FAST TRANSIENT STABILITY SOLUTIONS H.W. Dommel N.Sato Bonneville Power Administration Portland, Oregon ABSTRACT ‘Techniques are described for improving the speed of large transient stability studies without sacrificing accuracy. A fst iterative ‘method for solving the algebraic network equations, incuding the fect of generator saliency, is explained. A new technique for solving the differential equations with the implicit trapezoidal rule of integration is introduced. These two techniques can be combined nto one simultaneous solution, thereby eliminating. the problem of interface error between the differential and algebraic equation solutions of the traditional approach, INTRODUCTION ‘Transient stability studies now require considerable computer ‘me, especialy for solving large systems. This has motivated a satch {for faster solution techniques. Two possiblities for improving the speed of transient stability studies are: (a) Reduction of the total system to a smaller one, which could be solved fester, and (b) improvements in the numerical solution techniques. Reduction techniques produce equivalents which are normally ‘only approximations, Therefore, inspite of recent progres in finding ‘equivalents, the user must have a “feeling” for the problem when ‘equivalents are used. It should aio be noted that solutions based on syarity techniques become fear through reduction only t0 a certain point beeause of fillin caused by the reduction process! Judicious ‘we of equivalents can make stability solutions faster; however, this paper is concerned with the second posibity, improved numerical olution techniques, Improvements in numerical solution techniques can speed up the solution without placing any burden of judgment on the use, This paper describes experiments cated out at Bonneville Power Admini tration (BPA) over the lst few years with the objective of obtaining faster solutions for the stesdystate equations as well as for the differential equations which ars instability studies. Speeding up the solution of the steady-state equations has been accomplished by applying sparsity techniques and solution methods which require fewer iterations. Speeding up the solution of the ditteental ‘equations has been achieved with the implicit trapezoldal rule of integration, which has already been used siecetflly forthe solution of switching transients at BPA®, This method is numerically sable and accurate enough. It sbould be noted that explicit techniques, Including the Runge-Kutta methods, require higher accuracy primarily to insure numerical stability. As a consequence, sich techniques also require longer solution times, It e expected that implementation of these idess into BPA’s production program will make the solution bout five times faster. Brackets are used to indicate veetor and matrix. quantities Parentheses usually indicate a functional relationship of the valu for time A super dot is used to indicate derivatives with respect to time, UCTURE OF ‘STABILITY EQUATIONS ‘The initial conditions for a stabilty study are determined by a steady-state power flow solution. Thereafter, two sets of equations ‘must be solved simultaneously asa function of time, namely a sytem {ape 772.137 commended nd pond by he Powe Mesiog, New Yoik, WV slananry SoPebrony 4, 1972, Manos submited of steady-state equations (which describe the steady-state behavior of the network inching steedy-tate models of londs and the algebraic ‘equations of synchronous machines), a SalXps Ap or (ettx1y}=9 and a system of diferential equations (which describe the dynamic behavior of the machines and thele control citcults), TO ag Fm orfi] “Uh ty1.01 @) ix“ Jw) Sat ‘The structure of Bq. (1) will change at certain moments in time due to fault initialization, fault clearing, line switching, ete. Sch changes, which require resolutions without advancing the time, produce scontinultion in the value of the veetor(x]. No discontinuities ean appear in (y). ‘SLASSIFIOATION OF SOLUTION MErtiops On 8 digital computer, the values of (x] and [y] are computed at dlisrete points in time, thereby producing a sequence of snapshot pictures a intervals Qt. The sampling rte 1/A¢ determines the upper limit of the frequencies which will be adequately represented in (2) tnd (y]. As © consequence of the discretization, interpolation assumptions must normally be made on {x] to define the state between diserete points. In most cases linear interpolation is adequat ‘The methods for solving Eqs. (1) and (2) simultaneously fall roughly into three categories: (a) Alternating solution of Bas. (1) and (2), (b) elimination of [x], and (c) algbraiztion of Eq. (2) by implicit intogation. ‘The distinetion is not always clear. As an ‘example, the numerical solution of differential equations isin italt ‘based on an algebraization process. () Alternating Solution Method ‘This technique is probably used in most existing stability programs, The algorithm is roughly as follows: 1, Hetablah inital conditions by solving Eq. (1) at #0. 