Solves an interpolated scenario once per solver option set and returns runtime and objective for each, so back ends can be compared on the same linear programme.
Arguments
- scen
an interpolated scenario, from
energyRt::interpolate_model().- solvers
character vector of names in
energyRt::solver_options, or a named list of option sets. Defaults to the open solvers reported available byopen_solvers().- tol
relative objective difference treated as agreement.
- verbose
report each run as it completes.
Value
A data frame with solver, seconds, objective, status and
agrees, ordered by runtime. The reference for agrees is the first
successful run.
Details
Every run receives the same scenario, so the objectives are expected to
agree to solver tolerance. A disagreement larger than tol is reported in
the agrees column and is a signal to investigate the back end, not the
model.
Solving is sequential. A back end that fails records the error and does not stop the remaining runs.
Solver choice by model size
Relative performance depends strongly on model size, and the ordering reverses across the range. Measured on one workstation; treat the ratios rather than the absolute times as transferable.
| model | GLPK | CBC | HiGHS |
| 5 regions, 168 h | 10 s | 19 s | 24-47 s |
| 69 regions, 96 slices | 2158 s | 809 s | 79-82 s |
| 1035 regions, 96 slices | – | – | 18.5 min |
On the smallest models GLPK wins on start-up cost alone. From roughly a hundred thousand rows onward HiGHS dominates: at 69 regions it is 27 times faster than GLPK, and the gap widens with size.
At 1035 regions on a 96-slice sample the linear programme is 1.3 million
rows after presolve, solved by HiGHS interior point in about 16 minutes.
Extending the same model to 672 slices produces 9.4 million rows after
presolve, at which point the interior-point method fails to return a
solution; a first-order method (solver = "pdlp") or dual simplex is the
route to test at that size.
Julia carries a fixed start-up cost of 20-30 seconds per invocation for just-in-time compilation, which dominates small models and is amortised across a session. Pyomo starts in a few seconds. For a single small solve Pyomo is therefore quicker end to end even where Julia's solver call is faster.
Examples
if (FALSE) { # \dontrun{
scen <- energyRt::interpolate_model(pypsa_eur_5, name = "be")
benchmark_solvers(scen)
} # }