In the field of regional short-term combined forecasting of wind power, although deep learning methods can effectively learn the predictive features of each individual model, they tend to rely heavily on the training data distribution, which results in overfitting when the sample data size is small. Additionally, although clustering methods are used to improve the accuracy of regional-level forecasting, existing methods typically aim to minimize the dissimilarity among wind farms, such as geographic dissimilarity, without considering consistency with the forecasting objectives. To address these issues, this paper proposes a short-term wind power prediction method based on closed-loop clustering and multi-objective optimization. Initially, wind farms are divided into multiple clusters by the closed-loop clustering method. For each cluster, the Bootstrap method is utilized to randomly extract N training subsets with replacement from the original dataset. Subsequently, N convolutional neural networks are trained independently using these subsets. Finally, multi-objective optimization is employed to integrate the prediction results from the N convolutional neural networks. Case studies utilizing wind farm data from Inner Mongolia Autonomous Region, China, demonstrate that the proposed method reduces root mean square error by 33.81% compared with the long short-term memory model, by 24.08% compared with the convolutional neural network-based combined forecasting model, and by 14.05% compared to predictions based on the K-means clustering method.