Urban rail transit passenger flow prediction using large language model under multi-source spatiotemporal data fusion

With the continuous development of urban rail transit systems, passenger flow prediction has become a crucial component for optimizing operational scheduling. This paper proposes a metro passenger flow prediction method based on Large Language Model (LLM), incorporating hyperparameter tuning, pre-training, and multi-source spatiotemporal data fusion to enhance prediction accuracy. Firstly, the generative pre-trained transformer 2 (GPT-2) model is employed for passenger flow prediction, with fine-tuning strategies applied to adjust the pre-trained model to meet the specific requirements of the prediction task. During the fine-tuning process, some pre-trained parameters are frozen, and the Cosine Annealing Learning Rate strategy is used to gradually adjust the learning rate, effectively preventing overfitting and achieving efficient optimization. Secondly, the self-attention mechanism is utilized to fuse multi-source data, including air quality, weather conditions, and spatiotemporal passenger flow data, enhancing the model’s ability to capture complex fluctuations in passenger flow. The experimental results indicate that the proposed prediction model outperforms other models by achieving the lowest prediction error and demonstrating better capability in capturing complex passenger flow fluctuations. Ablation studies further validate the importance of multi-source data fusion in metro passenger flow prediction, as well as the effectiveness of fine-tuning in enhancing model performance. In addition, short-term prediction experiments investigate the model’s ability to handle immediate, near-future passenger flow forecasts. Furthermore, few-shot learning experiments analyze the impact of the training data sample ratio on prediction error and training efficiency. This paper offers a new perspective on applying LLM to metro passenger flow prediction.