This paper proposes a novel spatio-temporal graph neural network (STPGNN) for traffic flow forecasting, addressing the challenge of accurately predicting traffic flow at pivotal nodes, which have complex spatio-temporal dependencies. The method introduces a pivotal node identification module to detect nodes with high aggregation and distribution capabilities, and a pivotal graph convolution module to capture spatio-temporal dependencies around these nodes. A parallel framework is also proposed to extract spatio-temporal features on both pivotal and non-pivotal nodes. The model is evaluated on seven real-world traffic datasets, demonstrating its effectiveness and efficiency compared to state-of-the-art baselines. The results show that the proposed method outperforms existing methods in terms of mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE) on most datasets. The model's performance is further validated through case studies, where it successfully captures traffic flow aggregation and distribution at pivotal nodes. The method is also shown to be computationally efficient, with faster training and inference times compared to other models. The key contributions include the identification of pivotal nodes, the design of a pivotal graph convolution module, and the development of a parallel framework for spatio-temporal feature extraction. The results demonstrate that the proposed method significantly improves the accuracy of traffic flow forecasting, particularly for pivotal nodes.This paper proposes a novel spatio-temporal graph neural network (STPGNN) for traffic flow forecasting, addressing the challenge of accurately predicting traffic flow at pivotal nodes, which have complex spatio-temporal dependencies. The method introduces a pivotal node identification module to detect nodes with high aggregation and distribution capabilities, and a pivotal graph convolution module to capture spatio-temporal dependencies around these nodes. A parallel framework is also proposed to extract spatio-temporal features on both pivotal and non-pivotal nodes. The model is evaluated on seven real-world traffic datasets, demonstrating its effectiveness and efficiency compared to state-of-the-art baselines. The results show that the proposed method outperforms existing methods in terms of mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE) on most datasets. The model's performance is further validated through case studies, where it successfully captures traffic flow aggregation and distribution at pivotal nodes. The method is also shown to be computationally efficient, with faster training and inference times compared to other models. The key contributions include the identification of pivotal nodes, the design of a pivotal graph convolution module, and the development of a parallel framework for spatio-temporal feature extraction. The results demonstrate that the proposed method significantly improves the accuracy of traffic flow forecasting, particularly for pivotal nodes.