A Case Study of the Influence of Multifarious Factors on Traffic Flow Forecasting

摘要

Accurate traffic flow forecasting serves as the core foundation for intelligent transport systems to achieve efficient traffic management and optimized resource allocation, holding significant importance for alleviating congestion in modern cities. Influenced by the complex interplay of diverse heterogeneous factors such as meteorological conditions, temporal cycles, and unforeseen events, urban traffic flow data exhibits pronounced nonlinearity and random fluctuations, rendering high-precision forecasting exceptionally challenging. While mainstream deep learning models have advanced prediction accuracy, they struggle to quantify the specific contributions of different factors and often overlook significant multicollinearity issues among multi-source data. Addressing these challenges, this paper introduces several improvements: Firstly, it constructs a fused dataset incorporating multidimensional external factors such as meteorological conditions and events; secondly, it proposes an explainable prediction framework based on Random Forest with raw feature analysis by Principal Component Analysis (PCA); Thirdly, the SHAP game-theoretic method is introduced to achieve transparent attribution of prediction results. This paper first employs PCA to extract principal components from high-dimensional multi-source factors, effectively eliminating multicollinearity in the data. Subsequently, a robust random forest regression model is constructed for prediction. Based on four independent datasets from Beijing's TaxiBJ service, the proposed framework undergoes comprehensive performance validation and analysis. Results demonstrate that the model maintains a coefficient of determination consistently above 0.85 across all annual datasets, exhibiting outstanding predictive accuracy and robustness across temporal cycles. SHAP analysis further reveals a stable decision mechanism characterized by the first principal component driving periodicity, with subsequent components providing dynamic fine-tuning, successfully achieving traffic flow prediction that combines high precision with strong interpretability.