As renewable energy units are connected to the grid on a large scale, the variability of their output poses a severe challenge to the stable operation of the power grid. In order to accurately predict the short-term output power of wind turbine units, a short-term wind power prediction method is proposed based on ensemble empirical mode decomposition （EEMD） algorithm and attention-enhanced bidirectional gated recurrent unit （BiGRU）. Firstly, the EEMD algorithm is used for data preprocessing of the non-stationary original environmental feature sequences collected by sensors on-site, decomposing the nonlinear environmental feature sequences into multiple intrinsic modes and residuals to reduce the impact of random disturbance signals on the training accuracy of the composite model. Then, the Attention mechanism is introduced into the BiGRU network, assigning weight coefficients to the corresponding hidden layer states to delve deeper into potential information within the environmental mode feature sequences. An improved grey wolf algorithm is then used to optimize the initial hyperparameters of the BiGRU-Attention network post weight coefficient adjustment, establishing association matrices for each environment mode subsequence. Finally, using the trained EEMD-OBiGRU-Attention combined model, multidimensional temporal and spatial predictions are performed on the decomposed modal feature subsequences at each layer, and the reconstructed set of modal feature predictions is used to generate output results. Example results show that, compared to traditional prediction models, the EEMD-OBiGRU-Attention combined model has higher robust fitting accuracy in on-site short-term wind power prediction.