Wind power generation is significantly influenced by environmental factors, exhibiting high randomness and volatility, making it difficult to accurately predict the amount of electricity generated. To address this, an ultra-short-term wind power forecasting model combining multi-head self-attention, convolutional neural network, and long short-term memory (AM-CNN-LSTM) is proposed. The multi-head self-attention mechanism assigns weights based on the varying importance of historical wind power and meteorological data, generating key feature representations and mitigating interference. The convolutional neural network extracts complex spatial patterns and local dependencies between wind power and meteorological factors. The long short-term memory network captures long-term dependencies and dynamic changes in time series, enhancing forecasting accuracy. The results demonstrate that under the optimal time window of T=4 h, the model achieves mean absolute error, root mean square error, and mean absolute percentage error of 13.45 %, 19.48 % and 3.78 %, respectively, It outperforms comparative methods such as LSTM, CNN and CNN-LSTM. A comparison of forecasting errors for the optimal time window of different models over the next eight sample points reveals that the proposed model offers significant advantages in ultra-short-term forecasting accuracy and stability. Variable analysis indicates that the combined use of multiple meteorological variables significantly outperforms single-variable, uncovering the complex interactions and nonlinear relationships among meteorological factors.