Dataset Paper¶

A second scientific article introduces OptiWindNet RouteSets — a database of cable-routing solutions produced with OptiWindNet and released as an open benchmark for the wind farm cable-routing problem. The article is under review at Wind Energy Science; the preprint is open access:

Mauricio Souza de Alencar, Tuhfe Göçmen, Nicolaos A. Cutululis, OptiWindNet RouteSets: a solver-diverse benchmark dataset for the offshore wind-farm cable routing problem, Wind Energy Science Discussions [preprint], 2026, https://doi.org/10.5194/wes-2026-124, in review.

@Article{wes-2026-124,
  author = {Souza de Alencar, M. and G\"o\c{c}men, T. and Cutululis, N. A.},
  title = {OptiWindNet RouteSets: a solver-diverse benchmark dataset for the offshore wind-farm cable routing problem},
  journal = {Wind Energy Science Discussions},
  volume = {2026},
  year = {2026},
  pages = {1--13},
  url = {https://wes.copernicus.org/preprints/wes-2026-124/},
  doi = {10.5194/wes-2026-124},
}

The framework that produced the solutions is the subject of the Framework Paper.

Getting the database¶

Release 26.05-v4 is a single SQLite file, optiwindnet-routesets-r26.05-v4.sqlite, deposited on Zenodo under CC-BY-4.0:

Souza de Alencar, M. (2026). OptiWindNet RouteSets (26.05-v4) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.20053479

The download is optiwindnet-routesets-r26.05-v4.sqlite.xz (~28 MB compressed, ~88 MB decompressed). Nothing else is needed to read it — the optiwindnet package you already have provides the models.

What is in it¶

49 204 route sets over 12 814 distinct wind farm geometries, covering 13 954 distinct problem instances, where an instance is a (location, cable capacity) pair. Instances span 50–200 turbines and cable capacities of 2–12 turbines, in radial and branched topology. A solution with a proven optimality gap below 1 % is available for 63 % of the instances, and below 2 % for 79 %.

Several route sets may solve the same instance — that is deliberate, since the informative comparisons are paired ones in which a single method attribute differs.

The four solvers behind the route sets:

`creator`	backend	topology	route sets	%
`MILP.pyomo.cplex`	CPLEX 22.1.1, via pyomo	radial, branched	22 623	46
`baselines.hgs`	HGS-CVRP, via hybgensea	radial	13 857	28
`baselines.lkh`	LKH-3.0.14 binary	radial	12 166	25
`MILP.pyomo.gurobi`	Gurobi 12.0, via pyomo	radial, branched	558	1

The locations have a core of 98 built offshore wind farms — most of the existing ones in the covered turbine-count range — plus five hypothetical sites proposed by Cazzaro and Pisinger (2022), Yi et al. (2019) and Taylor et al. (2023). The remaining geometries were derived from those by the data augmentation in optiwindnet.augmentation: borders and substations are kept, turbines are re-scattered over the available area. Augmented names carry their donor, as !donor_name!digest, with suffixes recording transformations (.1_OSS and .1st_OSS reduce to a single substation, .solid removes obstacles).

Site families matter for machine learning

Variants sharing a donor are geometrically related, so a random split across geometries leaks between train and test. Split by site family instead — it separates within-family sensitivity to turbine placement from between-family generalization.

The four tables¶

nodeset: One row per distinct wind farm geometry, keyed by the SHA digest of that geometry, so identical geometries loaded from different files collapse into one row. All coordinates live in the VertexC binary blob as planar easting/northing (the UTM zone is not stored, so the geodesic position cannot be recovered from the file alone); the other fields say which coordinate plays which role — turbine count T, substation count R, and the B coordinates that only define borders and exclusion-zone polygons.
routeset: One row per solution. Solutions refer to coordinates by index into their nodeset row and carry no coordinates of their own. Each row references one nodeset, one method and one machine.
method: One row per distinct (solver_name, options, funhash) triple, describing the solver call.
machine: The computer that produced the solution.

