In order to improve the operation efficiency of offshore wind turbines, improve the prediction accuracy of short-term offshore wind power value, and meet the future development needs of offshore wind turbines, In this paper, the data of supervisory control and data acquisition (SCADA) system are input into the neural network as test materials to predict the power of offshore wind turbines. To fully reflect the correlation between different components of offshore wind turbines and the coupling between multi-state information of SCADA data.In order to improve the timeliness of power prediction and speed up short-term real-time prediction, this paper adopts random forest (RF) to select features from high-dimensional SCADA data. To simplify the bidirectional-long short erm memory (Bi-LSTM) recurrent neural network structure, accelerate neural network convergence, and improve training speed. The results of a practical offshore wind farm in China show that the proposed prediction method has the highest prediction accuracy. Compared with the other three prediction models, the root-mean-square error of evaluation index, the average absolute error and the average absolute percentage error are 17.60, 17.48 and 0.21%, respectively. The short-term offshore wind power forecasting method proposed in this paper is expected to provide decision-making guidance for the future development and planning layout of offshore wind power operation and maintenance.