Reliability Engineering and System St~fety 52 (1996) 65-75 ELSEVIER 0951-8320(95)00092-5 © 1996 Elsevier Science Limited Printed in Northern Ireland. All rights reserved 0951-8320]96/$15.IX) A Monte Carlo estimation of the marginal distributions in a problem of probabilistic dynamics P. E. Labeau* Service de Mdtrologie NuelOaire, Universit~ Libre de Bruxelles, Avenue F.D. Roosevelt, 50 1050 Bruxelles, Belgium (Received 2 May 1995; revised 3 July 1995; accepted 30 August 1995) Modelling the effect of the dynamic behaviour of a system on its PSA study leads, in a Markovian framework, to a development at first order of the Chapman-Kolmogorov equation, whose solutions are the probability densities of the problem. Because of its size, there is no hope of solving directly these equations in realistic circumstances. We present in this paper a biased simulation giving the marginals and compare different ways of speeding up the integration of the equations of the dynamics. © 1996 Elsevier Science Limited. 1 INTRODUCTION 2 PROBABILISTIC DYNAMICS Recently, researchers in reliability have stressed the need to incorporate in a usual safety study the dynamic behaviour of a system and its influence on the transitions likely to occur in an accidental transient. I-4 Indeed, each reachable c o m p o n e n t state corresponds to a given dynamics, and, in turn, the transitions between states depend on the transition rates which are influenced by the evolution of the physical variables describing the system. This feedback is quantified by the probability densities of being in a given state at a given time, with given values of the physical variables. Estimating these distributions is far from being an easy computational task, since we are faced with high-dimensional problems in realistic circumstances. We shall first remind readers how this dynamic concept has been modelled in the Markovian assumption, as well as mentioning some previous attempts to deal with it. We shall then present a general way of simulating the behaviour of the system. We will then show how to bias this algorithm and apply it on a benchmark. Different techniques for accelerating the integration of the equations of the dynamics will be given and c o m p a r e d on the same benchmark. A m o r e complicated application on a nuclear reactor transient will then be solved. Finally, we will give some concluding remarks. Let us consider how a system starting from state i behaves. The physical variables will evolve deterministically, according to the dynamics in state i d.f = f-(x- ) (1) from time t = 0 to t = tj, where a stochastic transition occurs. The system then changes its state to j, in which a similar deterministic evolution will take place until the next transition. To quantify this deterministic-stochastic process, we introduce lr(i, £, t), density probability that the system is in state i at time t with physical variables £, for given initial conditions. It has been shown 5 that rc(i,Y,t) obeys a d e v e l o p m e n t at first order of the C h a p m a n - K o l m o g o r o v equation, if we assume a Markovian f r a m e w o r k O~( i,£,t ) - + div,~f~(Y);c(i,£,t)) + Ai(Y)Ir(i,Y,t) Ot = ~p(j--->il£)rc(j,Y,t ) (2) j~i where A i ( £ ) a n d p ( j - - - ~ i l £ ) are the rate of transition out of state i and the transition rate from state j to state i, respectively. To solve system (2), the usual numerical schemes fail to work for realistic problems, since a great n u m b e r of states have to be considered and enough variables have to be taken in order to model the *Research Assistant (National Fund for Scientific Research, Belgium. 65 P. E. Labeau 66 system as neatly as possible. Therefore, alternate routes have been tested: simulation of functionals of the distributions by a Monte Carlo algorithm, 6'7 synthesis of the distributions from their moments, s9 ... We will now describe an algorithm to simulate the marginal distributions of such a problem. 