2 Predict how [x] behaves over the next time interval by using ‘20t0, first or second order extrapolation. 8. Solve Bq. (2) for [y(t)] by any numerical technique, using the polynomial obtained instep 2 asan interpolation formule 4. Solve Bq. (1) for [x(¢)] by any of the wellknown power flow solution techniques. st Engen Comes fhe IEEE ove ering Sty ox penn ote ERE We ‘a ptember 20,971; ade alae for ping Nove 19,1971 1643‘Use the computed value [x(t)] to correct the prediction over the interval from tAtto t Solve a. (2) for {y(¢)] by any numerical technique, using the Polynomial obtained in step 6 at an interpolation formela, Check differences of two su ditterences are step 4, 1ecesive solutions for [y(t]. It the negligible proceed to step 8. Otherwise go to ‘Advance by one time step and loop back to step 2 ‘Several variations are posible. Inthe version described above, the steady-state equations (1) must be tolved once, most of the time, since the differences in step 7 are normally sufficiently small on the fest pass. The prediction and correction process of the alternating tolution method should not be confused with predictorcomectot ‘methods for solving diterential equations if euch methods are sed, steps 9 and 6 would also involve another kind of predctioncorrection procea in ite, It is essential to include the check of step 7 to avoid so-called “Interface” errors. If this s not done, a careful assessment should be made about the acceptability of the accumulated erors. With the interface loop closed, the quality of the prediction (step 2) will determine how often the loop must be executed and, in effet, how long the study wil take, but it will not degrade the accuracy of tho solution because of corrections initiated instep 7, Thins not true for ‘open-loop solutions, where the quality of the predietion wil have decisive intuance on the accuracy. (0) Elimination of [x ‘This approach is used for small daturbance stability studles®, Por ‘mall vaviations {4.x}, (Ay) around an operating point, Eq (1) can mene (Bs) ~ (Ilo a withthe Jacobian mace 34) [35] in ‘lution point. Ex 2) cn ab Be Urn a [ai] = (alloy! + Bite] Both inet eqitions combined income ‘evaluated at the given around the opereting BE) [Belfton [0 wi liad) [al and after elimination of (x] by tiangulatzation, lax] ° [(Alreducet} [fava] [tad] [51 (Aleeduced “(4Y) @ ‘The eigenvalues of (Alraduced five the damping (reel part) and natural frequencies (imauinary part) of all posible modes of ‘osellations without having to make any saeumption about the specific (sell) disturbance, The system wil be stable a long as the real parts tall eigenvalues are negative (Cat method of Lyapunov). (6) Algebraization by Implicit Integration toxgl™* MOODY step solution of Ba, (2) est expresed In integral “tons toteanl+ Fy htxdader ry Tid imibadsdiemabaasdaii ail Implicit integration techniques use interpolation functions for the expressions under the integral. Interpolation means thatthe functions ‘must pass through the yet unknown points at time t, which most therfore be expresd 8 vartbles. In general, the solution may require iterations. If the diterential equations are linea, however, ‘then a direct solution becomes posible, Let the differentia equations (2) be linear ofthe form (TCA) +1810) (4a) Most ditterential equations ae linear in stability studies. There are some nonlinearities, however, but their inclusion poses no sedoce ‘problem, as shows later for the excitation system. Bq. (4a) can be rewritten asa step-by-step integration, teen] + (al) lars tay} oar Geo =teann «al tater At (4) ‘The simplest implicit intogration scheme isthe trapezoidal rule of integration.