| 28 Jul 2026
OptiWindNet RouteSets: a solver-diverse benchmark dataset for the offshore wind-farm cable routing problem
Abstract. Offshore wind-farm collection-system design is a cost-relevant combinatorial optimization problem whose difficulty grows exponentially with turbine count. This paper introduces OptiWindNet RouteSets, a database of cable-routing solutions to 13954 distinct problem instances based on built, proposed, and procedurally-generated farm layouts. The database is intended as an open benchmark for routing algorithms, a source of hard instances, a strong baseline for reinforcement-learning solvers, and a training corpus for supervised-learning models for the Wind Farm Cable-Routing Problem.
The solutions were produced using the optiwindnet Python package with three solvers: exact mathematical optimization and two meta-heuristics (inexact) – hybrid genetic search (HGS) and Lin–Kernighan–Helsgaun (LKH). Two network topologies are covered: radial (path-based) or branched (tree-based). Each solution contains the feasible network and its metadata (solver used, cable capacity, method configuration, route length, detour overhead, solver runtime, and, for exact runs, a proven optimality gap). The problem instances vary in: number of wind turbines (∈ [50, 200]), maximum cable capacity (∈ [2, 12]), and location geometry. For 63% of the problem instances, a solution with <1% gap is available (and 79% with <2% gap).
We analyze: (a) the total length of meta-heuristic solutions compared to their exact counterparts – HGS median increase is 0.0%, while LKH is 0.6%; (b) the instance difficulty as a function of capacity and turbine count – those two variables interact, the difficulty increases monotonically with count, but exhibits a count-dependent peak across capacities; (c) the length reduction of branched topology compared to radial – 0.34% median; (d) the detour-caused increase in length over the solver-optimized objective – 0.16% median.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Wind Energy Science.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
General comments
This manuscript presents OptiWindNet RouteSets, a database of 13,954 offshore wind-farm cable-routing problem instances and corresponding solutions obtained using three different optimization approaches. Overall, I find the dataset to be well constructed and potentially valuable to the wind-energy and optimization communities. In particular, the combination of real and procedurally generated farm layouts, different turbine counts and cable capacities, and solutions obtained with both exact and meta-heuristic methods makes the dataset relevant for several of the applications identified by the authors, including benchmarking routing algorithms and developing and evaluating learning-based approaches.
The manuscript provides a useful analysis of the dataset and its characteristics, and the reported results give confidence that the instances and solutions exhibit meaningful variation in difficulty and network characteristics. The availability of solutions from different solvers is also particularly useful for assessing solver performance and for constructing learning-based benchmarks.
I nevertheless see several opportunities to improve the manuscript and, more importantly, to make the scope, limitations, and future usefulness of the dataset clearer. In particular, I would encourage the authors to discuss more explicitly the assumptions underlying the instances and the extent to which these assumptions may constrain the applicability of the dataset. A more explicit discussion of these aspects would help potential users understand which applications the dataset is suitable for and which conclusions should not be generalized beyond the considered problem setting.
I also have a number of suggestions concerning the organization and presentation of the dataset, which I believe would make it substantially easier for future users to understand and reuse it.
Specific comments
Scope and representativeness of the dataset
I suggest specifying explicitly in the Introduction, or at least early in the manuscript, whether the real-world instances considered in the dataset correspond exclusively to bottom-fixed offshore wind farms. This clarification would also be useful in the discussion around the statement that cable-related costs account for approximately 11% of LCOE (around line 21), since the relevance of this figure depends on the type and configuration of offshore wind farm being considered.
Around line 71, I would encourage the authors to discuss in greater detail the limitation imposed by considering a single substation. In particular, it would be useful to explain how this assumption may affect the characteristics of the generated instances and potentially introduce a systematic bias into the benchmark dataset. This is especially relevant if the dataset is subsequently used for machine-learning applications. More generally, I recommend expanding the discussion of the limitations of the current dataset and identifying which extensions would be most useful for future versions.
Around line 118, I would appreciate a more detailed description of how the different classes of instances were generated. Please also explain the difference in the number of instances available for HGS-CVRP and LKH3, as this is potentially relevant when using the dataset to compare solver performance.
Around line 57, I suggest clarifying whether the UTM zone of the real-world instances could also be stored as part of the dataset metadata. If this information is available, including it would improve the geospatial reusability and interoperability of the dataset and would make it easier for users to reconstruct or interpret the spatial information associated with the instances.
Around line 176, I suggest providing a more detailed description of how the boundaries of the different instances are defined. In particular, it would be interesting to clarify whether the boundaries may be concave or convex and whether instances can contain holes or other internal excluded areas. Since these geometric characteristics may influence routing and detour length, it would also be valuable to discuss whether the authors have investigated any relationship between boundary geometry and the observed detour-caused increase in cable length.
Dataset organization and FAIRness
Section 2 would benefit from a visual representation of the hierarchical structure of the dataset. A schematic showing, for example, the relationship between farm layouts, problem instances, topology, cable capacity, solver, and solution-level metadata would make the dataset considerably easier to understand for new users.
I would also consider reorganizing Section 2 to follow more closely the structure of the dataset itself, in particular by aligning the discussion with the three tables subsequently introduced. This could make it easier for readers to move from the conceptual description of the dataset to its actual implementation and contents.
Presentation of the results
I suggest revising the final sentences of the abstract for improved readability and flow. In particular, the statement that the median increase of HGS relative to the exact solutions is 0.0% could be reformulated. A 0.0% median increase is perhaps better described as the median difference being 0.0% rather than as an increase, since the latter wording may imply that an increase is actually observed for the median case.
The shaded bands in Figures 3, 6, and 8 would benefit from an explicit legend explaining their meaning. Although their interpretation may be inferred from the text, labeling them directly in the figures would make the plots more self-contained.
Figure 5 may benefit from explicitly showing the mean line, since the mean is also discussed in the corresponding text. Including it in the figure would allow the reader to relate the quantitative discussion more directly to the visualization.
The caption of Figure 8 could be made more consistent with the caption of Figure 6, particularly in terms of terminology and description of the plotted quantities.
Around line 196, the increase in runtime appears to be associated with the range of approximately 80–100 turbines, whereas an earlier discussion refers to the range 70–110 turbines. I suggest making these ranges consistent, or explaining why different ranges are used. This would also be a good opportunity to expand the discussion of the observed runtime behavior and its relationship to turbine count and cable capacity.
Future development of the dataset
The discussion of alternative analyses that could be performed with the dataset is useful, and the analyses already presented address several important characteristics. Nevertheless, I would encourage the authors to expand the discussion of how the dataset itself could be extended. In particular, it would be useful to identify which additional instance characteristics would provide genuinely new information rather than simply increasing the number of instances. This could include, for example, additional spatial configurations, different numbers or arrangements of substations, alternative boundary geometries, or other relevant constraints on collection-system design. Such a discussion would help establish a roadmap for future versions of the benchmark and clarify how its coverage could be expanded in ways that are most useful to the community.
Language and presentation
The manuscript is generally clear, but I recommend a careful final grammar and language check. I did not notice significant typographical errors, although there are some instances where Oxford commas are omitted. This may of course reflect the authors' preferred style or the journal style, so I leave this primarily as a consistency check rather than a substantive concern.
I suggest rewriting the caption of Figure 1 to reduce nominalization and improve clarity. A more direct description of what is shown and how the reader should interpret the figure would improve its accessibility.
Overall, I consider these to be mostly improvements to the description, scope, and usability of an otherwise promising dataset rather than fundamental concerns about its scientific value.
Technical corrections