Abstract
Once the PV-panels—the PV-modules—of a solar plant are placed, its electrical design of a solar plant with the so-called distributed string-inverters topology couples six decisions: the partition of modules into PV-strings; the type and placement of the inverters; the assignment of strings to inverters; and the routing of DC-homerun and AC-feeder cables. Inverters, cables, and the trenches that carry them absorb a large share of a plant's construction cost, yet industry practice fixes the six decisions sequentially by hand, and no prior optimisation model covers all six decisions within a single program on the irregular sites where such plants are built. This work formulates the joint problem as two monolithic mixed-integer linear programs over discretised routing graphs: a string-indexed model and an inverter-type-indexed model. The second is an approximation: it gives up the upper bound on each individual string-to-inverter circuit's length, bounding instead its average over the strings of each inverter. It further contributes nine benchmark instances—operating plants of 1 to 10 MW, transcribed from engineering drawings and each annotated by hand with an engineer-style reference design—against which this and future layout algorithms can be evaluated. Solved directly, the more scalable of the two models produces a layout on only one of the nine plants within an hour. Fixing decision groups to the engineer's reference values one group at a time reveals the bottleneck: the partition of modules into PV-strings. A four-stage decomposition built on that diagnosis—the topology fixed by a single-commodity module flow, module-to-inverter paths as a multi-commodity flow, the full model under the frozen topology, and an unfrozen polish—produces complete layouts on all nine plants without any annotation, each priced 0.3–5.2% below the reference design under the model's objective. On the six smaller plants it reproduces the reference design's exact inverter and string counts in layouts that at plant scale are hard to distinguish from the hand designs, while routing 10–34% less DC-homerun cable (18.3 km in aggregate) and 2–14% less PV-string cable—savings hidden in terminal orientations and string-to-inverter pairings the eye does not resolve. On the three larger plants it instead revises the design, removing inverters or trenches at the price of longer homeruns—trades that improve the model's objective, though their worth under market prices remains open.