To enhance the accuracy of ultra-short-term photovoltaic power forecasting under meteorological disturbance scenarios, a hybrid model integrating an improved rime optimization algorithm (IRIME), a temporal convolutional network (TCN), a multi-strategy optimization temporal channel attention (MSOTCA) mechanism, and a bidirectional gated recurrent unit (BiGRU) was proposed. Firsty, key meteorological features were selected using Pearson correlation analysis and fed into the TCN for preliminary temporal feature extraction. Next, an MSOTCA mechanism was designed by integrating multi-scale temporal information, meteorological ramp enhancement, and residual connections to strengthen the model’s dynamic responsiveness to complex disturbance patterns. Subsequently, an IRIME algorithm enhanced with a meteorological-fluctuation-aware dynamic convergence factor was developed to automatically optimize the model hyperparameters. The BiGRU was then employed as the temporal decision layer to output point forecasts. Finally, prediction intervals were generated via Monte Carlo sampling to quantify uncertainty. Experiments were conducted using measured data from photovoltaic power stations in Yanggao, Shanxi and Hami, Xinjiang. The results demonstrate that compared to benchmark models such as Transformer and Informer, the proposed model achieves average reductions in mean absolute error (MAE) and root mean square error (RMSE) ranging from 19.20% to 59.54% and from 9.22% to 49.57%, respectively. These results indicate superior prediction accuracy and stability, validating the effectiveness and robustness of the proposed method under meteorological disturbance scenarios.