Accurate wind power forecasting is an effective means of improving the stability of wind power grid integration and the economic performance of wind farms. Aiming at the prominent challenge that the complexity and randomness of natural meteorological characteristics make wind power difficult to predict accurately, this paper proposes an ultra-short-term wind power forecasting method based on multidimensional data fusion and the joint optimization of a convolutional neural network-bidirectional long short-term memory network (CNN-BiLSTM). The method consists of two main phases. First, in the input data processing phase, a new multidimensional feature data is constructed to improve the accuracy of the prediction model, which combines the key meteorological factors selected by principal component analysis (PCA) with the wind power intrinsic modal components obtained via optimized variational mode decomposition (OVMD). Second, in the joint optimization phase of the forecasting model, a cascaded hybrid forecasting model integrating CNN and BiLSTM is constructed, and the red-billed blue magpie optimizer (RBMO) is employed to jointly optimize the CNN and BiLSTM models. This allows the complementary advantages of the two models to be fully exploited, further enhancing forecasting accuracy. Comparative analyses of wind power forecasting results demonstrate that the proposed PCA-OVMD-RBMO-(CNN-BiLSTM) method achieves higher prediction accuracy than other benchmark forecasting methods.