ExaOpt

ExaOpt extends ExaPF.jl by implementing the optimization routines required to solve the optimal power flow (OPF) problems. The routines come in two flavors:

in src/Evaluators/, various optimization evaluators implement the different callbacks required in the optimization algorithm (objective, gradient, Hessian).
in src/Algorithms/, an Augmented Lagrangian algorithm is implemented, targeting primarily the resolution of large-scale OPF problems on GPU architectures.

Installation

ExaOpt is currently unregistered. To install it, you should be a member of the exanauts organization.

add "git@github.com:exanauts/ExaPF-Opt.git"

To check that everything is working as expected, please run

test ExaOpt

By default, this command tests all the Evaluators implemented in ExaOpt on the CPU and, if available, on CUDA GPUs.

Quickstart

ExaOpt's cornerstone is the ReducedSpaceEvaluator, allowing to run the optimization algorithm directly in the reduced space induced by the power flow equations. Instantiating a new ReducedSpaceEvaluator from a MATPOWER file simply amounts to

nlp = ExaOpt.ReducedSpaceEvaluator("case57.m")

An initial optimization variable can be computed as

u = ExaPF.initial(nlp)

The variable u is the control that will be used all throughout the optimization. Then, the function update! solves the power flow and updates accordingly all the structures inside nlp:

ExaPF.update!(nlp, u)

Once the power flow equations solved, the other callbacks can be evaluated as well:

ExaPF.objective(nlp, u) # objective
ExaPF.gradient(nlp, u)  # reduced gradient
ExaPF.jacobian(nlp, u)  # reduced Jacobian
ExaPF.hessian(nlp, u)   # reduced Hessian

Use the ReducedSpaceEvaluator on GPU accelerators

ExaPF.jl is using KernelAbstractions to implement all its core operations. Hence, deporting the computation in ExaPF on GPU accelerators is straightforward. ExaOpt inherits this behavior and all evaluators can be instantiated on GPU accelerators, simply as

using CUDAKernels # Load CUDA backend for KernelAbstractions
nlp = ExaOpt.ReducedSpaceEvaluator("case57.m"; device=CUDADevice())

When doing so, all kernels are instantiated on the GPU to avoid memory transfer between the host and the device. The sparse linear algebra operations are handled by cuSPARSE, and the sparse factorizations are performed using cusolverRF. When deporting the computation on the GPU, the reduced Hessian can be evaluated in parallel.

Batch evaluation of the reduced Hessian

Instead of computing the reduced Hessian one Hessian-vector product after one Hessian-vector product, the Hessian-vector products are directly evaluated in batch in this case. To activate the batch evaluation for the reduced Hessian, please specify the number of Hessian-vector products to perform in one batch as

nlp = ExaOpt.ReducedSpaceEvaluator("case57.m"; device=CUDADevice(), nbatch_hessian=8)

Note that on large instances, the batch computation could be quite heavy on the GPU's memory.

MOI wrapper

ExaPF implements a wrapper to MathOptInterface that allows to solve the optimal power flow problem directly in the reduced space induced by the power flow equations:

nlp = ExaPF.ReducedSpaceEvaluator("case57.m")
optimizer = Ipopt.Optimizer() # MOI optimizer
# Use LBFGS algorithm, as reduced Hessian is not available by default!
MOI.set(optimizer, MOI.RawParameter("hessian_approximation"), "limited-memory")
MOI.set(optimizer, MOI.RawParameter("tol"), 1e-4)
# Solve reduced space problem
solution = ExaPF.optimize!(optimizer, nlp)

sshin23 / ExaPF-Opt