为解决现有高速公路交通量预测方法在捕捉动态时空依赖关系方面的不足，提出了一种融合信息几何方法与注意力机制的新型高速路网交通量预测模型。首先，利用信息几何方法量化ETC门架之间的动态数据分布差异。然后，利用注意力机制来捕获交通的动态空间依赖关系。最后，结合一种堆叠的双向递归层结构，提出了一种长时间跨度的并行子模型算法，即基于信息几何方法（Information Geometry）和注意力机制（Attention Mechanism）优化的图卷积神经网络（GCN）结合堆叠双向单向长短期记忆神经网络（sbuLSTM）的组合模型—IGAGCN-sbuLSTM。采用该模型对100多条路段、3 000多处门架近7亿条高速公路ETC门架系统数据进行分析，结果显示：与LSTM、GCN、GCN-LSTM、ASTGCN等现有4种模型相比，在10 min时间尺度下，IGAGCN-sbuLSTM组合模型的平均绝对误差（MAE）分别降低了2.39, 3.72, 1.02, 1.46，均方根误差（RMSE）分别降低了3.25, 4.32, 2.05, 5.65，平均绝对百分比误差（MAPE）分别降低了5.49%, 12.54%, 1.56%, 0.5%。研究表明，IGAGCN-sbuLSTM模型在预测精度和不同时间间隔的预测性能上均优于现有的单一捕获特性模型及其他常用的组合模型，可广泛应用于高速公路收费、车速等数据的预测分析。

In order to address the shortcomings of existing expressway traffic volume prediction methods in capturing dynamic spatio-temporal dependencies, the paper proposed a novel traffic volume prediction model for expressway networks that incorporates information geometry methods and attention mechanisms. Initially, an information geometry method was used to quantify the difference in dynamic data distribution between ETC gantries. Subsequently, an attention-based mechanism was devised to encapsulate the dynamic spatial dependencies inherent in traffic patterns. In conclusion, by amalgamating a configuration of stacked bidirectional recurrent layers, a long-time span parallel sub-model algorithm named IGAGCN-sbuLSTM was proposed. This acronym denoted a synergistic model that combined GCN(Graph Convolutional Neural Network), which was optimized using Information Geometry Approach and Attention Mechanism, with sbuLSTM(Stacked Bidirectional Unidirectional Long Short-Term Memory Neural Network). Leveraging over 100 road segments, surpassing 3 000 gantries from nearly 700 billion entries in the expressway ETC gantry system database, experimental outcomes demonstrated that the IGAGCN-sbuLSTM algorithm, when benchmarked against existing models such as LSTM, GCN, GCN-LSTM, ASTGCN, registered a reduction in MAE (Mean Absolute Error) by 2.39, 3.72, 1.02, and 1.46 respectively on a 10-minute time scale. The RMSE (Root Mean Square Error) diminished by 3.25, 4.32, 2.05, and 5.65, while the MAPE (Mean Absolute Percentage Error) decreased by 5.49%, 12.54%, 1.56%, and 0.5% respectively. These results affirm that the IGAGCN-sbuLSTM model surpasses its predecessors with single capture characteristic and other prevalent combined models in terms of predictive accuracy and capability to handle expanding time intervals, rendering it highly applicable for predictions analysis in areas such as expressway tolling and vehicular speeds.