Intermittent sunlight leads to significant fluctuations in photovoltaic power generation, resulting in low accuracy in predicting the power output. In this study, a new ultra-short-term photovoltaic power generation forecasting model based on successive variational mode decomposition (SVMD), Bayesian optimization (BO) algorithm, and bidirectional temporal convolutional network (BiTCN) is proposed to improve prediction accuracy. Initially, the raw photovoltaic power generation data is decomposed into multiple power components and a power residual using SVMD to obtain multiple sequences with small fluctuations. Subsequently, the improved BiTCN replaces the temporal convolutional network (TCN) to perform bidirectional feature extraction and prediction of the SVMD decomposition results with low latency. Then, BO algorithm is used to search for BiTCN hyperparameters efficiently, so as to improve the prediction accuracy of each power component and power residual. Finally, the predicted results are summed and reconstructed to achieve ultra-short-term photovoltaic power generation prediction. Experiments demonstrate that the proposed model achieves a 35.18% reduction in root mean square error (RMSE) and a 4.82% increase in coefficient of determination compared to the single TCN model.