Traffic flow prediction is an important component of the Intelligent Transportation System (ITS). Currently, traffic flow prediction still has the following limitations: insufficient attention to the spatial correlation in the traffic flow prediction process, mainly focusing on traffic flow prediction at individual scale patterns, and rarely considering the interference of external factors such as weather on traffic flow prediction. To address these issues, this study proposes a traffic flow prediction method that combines multi-scale spatiotemporal graph convolution and large language models. This method first divides the traffic data into trend sequences, designs a spatiotemporal embedding layer including attention modules and convolutional modules in the spatial and temporal dimensions to extract the spatial and temporal features of the traffic data. The extracted spatiotemporal features are input into a large language model (LLM) with a dynamically frozen partial attention mechanism, and the parameter-efficient fine-tuning (PEFT) technique is applied to make the model better adapt to traffic flow prediction tasks with dynamic spatiotemporal correlations. Additionally, considering the non-stationarity of traffic data, a multi-layer perceptron module is used externally to enhance the prediction of traffic flow. Finally, the effectiveness of the model is verified on two real datasets, PEMS04 and PEMS08. The results show that the average MAE of the prediction error accuracy index of the proposed model is 1.05% and 1.87% higher than that of the suboptimal model OFA.