1 2 static char help[] = "Finds optimal parameter P_m for the generator system while maintaining generator stability.\n"; 3 4 /*F 5 6 \begin{eqnarray} 7 \frac{d \theta}{dt} = \omega_b (\omega - \omega_s) 8 \frac{2 H}{\omega_s}\frac{d \omega}{dt} & = & P_m - P_max \sin(\theta) -D(\omega - \omega_s)\\ 9 \end{eqnarray} 10 11 F*/ 12 13 /* 14 This code demonstrates how to solve a ODE-constrained optimization problem with TAO, TSEvent, TSAdjoint and TS. 15 The problem features discontinuities and a cost function in integral form. 16 The gradient is computed with the discrete adjoint of an implicit theta method, see ex3adj.c for details. 17 */ 18 19 #include <petsctao.h> 20 #include <petscts.h> 21 #include "ex3.h" 22 23 PetscErrorCode FormFunctionGradient(Tao, Vec, PetscReal *, Vec, void *); 24 25 PetscErrorCode monitor(Tao tao, AppCtx *ctx) 26 { 27 FILE *fp; 28 PetscInt iterate; 29 PetscReal f, gnorm, cnorm, xdiff; 30 TaoConvergedReason reason; 31 32 PetscFunctionBeginUser; 33 PetscCall(TaoGetSolutionStatus(tao, &iterate, &f, &gnorm, &cnorm, &xdiff, &reason)); 34 35 fp = fopen("ex3opt_conv.out", "a"); 36 PetscCall(PetscFPrintf(PETSC_COMM_WORLD, fp, "%" PetscInt_FMT " %g\n", iterate, (double)gnorm)); 37 fclose(fp); 38 PetscFunctionReturn(0); 39 } 40 41 int main(int argc, char **argv) 42 { 43 Vec p; 44 PetscScalar *x_ptr; 45 PetscMPIInt size; 46 AppCtx ctx; 47 Tao tao; 48 KSP ksp; 49 PC pc; 50 Vec lambda[1], mu[1], lowerb, upperb; 51 PetscBool printtofile; 52 PetscInt direction[2]; 53 PetscBool terminate[2]; 54 Mat qgrad; /* Forward sesivitiy */ 55 Mat sp; /* Forward sensitivity matrix */ 56 57 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 58 Initialize program 59 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 60 PetscFunctionBeginUser; 61 PetscCall(PetscInitialize(&argc, &argv, NULL, help)); 62 PetscFunctionBeginUser; 63 PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size)); 64 PetscCheck(size == 1, PETSC_COMM_WORLD, PETSC_ERR_WRONG_MPI_SIZE, "This is a uniprocessor example only!"); 65 66 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 67 Set runtime options 68 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 69 PetscOptionsBegin(PETSC_COMM_WORLD, NULL, "Swing equation options", ""); 70 { 71 ctx.beta = 2; 72 ctx.c = 10000.0; 73 ctx.u_s = 1.0; 74 ctx.omega_s = 1.0; 75 ctx.omega_b = 120.0 * PETSC_PI; 76 ctx.H = 5.0; 77 PetscCall(PetscOptionsScalar("-Inertia", "", "", ctx.H, &ctx.H, NULL)); 78 ctx.D = 5.0; 79 PetscCall(PetscOptionsScalar("-D", "", "", ctx.D, &ctx.D, NULL)); 80 ctx.E = 1.1378; 81 ctx.V = 1.0; 82 ctx.X = 0.545; 83 ctx.Pmax = ctx.E * ctx.V / ctx.X; 84 ctx.Pmax_ini = ctx.Pmax; 85 PetscCall(PetscOptionsScalar("-Pmax", "", "", ctx.Pmax, &ctx.Pmax, NULL)); 86 ctx.Pm = 1.06; 87 PetscCall(PetscOptionsScalar("-Pm", "", "", ctx.Pm, &ctx.Pm, NULL)); 88 ctx.tf = 0.1; 89 ctx.tcl = 0.2; 90 PetscCall(PetscOptionsReal("-tf", "Time to start fault", "", ctx.tf, &ctx.tf, NULL)); 91 PetscCall(PetscOptionsReal("-tcl", "Time to end fault", "", ctx.tcl, &ctx.tcl, NULL)); 92 printtofile = PETSC_FALSE; 93 PetscCall(PetscOptionsBool("-printtofile", "Print convergence results to file", "", printtofile, &printtofile, NULL)); 