3 ANALOGUE MARGINALS SIMULATION OF T H E Looking back to the description of the behaviour of the system given in Section 2, one can easily find how to simulate it. Since its evolution is m a d e of a succession of deterministic walks in one state and of stochastic transitions, a history consists, after having sampled the initial state, of repeating the two following steps until the mission time is reached: m 1. sample the time of the next transition from an exponential c.d.f, and integrate the equations of the dynamics up to this instant; 2. sample the new state from the probabilities p(i--->j]£) that the next state will be j if the ,~(~) system leaves state i in £. T o obtain the marginal distributions, one simply has to divide the range of each variable in n~ intervals, and to associate a counter c(I, m) to the m t h interval of t h e / t h variable, for each state and each time rk one is interested in. When the system reaches one rk, in state i at point £, the corresponding counters are increased by one. The distributions are then obtained by averaging the results on all the histories. One can notice that the total distributions could be obtained in the same way, but for a high n u m b e r of variables, the amount of counters would quickly b e c o m e too large. However, the simulation of bimarginal distributions does not lead to this problem and gives interesting correlations between important variables. If much information can be obtained in one simulation, this algorithm presents an important weakness: some states have a very low probability, and the tails of the distributions in these states correspond to very rare events. These are unlikely to occur in our analogue game, except if the n u m b e r of histories is considerably increased, leading to prohibitively large computational times. We must thus look for a way of biasing the simulation. 4 I M P R O V E M E N T OF T H E S I M U L A T I O N USING BIASING TECHNIQUES problem leads to m a n y useless histories, and therefore to an important waste of computation time. If the probabilistic laws can be modified in order to increase the occurence rate of the interesting events, the statistical accuracy of the results will be considerably improved for a given n u m b e r of histories. In order to ensure the unbiasedness of the estimation, correcting factors, called statistical weights and defined as the ratio of the analogue probability to the modified one, have to be used: contributions to the score have to be multiplied by the current weight of the history. ~l~z Let us go back to the simulation of the marginals. It is well-known that the m o r e precise the purpose of the Monte Carlo estimation is, the more effective the calculation is. We restrict ourselves to the computation of all the marginals, in one state i* at one time T. Therefore, the biased scheme must aim at allowing only transitions to states from which state i* can be reached in a finite n u m b e r of steps and forcing the system to reach state i* before T. In the following, we develop this general idea, assuming the system is non-repairable. For the sake of simplicity, the transition rates are taken as constant, even though the same treatment is still applicable in the general case. From the state graph, we can find from which states we can start a history that could reach state i*. Let I be the set of these states, and w0 = ~ic~c(i, 0), where 7r(i, 0) is the initial probability of being in state i. If we sample the initial state from set I without modifying the initial distribution, we introduce a weight w0. Again from the state graph, we can determine for state i of I which of its first successors are part of a path leading to i*. Let succ(i) be the set of these states. When a transition out of state i occurs, we force the new state to be part of succ(i). If we do not change the transition rates and if A* = ~j~,,,,~i~p(i--~ Ai j), another weight w,(i)= AT has to be taken into account. This step is repeated for all transitions until state i* is reached. Let us illustrate this scheme on the simple state graph given in Fig. 1. If we are interested in the distributions in state 4, it is useless to start histories from states 3 and 5. Therefore 1 : {1,2,4}. It is also easily seen that succ(1) = {2, 4} and succ(2) - {4}. BY ~ J 3 If one is interested in the simulation of very rare events, sampling from the probabilistic laws of the Fig. 1. A M o n t e Carlo estimation o f the marginal distributions in a p r o b l e m ofprobabilistic dynamics It has to be noticed that, even with this selection of the useful paths, the drawback of the analogue simulation can still be encountered. Indeed, the probabilities to sample these paths can have very different orders of magnitude. In this circumstance, some parts of the distributions could not be accurate; biasing the transition rates to the different authorized states, for a given transition rate out of the current state, would solve this problem, a second weight having to be introduced. U p to now, we have selected the histories leading to state i*. But we still have to reach it before time T. To do so, we force all the transitions to happen before T. Let j ( k ) be the kth state of a path leading to i* and tick) the time spent in it. In the Markovian hypothesis, the analog c.d.f, of the time of the transition out of state j ( k ) is exponential: Fk(tj(k)) = 1 - - e ~,~,,k~. If k--I t = 2~ tj(t) is the time elapsed from the beginning /--I of the history to the transition to state j ( k ) , then the c.d.f, forcing the next transition to take place before T is F~ti(k)) = I 1 - e Atl'~)li(k) - - e a,k,(T-~,) if tj(k) < T - t otherwise. 