交通流量预测和到达时间预测（ETA）是交通预测中的重要任务，也是两个不可分割的组成部分。特别是对于企业而言，交通流量和到达时间的预测直接关系到生产活动的顺利进行，其合理准确的预测可以帮助企业规划最佳货运路线，提前安排货物的运输和交货时间，从而有效地降低成本、提高企业的工作效率、提高客户的满意度。本文的研究工作基于企业的真实需求展开。本文的主要工作与创新之处为：首先，基于历史信息和扩散衰减图卷积，提出了一种基于交通流分解的交通流量预测模型。考虑道路时间维度的关联，模型建立了随时间变化的邻接矩阵，并采用带有衰减系数的扩散卷积和时序建模分别对外部输入的流量和内部产生的流量进行处理。所提方法结合公开数据集 PEMS04 和 PEMS08 进行测试。针对 PEMS04数据集，预测第 15 分钟流量的 MAE 为 17.14，第 30 分钟流量的 MAE 为 18.01，第1 小时流量的 MAE 的为 19.33。针对 PEMS08 数据集，预测第 15 分钟流量的 MAE为 13.07，第 30 分钟流量的 MAE 为 13.96，第 1 小时流量的 MAE 的为 15.34。其次，为了提取到相距较远的路段之间的历史信息，提出了一种基于历史自注意力机制的交通流量预测模型。历史自注意力机制能够将不同路段的实时信息和历史信息进行交互。此外，模型基于路段流量相似性进行流量聚类，建立了路段的邻接掩码矩阵。结合公开数据集 PEMS04 和 PEMS08 对所提方法的可行性进行了验证。针对 PEMS04 数据集，预测第 15 分钟流量的 MAE 为 17.24，第 30 分钟流量的 MAE 为 17.92，第 1 小时流量的 MAE 的为 18.99。针对 PEMS08 数据集，预测第 15 分钟流量的 MAE 为 12.71，第 30 分钟流量的 MAE 为 13.80，第 1 小时流量的 MAE 的为 15.05。最后，针对企业生产环境中的到达时间预测问题，提出了一种基于交通流量与时序模型的到达时间预测方法。在交通流量预测模型的基础上，通过滑动窗口和自注意力分别提取相邻路段和非相邻路段的信息，并加入离散特征提取模块，最后加入交通流量预测的辅助损失，模型可以同时预测交通流量以及到达时间。该方法有较强的适配性，既可以用于具有历史交通流量的数据，也可以用于无历史流量的企业数据，针对滴滴比赛数据集所提方法将 MAPE 降低到了 12.23%。针对满帮抽取的数据集，所提方法也取得了 21.74% 的 MAPE，效果良好。

Traffic flow prediction and estimated-time-of-arrival (ETA) are important tasks in traffic prediction and are an inseparable integral. Especially for enterprises, the prediction of traffic flow and arrival time directly affects the normal running of production activities. Reasonable traffic flow prediction and accurate ETA can help enterprises plan the best freight route, arrange the transportation and delivery time of goods in advance, effectively reduce costs, improve work efficiency, and enhance customer satisfaction. The research work of this thesis is based on the needs of enterprises.The main contributions and innovations of this thesis are given as follows:First, this thesis proposes a traffic flow decomposition based prediction approach based on historical information and diffusion attenuation graph convolution. To consider the correlation of the road temporal dimension, an adaptive adjacency matrix that varies with time is established, and diffusion convolution and time series modeling with atten- uation coefficients are used to process the external input flow and the internal generated flow, respectively. The proposed approach is tested on the public datasets of PEMS04 and PEMS08. On the PEMS04 dataset, the MAE for predicting the 15th-minute flow is 17.14, the MAE for predicting the 30th-minute flow is 18.01, and the MAE for predict- ing the 1st-hour flow is 19.33, and on the PEMS08 dataset, the MAE for predicting the 15th-minute flow is 13.07, the MAE for predicting the 30th-minute flow is 13.96, and the MAE for predicting the 1st-hour flow is 15.34.Second, in order to extract historical information between road segments that are far apart, this thesis proposes a traffic flow prediction approach based on historical self- attention mechanism. The historical self-attention mechanism can interact real-time in- formation with historical information of different road segments. In addition, our new approach performs flow clustering based on flow similarity to set up an adjacency mask matrix for segments. The feasibility of the proposed approach is verified on the public datasets of PEMS04 and PEMS08. On the PEMS04 dataset, the MAE for predicting the 15th-minute flow is 17.24, the MAE for predicting the 30th-minute flow is 17.92, and the MAE for predicting the 1st-hour flow is 18.99, and on the PEMS08 dataset, the MAE for predicting the 15th-minute flow is 12.71, the MAE for predicting the 30th-minute flow is 13.80, and the MAE for predicting the 1st-hour flow is 15.05. Finally, for the ETA problem that exists in the enterprise production, this thesis pro- poses an ETA approach based on traffic flow prediction and time series modeling. Based on the traffic flow prediction model, our approach has extraction of information from ad- jacent and non-adjacent road segments using sliding windows and self-attention, and adds a discrete feature extraction module. Furthermore, the auxiliary loss of traffic flow pre- diction is added, and the model can simultaneously predict traffic flow and arrival time. The proposed approach has strong adaptability and can be available to either data with historical traffic flow or enterprise data without any historical flow. The experimental results demonstrate that the new approach decreases the MAPE to 12.23% on the Didi competition dataset and achieves a MAPE of 21.74% on the extracted dataset from Full Truck Alliance, which has a good performance.