PowerSimulations.jl
PowerSimulations.jl copied to clipboard
NREL-Sienna
Performance disparity in StandardUC between Model and Simulation
When we solve the StandardUnitCommitment model using just the Problem framework, we implement one formulation of the duration constraints that use the initial condition of the time spent in a particular state (on/off) to decide the number of variables to add to the model. This is incompatible with in-memory updates using ParameterJuMP (can only handle RHS), so we implement a different formulation that keeps track of the state on the RHS. This has been discussed in the literature here and here
I noted that depending on the solver, this could have large performance consequences. Presumably, because the solvers perform different internal substitutions. The tradeoff here is the re-creating the UC problem for every solve vs. in-memory update. There is no straightforward rule of thumb since these outcomes are heavily dependent on the model's parameters, the number of units, and the solver heuristics.
Using the ExtremeSolarTexas case, we can see the following outcomes.
With CPLEX not using the parameters we get:
version identifier: 20.1.0.0 | 2020-11-10 | 9bedb6d68
CPXPARAM_Emphasis_MIP 1
CPXPARAM_MIP_Tolerances_MIPGap 0.01
Tried aggregator 3 times.
MIP Presolve eliminated 23868 rows and 12408 columns.
MIP Presolve added 108 rows and 0 columns.
MIP Presolve modified 6092 coefficients.
Aggregator did 1503 substitutions.
Reduced MIP has 103185 rows, 124689 columns, and 1100035 nonzeros.
Reduced MIP has 34301 binaries, 0 generals, 0 SOSs, and 0 indicators.
Presolve time = 1.39 sec. (924.73 ticks)
Probing fixed 1501 vars, tightened 400 bounds.
Probing changed sense of 242 constraints.
Probing time = 2.49 sec. (190.52 ticks)
Cover probing fixed 0 vars, tightened 63 bounds.
Tried aggregator 2 times.
Detecting symmetries...
MIP Presolve eliminated 3395 rows and 2959 columns.
MIP Presolve added 2 rows and 0 columns.
MIP Presolve modified 185 coefficients.
Aggregator did 2 substitutions.
Reduced MIP has 99790 rows, 121728 columns, and 1080671 nonzeros.
Reduced MIP has 32708 binaries, 0 generals, 0 SOSs, and 0 indicators.
Presolve time = 1.37 sec. (1071.98 ticks)
Probing time = 0.24 sec. (35.56 ticks)
Cover probing fixed 0 vars, tightened 63 bounds.
Clique table members: 565922.
Tightened 63 constraints.
MIP emphasis: integer feasibility.
MIP search method: dynamic search.
Parallel mode: deterministic, using up to 8 threads.
Root relaxation solution time = 33.34 sec. (10575.89 ticks)
Nodes Cuts/
Node Left Objective IInf Best Integer Best Bound ItCnt Gap
0 0 3.09365e+07 5141 3.09365e+07 41
0 0 3.21964e+07 2875 Cuts: 8993 33612
* 0+ 0 3.25812e+07 3.21964e+07 1.18%
0 0 3.23820e+07 777 3.25812e+07 Cuts: 8764 40919 0.61%
GUB cover cuts applied: 2
Clique cuts applied: 1255
Implied bound cuts applied: 9557
Flow cuts applied: 2154
Mixed integer rounding cuts applied: 2520
Zero-half cuts applied: 27
Lift and project cuts applied: 82
Gomory fractional cuts applied: 76
Root node processing (before b&c):
Real time = 193.43 sec. (66501.28 ticks)
Parallel b&c, 8 threads:
Real time = 0.00 sec. (0.00 ticks)
Sync time (average) = 0.00 sec.
Wait time (average) = 0.00 sec.
------------
Total (root+branch&cut) = 193.43 sec. (66501.28 ticks)
While when solving as part of the simulations with parameters, we get this from the solver:
Version identifier: 20.1.0.0 | 2020-11-10 | 9bedb6d68
CPXPARAM_Emphasis_MIP 1
CPXPARAM_MIP_Tolerances_MIPGap 0.01
Tried aggregator 3 times.
MIP Presolve eliminated 24854 rows and 12495 columns.
