Abstract: Accurate prediction of wind power is an important means to realize friendly grid connection of large-scale offshore wind power. Large offshore wind farms have many units with different states. The influence of unit state, wake and space-time characteristics on wind power prediction cannot be ignored. Based on long short-term memory-temporal convolutional network (LSTM-TCN), an ultra-short-term power prediction method for offshore wind power was proposed in this paper, which considered the unit state, the wake of wind turbines and the spatial distribution characteristics of wind farms. Firstly, the influence of unit state and wake data on power prediction was analyzed, and then the deep learning prediction model of wind turbine operation data was established based on LSTM, which realized the mapping of unit health state to operation data, and continuously corrected the unit health state through real-time rolling of data. On this basis, the improved LSTM-TCN model was added with the modules of attention enhancement and random spatial characteristics weakening. Compared with TCN algorithm and LSTM algorithm, the proposed method could improve the accuracy of wind power prediction, especially for the common sudden change of wind speed at sea. This method improved the problem that TCN algorithm over-fits spatial characteristics. Based on the accurate prediction, it could be further used for coordinated optimization control of units in large-scale offshore wind farms, and thus improving the reliability of offshore wind power output.