Non robust results

Q&A
, , ,
1

Hi all,

I am facing a problem to get robust results using oemof-solph 0.5.2 and gurobi 11.0.1 on Windows via WinPyhton, but also have the problem via Anaconda. I have a fairly large energy model with the following components:

  • 4x buses, like electricity, CH4, FT, H2
  • 13x flexible sources (dispatch), like gas power plants
  • 15x fixed sources (dispatch), like pv
  • 8x converters (dispatch), like electrolysis
  • 5x flexible sinks (dispatch), like export
  • 123x fixed sinks (dispatch), like various industrial processes
  • 4 storages (invest for input-power, capacitiy and output-power in combination with exist), like H2-, CH4, FT, battery storage

Running the energy model in Spyder gives me some results, running it again in Spyder without closing it before gives me the same results. However, if I close Spyder and open it again I get slightly different results. I have already tried setting the gurobi parameter “seed”, but it did not help.

I am solving the energy model with following code line:

modell.solve(solver="gurobi", solve_kwargs={"tee": True, "logfile": os.path.join(os.getcwd(), modell_logfile)}, cmdline_options={"Seed": 123})

Has anyone else experienced this problem?

Thanks in advance for any help!

Kind regards,
Alex

2

Hi @Alexander,

What does “slightly different” mean? If I understand the Gurobi documentation correct, is stops if it approaches the optimum to 0.1 % (see MIPGap - Gurobi Optimization). Is this the order of magnitude of the differences you observe?

Cheers,
Patrik

3

Hi @pschoen,

Thanks for your reply. Indeed, it is less than 0.1%. But why do I only get this difference when I close and open Spyder? How can I adjust it to get reproducible results? I do not change any input data. The idea is that I want to use these optimisation results for another run with oemof. So with each iteration, the overall magnitude of the differences would increase.

By the way, if I don’t take energy storages into account, then I don’t have any differences.

Best regards,
Alex

4

Another interesting point is that the optimal objective value is always the same. However, the use of the components is slightly different. It seems that there are more minima. Can I tell gurobi to use the same method/starting points to find the same minima even after restarting spyder?

5

Maybe, it is a single wide minimum. (This is more likely for a linear problem. But with integer variables I won’t exclude more than one minimum.) However, I’m sorry but I have no idea. I never really used Gurobi.