Extreme Performance OPF Solvers for Large Number of Time Periods
Structure exploiting, data-compression algorithms for extreme scalability and low memory footprint
Slide 2
Slide 3
Multiperiod OPF Problems
Multiperiod optimal power flow problems including storage devices and generator ramping
SC-OPF Problems
Security constrained optimal power flow problems for very large number of contingencies
Stochastic OPF Problems
Stochastic OPF problems with thousands of scenarios
About
Beltistos is a suite of high-performance OPF algorithms including extremely scalable and low memory multiperiod OPF (MPOPF) and security-constrained OPF (SCOPF) solvers. Beltistos adopts selected algorithms implemented in IPOPT adjusted specifically for the nature of the OPF problems. The MPOPF and SCOPF solution algorithms implement structure exploiting and data compression algorithms designed for the particular structure of the MPOPF problems. Beltistos-mem is a variant of the MPOPF algorithm which sacrifices redundant computation in favor of memory efficiency. Beltistos relies on PARDISO linear solver, providing improved performance and accuracy by utilizing features such as reordering, scaling and pivoting.
Examples
OPF Performance Comparison
Time for the solution of various OPF benchmarks using various commercial and open source optimizers is shown below.
Beltistos is the fastest optimizer for all large-scale cases. MIPS-PARDISO was slower by a factor of 4,
while KNITRO and IPOPTPARDISO were up to 5.5 times slower. MIPS with the default linear solver and FMINCON were slower up to a factor of 35 or 40, respectively.
Concerning the number of iterations, IPOPT-PARDISO and MIPS-PARDISO perform up to 2.5 times more iterations compared to the Beltistos.
Regarding the memory requirements, MIPS-PARDISO is the most efficient optimizer for roughly 50% of the benchmark cases, very closely followed by Beltistos and IPOPT-PARDISO.
MIPS with the default linear solver required up to 7 times more memory while solving the largest benchmark. More details about the benchmarks can be found at the following reference: Technical report.
MIPS
MIPS-Pardiso
IPOPT-Pardiso
IPOPT-MA57
Beltistos
FMINCON
KNITRO
MPOPF Comparision - case118 Time and Memory
The average time per iteration for number of time periods N=3,600 up to N=8,760 corresponding to one year with a time step size corresponding to one hour is shown below.
It is worth noting that BELTISTOS-mem, the memory saving approach of Beltistos, is slightly slower than the normal mode of BELTISTOS but it is still
almost four orders of magnitude faster than IPOPT and KNITRO. It also needs approximately two orders of magnitude less memory than IPOPT.
More details about the benchmarks can be found at the following reference: MPOPF IEEE paper.
Beltistos
Beltistos-mem
IPOPT
KNITRO
MPOPF Performance Comparison
The average time per iteration for number of time periods N=600 up to N=4,800 is shown below for different power grid networks.
Beltistos
Beltistos-mem
IPOPT
SCOPF Performance Comparison
The average time per iteration is shown for the SCOPF benchmarks with different number of (artifitial) contingency scenarios.
Serial IPOPT and parallel implementation of Beltistos with distributed memory using different number of MPI proccesses is shown.
More details about the benchmarks can be found at the following reference: SCOPF IEEE paper.