The accuracy of short-term photovoltaic (PV) power forecasting is vital for the secure operation of power grids and the efficient integration of renewable energy. To address the complex frequency components in PV power series and the challenge of jointly extracting spatiotemporal dependencies, this study proposes an ADM-DGCN-BiLSTM model that combines an adaptive decomposition module (ADM), a dynamic graph convolutional network (DGCN), and a bidirectional long short-term memory network (BiLSTM). First, the ADM employs a fast Fourier transform (FFT) to dynamically analyze the spectral characteristics of PV power sequences and adaptively decomposes the series into three key components: periodic, trend, and residual components. In addition, the impact of dominant frequency estimation accuracy on decomposition quality and overall forecasting performance is further investigated. Second, the spatial stream employs the DGCN to learn time-varying spatial correlations among multidimensional meteorological variables, while the temporal stream adopts the BiLSTM to capture power time-series variation characteristics at a single site. Finally, the outputs of the two streams are fused to generate the final power forecast. Experimental results demonstrate that the ADM-DGCN-BiLSTM model markedly enhances short-term PV power forecasting precision and significantly outperforms conventional methods, offering a new technical pathway for efficient PV power prediction.