With the increasing penetration of photovoltaic (PV) power generation, accurate prediction of PV output is essential for grid stability and power accommodation. However, traditional time series forecasting methods exhibit limited capability in characterizing the complex dynamics of PV power, while existing graph neural network based spatiotemporal forecasting models still face certain challenges in capturing long-term temporal dependencies and global spatial correlations, particularly in multi-step forecasting where error accumulation easily occurs. To address these limitations, an improved spatiotemporal graph neural network model, termed MTGNN-BiMamba2-NL, is proposed. Within the multivariate time series forecasting framework based on graph neural networks (MTGNN), the temporal convolution module is replaced by a bidirectional state space model (BiMamba2) to enhance long-term temporal feature extraction. Meanwhile, a Non-local (NL) attention is incorporated after the graph convolution layers to strengthen the modeling of global spatial correlations among sites. Experiments were conducted on two real-world PV datasets. The results demonstrate that the proposed model effectively alleviates error accumulation in multi-step forecasting and achieves superior prediction accuracy and stability compared with existing baseline methods.