static char help[] = "Finds optimal parameter P_m for the generator system while maintaining generator stability.\n"; /*F \begin{eqnarray} \frac{d \theta}{dt} = \omega_b (\omega - \omega_s) \frac{2 H}{\omega_s}\frac{d \omega}{dt} & = & P_m - P_max \sin(\theta) -D(\omega - \omega_s)\\ \end{eqnarray} F*/ /* This code demonstrates how to solve a ODE-constrained optimization problem with TAO, TSEvent, TSAdjoint and TS. The problem features discontinuities and a cost function in integral form. The gradient is computed with the discrete adjoint of an implicit theta method, see ex3adj.c for details. */ #include #include #include "ex3.h" PetscErrorCode FormFunctionGradient(Tao, Vec, PetscReal *, Vec, void *); PetscErrorCode monitor(Tao tao, AppCtx *ctx) { FILE *fp; PetscInt iterate; PetscReal f, gnorm, cnorm, xdiff; TaoConvergedReason reason; PetscFunctionBeginUser; PetscCall(TaoGetSolutionStatus(tao, &iterate, &f, &gnorm, &cnorm, &xdiff, &reason)); fp = fopen("ex3opt_conv.out", "a"); PetscCall(PetscFPrintf(PETSC_COMM_WORLD, fp, "%" PetscInt_FMT " %g\n", iterate, (double)gnorm)); fclose(fp); PetscFunctionReturn(PETSC_SUCCESS); } int main(int argc, char **argv) { Vec p; PetscScalar *x_ptr; PetscMPIInt size; AppCtx ctx; Tao tao; KSP ksp; PC pc; Vec lambda[1], mu[1], lowerb, upperb; PetscBool printtofile; PetscInt direction[2]; PetscBool terminate[2]; Mat qgrad; /* Forward sesivitiy */ Mat sp; /* Forward sensitivity matrix */ /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Initialize program - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ PetscFunctionBeginUser; PetscCall(PetscInitialize(&argc, &argv, NULL, help)); PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size)); PetscCheck(size == 1, PETSC_COMM_WORLD, PETSC_ERR_WRONG_MPI_SIZE, "This is a uniprocessor example only!"); /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Set runtime options - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ PetscOptionsBegin(PETSC_COMM_WORLD, NULL, "Swing equation options", ""); { ctx.beta = 2; ctx.c = 10000.0; ctx.u_s = 1.0; ctx.omega_s = 1.0; ctx.omega_b = 120.0 * PETSC_PI; ctx.H = 5.0; PetscCall(PetscOptionsScalar("-Inertia", "", "", ctx.H, &ctx.H, NULL)); ctx.D = 5.0; PetscCall(PetscOptionsScalar("-D", "", "", ctx.D, &ctx.D, NULL)); ctx.E = 1.1378; ctx.V = 1.0; ctx.X = 0.545; ctx.Pmax = ctx.E * ctx.V / ctx.X; ctx.Pmax_ini = ctx.Pmax; PetscCall(PetscOptionsScalar("-Pmax", "", "", ctx.Pmax, &ctx.Pmax, NULL)); ctx.Pm = 1.06; PetscCall(PetscOptionsScalar("-Pm", "", "", ctx.Pm, &ctx.Pm, NULL)); ctx.tf = 0.1; ctx.tcl = 0.2; PetscCall(PetscOptionsReal("-tf", "Time to start fault", "", ctx.tf, &ctx.tf, NULL)); PetscCall(PetscOptionsReal("-tcl", "Time to end fault", "", ctx.tcl, &ctx.tcl, NULL)); printtofile = PETSC_FALSE; PetscCall(PetscOptionsBool("-printtofile", "Print convergence results to file", "", printtofile, &printtofile, NULL)); ctx.sa = SA_ADJ; PetscCall(PetscOptionsEnum("-sa_method", "Sensitivity analysis method (adj or tlm)", "", SAMethods, (PetscEnum)ctx.sa, (PetscEnum *)&ctx.sa, NULL)); } PetscOptionsEnd(); /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Create necessary matrix and vectors - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ PetscCall(MatCreate(PETSC_COMM_WORLD, &ctx.Jac)); PetscCall(MatSetSizes(ctx.Jac, 2, 2, PETSC_DETERMINE, PETSC_DETERMINE)); PetscCall(MatSetType(ctx.Jac, MATDENSE)); PetscCall(MatSetFromOptions(ctx.Jac)); PetscCall(MatSetUp(ctx.Jac)); PetscCall(MatCreate(PETSC_COMM_WORLD, &ctx.Jacp)); PetscCall(MatSetSizes(ctx.Jacp, PETSC_DECIDE, PETSC_DECIDE, 2, 1)); PetscCall(MatSetFromOptions(ctx.Jacp)); PetscCall(MatSetUp(ctx.Jacp)); PetscCall(MatCreateVecs(ctx.Jac, &ctx.U, NULL)); PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 1, 1, NULL, &ctx.DRDP)); PetscCall(MatSetUp(ctx.DRDP)); PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 2, 1, NULL, &ctx.DRDU)); PetscCall(MatSetUp(ctx.DRDU)); /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Create timestepping solver context - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ PetscCall(TSCreate(PETSC_COMM_WORLD, &ctx.ts)); PetscCall(TSSetProblemType(ctx.ts, TS_NONLINEAR)); PetscCall(TSSetType(ctx.ts, TSCN)); PetscCall(TSSetRHSFunction(ctx.ts, NULL, (TSRHSFunctionFn *)RHSFunction, &ctx)); PetscCall(TSSetRHSJacobian(ctx.ts, ctx.Jac, ctx.Jac, (TSRHSJacobianFn *)RHSJacobian, &ctx)); PetscCall(TSSetRHSJacobianP(ctx.ts, ctx.Jacp, RHSJacobianP, &ctx)); if (ctx.sa == SA_ADJ) { PetscCall(MatCreateVecs(ctx.Jac, &lambda[0], NULL)); PetscCall(MatCreateVecs(ctx.Jacp, &mu[0], NULL)); PetscCall(TSSetSaveTrajectory(ctx.ts)); PetscCall(TSSetCostGradients(ctx.ts, 1, lambda, mu)); PetscCall(TSCreateQuadratureTS(ctx.ts, PETSC_FALSE, &ctx.quadts)); PetscCall(TSSetRHSFunction(ctx.quadts, NULL, (TSRHSFunctionFn *)CostIntegrand, &ctx)); PetscCall(TSSetRHSJacobian(ctx.quadts, ctx.DRDU, ctx.DRDU, (TSRHSJacobianFn *)DRDUJacobianTranspose, &ctx)); PetscCall(TSSetRHSJacobianP(ctx.quadts, ctx.DRDP, DRDPJacobianTranspose, &ctx)); } if (ctx.sa == SA_TLM) { PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 1, 1, NULL, &qgrad)); PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 2, 1, NULL, &sp)); PetscCall(TSForwardSetSensitivities(ctx.ts, 1, sp)); PetscCall(TSCreateQuadratureTS(ctx.ts, PETSC_TRUE, &ctx.quadts)); PetscCall(TSForwardSetSensitivities(ctx.quadts, 1, qgrad)); PetscCall(TSSetRHSFunction(ctx.quadts, NULL, (TSRHSFunctionFn *)CostIntegrand, &ctx)); PetscCall(TSSetRHSJacobian(ctx.quadts, ctx.DRDU, ctx.DRDU, (TSRHSJacobianFn *)DRDUJacobianTranspose, &ctx)); PetscCall(TSSetRHSJacobianP(ctx.quadts, ctx.DRDP, DRDPJacobianTranspose, &ctx)); } /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Set solver options - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ PetscCall(TSSetMaxTime(ctx.ts, 1.0)); PetscCall(TSSetExactFinalTime(ctx.ts, TS_EXACTFINALTIME_MATCHSTEP)); PetscCall(TSSetTimeStep(ctx.ts, 0.03125)); PetscCall(TSSetFromOptions(ctx.ts)); direction[0] = direction[1] = 1; terminate[0] = terminate[1] = PETSC_FALSE; PetscCall(TSSetEventHandler(ctx.ts, 2, direction, terminate, EventFunction, PostEventFunction, &ctx)); /* Create TAO solver and set desired solution method */ PetscCall(TaoCreate(PETSC_COMM_WORLD, &tao)); PetscCall(TaoSetType(tao, TAOBLMVM)); if (printtofile) PetscCall(TaoMonitorSet(tao, (PetscErrorCode (*)(Tao, void *))monitor, (void *)&ctx, NULL)); /* Optimization starts */ /* Set initial solution guess */ PetscCall(VecCreateSeq(PETSC_COMM_WORLD, 1, &p)); PetscCall(VecGetArray(p, &x_ptr)); x_ptr[0] = ctx.Pm; PetscCall(VecRestoreArray(p, &x_ptr)); PetscCall(TaoSetSolution(tao, p)); /* Set routine for function and gradient evaluation */ PetscCall(TaoSetObjectiveAndGradient(tao, NULL, FormFunctionGradient, (void *)&ctx)); /* Set bounds for the optimization */ PetscCall(VecDuplicate(p, &lowerb)); PetscCall(VecDuplicate(p, &upperb)); PetscCall(VecGetArray(lowerb, &x_ptr)); x_ptr[0] = 0.; PetscCall(VecRestoreArray(lowerb, &x_ptr)); PetscCall(VecGetArray(upperb, &x_ptr)); x_ptr[0] = 1.1; PetscCall(VecRestoreArray(upperb, &x_ptr)); PetscCall(TaoSetVariableBounds(tao, lowerb, upperb)); /* Check for any TAO command line options */ PetscCall(TaoSetFromOptions(tao)); PetscCall(TaoGetKSP(tao, &ksp)); if (ksp) { PetscCall(KSPGetPC(ksp, &pc)); PetscCall(PCSetType(pc, PCNONE)); } /* SOLVE THE APPLICATION */ PetscCall(TaoSolve(tao)); PetscCall(VecView(p, PETSC_VIEWER_STDOUT_WORLD)); /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Free work space. All PETSc objects should be destroyed when they are no longer needed. - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ PetscCall(MatDestroy(&ctx.Jac)); PetscCall(MatDestroy(&ctx.Jacp)); PetscCall(MatDestroy(&ctx.DRDU)); PetscCall(MatDestroy(&ctx.DRDP)); PetscCall(VecDestroy(&ctx.U)); if (ctx.sa == SA_ADJ) { PetscCall(VecDestroy(&lambda[0])); PetscCall(VecDestroy(&mu[0])); } if (ctx.sa == SA_TLM) { PetscCall(MatDestroy(&qgrad)); PetscCall(MatDestroy(&sp)); } PetscCall(TSDestroy(&ctx.ts)); PetscCall(VecDestroy(&p)); PetscCall(VecDestroy(&lowerb)); PetscCall(VecDestroy(&upperb)); PetscCall(TaoDestroy(&tao)); PetscCall(PetscFinalize()); return 0; } /* ------------------------------------------------------------------ */ /* FormFunctionGradient - Evaluates the function and corresponding gradient. Input Parameters: tao - the Tao context X - the input vector ptr - optional user-defined context, as set by TaoSetObjectiveAndGradient() Output Parameters: f - the newly evaluated function G - the newly evaluated gradient */ PetscErrorCode FormFunctionGradient(Tao tao, Vec P, PetscReal *f, Vec G, PetscCtx ctx0) { AppCtx *ctx = (AppCtx *)ctx0; PetscInt nadj; PetscReal ftime; PetscInt steps; PetscScalar *u; PetscScalar *x_ptr, *y_ptr; Vec q; Mat qgrad; PetscFunctionBeginUser; PetscCall(VecGetArrayRead(P, (const PetscScalar **)&x_ptr)); ctx->Pm = x_ptr[0]; PetscCall(VecRestoreArrayRead(P, (const PetscScalar **)&x_ptr)); /* reinitialize the solution vector */ PetscCall(VecGetArray(ctx->U, &u)); u[0] = PetscAsinScalar(ctx->Pm / ctx->Pmax); u[1] = 1.0; PetscCall(VecRestoreArray(ctx->U, &u)); PetscCall(TSSetSolution(ctx->ts, ctx->U)); /* reset time */ PetscCall(TSSetTime(ctx->ts, 0.0)); /* reset step counter, this is critical for adjoint solver */ PetscCall(TSSetStepNumber(ctx->ts, 0)); /* reset step size, the step size becomes negative after TSAdjointSolve */ PetscCall(TSSetTimeStep(ctx->ts, 0.03125)); /* reinitialize the integral value */ PetscCall(TSGetQuadratureTS(ctx->ts, NULL, &ctx->quadts)); PetscCall(TSGetSolution(ctx->quadts, &q)); PetscCall(VecSet(q, 0.0)); if (ctx->sa == SA_TLM) { /* reset the forward sensitivities */ TS quadts; Mat sp; PetscScalar val[2]; const PetscInt row[] = {0, 1}, col[] = {0}; PetscCall(TSGetQuadratureTS(ctx->ts, NULL, &quadts)); PetscCall(TSForwardGetSensitivities(quadts, NULL, &qgrad)); PetscCall(MatZeroEntries(qgrad)); PetscCall(TSForwardGetSensitivities(ctx->ts, NULL, &sp)); val[0] = 1. / PetscSqrtScalar(1. - (ctx->Pm / ctx->Pmax) * (ctx->Pm / ctx->Pmax)) / ctx->Pmax; val[1] = 0.0; PetscCall(MatSetValues(sp, 2, row, 1, col, val, INSERT_VALUES)); PetscCall(MatAssemblyBegin(sp, MAT_FINAL_ASSEMBLY)); PetscCall(MatAssemblyEnd(sp, MAT_FINAL_ASSEMBLY)); } /* solve the ODE */ PetscCall(TSSolve(ctx->ts, ctx->U)); PetscCall(TSGetSolveTime(ctx->ts, &ftime)); PetscCall(TSGetStepNumber(ctx->ts, &steps)); if (ctx->sa == SA_ADJ) { Vec *lambda, *mu; /* reset the terminal condition for adjoint */ PetscCall(TSGetCostGradients(ctx->ts, &nadj, &lambda, &mu)); PetscCall(VecGetArray(lambda[0], &y_ptr)); y_ptr[0] = 0.0; y_ptr[1] = 0.0; PetscCall(VecRestoreArray(lambda[0], &y_ptr)); PetscCall(VecGetArray(mu[0], &x_ptr)); x_ptr[0] = -1.0; PetscCall(VecRestoreArray(mu[0], &x_ptr)); /* solve the adjont */ PetscCall(TSAdjointSolve(ctx->ts)); PetscCall(ComputeSensiP(lambda[0], mu[0], ctx)); PetscCall(VecCopy(mu[0], G)); } if (ctx->sa == SA_TLM) { PetscCall(VecGetArray(G, &x_ptr)); PetscCall(MatDenseGetArray(qgrad, &y_ptr)); x_ptr[0] = y_ptr[0] - 1.; PetscCall(MatDenseRestoreArray(qgrad, &y_ptr)); PetscCall(VecRestoreArray(G, &x_ptr)); } PetscCall(TSGetSolution(ctx->quadts, &q)); PetscCall(VecGetArray(q, &x_ptr)); *f = -ctx->Pm + x_ptr[0]; PetscCall(VecRestoreArray(q, &x_ptr)); PetscFunctionReturn(PETSC_SUCCESS); } /*TEST build: requires: !complex !single test: args: -viewer_binary_skip_info -ts_type cn -pc_type lu -tao_monitor test: suffix: 2 output_file: output/ex3opt_1.out args: -sa_method tlm -ts_type cn -pc_type lu -tao_monitor TEST*/