Implement Hawaii 37-bus case

[pelesh]

Summary

Implement phasor domain dynamics example simulating 37-bus Hawaii synthetic grid model.

Rationale

This is a medium-size test case, which was running too slow with dense linear solver. With sparse Jacobian evaluation available now, simulating this case would help verify the new infrastructure in GridKit.

Description

The build framework and JSON input file already exist in GridKit. The mini-app that runs the simulation needs to be implemented within the placeholder file hawaiiJSON.cpp, preferably using new monitored variable feature.

Additional information

No response

[abirchfield]

Branch I'm working on is here

Attaching files for debugging:

• Printout of Jacobian in coordinate form based on model->getJacobian()
• Printout of Jacobian from SUNSparseMatrix_Print after jacobian is sent to Sundials
• Sundials log messages

The above are attached for both the successful run (examples/PhasorDynamics/Small/TenGen/Genrou/TenGenGenrou) and the failed run (examples/PhasorDynamics/Medium/Hawaii/hawaiiJson)

sundials-messages-failed.txt
jacdump-coo-failed.txt
jacdump-sun-failed.txt

sundials-messages-success.txt
jacdump-coo-success.txt
jacdump-sun-success.txt

[pelesh]

It seems there is something wrong with the Jacobian sparsity pattern. I instrumented SUNDIALS code and the failure is at KLU symbolic factorization:

32: Input file: "/Users/55y/src/gridkit/build-clang/examples/PhasorDynamics/Medium/Hawaii/hawaii.json"
32: Configuring linear solver sparse.
32: SUNLinSolSetup_KLU called
32: KLU symbolic analysis failed
32: idaLsSetup: SUNLinSolSetup returned flag = NONE

[pelesh]

The issue seems to be that duplicate entries are not removed from the Jacobian. See for example row 1 of the Hawaii case Jacobian:

row 1 : locations 22 to 43
  0:  3.249804957419269e+02  1: -5.627980198922352e+01  2: -8.067426042704465e+00  2: -8.067426042704465e+00  2: -8.067426042704465e+00  3:  3.154852063304107e-01  3:  3.154852063304107e-01  3:  3.154852063304107e-01  8: -4.658580502831715e+01  8: -4.658580502831715e+01  9:  7.573841012589218e+00  9:  7.573841012589218e+00  42: -4.040135416888868e+01  43:  6.250819699319226e+00  48: -6.124943639507730e+01  48: -6.124943639507730e+01  49:  1.120463800365701e+01  49:  1.120463800365701e+01  68: -2.236920779906800e+01  68: -2.236920779906800e+01  69:  5.762784319210305e+00  69:  5.762784319210305e+00  

If you look closer, it seems that some Jacobian entries are just repeated.

[nkoukpaizan]

Good catch @pelesh. I got distracted by the nan values (which I think come from the limits) and didn't look closer.

Looking closer, the Jacobian assembly is doing what I expect it to do. In the hawaii.json file, there are branch devices that are connecting the same buses. For examples:

• Branches with ids, BR_1_2_1, BR_1_2_2 and BR_1_2_3 all connect buses 1 and 2.
• Branches with ids, BR_1_5_1 and BR_1_5_2 all connect buses 1 and 5.

This is not something I am accounting for. I was assuming branch connections between buses are unique.

[nkoukpaizan]

I just pushed a branch (https://github.com/ORNL/GridKit/tree/nicholson/hawaii) with the following changes:

• A workaround to add Jacobian contributions from components to support parallel branches. I'll have to think about efficient ways to do this (e.g., in parallel), because each branch owns the off diagonal contributions and is unaware of other branches. But for most components, checking this addition is wasteful.
• An implementation of my proposal for limits in Question: behavior of sigmoid and Indicators #320, which removes nan values.
• Added monitors to hawaii.json to generate a mon.csv file.

These allowed me to fully run the hawaii example to completion.
@abirchfield, please try this and see if it helps move your work forward.