Abstract: With the increasing penetration of renewable energy in power grids, the strong randomness and fluctuation of photovoltaic output have brought great challenges to the safe and stable operation of new power systems. Considering that traditional deterministic forecasting methods are difficult to effectively quantify prediction uncertainty, and that conventional Transformer models based on time-step attention tend to lose the physical coupling relationships among variables when processing multivariate long sequences, this paper proposes an ultra-short-term probabilistic forecasting framework that integrates Variational Mode Decomposition (VMD), Patch-inverted Transformer (Patch-iTransformer), and Flow Matching. First, Pearson correlation coefficient is used to select highly correlated meteorological features, including global irradiance, temperature, and humidity, and cyclic time encoding is introduced to enhance temporal information representation. Second, VMD is applied to decompose the non-stationary power sequence into multiple modal components, thereby reducing fluctuation interference and improving data stationarity. Third, a two-stage decoupled Flow Matching generative model based on “point forecasting-residual fitting” is constructed. On the basis of using Patch-iTransformer to overcome the limitation of time-step-based modeling and perform deterministic baseline point forecasting, Flow Matching is employed to fit the prediction residual distribution, and an ordinary differential equation (ODE) solver is used to realize the evolution from Gaussian noise to the real residual space. In addition, strict gradient blocking and residual scaling alignment techniques are introduced to effectively solve the signal-to-noise imbalance caused by extremely small power residuals. Subsequently, an Optuna Bayesian optimization framework based on the Tree-structured Parzen Estimator (TPE), combined with the Asynchronous Successive Halving Algorithm (ASHA) pruning mechanism, is adopted to jointly optimize the VMD decomposition parameters and the hyperparameters of Patch-iTransformer, thereby enhancing both decomposition and forecasting performance. Finally, probabilistic prediction intervals of photovoltaic output are generated based on multi-round sampling and interval calibration. Experimental results on measured photovoltaic data from Yunnan show that the proposed model performs well in both point forecasting accuracy and interval prediction sharpness, with MAE reduced to 0.1617, RMSE of 0.3946, MAPE of only 5.3744%, and R2 reaching 0.9989. Under the high reliability level of 96.4583% PICP, MPIW is narrowed to 1.2328 and PINAW is 3.4973%, which can provide reliable decision support for refined photovoltaic grid-connected dispatching and risk management.