Abstract: To address the issue of insufficient extraction of spatio-temporal correlation features by traditional prediction models, a two-stage short-term wind power prediction framework considering spatio-temporal feature fusion and multi-factor error correction was proposed. Firstly, the original wind power sequence was denoised and decomposed by symplectic geometry packet with signal fidelity characteristics, while continuous time-series feature values were obtained and integrated with the operation sequence of wind turbines to form a time-series feature matrix. Secondly, a graph convolutional network with spatial modeling capability was introduced, by which the time-series feature matrix and the spatial graph structure of the wind farm were integrated into a spatio-temporal matrix. Finally, the spatio-temporal matrix was predicted using Transformer-bidirectional long short-term memory(BiLSTM) model optimized by Newton-Raphson-based optimizer. A dynamic error correction mechanism was established through the prediction model, and the original prediction results were corrected based on the predicted error values, forming a prediction system with self-correction capabilities. Results demonstrate that the root mean square error and mean absolute error of this model are reduced by an average of 41.4% and 36.4%, respectively, compared with other combined comparison models, and the coefficient of determination is increased by an average of 6%. It performs optimally in the datasets of spring, summer, autumn, and winter, as well as in 72-hour continuous prediction, significantly improving the accuracy of wind energy prediction and providing reliable support for wind power grid connection.