基于大语言模型的缺失数据交通流预测

Large Language Model-Based Missing Traffic Flow Prediction

摘要: 交通流预测在智能交通系统中具有重要意义。近年来，交通预测模型多基于复杂的深度学习架构，然而，这些模型往往缺乏对输入到输出过程及其结果的直观解释。由于交通数据本身的复杂性、模型不透明性以及数据收集过程中常见的缺失问题，缺失数据的处理和预测结果的可解释性仍然面临重大挑战。针对这一问题，我们提出了一种基于大语言模型(LLM-MTFP)的可解释性缺失数据交通预测模型。具体而言，该模型通过将多模式交通数据转换为自然语言描述，利用大语言模型捕捉复杂的时空特征和外部因素，并基于语言指令进行微调。我们在加利福尼亚州多模态数据集上进行了实验验证。结果表明，本文提出的方法在预测准确性方面优于基线模型，并且能够提供可靠的解释，验证了大语言模型在交通流预测中的应用潜力。

Abstract: Traffic flow prediction holds significant importance in intelligent transportation systems. In recent years, traffic prediction models have largely been based on complex deep learning architectures. However, these models often lack intuitive explanations of the input-output processes and their results. Due to the inherent complexity of traffic data, model opacity, and common data missing issues during the data collection process, handling missing data and ensuring the interpretability of prediction results remain major challenges. To address this problem, we propose an explainable missing data traffic prediction model based on large language models (LLM-MTFP). Specifically, the model converts multimodal traffic data into natural language descriptions, leverages large language models to capture complex spatiotemporal features and external factors, and fine-tunes the model based on linguistic instructions. We conducted experimental validations on the California multimodal dataset. The results demonstrate that the proposed method outperforms baseline models in prediction accuracy and can provide reliable explanations, verifying the application potential of large language models in traffic flow prediction.