Abstract: To address the challenges of wind speed representation distortion and insufficient dynamic response in ultra short term wind power prediction, thispaper proposes a prediction method integrating high-fidelity inflow wind speed features with the TimeMixer architecture. First, considering the wind farm’s terrain characteristics, turbine layout, and wake interference effects, a spatially corrected wind field is constructed to estimate the undisturbed inflow wind speed at hub height for each turbine. Second, using this inflow wind speed as the primary input, a TimeMixer-based model is developed, incorporating a dual-channel decomposition mechanism for trend and high-frequency fluctuation components, and employing a multi-scale feature fusion strategy to enhance the model’s ability to jointly capture high-frequency fluctuations and low-frequency trends in power evolution. Finally, Real-world operational data from a mountainous wind farm in Qujing, Yunnan is used to validate the prediction method. Furthermore, we also compared the proposed model with nacelle wind speed model. The results show that the model with inflow wind speed as input has 18.75%, 25%, 7.69% and 6.37% higher than the nacelle wind speed model in terms of root mean square error (E_rmse), mean absolutely error (E_mae), r and Q_R respectively. Compared with traditional TCN and LSTM, TimeMixer performs better in trend tracking and fluctuation response. This method improves ultra-short-term forecasting accuracy and offers practical technical support for the stable operation of power systems with high renewable penetration.