94 ctx.sa = SA_ADJ; 95 PetscCall(PetscOptionsEnum("-sa_method", "Sensitivity analysis method (adj or tlm)", "", SAMethods, (PetscEnum)ctx.sa, (PetscEnum *)&ctx.sa, NULL)); 96 } 97 PetscOptionsEnd(); 98 99 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 100 Create necessary matrix and vectors 101 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 102 PetscCall(MatCreate(PETSC_COMM_WORLD, &ctx.Jac)); 103 PetscCall(MatSetSizes(ctx.Jac, 2, 2, PETSC_DETERMINE, PETSC_DETERMINE)); 104 PetscCall(MatSetType(ctx.Jac, MATDENSE)); 105 PetscCall(MatSetFromOptions(ctx.Jac)); 106 PetscCall(MatSetUp(ctx.Jac)); 107 PetscCall(MatCreate(PETSC_COMM_WORLD, &ctx.Jacp)); 108 PetscCall(MatSetSizes(ctx.Jacp, PETSC_DECIDE, PETSC_DECIDE, 2, 1)); 109 PetscCall(MatSetFromOptions(ctx.Jacp)); 110 PetscCall(MatSetUp(ctx.Jacp)); 111 PetscCall(MatCreateVecs(ctx.Jac, &ctx.U, NULL)); 112 PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 1, 1, NULL, &ctx.DRDP)); 113 PetscCall(MatSetUp(ctx.DRDP)); 114 PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 2, 1, NULL, &ctx.DRDU)); 115 PetscCall(MatSetUp(ctx.DRDU)); 116 117 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 118 Create timestepping solver context 119 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 120 PetscCall(TSCreate(PETSC_COMM_WORLD, &ctx.ts)); 121 PetscCall(TSSetProblemType(ctx.ts, TS_NONLINEAR)); 122 PetscCall(TSSetType(ctx.ts, TSCN)); 123 PetscCall(TSSetRHSFunction(ctx.ts, NULL, (TSRHSFunction)RHSFunction, &ctx)); 124 PetscCall(TSSetRHSJacobian(ctx.ts, ctx.Jac, ctx.Jac, (TSRHSJacobian)RHSJacobian, &ctx)); 125 PetscCall(TSSetRHSJacobianP(ctx.ts, ctx.Jacp, RHSJacobianP, &ctx)); 126 127 if (ctx.sa == SA_ADJ) { 128 PetscCall(MatCreateVecs(ctx.Jac, &lambda[0], NULL)); 129 PetscCall(MatCreateVecs(ctx.Jacp, &mu[0], NULL)); 130 PetscCall(TSSetSaveTrajectory(ctx.ts)); 131 PetscCall(TSSetCostGradients(ctx.ts, 1, lambda, mu)); 132 PetscCall(TSCreateQuadratureTS(ctx.ts, PETSC_FALSE, &ctx.quadts)); 133 PetscCall(TSSetRHSFunction(ctx.quadts, NULL, (TSRHSFunction)CostIntegrand, &ctx)); 134 PetscCall(TSSetRHSJacobian(ctx.quadts, ctx.DRDU, ctx.DRDU, (TSRHSJacobian)DRDUJacobianTranspose, &ctx)); 135 PetscCall(TSSetRHSJacobianP(ctx.quadts, ctx.DRDP, DRDPJacobianTranspose, &ctx)); 136 } 137 if (ctx.sa == SA_TLM) { 138 PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 1, 1, NULL, &qgrad)); 139 PetscCall(MatCreateDense(PETSC_COMM_WORLD, PETSC_DECIDE, PETSC_DECIDE, 2, 1, NULL, &sp)); 140 PetscCall(TSForwardSetSensitivities(ctx.ts, 1, sp)); 141 PetscCall(TSCreateQuadratureTS(ctx.ts, PETSC_TRUE, &ctx.quadts)); 142 PetscCall(TSForwardSetSensitivities(ctx.quadts, 1, qgrad)); 143 PetscCall(TSSetRHSFunction(ctx.quadts, NULL, (TSRHSFunction)CostIntegrand, &ctx)); 144 PetscCall(TSSetRHSJacobian(ctx.quadts, ctx.DRDU, ctx.DRDU, (TSRHSJacobian)DRDUJacobianTranspose, &ctx)); 145 PetscCall(TSSetRHSJacobianP(ctx.quadts, ctx.DRDP, DRDPJacobianTranspose, &ctx)); 146 } 147 148 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 149 Set solver options 150 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 151 PetscCall(TSSetMaxTime(ctx.ts, 1.0)); 152 PetscCall(TSSetExactFinalTime(ctx.ts, TS_EXACTFINALTIME_MATCHSTEP)); 