67 Since f b,~7r(i*, xt, T)dxt = tr(i*, T), we can deduce: ni c ( l , m ) zc(i*,x,,,,,T) ~ - b ~ - a~ nh (6) and we have thus achieved our goal. T o build this algorithm, we made the assumption that the system was not repairable. As a consequence, we supposed that when state i* was reached, the system had to evolve in it until time T to contribute to the score. It may not be true in the repairable case, when a state can be reached after an infinite n u m b e r of transitions and after having already been in this state previously. However, a score due to m a n y transitions will be very small, and could even become smaller than the statistical error on the estimation. Therefore, we can adopt the following rule-of-thumb. Let K ....... be the maximal n u m b e r of transitions considered when reparation is active. If state i* is reached in less than K, ...... steps, then the probability of staying in i* till time T is calculated and a r a n d o m n u m b e r decides whether or not the system will keep evolving in this state until time T. If K ....... transitions have already taken place, the history is brought to state i* in as few transitions as possible. (3) The weight associated to this bias is w f ( j ( k ) ) = 1-e Ai,~,(7 t). 5 APPLICATION So state i* is reached before T. But the system must not leave it before T. We thus want the next transition to occur after a time t + ti. > T, i.e., t* corresponds to the conditional c.d.f.: Consider a two-dimensional three-state benchmark, whose dynamics are: dx dtt = a i x fea,(T-t)(1-e [0 otherwise E.(t,.) a,h.)ifti.> T-t , x(O) = xo,y(O) (4) and to the new weight w * = e -~'~r-'). The total weight of an history reaching i* in K transitions is: dy = y , , , i = 1...3 (7) = bix + ci y K W = wow* ~ w,(j(k))wr(j(k)). (5) k-I The evolution of the physical variables is calculated throughout the history, and we only have to determine at its end which counters c(l, m ) have to be increased by W. If nh histories are played, and if Wh is the final weight of the hth history, then the m e a n weight (W) = 1 ~ Its state graph is given in Fig. 2. The distributions in state 39 are given in the Appendix. With the data of Table 1, we obtain: 3 (x-5/3e'/2 tc(3,x,t) = 1l/Seg'/m) x 3 (x 5/3e'/2 - x-7/5e t/l°) e 3t/4 ~ x ~ e t (8) e t ~ x ~ e 3t/2 Wh is an unbiased estimate of ]r(i*,T), the nh h- I probability of being in state i* at time T. Let [a~, b~] be the range of the lth variable and xt,, = a~ + m - --, 0 otherwise m = 1...n~ the centres of the Y/i n~ intervals to which are associated the counters c ( l , m ) . These values are proportional to ~(i*, x~,,, T), the value in x~,,, of the lth marginal distribution in state i* at time T. 1 3 Fig. 2. 3 68 P. E. Labeau 0.01 4 Table 1 Monte Codo eimulotion '~' " " n 0.012 a, = 1.5 a~= 1.0 a3 = 0.75 At = 0 . 5 b] = 1.0 C1 0.5 b2 = 0.75 C2 = 0.25 b3 = 0.0 C~= 0.5 A2 = 0.3 X~,= y(, = 1.0 0.01 0 = 0.008 cx and 3 ( y N2etl4 -- ~15e3'11°) ]'C(3,y,t) z~N(y-3/2et/4 1,0 y e t12 <-- ct~y~ 7/5ct/lO) 0.004 <-- e ~ Y 3t/2 0.006 0.002 ( 9 ) 0.000 otherwise 1.04 if ~ ( i , 0 ) = 6i~. T h e s e distributions are n o r m a l i z e d to n(3,t), the probability of being in state 3 at time t. Figures 3 - 6 c o m p a r e the theoretical and simulated distributions, for T = 0.1 and T =2. 5 × 104 histories have b e e n played, and the ranges of both variables were divided into 40 intervals. F o r a given n u m b e r of histories, there is a balance to find b e t w e e n the n u m b e r of intervals and the statistical accuracy of the results, but this choice has no influence on the c o m p u t a t i o n time. In o r d e r to estimate the quality of the biased game, an a n a l o g u e simulation was p e r f o r m e d for T = 0 . 1 . T h e c o m p u t a t i o n time n e e d e d for this case was m u c h smaller, since a history is finished without any score w h e n a transition time greater than T is sampled. T h e results are completely different to the theoretical solutions, since very few histories are useful. A second a n a l o g u e simulation with 1.5 × 106 histories has been realized, leading to a similar c o m p u t a t i o n time as for the biased m e t h o d with 5 × 104 histories. O n c e again, the distributions o b t a i n e d in this application are much less accurate. Figures 7 and 8 c o m p a r e the results of these different tests, while Figs 9 and 10 give the s t a n d a r d deviations of the estimates. The superiority of the n o n a n a l o g u e g a m e is obvious. 