Accurate photovoltaic (PV) power forecasting is essential for ensuring grid stability and economical power system operation. However, the influence of weather stochasticity has resulted in progressively intricate patterns of PV power generation, presenting significant challenges to existing forecasting methods. This study proposes a parallel dendritic learning framework that integrates graph convolution, long short-term memory, and the dendritic neural model to address the above issue. Firstly, variational mode decomposition is employed to decompose the PV power series, capturing underlying dynamics and detailed time-frequency characteristics. Then, multivariate PV time series are represented as a graph, where feature relationships are modeled as edges, and graph convolution is applied to extract informative representations. Meanwhile, long short-term memory is employed to capture temporal dependencies. Finally, inspired by the nonlinear excitatory behavior of biological neurons, the dendritic neuron model is incorporated, leading to a parallel dendritic learning framework for deep feature integration. Extensive experiments on two public datasets demonstrate that the proposed model consistently outperforms six baseline methods across four evaluation metrics. Ablation studies further indicate that the integration of dendritic learning effectively enhances model performance for accurate PV power forecasting.