加州第七大区高速交通流量数据集（PEMS07）以加州交通局第七片区（洛杉矶 + 文图拉县）高速公路路网为研究区域，基于 Caltrans PeMS 系统采集的交通流量、车速、占有率数据，统一筛选、清洗、标准化后形成大规模路网交通预测基准数据集。数据集包含 883 个环路检测器，时间范围为 2017 年 5 月 1 日至 2017 年 8 月 31 日，时间分辨率为 5 分钟，共 28224 个时间步长，完整覆盖 4 个月全天交通时序。原始采集包含流量、车速、车道占有率 3 项交通特征，实验主流采用单流量特征建模；原始缺失值经时序线性插值填充，成品数据集无缺失。按照 60%、20%、20% 的时序比例划分为训练集、验证集和测试集。基于道路路网距离采用阈值化高斯核构建传感器间加权邻接矩阵。该数据集是当前节点规模最大的公开交通时空基准数据集，广泛用于大规模路网时空图神经网络、长时序交通流量预测、路网异质性建模等模型性能验证。
", "ds_source": "交通状态数据：源自加州交通局 Caltrans PeMS 官方监测系统（https://pems.dot.ca.gov/），由 Yu 等人（2018）在 STGCN 开源项目中从 District 7 原始海量检测器中筛选有效站点并完成全流程预处理后公开。数据集开源代码与下载资源地址：https://github.com/Davidham3/STSGCN
", "ds_process_way": "一、数据采集
\n从 Caltrans PeMS 业务运营系统获取 District 7 片区原始环路检测器数据。PeMS 系统以 30 秒高频实时采集加州全域高速公路车流量、平均车速、车道占有率三类交通参数，由加州交通局统一运维、存储原始监测数据。
\n二、数据筛选与预处理
\nYu 等人（2018）筛选 District 7 片区内 883 台长期稳定运行、故障率极低的环路检测器，覆盖洛杉矶县、文图拉县全域主干高速路网；剔除长期离线、数据异常的失效传感器。原始 30 秒高频数据统一聚合为 5 分钟时间窗口均值时序；原始数据存在约 0.45% 缺失值，采用时序线性插值结合邻近车道空间信息完成填充，标准化发布版本无空缺值。完整保留流量、车速、占有率三类原始观测特征供研究者按需选用。
\n三、数据集划分
\n全部时序数据严格按照时间先后顺序切分，以 60%、20%、20% 比例划分为训练集、验证集、测试集，全程不打乱时序，贴合真实交通预测离线训练场景。
\n四、数据标准化
\n采用 Z-score 标准化方法，仅在训练集上计算特征均值与标准差，再统一映射至验证集、测试集，严格规避跨时序数据泄露问题。
\n五、传感器图构建
\n计算任意两台检测器间实际道路路网距离（非欧式直线距离），采用阈值化高斯核生成加权邻接矩阵：
\nW_ij = exp(-dist(v_i, v_j)² / σ²) 若 dist(v_i, v_j) ≤ κ，否则为0
\n其中σ为距离标准差，κ为阈值。生成 883×883 对称邻接距离矩阵与时序数据配套发布，用于建模路网空间交通传播依赖关系。
", "ds_quality": "本数据集的底层监测数据源自Caltrans PeMS官方标准化交通监测平台，经二次严格筛选仅保留长期稳定、故障记录少的883台检测器，从源头上剔除了低质量采集站点以保障原始数据源的可靠性。针对原始少量的缺失值，通过时空联合插值完成了完整修复，使集成数据成品的缺失率降至0%，无需额外补全操作即可直接投入大规模路网模型训练。在数据标准化流程中，Z-score归一化参数严格仅由训练集计算得出，验证集与测试集仅进行同步变换，从而有效杜绝了测试阶段的未来信息泄露，确保了实验对比结果的公平性。同时，数据集完成了图结构完整性校验，配套的距离矩阵为与传感器节点数完全匹配的883×883标准方阵，节点空间关联关系完整且无重复或行列缺失。此外，该数据在时序连续性上具备强力保障，完整覆盖了长达4个月全天24小时的时序信息，全面囊括工作日、周末及节假日等全场景交通周期规律，能够充分且严谨地验证模型对周期性与突变车流的拟合与泛化能力。
", "ds_acq_start_time": null, "ds_acq_end_time": null, "ds_acq_place": "美国加利福尼亚州第七交通大区（洛杉矶县 + 文图拉县）高速公路路网", "ds_acq_lon_east": -114.11, "ds_acq_lat_south": 33.7, "ds_acq_lon_west": -116.68333333333334, "ds_acq_lat_north": 34.85, "ds_acq_alt_low": null, "ds_acq_alt_high": null, "ds_share_type": "open-access", "ds_total_size": 1196229618, "ds_files_count": 0, "ds_format": "NPZ 、CSV ", "ds_space_res": "", "ds_time_res": "", "ds_coordinate": "无", "ds_projection": "", "ds_thumbnail": "296b5fac-21d3-4619-b898-932b864b8546.jpeg", "ds_thumb_from": 0, "ds_ref_way": "", "paper_ref_way": "", "ds_ref_instruction": "本数据集为 STGCN 论文（Yu et al., IJCAI 2018）公开的 PEMS07 大规模路网交通基准数据集。底层原始监测数据来源于加州交通局 Caltrans PeMS 系统（https://pems.dot.ca.gov/）。标准化清洗后的数据集可通过 STSGCN 开源仓库（https://github.com/Davidham3/STSGCN）或官方网盘镜像开源获取。