Traditionally, offshore wind power forecast research has focused on predicting power curves without extensive analysis of their variations. However, it is crucial to recognize that distinct types of offshore wind power output curves may lead to different operating modes of the grid. In this study, we address this gap by investigating the trends and patterns of offshore wind power output within a day to achieve a better understanding of grid operation modes. We propose an approach that integrates Long Short-Term Memory (LSTM) and Convolutional Gated Recurrent Unit (GRU) prediction models with K-means clustering for the prediction and categorization of offshore wind power. Specifically, we utilize LSTM and GRU models to predict offshore wind output, followed by a comparison of the effectiveness of these two prediction methods. Subsequently, we apply K-means clustering to categorize the daily predicted curves into distinct groups. By analyzing the curves of different categories, we can identify types of curves suitable for different grid operating modes. Finally, we validate the effectiveness of this methodology through verification analysis using actual offshore wind power data from Zhuhai, providing insights to enhance grid operation.