing, making accurate forecasting essential for grid stability. This paper proposes a prediction model that integrates improved K-means clustering with DCNN-LSTM-Transformer weight allocation. Initially, the improved K-means algorithm is employed to classify weather types from the sample data. Atrous convolution is then used to extract and encode features from numerical weather prediction data. These encoded features are input into both LSTM and Transformer sequence prediction models to forecast power output for the next 24 h. The final prediction combines the results of these models, weighted according to their mean absolute error during training. Case studies indicate that the proposed method significantly reduces RMSE compared to using LSTM or Transformer models alone, demonstrating its effectiveness in improving offshore wind power forecasting accuracy. This approach aids in better scheduling and operational planning, maximizing wind resource utilization and enhancing the efficiency of offshore wind power generation.