开展学术研究、模型实验时请引用 Yu 等人 2018 年 IJCAI 论文。", "ds_from_station": "", "organization_id": "9ecaaa78-39e9-411e-9f24-274e12aa643f", "ds_serv_man": "张新民", "ds_serv_phone": "15869040017", "ds_serv_mail": "xinminzhang@zju.edu.cn", "doi_value": "", "subject_codes": [ "410" ], "quality_level": 0, "publish_time": "2026-07-09 10:58:54", "last_updated": "2026-07-09 10:58:54", "protected": false, "protected_to": "2028-06-30 00:00:00", "lang": "zh", "cstr": "", "i18n": { "en": { "title": "California Region 7 High-Speed Traffic Data Set (PEMS07)", "ds_format": "NPZ , CSV", "ds_source": "Traffic monitoring data is collected from the official Caltrans Performance Measurement System (https://pems.dot.ca.gov/). The original District 7 traffic data was explored in STGCN (IJCAI 2018). The complete standardized PEMS07 dataset is screened, preprocessed and open-sourced via the STSGCN open-source project. Official dataset repository: https://github.com/Davidham3/STSGCN", "ds_quality": "1. Source Reliability\r\nRaw data comes from officially maintained Caltrans PeMS platform. Valid detectors with long-term stable operation are selected to guarantee data validity.\r\n2. Missing Value Processing\r\nAll missing records are filled scientifically, and the released dataset has zero missing value for direct model training.\r\n3. Standardization Standard\r\nNormalization parameters are isolated within the training set to avoid future information leakage.\r\n4. Graph Structure Integrity\r\nThe released adjacency matrix is an 883×883 standard square matrix, matching the total number of detector nodes completely.\r\n5. Time-series Completeness\r\nThe 4-month continuous dataset covers working days, weekends and seasonal traffic patterns, supporting comprehensive model validation.", "ds_ref_way": "", "ds_abstract": "California's Seventh Region High-Speed Traffic Flow Data Set (PEMS07) takes the California Transportation Authority's Seventh Region (Los Angeles + Ventura County) highway network as the research area. It is based on traffic flow, speed, and occupancy data collected by the Caltrans PeMS system. Data, unified screening, cleaning, and standardization to form a large-scale road network traffic prediction benchmark data set. The data set contains 883 loop detectors, with a time range from May 1, 2017 to August 31, 2017, a time resolution of 5 minutes, a total of 28224 time steps, completely covering the four-month all-day traffic sequence. The original collection includes three traffic characteristics: flow rate, speed, and lane occupancy. The mainstream of the experiment uses single flow feature modeling; the original missing values are filled in through time series linear interpolation, and the finished data set has no missing. It is divided into training set, verification set and test set according to the time series ratios of 60%, 20%, and 20%. Based on road network distances, thresholded Gaussian kernels are used to build a weighted adjacency matrix between sensors. This dataset is the largest open traffic spatio-temporal benchmark dataset with current node size. It is widely used for model performance verification such as large-scale road network spatio-temporal map neural networks, long-time series traffic flow prediction, and road network heterogeneity modeling.