* It is based on the assumption that [x} and [3] very linearly over the interval from t-At tot, Then Eq. (48) becomes (eo) = toon + ca [Lota +t] +10) [ran + tx] Which ie simply sytem of tna lhe equations [ta-Seal]oor-# toner recon ‘The vector on the right-hand side is known from the values already ‘computed atthe preceding time step: totam tayeca01 fore taljowean 9 101 dvs it ony ma) en Be sed dizectiy with Bq. (1) as one complete system of steady.etate ‘equations (6a) NUMERICAL STABILITY OF THE [TRAPEZOIDAL RULE OF INTEGRATION ‘The trapezoidal rule of integration i numerically stable, ‘To ilustrate the problem of numerical stability, it will be assumed that « {ast oscillation in a control eieult produces “2ippes” of very smell ‘magnitudes which do not have any influence on the overall behavior of the cireuit. Such # mode of oscllation could be described by 3+ y «0, wh (0) ~0, 3(0)=J04 (6s) ‘The, exact solution for this rippe is y = 10“in(); ite amplitude of 10-4 wil be considered as very small by definition, Bq. (Ga) must be ewritien as a system of Ciest-order differential equations in order to spply any ofthe numerical solution techni bel ll ia, 10 ‘The exact step-by-step solution se 1 (@) 109 Es] : sone] wiaay aye 3] with the transition matrix efAJ4t, Application of the trapezoidal rule to Bq. (6b) gives 2 Ft 2 eT Gyeaey er] 2 a oe} 1S | ae 1-2 ean « ‘There are also higher-order generalizations of the trapezoidal rule4, 1644 Cava kk hn sed dT oy ahd ole Al dlics (ag Ad sd RARIt can be shown that y(t) + y9(t) = y2(eae) + y2(e04) tn Ex (7) for any choice oft. Tor, he nrges i started Ph te cae nl conn $40) 389) -108, the ation for y will always lie between -10™ and +10°, evn for stepewidthe which are much larger than one cycle of the osilation. In other words, the trapezoidal rule “cuts serose” oscillations which are very fast but of negligible amplitude, without any danger of numerical inatebility epic intogation techniques, which include Runge-Kutta math- ods, ae inherently unstable, They requize a tepwidth tailored tothe ihest frequency or smallest time constant’ (rule of thumb: AUS Thai) even though this mode may produce only negligible tiple, "Eh the overall ebavor determined by larg tine contenes asin socalled “tytn” Azaig te conetonl orthade Ranga mat a (0) iba to ound ora ‘tion mit nc oe ans ead Aan coy aca +E ato 2 rast Sait, igh tee gt on ales? loo raef [ove Ae® 1-082. 24)" | socal Peg hc «aon eal eae eames We eae al cmd scourate at fist than the trapezoidal rue, but vende to lose the amplitude later on (Fig. 1). «07! Fig. 1. Numezieal solution of J +y =0 (a) Braet, (b) Runge-Kutta, (6) trapezoidal rule ‘This is not serious since the ripple is assumed to be unimportant the fit place. If the step-width is increased, however, to At> cycles (At 70.4502 eycles), then the amplitude wil eventually grow to infinity. This is ilusrated in table T for At «1 eye, tineyees z 2 0.0088 n---= 0.018 0.028 Fig, 2. Tost example for trapezoidal rule. (@) Control ciruit.(b) Solution for sep-function input u= 02. ‘The plots of Fig. (2) were obtained with the trapezoidal rte using dittrentstop-widthe. The results are acceptable with a step-width as large as 0.012, which is 500 times lager than the approximate stapewidth 1 Tyg required forthe Runge-Kutta method. Block dlapams with sal tine constants ae often described by scotia otto v2 2 eka ® fs fn ig (20), 17 in vey sal, then the block ia busily 2 talipl iemeo yor Ky, a Ut explicit integration techniques are used in such cases, then it ‘becomes necessary to set small time constants equal to zero to avoid ‘merical instability. This is botheriome to the user because the problem manifests ital quite often only after the algorithm has {hiled to give solution. On the other hand, the programmer hesitates to build the decision into the program because It is difficult to draw the line between small and not so small ime constants forall posible situations. This problem does not aise withthe trapezoidal rule since ‘Table L. Numerical solution of Eq (6a) with At= 1 eyele ‘The fact that the trapezoidal rule of integration is stable oven it the step-width is much larger than the smallest time constant ia of bractical importance for stability studies. Fig. (2a) shows a hypothet- ‘eal control circut with 2 smallest time constant of 0.0001 s, and « ‘ep funetion input, which i regarded asthe most saver test. exact ° ° ° trapezoidal 0.58 104 0.94 104 0.96 10 0.63 104 0.06 104 0.53 104 rule Runge-Kutta 0.006 02 a8 5906800 2,600,000 ERY SMALL TIME CONSTANTS the limit case TO turns out to be identical to the algebraic equation am ZOD AL RULE (10). Applying the trapezoidal rle to Ba. (9) gives (+B) volt) Kyy(0) =~ Dygle.ae) + Kyy tse) snd for Tat ya(0)- Ky =-/o(t0)-Ky(40)) an 1665, 7 Weds @AL- | bicggecdia 40) PTT et te ter phate