6 ACCELERATION OF T H E I N T E G R A T I O N For each history, the evolution of the physical variables has to be c o m p u t e d . Obviously, any effort to speed up the integration of the equations of the 1.06 1.08 1,10 1.12 1,16 1.18 Fig. 4. Marginal distribution of the y variable in state 3 t 0.1: integration with RK4. d y n a m i c s will positively affect the c o m p u t a t i o n time. 13 W e have tested different ways of achieving this p u r p o s e and r e p o r t the results in the following. T h e first idea is very simple. Since the range of each variable is divided into intervals, the accuracy on the solution of the equations of the dynamics must not be very high. So one can think of replacing the usual R K 4 scheme 7 by a R K 2 which, obtaining the solution in two steps rather than four, will be twice as fast. Notice that, even for the estimation of o t h e r reliability characteristics, such as m e a n time to failure, generalized nonreliability, etc, there is no n e e d of a very precise integration scheme, and using a m e t h o d of o r d e r 2 in place of a m e t h o d of o r d e r 4 is not a problem. T h e second trick is based on the following constatation. C o n s i d e r a batch of rn histories starting from the same state, with the same initial values of the variables. Let tl < r e < . . . <t,,, be the first transition times of each of these runs. It m e a n s that the integration of the equations giving the variables b e t w e e n 0 and t, will be p e r f o r m e d m times, the one b e t w e e n tL and t2 ( m - 1 ) times, and so on. This repetition of calculations results in a possibly i m p o r t a n t waste of c o m p u t a t i o n time. This weakness can be o v e r c o m e by realizing a preliminary stage 0.04 0.020 0.016 1.14 Y -- - - Monte Cork) i;;mulot;on , --" At~olyti¢ol solut;on --, 0.016 Monte Cork) slmulot;on Solut;on Anolyt;¢OI 0.03 0.014 0.012 x ,4 "~ • ". 0.010 FI 0.02 o. 0.008 0.006 0.01 0.004 0.002 0.000 o.oo i 1.07 , i 1.08 , | 1.09 . • 1.10 . , 1.11 - , 1.12 - , 1.13 - • 1,14 - , 1.15 - , 1.16 - , 1.17 * | i 4 6 8 , ! 10 . . . 12 . . 14 . 16 . . 18 _ . 20 . I 22 x X Fig. 3. Marginal distribution of the x variable in state 3 t = 0. l: integration with RK4. Fig. 5. Marginal distribution of the x variable in state 3 t = 2: integration with RK4. A Monte Carlo estimation o f the marginal distributions in a problem ofprobabilistic dynamics 0.030 -.- • Monte Carlo simulotion Anolytl¢ OI solution 69 0.04 ~ 0.03 :: o n o l o g u e : .50000 h i s t o r i e s . . . . onologue : 1~ ) 0 0 0 0 histories - - - biosed : 5 0 0 0 0 h i s t o r i e s : 0.02 : theoretical .... 0.025 0.020 -.c: >" 0.01 5 >'. w3 ~• Q. :: . 0.010 o.oo°"°1 0.005 0.000 : ~ II J, .3 . , 5 _ , . | 7 . g , . 11 i . , 13 . 15 o 17 . i 19 . l 1.04 ~:: i: i 21 : , • 1.06 C)8 . :: ' I, I , I 1.10 , 1,12 I , 1.14 i , 1,16 i 1.18 Y Y Fig. 6. Marginal distribution of the y variable in state 3 t = 2: integration with RK4. before the simulation: the evolution of the variables in all states of I whose initial probabilities are non zero is calculated throughout the mission time, and the results are m e m o r i z e d at some predefined times. Then, when the first transition time r is sampled in a new history, one can easily find in which time interval it lies, and only compute the values of the variables from the beginning of this interval to r, avoiding useless calculations. To p e r f o r m this memorization, we assume that the initial values of the variables are perfectly known. D o e s it correspond to a realistic circumstance? An important application of probabilistic dynamics is to quantify the behaviour of a system, when an 'accident' at t = 0 induces a sequence of possibly dangerous events. If the regulation of the system is assumed reliable before the accident, the dispersion about the initial values of the variables can be neglected. Memorizing the evolution of the variables in each interesting state is therefore practically useful. The third way of speeding up the integration is to use an adaptable time step. Although elaborated time steps policies can be usedJ 4 we have a simple heuristic one: every n, time steps, E, the relative variation of 0.04 Fig. 8. Comparison