EVRP-TW-D

EVRP-TW-D is a real-geography, semi-synthetic dataset and benchmark suite for the Electric Vehicle Routing Problem with Time Windows (EVRPTW). This release, Geo-AC-v1.1 / NA-US-20, provides 20 North American service territories constructed from public geospatial data and a fixed evaluation split of 1,600 operating-day EVRPTW instances. The dataset contains two main components. First, the source_data directory provides normalized geospatial inputs for each service territory, including road nodes and edges, community-level customer seed locations, generated latent customer locations, public charging station candidates, and depot candidates. These layers are derived from public data sources, including Census TIGER/Line boundaries, ACS occupied housing-unit counts, OSM/OSMnx and TIGER road networks, NREL/AFDC public EV charging stations, and open warehouse/logistics/industrial depot candidate sources. Second, the eval_standard_20 directory provides a fixed evaluation set with 20 instances per territory and per customer scale. The evaluation scales are Cus5, Cus15, Cus50, and Cus100, giving 20 × 4 × 20 = 1,600 EVRPTW instances. Each instance represents one operating day within a service territory, where active customers and charging stations are sampled from the underlying real-geography service territory. This release is intended for benchmarking EVRPTW solvers under consistent spatial, operational, and data-generation semantics. The benchmark should be described as real-geography, semi-synthetic: customer locations, road networks, charging station candidates, and depot candidates are geography-driven, while daily demand, service time, time windows, and instance activation are generated by the EVRPTW benchmark pipeline. Future releases may add larger training splits, additional North American territories, exact-solver reference snapshots, certified optimal solutions where available, and best-known solution records for larger instances.