MIP Presolve added 108 rows and 0 columns.
MIP Presolve modified 4630 coefficients.
Aggregator did 1617 substitutions.
Reduced MIP has 102085 rows, 124488 columns, and 1093674 nonzeros.
Reduced MIP has 34100 binaries, 0 generals, 0 SOSs, and 0 indicators.
Presolve time = 1.03 sec. (735.24 ticks)
Probing fixed 1270 vars, tightened 400 bounds.
Probing changed sense of 41 constraints.
Probing time = 2.71 sec. (185.18 ticks)
Cover probing fixed 0 vars, tightened 30 bounds.
Tried aggregator 2 times.
Detecting symmetries...
MIP Presolve eliminated 2658 rows and 2728 columns.
MIP Presolve added 2 rows and 0 columns.
MIP Presolve modified 186 coefficients.
Aggregator did 2 substitutions.
Reduced MIP has 99427 rows, 121758 columns, and 1078905 nonzeros.
Reduced MIP has 32738 binaries, 0 generals, 0 SOSs, and 0 indicators.
Presolve time = 1.49 sec. (1067.84 ticks)
Probing fixed 30 vars, tightened 0 bounds.
Probing time = 1.51 sec. (174.14 ticks)
Cover probing fixed 0 vars, tightened 151 bounds.
Clique table members: 629960.
Tightened 151 constraints.
MIP emphasis: integer feasibility.
MIP search method: dynamic search.
Parallel mode: deterministic, using up to 8 threads.
Root relaxation solution time = 37.57 sec. (10960.13 ticks)
Nodes Cuts/
Node Left Objective IInf Best Integer Best Bound ItCnt Gap
0 0 7817834.6638 3289 7817834.6638 42
0 0 8773510.4182 1597 Cuts: 7767 33588
* 0+ 0 3.58045e+07 8773510.4182 75.50%
0 0 8918337.8898 756 3.58045e+07 Cuts: 5708 46617 75.09%
0 0 8965520.7979 527 3.58045e+07 Cuts: 2188 52690 74.96%
* 0+ 0 2.57355e+07 8965520.7979 65.16%
0 0 -1.00000e+75 0 2.57355e+07 8965520.7979 52690 65.16%
0 2 8965520.7979 527 2.57355e+07 8965520.7979 52690 65.16%
Elapsed time = 266.30 sec. (100718.31 ticks, tree = 0.02 MB, solutions = 2)
1 3 8993198.9894 611 2.57355e+07 8966445.0306 55322 65.16%
2 4 8994688.6932 601 2.57355e+07 8966445.0306 56248 65.16%
3 5 8994966.9839 607 2.57355e+07 8966445.0306 57024 65.16%
4 6 8995193.6979 593 2.57355e+07 8966445.0306 62171 65.16%
5 7 8996435.1007 597 2.57355e+07 8966445.0306 68279 65.16%
6 8 9004228.9841 600 2.57355e+07 8966445.0306 73001 65.16%
7 9 9005030.8684 592 2.57355e+07 8966445.0306 76476 65.16%
8 10 9005422.4586 590 2.57355e+07 8966445.0306 76569 65.16%
9 11 9005591.5063 584 2.57355e+07 8966445.0306 76854 65.16%
* 10+ 1 2.56001e+07 8966445.0306 64.97%
We should find a way to reduce the performance hit of the formulation that enables RHS updating or a new formulation that will allow RHS update without such a penalty.
@jd-lara is there any advantage to updating the RHS of a constraint on a varon variable instead of fixing/unfixing it?
@alefcastelli and I think we have a fix in the formulation for this -- we just want to implement it in the way that is most preferred.
@jd-lara is there any advantage to updating the RHS of a constraint on a
varonvariable instead of fixing/unfixing it?@alefcastelli and I think we have a fix in the formulation for this -- we just want to implement it in the way that is most preferred.
I don't think we have tried the approach to fix/unfix in the update code vs setting the value. I would have to check the implementation of the parameter update. My only concern would be if I the model add additional x = val constraints which is kinda solver dependent. We can try it out.
After careful study, this issue couldn't be reproduce adequately, it's been open since 2021 and there have been a lot of changes already to the code.