of the results obtained by an analogue and a biased simulation: marginal distribution of the y variable. £(t), is calculated. If E < E,,,i,,, the time step is doubled, while it is divided by 2 if E > E,...... . These methods have been applied to the benchmark presented in the previous section, on a IBM RISC 6000 workstation. Table 2 gives the different c o m p u t a t i o n times, using the R K 4 and the R K 2 m e t h o d s , with or without m e m o r i z a t i o n . Since there are no observable differences in the results obtained for the distributions, the reduction of the computation times is a direct benefit. If the R K 2 scheme logically results in a twice as fast calculation, one can see the obvious advantage of memorizing the dynamic evolution of the variables. O f course, if this m e t h o d looks quite attractive for a system with a reduced n u m b e r of transitions, the gain could b e c o m e m u c h smaller for larger systems. Anyway, we believe it is still w o r t h using it in such circumstances. If there are m a n y possible initial states, one can limit the application of the m e m o r i z a t i o n to those states whose initial probability is larger than a predefined threshold. T h e n , the same p r o b l e m has b e e n run with a time step h constrained to evolve in the interval [h°/4, 4h°], where h ° is the initial value of h. Table 3 contains the theoretlco~ onologue : 5 0 0 0 0 histories - - onologue : 1 5 0 0 0 0 0 h i s t o r i e s - - - blo~KI : 5 0 0 0 0 histories .... . 1 onologue ~0 J ~ .... : 50000 h;stories onologue : 1500000 histories 0.03 . = 0.02 : 11" .~ . : .~ 0.7 "~ 0.5 " 0.6 o O.Ol i ~ : :t . . '- 0.00 1.07 1.08 1.09 1.10 1.11 1.12 1,13 1.14 1.15 1,16 1.17 X Fig. 7. Comparison of the results obtained by an analogue and a biased simulation: marginal distribution of the x variable. 0.4 0.3 N 0.2 ii 0.1 j ~ 0.0 ~ | [" . . . . . . . . . . . . . . 1.07 1.08 1.0g 1.10 1.11 1.12 1.13 1.14 1.15 1.16 1.17 M Fig. 9. Comparison of the standard deviation obtained by an analogue and a biased simulation: x variable. P. E. Labeau 70 onologue : 5 0 0 0 0 histories onologue : 1 5 0 0 0 0 0 histories bioesd : 5 0 0 0 0 histories 1.0 0.9 0.8 g 0.7 i~ 0.6 t V :, 0.5 0.4 0.2 0.1 0.0 Table 3 T Integration scheme 0.1 0.1 2.0 2.0 RK4 RK2 RK4 RK2 e,,,,, - 0 . 2 5 ( % ) : histories I 1 .o4 1.1~ , .',0 1 .o8 ,.;2 ,.;, ,.',6 No Memorization memorization (nbt = 50) 13"3 7"3 3'44"4 1 '49"5 e ...... = 2 . 5 ( % ) : 7"9 4"8 2' 12"5 l '07"4 h"=10 s: 5×104 ,.;8 Fig. 10. Comparison of the standard deviation obtained by an analogue and a biased simulation: y variable. computation times for some of the cases mentioned in Table 2. Comparing the initial method (RK4, no memorization, constant time step) to the last one (RK2, memorization, adaptable time step), it can be seen that one order of magnitude has been gained in the computation time. Finally, we have checked how the choice of p a r a m e t e r s e,,,i,, and e,...... could affect the results. Indeed, a too fast integration would lead to unprecise results, while a too small time step would needlessly lengthen the computation time. Several attempts have been made with different values of e,,,,, and E,...... ; Table 4 gathers the corresponding results. It has to be noticed that these acceleration techniques do not influence the variance, but only the time per history. Therefore, the efficiency of the simulation, defined as usual by (o-2T) ~, where ~r2 is the variance of the estimate and T the computation time per history, is increased in the same proportion as the integration is speeded up. simplified and slightly modified version of the problem of a transient in the fast reactor Europa. 7~~ We study the effects of a reactivity slope on the reactor. Five physical variables describe the dynamics: the power P(t), the sodium t e m p e r a t u r e T(t), the fuel t e m p e r a t u r e T,(t), the.flow of coolant G(t) and the angular speed of the primary p u m p w(t). A supplementary variable y(t) is used to model the progressive introduction of antireactivity due to the scram, which is triggered on by the crossing of either a power threshold or a coolant t e m p e r a t u r e threshold. Three failures are considered: the p u m p m o t o r stops, the coolant t e m p e r a t u r e sensor fails, the power sensor fails. The system has thus 8 different states, denoted (i,j,k): i = 1 if the p u m p m o t o r works, and 0 else, similarly, j and k give the state of the coolant temperature sensor and the power sensor respectively. The state graph is given in Fig. 11. We use a point kinetics model without delayed neutrons for the evolution of the power: dP(t) _ p(t) - - dt A (10) P(t). 