", "ds_time_res": "", "ds_acq_place": "Highway Network of California Transportation District 7, USA (Los Angeles County & Ventura County)", "ds_space_res": "", "ds_projection": "", "ds_process_way": "1. Data Collection\r\nRaw traffic data is acquired from Caltrans PeMS official platform. The system collects highway traffic flow, average speed and lane occupancy every 30 seconds, operated and maintained by California Department of Transportation.\r\n2. Sensor Filtering and Preprocessing\r\nThe STSGCN team reserved 883 stable loop detectors covering major freeways in Los Angeles County and Ventura County, and eliminated malfunctioned and abnormal sensors. Raw 30s sampling data is aggregated into 5-minute average traffic records. Original 0.45% missing values are complemented by temporal interpolation and spatial correlation fusion without data leakage.\r\n3. Dataset Partition\r\nAll time-series samples are divided in chronological order without shuffling, following the split ratio: 60% training set, 20% validation set, 20% test set.\r\n4. Data Normalization\r\nZ-score normalization is adopted. Mean value and standard deviation are calculated only on the training set, and applied to validation and test sets uniformly.\r\n5. Graph Topology Construction\r\nThe weighted adjacency matrix is constructed by road network distance rather than Euclidean distance via thresholded Gaussian kernel: W_ij = exp(-dist(v_i, v_j)^2 / sigma^2) if dist(v_i, v_j) <= kappa, else 0. Where σ denotes distance standard deviation, κ denotes distance threshold. The 883×883 adjacency matrix is released together with time-series data for modeling traffic spatial diffusion patterns.", "ds_ref_instruction": "The standardized PEMS07 dataset is officially released by the STSGCN project (Guo et al., AAAI 2020). The original District 7 highway traffic data was firstly applied for traffic forecasting research by STGCN (Yu et al., IJCAI 2018). Raw traffic data is sourced from Caltrans PeMS official system (https://pems.dot.ca.gov/). Processed datasets can be downloaded from the STSGCN GitHub repository. Researchers shall cite both two papers for academic use." } }, "submit_center_id": "ncdc", "data_level": 0, "recommendation_value": 0, "license_type": "https://creativecommons.org/licenses/by/4.0/", "doi_reg_from": "reg_local", "cstr_reg_from": "reg_local", "doi_not_reg_reason": null, "cstr_not_reg_reason": null, "is_paper_in_submitting": false, "belong_to_nieer": false, "ds_topic_tags": [ "交通流量预测", "短时交通预测", "时空图预测", "大规模路网建模", "图神经网络", "交通基准数据集" ], "ds_subject_tags": [ "工程与技术科学基础学科" ], "ds_class_tags": [], "ds_locus_tags": [], "ds_time_tags": [], "ds_contributors": [ { "true_name": "张新民", "email": "xinminzhang@zju.edu.cn", "work_for": "浙江大学", "country": "中国" } ], "ds_meta_authors": [ { "true_name": "Bing Yu", "email": "bingyu@uchicago.edu", "work_for": "University of Chicago", "country": "中国" }, { "true_name": "Haoteng Yin", "email": "hyin@uchicago.edu", "work_for": "University of Chicago", "country": "中国" }, { "true_name": "Zhanxing Zhu", "email": "zhanxing.zhu@gmail.com", "work_for": "University of Chicago", "country": "中国" } ], "ds_managers": [ { "true_name": "张新民", "email": "xinminzhang@zju.edu.cn", "work_for": "浙江大学", "country": "中国" } ], "category": "其他" }