Learning sos2

[lisazeyen]

Closes # (if applicable).

Endogenous technology learning

This PR should include endogenous investment cost reductions based on installed capacities. Cumulative cost curves of one technology are piecewised linearised using SOS2 constraints. This results in an MIP problem. Currently endogenous learning is only implemented for pyomo=False and solver_name="gurobi".

Checklist

• Code changes are sufficiently documented; i.e. new functions contain docstrings and further explanations may be given in doc.
• Unit tests for new features were added (if applicable).
• Newly introduced dependencies are added to environment.yaml, environment_docs.yaml and setup.py (if applicable).
• A note for the release notes doc/release_notes.rst of the upcoming release is included.
• I consent to the release of this PR's code under the MIT license.

[lisazeyen]

@fneum and @FabianHofmann I am not sure if it is worth it to pull this into the main pypsa code since it is only implemented for gurobi. Another option could be that we just include into pypsa the option to write SOS2 constraints with gurobi and I keep the learning script in my private repository. One could run the learning than just as an extra_functionality with skip_objective=True. Or we start to make a collection with helpful extra_functionalities and store the learning script there. What do you think?

[fneum]

I'm tending towards including learning functionality in PyPSA. It's quite a nice distinguishing feature. Would it be an option to include SOS2 for Gurobi and MILP formulation for other solvers?

[coroa]

@lisazeyen I just checked the CPLEX doc at https://www.ibm.com/docs/en/icos/12.8.0.0?topic=representation-mip-features-in-lp-file-format to find that their SOS2 constraint representation is compatible to Gurobi's, so that you at least capture both of them. What's missing here? Apart from porting it to linopy? :)

[lisazeyen]

@lisazeyen I just checked the CPLEX doc at https://www.ibm.com/docs/en/icos/12.8.0.0?topic=representation-mip-features-in-lp-file-format to find that their SOS2 constraint representation is compatible to Gurobi's, so that you at least capture both of them. What's missing here? Apart from porting it to linopy? :)

Ah great, thanks for checking! Then we can just add a check for the solver and allow learning only for gurobi and cplex. I haven't done the MIP formulation for the other solvers but we could also pull it in now and do the implementation for the others at a later point.