7 A SIMPLIFIED M O D E L OF A FAST R E A C T O R T R A N S I E N T I N D U C E D BY A RAMP T h r e e p h e n o m e n a are opposed to the increase of reactivity: the control rods when safety thresholds are reached, the negative t e m p e r a t u r e coefficient of the m o d e r a t o r and the D o p p l e r effect. Therefore, the reactivity has the following form In this section, we apply our scheme to model a p(t) = a t - A , , , y ( t ) - The meaning of the different p a r a m e t e r s is given in Table 5. Table 2 T Integration scheme computation time No memorization 0.1 0.1 2.0 2.0 RK4 RK2 RK4 RK2 a s ( T ( t ) - To)-aD(Tc(t)--T~,,). Table 4 memorization nbt-10 nbt=20 nbt-50 e,,,,,(%) e,,,,,~(% ) computation time 25"4 12"8 8'55"5 4'11"0 24"6 12"4 8'32"3 3'58"5 24"1 12"1 8'19"4 3'52"8 0.25 0.25 0.1 0.2 2.5 1.0 1.0 1.0 1'07"4 1'06"5 2' 10"3 1'06"7 44"3 21"2 14'01"5 6'31"0 nbt = n u m b e r of memorization times: 5,10 4 histories h " - 10 3: T - 2.0: nbt = 50:5 × 104 histories A Monte Carlo estimation o f the marginal distributions in a problem ofprobabilistic dynamics .... //~(l ' (L0a) J,l) ' F r o m the e n e r g y conservation law, we obtain an evolution e q u a t i o n for the s o d i u m t e m p e r a t u r e : ~ G / (l,0,0) x, d(T(t)- t dt " Xp ~ 71 Ar ] T,,) P(t) - t 2 A G ( t ) r CR { 1 [ 1 dG(t) G-(t) dtt 1] +-r 1: working × (T(t) - L)}H(Ti~,b -- T(t)) (11) (0.0A) Stat, e (i.3.14) i j k slate of t h e where A is the section of the channel, T,, the coolant r e a c t o r inlet t e m p e r a t u r e , Te~ the sodium boiling point, CR and CF the specific heats of the coolant and the fuel, respectively, and r a time constant related to pump state of the coolant Lemperature sensor 'dale of I h e power sensor CF. Fig. 11. Table 5 ol S coolant reactor inlet temperature sodium boiling temperature sodium specific heat initial sodium temperature temperature threshold sodium temperature coefficient CF fuel specific heat R thermal resistance by length unit time constant of the fuel heating initial fuel temperature Doppler coefficient L T:.\ CR T(O) Tma~ "t" Too ol D P(0) 653(K) 1156(K) 1629 - 0.8329T(J/kgK) 873(K) 923(K) 2 × 1 0 ~(K ') (12.54 + 0.017T,. - 0.117.10 4T2 +0.307 × lO-ST~)15.4894(J/kgK) 0.62 × 10 ~(K/W) 0.9031 × 10 ~'(s) 1893(K) 1.3 × 10 5(K ') 2000(MW) 2100(MW) 3.98 X 10 2(s) 0.09 2 X 10 2(s ~) t* initial power power threshold neutron production lifetime Antireactivity of the scram reactivity slope time constant of the scram c(o) A initial flow section of the channel 4000(kg/m 2 s) 1.26(m 2) first time constant of the pump 201.94(s ') second time constant of the pump 201.99(s ') Pmar A A~,r O/ 1 2(s) lo 1 t,, m~,gC2 L m,, L g gOo K CM I ~/mmp AT Ap 5.56 × 10 6(kgm 2) sodium specific mass length of the circuit acceleration of gravity initial angular speed friction constant pump torque moment of inertia 6000(s ') 10(kgm 2 s i) 60000(Nm) 10(kgm 2) failure rate of the pump failure rate of the temperature sensor failure rate of the power sensor 10-~(s ') 4 x 1 0 ~(s ') 4 x 10 6(s i) P. E. Labeau 72 The evolution law for the fuel t e m p e r a t u r e is obtained by averaging on the fuel section the heat transfer equation, which leads to: dT~.(t)_ dt RP(t) r T~.(t) - T,, (12) r where R is the thermal resistance of the fuel by length unit. The flow of coolant is increased because of the manometric height of the p u m p H , and decreased because of friction. If we assume H = C I G 2 + C2w 2, we get: d G(t) _ [(G(t))2 1 m,,gC2 (w(t)2 _ coO) + LG(O) 1 t e m p e r a t u r e at short times, for which the situation is more complicated, the different marginals show two peaks. They represent the two situations the system can be in: either the scram was triggered on before the power sensor failed or the failure of this c o m p o n e n t occured first. Since the control rods would normally (i.e. without failures) fall into the core at t = 0.45s, the peak corresponding to the evolution of the variables without scram logically appears to be predominant at t = 0.5s, but becomes less and less important at larger times. The peaks observed are broadened because of the failure of the p u m p motor, which influences slightly the values of the variables at short times. Table 6 gives the computation times for the different trials. Let us finally notice that this application could not have been run with the synthesis method, s'9 since in its current development it asks for dynamics which are quadratic forms in the physical