Do not query methods by primary key

funhash hashes the solver function’s bytecode, so it drifts across OptiWindNet versions: the same solver call may be recorded under several method rows. Match on solver_name and the options (key, value) pairs instead of on digest.

Objective, length and gap¶

The optimizer minimizes the sum of selected link lengths in the available-links graph, which abstracts away some geometric detail. The best modeled length it found is misc['objective'], and, for MILP runs, the best dual bound is misc['bound'] — together they give the proven relative gap misc['relgap'] for that length objective. Length and cost are often strongly correlated, so a small length gap is a useful indication of cost quality, but it does not bound the gap to the minimum possible cost; see Objective and reported cost.

The route set itself is the embedded network: the cable routes in the plane that implement that connectivity, after feeder detours are added. Its total length is the length field, which may exceed the objective; detextra is the relative increase between the two, so length = (1 + detextra) * objective. The pre-detour graph is recoverable from the route set and is therefore not stored.

Reading the database¶

The tables are peewee models, so the database is queried through RouteSet and its siblings, and a row is turned back into a routed graph by G_from_routeset():

from optiwindnet.db import RouteSet, G_from_routeset, database_connection
from optiwindnet.plotting import gplot

with database_connection('optiwindnet-routesets-r26.05-v4.sqlite'):
    routeset = (
        RouteSet.select()
        .where((RouteSet.handle == 'dudgeon') & (RouteSet.capacity == 5))
        .order_by(RouteSet.length)
        .first()
    )
    print(routeset.creator, routeset.T, routeset.length, routeset.misc['relgap'])
    G = G_from_routeset(routeset)

gplot(G)

A location alone, without a solution, comes from L_from_nodeset(). Both functions rebuild the graph metadata the rest of the library expects — see Graph representations — which a raw SQL query would leave behind.

What the article reports¶

Four analyses over the release, each on paired instances:

Meta-heuristic solution quality. Against the shortest MILP-radial route set of the same instance, and with 120 s of single-core time, HGS-CVRP has a median length increase of 0.0 % (90th percentile 0.225 %) and LKH-3 of 0.555 % (90th percentile 2.187 %).
Problem difficulty. Read through the MIP gap and runtime, difficulty rises monotonically with turbine count, while its dependence on capacity peaks at intermediate values and depends on the count. Median runtime goes from seconds to the 6 h cap between 70 and 110 turbines. Branched instances are somewhat harder than radial ones.
Branched vs. radial topology. Over 8 696 matched instances, branching shortens the network by a median 0.34 % (mean 0.64 %), and is shorter in 6 741 of them. The advantage grows with cable capacity and shrinks with turbine count, with enough spread to warrant deciding case by case.
Detour impact. The length added by feeder detours over the solver-optimized objective has a median of 0.171 % for radial and 0.154 % for branched, largest at capacities 3–6.

Reproducing the analyses¶

The four notebooks that produce the article’s figures, together with the SQL access layer they use, are archived separately:

Souza de Alencar, M. (2026). Code and Computational Artifacts for the PhD Thesis “Wind Farm Collection System Optimization for Integrated Design”. Zenodo. https://doi.org/10.5281/zenodo.20140812

Download optiwindnet_routesets.zip from that record. It holds case1_metaheuristic_solution_quality.ipynb, case2_problem_difficulty.ipynb, case3_branched_vs_radial_topology.ipynb and case4_detour_impact.ipynb — one per analysis above, independent of each other — plus db_access.py, which fetches and verifies the database on first run and pushes the projection, filtering and best-per-instance selection down into SQL. The archive also ships the figures, so it can be read without running anything.

How to cite¶

Cite the article for the analyses and the Zenodo record for the data:

Souza de Alencar, M., Göçmen, T., and Cutululis, N. A.: OptiWindNet RouteSets: a solver-diverse benchmark dataset for the offshore wind-farm cable routing problem, Wind Energ. Sci. Discuss. [preprint], https://doi.org/10.5194/wes-2026-124, in review, 2026.

Souza de Alencar, M. (2026). OptiWindNet RouteSets (26.05-v4) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.20053479

The software itself is cited as described in How to Cite.