Appropriate solve parameter for large model

[fg]

Alright, thanks for the hints!
Yesterday, I did model runs all day long, comparing the run durations  (CPLEX time) between the former and your alternative suggestion for the milestone years, by letting the model run different ending years:

   
(One attempt up to the ending year 2050 is still underway on the workstation...)

The alternative milestone year definition seems to be better in some cases and worse in others, but since it seems to be a more straightforward definition, I will keep it for now.
I also checked and confirm that OBLONG is activated and tried explicitly the modified objective function definition, which did not show any difference.

Right now, I'm experimenting with different CPLEX parameters such as setting barepcomp to a somewhat higher value (1e-6, 1e-4, ...), but it also seems to have a rather small effect.

Yeah, so I would agree with you, it seems that I'm facing numerical problems. Therefore, another thing that came into my mind, would be rounding the AF and COM_FR values to, say, 5 or 6 digits behind the decimal point (right now they are not rounded at all and could, for example, be 0.170984708319).
Do you think that might mitigate numerical problems?

If nothings helps, I would enable a time-stepped solution or, as the ultimate last resort, decrease the temporal resolution of the model...

[Antti-L]

Well, there are also some other things to check for improving numerical stability, for example:

• Remove dummy import processes from your model if you have such (e.g. NCAP_START(IMP*Z)=2200)
• Make sure you don't have any other artificially high costs in your model (to prevent some technology from entering the solution), and use zero bounds or NCAP_START instead

Even if the dummy imports are not active in the solution, they can cause poor numerical stability for the barrier algorithm if the processes are available in the model.

[fg]

Awesome, Antti! That was the major cause of the instability - I deactivated all dummy imports and the run times dropped tremendously both on the desktop and the workstation - have a look:    

Very big thank you for that!
Now I can even think of increasing the temporal resolution and/or adding more fancy future technologies...

[Antti-L]

Great, thanks for the very good news!