<p>Accurate traffic flow prediction is essential for optimizing traffic management departments’ decision-making and improving road utilization efficiency. Previously, models focused on improving accuracy but demanded significant resources for large-scale node prediction, constraining their real-time applicability. To achieve an equilibrium between prediction accuracy and hardware resource consumption, we propose a Spatial-Temporal Fusion Residual Convolutional Network (STFRCN) for traffic flow prediction. STFRCN consists of a data embedding layer, a time fusion gated tanh unit (TF-GTU), a residual interactive convolution block (RICB), and an output layer. The data embedding layer consolidates diverse features along the time dimension, and then the TF-GTU further fuses and filters these features, thus effectively extracting spatial-temporal dependencies via the RICB. STFRCN is predominantly composed of convolution, thereby endowing the model with reduced computational costs. It sustains the model’s accuracy in predicting large-scale nodes while retaining high temporal efficiency. This lightweight convolution-based design bequeaths the model with diminished hardware resource consumption and affords it a broader application scope. Experiments are conducted on four real-world traffic flow datasets to validate the superiority of our model in large-scale node prediction. The effectiveness of each module is validated through ablation experiments. Ultimately, STFRCN achieves the highest prediction accuracy on the PEMS07 dataset, which has the largest number of nodes, while consuming less than 10% of the memory resources required by the most accurate baseline model.</p>