To address the challenges of uncertainty

volatility

and non-stationarity of wind speed in wind power forecasting

this paper integrates physical mechanisms such as wake effects and wind energy conversion with neural network techniques. We propose an ultra-short-term wind power forecasting method considering the dynamic spatiotemporal characteristics of wind speed and the wind speed-power conversion mechanism. First

historical data

including the wind speed and pitch angle measurements at each wind turbine hub within the sliding window

are used to construct an input feature map. A deep residual contraction network is employed for denoising and extracting spatial distribution characteristics of wind speed

thereby reflecting how wake effects influence wind speed distribution in the wind farm under varying meteorological conditions such as wind direction. Subsequently

a bidirectional long short-term memory (BiLSTM) network links these spatial wind speed features with meteorological forecast data

capturing the dynamic temporal evolution patterns of wind speed over time and enabling predictions of hub wind speeds for future periods. Finally

the predicted wind speed of each turbine is fed into an embedded neural network developed on the basis of the wind-speed-to-power conversion mechanism to obtain the predicted power output for both individual turbines and the entire wind farm. Practical results show that the proposed method can be adopted to improve the stability and accuracy of the ultra-short term wind power prediction model.