153 PetscCall(TSSetTimeStep(ctx.ts, 0.03125)); 154 PetscCall(TSSetFromOptions(ctx.ts)); 155 156 direction[0] = direction[1] = 1; 157 terminate[0] = terminate[1] = PETSC_FALSE; 158 PetscCall(TSSetEventHandler(ctx.ts, 2, direction, terminate, EventFunction, PostEventFunction, &ctx)); 159 160 /* Create TAO solver and set desired solution method */ 161 PetscCall(TaoCreate(PETSC_COMM_WORLD, &tao)); 162 PetscCall(TaoSetType(tao, TAOBLMVM)); 163 if (printtofile) PetscCall(TaoSetMonitor(tao, (PetscErrorCode(*)(Tao, void *))monitor, (void *)&ctx, PETSC_NULL)); 164 /* 165 Optimization starts 166 */ 167 /* Set initial solution guess */ 168 PetscCall(VecCreateSeq(PETSC_COMM_WORLD, 1, &p)); 169 PetscCall(VecGetArray(p, &x_ptr)); 170 x_ptr[0] = ctx.Pm; 171 PetscCall(VecRestoreArray(p, &x_ptr)); 172 173 PetscCall(TaoSetSolution(tao, p)); 174 /* Set routine for function and gradient evaluation */ 175 PetscCall(TaoSetObjectiveAndGradient(tao, NULL, FormFunctionGradient, (void *)&ctx)); 176 177 /* Set bounds for the optimization */ 178 PetscCall(VecDuplicate(p, &lowerb)); 179 PetscCall(VecDuplicate(p, &upperb)); 180 PetscCall(VecGetArray(lowerb, &x_ptr)); 181 x_ptr[0] = 0.; 182 PetscCall(VecRestoreArray(lowerb, &x_ptr)); 183 PetscCall(VecGetArray(upperb, &x_ptr)); 184 x_ptr[0] = 1.1; 185 PetscCall(VecRestoreArray(upperb, &x_ptr)); 186 PetscCall(TaoSetVariableBounds(tao, lowerb, upperb)); 187 188 /* Check for any TAO command line options */ 189 PetscCall(TaoSetFromOptions(tao)); 190 PetscCall(TaoGetKSP(tao, &ksp)); 191 if (ksp) { 192 PetscCall(KSPGetPC(ksp, &pc)); 193 PetscCall(PCSetType(pc, PCNONE)); 194 } 195 196 /* SOLVE THE APPLICATION */ 197 PetscCall(TaoSolve(tao)); 198 199 PetscCall(VecView(p, PETSC_VIEWER_STDOUT_WORLD)); 200 201 /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 202 Free work space. All PETSc objects should be destroyed when they are no longer needed. 203 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */ 204 PetscCall(MatDestroy(&ctx.Jac)); 205 PetscCall(MatDestroy(&ctx.Jacp)); 206 PetscCall(MatDestroy(&ctx.DRDU)); 207 PetscCall(MatDestroy(&ctx.DRDP)); 208 PetscCall(VecDestroy(&ctx.U)); 209 if (ctx.sa == SA_ADJ) { 210 PetscCall(VecDestroy(&lambda[0])); 211 PetscCall(VecDestroy(&mu[0])); 212 } 213 if (ctx.sa == SA_TLM) { 214 PetscCall(MatDestroy(&qgrad)); 215 PetscCall(MatDestroy(&sp)); 216 } 217 PetscCall(TSDestroy(&ctx.ts)); 218 PetscCall(VecDestroy(&p)); 219 PetscCall(VecDestroy(&lowerb)); 220 PetscCall(VecDestroy(&upperb)); 221 PetscCall(TaoDestroy(&tao)); 222 PetscCall(PetscFinalize()); 223 return 0; 224 } 225 226 /* ------------------------------------------------------------------ */ 227 /* 228 FormFunctionGradient - Evaluates the function and corresponding gradient. 