To address the issues of data correlation and error correction adaptability in wind power forecasting models, an ultra-short-term wind power prediction method based on hyperparameter optimization and error correction unit switching mechanism is proposed. First, a spatiotemporal attention gated network (STAGN) forecasting model is developed, and hyperparameter optimization is carried out using the improved Kepler optimization algorithm. Second, an error correction adaptive unit is constructed by considering the nonlinear correlation between wind farm data and forecasting errors. Meanwhile, the temporal variation characteristics of wind speed are explored to construct a deep learning unit. On this basis, the error correction unit switching strategy based on the wind speed matrix gradient is proposed. Finally, the model is applied to power forecasting in an actual wind farm and compared with other models. The results show that the proposed method outperforms others in terms of forecasting accuracy and maintains high forecasting accuracy in wind farms with highly variable wind speeds, verifying its accuracy and applicability.