variables. (13) 8 CONCLUSIONS where m~ is the sodium specific mass, g the gravity acceleration, L the length of the circuit, and to and t~, two time constants. The variation of the angular speed of the p u m p comes directly from the equation of the kinetic m o m e n t . If I is the m o m e n t of inertia of the pump, CM its torque, and K a friction constant, we have /do(t) dt = CM61' - K w ( t ) (14) with 6/, = 1 if the p u m p m o t o r works, and 0 else. Finally, we assume an exponential law for the introduction of the antireactivity of the control rods, with a time constant equal to their fall time in the core dy(t) dt- ? (1 - y ( t ) ) (15) where c~,,, switches from 0 to 1 when the scram is triggered on. The difference between the states appear in 6p and in the time when a~r becomes 1. The numerical values of the coefficients are gathered in Table 5, while Figs 12, 13 and 14 give the marginal distributions of the three first variables at different times in state (0, 1,0), i.e., when the p u m p m o t o r and the power sensor have failed. Let us remind ourselves that these distributions are normalized to the probability of being in the corresponding state at the time of study. The interpretation of the results is quite easily done. Except perhaps for the distribution of the sodium Simulation seems to be the only practical way of dealing with the size of realistic problems of probabilistic dynamics. Indeed, too many assumptions have to be done in semi-analytical models, narrowing the range of applications likely to be treated by this approach. Moreover, Monte Carlo simulation can deal with a non Markovian framework, when the assumptions made in probabilistic dynamics are too restrictive. To estimate the marginal distributions, an analogue Monte Carlo game appears to be efficient. Therefore, we have developed a biased scheme to obtain all the marginals in one state i* at one time T. The histories played follow paths of the state graph leading to this state, and all the transitions are forced to occur before T. Every history thus contributes to the score. The scheme was first worked out for a non repairable system, but a way to adapt it to the repairable case was suggested, yet not fully demonstrated. The p r o g r a m was tested on a simple two-dimensional three-state benchmark, and the agreement between theoretical and numerical results was satisfactory, even though the m e a n n u m b e r of histories by estimation point was quite low. Then, different ways of accelerating the integration of the equations of the dynamics were introduced. They are based on two constatations: the variables need not to be known with a very high accuracy, and many identical calculations are needlessly repeated in a simulation. The implementation of these tricks in the algorithm allowed us to m a k e the simulation up to ten times faster, when applied to the benchmark. We believe the acceleration techniques can also be used in the estimation of more important characteristics (e.g., Morginol distribution of In(P/Po) t-0.5 s Morginol d i s t r i b u t i o n o f I n ( P / P o ) t-2s 9.OOE- 8 4.0o(-9 B.OOE-B 7.00E-B 3.ooE-9 6.00£ -8 5.00( - 8 2.ooE-9 • 4.00E-B ~.~-8 2.~-8 0.~0 ~ ! . ~ - 8 , 1.00(-9 l 0.(x~0 • i o.06o2 o.0607 , i , 0.0612 i , 0.0617 i , i , i o.0622 o.o627 0 . ~ 3 2 -0.2'-0.1 In(p/po) MorQinol d i s t r i b u t i o n o f I n ( P / P o ) t=l s . , . , _ , _ , . , _ , . , . A , ' 010 ' 0.1 0.2 O~ 0.4 0.5 0,6 0.7 0,8 0.9 In(P/Po) . , . 1.0 Morginol d i s t r i b u t i o n o f I n ( P / P o ) t:3s 6.00£-9 5.00E-9 5.00E-9 4.00( -9 4.00E-9 3.00E-9 3.00E-9 2,00( -9 2.00E-9 l 00( -9 1.00E-9 0.00[0 O.OOEO , i -0.5-0.3-0.1 . . . . . . . . . . . . . . . . . ! 0.08 0.10 0.12 0.14 0.16 0.18 0.20 0.22 0.24 0.26 In(P/Po) , I , . . . . 0.1 0.3 , 0.5 - 0.7 , 0.9 . . . . . . . . 1.1 1.3 1.5 1.7 ,.(P/Po) Fig. 12. Morginol distribution of (T/To) t-O.Ss Morginol d i s t r i b u t i o n o f ( T / T o ) t-2s 1.80E-8 9.00E-9 1.60£-8 B.00£-9 1.40£-8 7.00E-9 1.20E-8 6.00F-9 1.00E-8 5.00E-9 8.00E-9 4.00E-9 6.00E-g 3.00£-9 4.00£-9 2.00E-9 2.00(-9 1.00E-9 O.OOEO 1.0005 0.00£0 1.0010 1.0015 1.0020 1.0025 1.0030 1.0035 1.00 !.01 (T/To) Morcjinol d i s t r i b u t i o n o f ( T / T o ) t:l 1.02 !.03 1.04 1.05 1.08 1.07 1.08 1.09 (T/To) Morglnol d i s t r i b u t i o n o f ( T / T o ) t-3s S 140E8 1 .I OE-B I .OOE-8 g.OOE-9 B.00£-9 7.00E-9 6.00E-g 5.00E-9 4.00E-9 5.001[-9 1.20£-II 1.00£-8 ILIXX-9 6.00E-9 4,00[-g 2.00E-9 I .OOE-9 O.OOEO 2.lX~-9 O.OO(O i 1.003 1.005 1.007 1.001) 1.011 1,01S 1.015 (T/PI'O) Fig. 13. 0.95 1.00 i i 1.05 i i 1.10 (T/To) i 1. ! 