229 230 Input Parameters: 231 tao - the Tao context 232 X - the input vector 233 ptr - optional user-defined context, as set by TaoSetObjectiveAndGradient() 234 235 Output Parameters: 236 f - the newly evaluated function 237 G - the newly evaluated gradient 238 */ 239 PetscErrorCode FormFunctionGradient(Tao tao, Vec P, PetscReal *f, Vec G, void *ctx0) 240 { 241 AppCtx *ctx = (AppCtx *)ctx0; 242 PetscInt nadj; 243 PetscReal ftime; 244 PetscInt steps; 245 PetscScalar *u; 246 PetscScalar *x_ptr, *y_ptr; 247 Vec q; 248 Mat qgrad; 249 250 PetscCall(VecGetArrayRead(P, (const PetscScalar **)&x_ptr)); 251 ctx->Pm = x_ptr[0]; 252 PetscCall(VecRestoreArrayRead(P, (const PetscScalar **)&x_ptr)); 253 254 /* reinitialize the solution vector */ 255 PetscCall(VecGetArray(ctx->U, &u)); 256 u[0] = PetscAsinScalar(ctx->Pm / ctx->Pmax); 257 u[1] = 1.0; 258 PetscCall(VecRestoreArray(ctx->U, &u)); 259 PetscCall(TSSetSolution(ctx->ts, ctx->U)); 260 261 /* reset time */ 262 PetscCall(TSSetTime(ctx->ts, 0.0)); 263 264 /* reset step counter, this is critical for adjoint solver */ 265 PetscCall(TSSetStepNumber(ctx->ts, 0)); 266 267 /* reset step size, the step size becomes negative after TSAdjointSolve */ 268 PetscCall(TSSetTimeStep(ctx->ts, 0.03125)); 269 270 /* reinitialize the integral value */ 271 PetscCall(TSGetQuadratureTS(ctx->ts, NULL, &ctx->quadts)); 272 PetscCall(TSGetSolution(ctx->quadts, &q)); 273 PetscCall(VecSet(q, 0.0)); 274 275 if (ctx->sa == SA_TLM) { /* reset the forward sensitivities */ 276 TS quadts; 277 Mat sp; 278 PetscScalar val[2]; 279 const PetscInt row[] = {0, 1}, col[] = {0}; 280 281 PetscCall(TSGetQuadratureTS(ctx->ts, NULL, &quadts)); 282 PetscCall(TSForwardGetSensitivities(quadts, NULL, &qgrad)); 283 PetscCall(MatZeroEntries(qgrad)); 284 PetscCall(TSForwardGetSensitivities(ctx->ts, NULL, &sp)); 285 val[0] = 1. / PetscSqrtScalar(1. - (ctx->Pm / ctx->Pmax) * (ctx->Pm / ctx->Pmax)) / ctx->Pmax; 286 val[1] = 0.0; 287 PetscCall(MatSetValues(sp, 2, row, 1, col, val, INSERT_VALUES)); 288 PetscCall(MatAssemblyBegin(sp, MAT_FINAL_ASSEMBLY)); 289 PetscCall(MatAssemblyEnd(sp, MAT_FINAL_ASSEMBLY)); 290 } 291 292 /* solve the ODE */ 293 PetscCall(TSSolve(ctx->ts, ctx->U)); 294 PetscCall(TSGetSolveTime(ctx->ts, &ftime)); 295 PetscCall(TSGetStepNumber(ctx->ts, &steps)); 296 297 if (ctx->sa == SA_ADJ) { 298 Vec *lambda, *mu; 299 /* reset the terminal condition for adjoint */ 300 PetscCall(TSGetCostGradients(ctx->ts, &nadj, &lambda, &mu)); 301 PetscCall(VecGetArray(lambda[0], &y_ptr)); 302 y_ptr[0] = 0.0; 303 y_ptr[1] = 0.0; 304 PetscCall(VecRestoreArray(lambda[0], &y_ptr)); 305 PetscCall(VecGetArray(mu[0], &x_ptr)); 306 x_ptr[0] = -1.0; 307 PetscCall(VecRestoreArray(mu[0], &x_ptr)); 308 309 /* solve the adjont */ 310 PetscCall(TSAdjointSolve(ctx->ts)); 311 312 PetscCall(ComputeSensiP(lambda[0], mu[0], ctx)); 313 PetscCall(VecCopy(mu[0], G)); 314 } 315 316 if (ctx->sa == SA_TLM) { 317 PetscCall(VecGetArray(G, &x_ptr)); 318 PetscCall(MatDenseGetArray(qgrad, &y_ptr)); 319 x_ptr[0] = y_ptr[0] - 1.; 320 PetscCall(MatDenseRestoreArray(qgrad, &y_ptr)); 321 PetscCall(VecRestoreArray(G, &x_ptr)); 322 } 323 324 PetscCall(TSGetSolution(ctx->quadts, &q)); 325 PetscCall(VecGetArray(q, &x_ptr)); 326 *f = -ctx->Pm + x_ptr[0]; 327 PetscCall(VecRestoreArray(q, &x_ptr)); 328 return 0; 329 } 330 331 /*TEST 332 333 build: 334 requires: !complex !single 335 336 test: 337 args: -viewer_binary_skip_info -ts_type cn -pc_type lu -tao_monitor 338 339 test: 340 suffix: 2 341 output_file: output/ex3opt_1.out 342 args: -sa_method tlm -ts_type cn -pc_type lu -tao_monitor 343 TEST*/ 344