5 i i i 1.20 1.25 P. E. Labeau 74 Morginol distribution of (Tc/Tco) t-2 s Morginol distribution of ( T c / T c o) t-0.5 s 5.00(-6 3.00( - 8 4.001[-6 2.50E-8 2.00E-8 3.00(-6 1.50E-8 2.00£-6 1.00E-8 1.00(-6 5.00E-9 0.00(0 O.OOEO 18 . . .20. . .22 24 '''8 2 . 28 . . . 30 . , . , . , . , . , ." j . • . , A . , 1.00 1.02 1.04 1,06 1.06 1.10 1.12 1.14 1.16 (Tc/rco) Morginol distribution of ( T c / T c o ) . .34. . 36 A ((Tc/Tco)-1.002)-I e+6 Morginol distribution of ( T c / T c o ) 1-3s t-1 S 1.80E-8 g.OO(-8 1.60E-B B.O0( - 8 1.40E-8 7.00(-8 6.00(-8 1.20E-8 5.00( - 8 1.00(-8 4.00( - 8 8.00E-g 3.00(-8 6.00E-9 2.00(-0 4.00(-9 1.00( - 8 2.00E-9 O.OOEO O.OOEO ' " ' - ' - ' - I , I i I | I I I 1.0o9 I.OLO 1 .ol I 1 .o12 1.ol 3 1.o14 t .ol s 1.016 1 .ol 7 (re/To o ) 0.9 1.0 1.1 i I i 1.2 I 1.3 i I 1.4 i I 1.5 (T¢/Tco) Fig. 14. Table 6 t computation time 0.5 1.0 2.0 3.0 3'17" 6'22" 12'43" 19'00" h" = 2.5.10 3: nbt = 100:5 × 104 histories generalized unreliability, m e a n time to failure...). T h e y constitute a simple alternative to the use of neural networks, 6''~ for achieving this integration. Their efficiency, as well as the parallelization of the simulation, is one of the keys to d y n a m i c P S A for large industrial systems. REFERENCES 1. Siu, N., Risk assessment for dynamic systems: an overview. Reliab. Engng Syst. Safety. 4 3 (1994) 43-73. 2. Devooght, J. & Smidts, C., Probabilistic dynamics: the mathematical and computing problems ahead. In Reliability and safety assessment for dynamic process systems, N A T O ASI Series, Series F, vo1.120, Springer-Verlag, Berlin, 1994. 3. Marseguerra, M. & Zio, E., Towards dynamic PSA via Monte Carlo methods. Proc. of Esrel'93, Elsevier, Amsterdam, 1993 415-427. 4. Devooght, J., lzquierdo, J. M. & Mel~ndez, E., Relationships between probabilistic dynamics, dynamic event trees and classical event trees. Reliab. Engng Syst. Safety (1995) (in press). 5. Devooght, J. & Smidts, C., Probabilistic reactor dynamics, l.The theory of continuous event trees. Nucl. Sci. Engng, 111 (1992) 229-240. 6. Marseguerra, M. & Zio, E., Improving the efficiency of Monte Carlo methods in PSA by using neural networks. Proc. of PSAM 11 (1994) 025.1-025.8. 7. Smidts, C. & Devooght, J., Probabilistic reactor dynamics, II.A Monte Carlo study of a fast reactor transient. Nucl. Sci. Engng, 111 (1992) 241-255. 8. Devooght, J. & Labeau, P. E., Moments of the distributions in probabilistic dynamics. Ann. Nacl. Engng. 22 (1995) 97-108. 9. Labeau, P. E. & Devooght, J., Synthesis of multivariate distributions from their moments for probabilistic dynamics. Ann. Nucl. Engng, 22 (1995) 109-124. 10. Lewis, E. E. & Bohm, F., Monte Carlo simulation of Markov unreliability models. Nucl. Engng Des. 77 (1984) 49-62. 11. Kalos, M, H. & Whitlock, P. A., Monte Carlo methods, Volume I: Basics, Wiley-Interscience, New York, 1986. 12. Spanier, J. & Gelbard, E. M., Monte Carlo principles and neutron transport problems, Addison-Wesley Publishing Company, Reading, 1969. 13. Marseguerra, M. & Zio, E., Monte Carlo approach to PSA for dynamic process systems, accepted for publication in Reliab. Engng Syst. Safety. A Monte Carlo estimation of the marginal distributions in a problem 14.Press, W. H., Flannery, B. P., Teukolsky, S. A. & W. T., Numerical recipes. The art of scientific computing (Fortran version), Cambridge University Press, Cambridge, 1990, chap.16. 1.5. Smidts, C., Simulation des sequences industrielles accidentelles prenant en compte le facteur humain. Application au domaine des centrales nucleaires, Ph.D thesis, Universite Libre de Bruxelles, 1991, chap. IX. 16.Marseguerra, M., Nutini, M. & Zio, E., Approximate physical modelling in dynamic PSA using artificial neural networks. Reliab. Engng Syst. Safety 45 (1994) In-< Vetterling, 47-56. of probabilistic dynamics H min ( 0 Marginal ?, x0 x x0 i Y-- “2-0.2 ep”? distributions Yo - b,x a3 - cJ b3xo 4 - cj if b3 = 0 -A1 Ir(3,x,t) = 1 azhl - a,Az - a3(A, - AZ); APPENDIX ‘r~A,~~,,A~~L’.;(*,~h~) Results of the benchmark Assumptions X H(x - x,le”~‘)H(xOeU~’- x) a, # ci, i = 1...3, a, > a2 > a3, a3 > c3 rr?~,~‘,,*~~uj(*,-AZ, b, =- b2 Yo -_= x0 al-cl a2 - c2 ’ Total distribution in state 3: <r,A,-“,h~ AlAze m,x,y,t) X H(x - x,,e”“)H(x,,e”” c-1 (a~- ad(a3- c.h(y.- ,““) Al r,,m<j n(3,y,t) = Y-- yo- - x) = bxx a3 - cj b3xo b+ a3 - r,?*,~‘,,~~-~:(h,~A2, (C81~ClKC~?~‘;) } “~*,~“,*2~‘i(*,~*2, cl,-<; a3 - c3 b2xo - yo- e_<x,) X H( y - y,,e’;‘)H( y,)e”?’ - y) H(x - x(,e”l’)H(x,,e”+ - x) H c4A1 -A2)+a,A2-a2A,y [ [ 1 _ (; a3 - c3 Y-- 1 cj X ; (1 X H(y - y,Jen~‘)H(y,je”l’ - 1. y) 75
© Copyright 2024 Paperzz