Calibration and Reliability in Groundwater Modelling (Proceedings of the ModelCARE 96 Conference held at Golden, Colorado, September 1996). IAHS Publ. no. 237, 1996. 523 Accuracy of uncertainty estimation using inverse firstorder reliability analysis M. A. BAILEY Cooperative Research Centre for Catchment Hydrology, Department of Civil Engineering, Monash University, Clayton, 3168, Australia Goulburn-Murray Water, Tatura, 3616, Australia L. D. CONNELL Cooperative Research Centre for Catchment Hydrology, Department of Civil Engineering, Monash University, Clayton, 3168, Australia R. J. NATHAN Sinclair Knight Men Pty Ltd, 590 Orrong Road, Armadale, 3143, Australia Abstract An inverse reliability procedure based on the first-order reliability method (FORM) relevant to hydrological modelling is presented. FORM and the inverse reliability procedure are compared to Monte Carlo and mean-value first-order methods for reliability analysis of two-dimensional contaminant transport in porous media. Both FORM and the inverse procedure provide relatively accurate predictions of reliability at parameter coefficients of variation up to 1.0 with significantly less computational effort than Monte Carlo simulation. Increased parameter variation introduced prediction overestimation in the two procedures, and created some instability in both the FORM and inverse procedure algorithms. The inverse reliability procedure was observed to offer greater stability than the FORM algorithm in this application. INTRODUCTION While hydrological models are useful tools for prediction and decisionmaking, even the most sophisticated models are based on significant simplifications of the complex interactions and spatial and temporal behaviour that occurs in the natural environment (Beven, 1989). Inaccuracies in processes conceptualization, incomplete knowledge of parameter behaviour and initial and boundary conditions, and numerical truncation combine to introduce significant error in model output (Guymon, 1994). In order to properly interpret model results, an indication of the uncertainty associated with the output is desirable. Monte Carlo-based procedures are popular for the analysis of prediction uncertainty under uncertain parameter information (e.g. Freeze, 1975; Binley et al., 1991). While these procedures are generally robust, they are computationally intensive, especially for low probability outcomes (McLaughlin & Wood, 1988). As a result, there has been considerable interest in alternatives which could offer computational savings. Dettinger & Wilson (1981), Townley & Wilson (1985) and Connell (1995) have described computationally efficient techniques based on second moment statistical information. Other procedures used include perturbation (e.g. Tang & Pinder, 1977), and two-point estimation (e.g. Yen & Guymon, 1990). 524 M. A. Bailey et al. More recently, reliability methods have attracted interest for use in uncertainty analysis of hydrological modelling (e.g. Sitar et ai, 1987; Melching, 1992; Cawlfield & Wu, 1993; Jang et al., 1994). These methods, originally derived for structural engineering problems, are used to estimate the probability of exceedance of a target value or of some other failure state. The primary benefit of these approaches is believed to lie in accuracy comparable with other rigorous techniques such as Monte Carlo analysis while requiring significantly less computational effort. However, the approach taken for reliability analysis is not appropriate for many hydrological applications. Most hydrological applications require the model result associated with a given probability. The standard reliability approach applied to hydrological problems to date has been to calculate the probability for a given outcome (Melching et al., 1990). This paper presents and applies an inverse reliability procedure based on the first-order reliability method (FORM). The performance and stability of both FORM and its inverse procedure with respect to increasing parameter variability is also examined. RELIABILITY METHODS Analysis of prediction uncertainty considers model output to be a function of deterministic parameters and a set of random variables with known statistical properties. Uncertainty or error in the modelling procedure is generated by these random parameters. Reliability methods are based on defining the probability that a target level is exceeded: PF = P[g(X) < 0] (1) where PF denotes the probability of failure, g(X) is known as the performance function and involves the model of interest, and X is a vector of random model parameters. The probability of failure defined by equation (1) is defined exactly as the integral of the j oint probability distribution function (PDF) of X in the region where g(X) is < 0 (Jang et al., 1994). This integral is usually impractical to evaluate due to its complexity, or more commonly, limited knowledge of the behaviour of the uncertain parameters (Melching, 1992). In practical implementations of reliability methods, the integral is approximated. First-order reliability method One approach to approximating the probability integral is the first-order reliability method, FORM (Madsen et al. ,1986; Sitar et al., 1987). The first step in the procedure is the transformation of the random variables, X, of the model to uncorrelated standard normal deviates, U, with zero mean and unit variance. With second moment statistical information available, the transformation is summarized by (Sitar et al., 1987): U = L~lD-\X-M) (2) where D = diag[o-,] is a diagonal matrix of the standard deviations, L = lower triangular Accuracy of uncertainty estimation using inverse first-order reliability analysis 525 (Cholesky) decomposition of the correlation matrix R = \py\ such that R = LU and M = mean vector of the variables of X. The performance function is then expressed in standard space by: G{U) = g{DLP+M) = g{X) (3) Der Kiureghian & Liu (1986) and Sitar et al. (1987) discuss incorporation of marginal distribution information for the parameters into the transformation. Hasofer & Lind (1974) introduced the concept of the reliability index (/?) and demonstrated its equivalence to the distance from the origin to the nearest point on the limit state surface defined by the performance function in standard space. The reliability index is obtained by linearizing the performance function at the point on the limit state surface nearest the origin (called the design point, denoted by u* in standard space and JC* in original space), and defined by: p = a*u* (4) where a* is a line vector describing the unit normal at the design point on the failure region directed towards region of failure (Sitar et al., 1987). The first-order approximation of the probability of failure is found by assuming the performance function in standard space to be normally distributed, and is given by: PF = *(-P) (5) where $(2) is the standard normal cumulative probability evaluated at i. Algorithm for determining the design point The design point and hence reliability index is found from solution of a constrained optimization problem to minimize the distance from the origin subject to the point being on the limit state surface (Zhang & Der Kiureghian, 1994): p = mm{\U\\G{U) = 0} (6) Various algorithms are available for the solution of this problem (e.g. Liu & Der Kiureghian, 1991). The Hasofer & Lind-Rackwitz & Fiessler (HL-RF) algorithm (Hasofer & Lind, 1974; Rackwitz & Fiessler, 1978) is used in both structural and hydrological applications (e.g. Madsen et al., 1986; Sitar et al, 1987; Jang et al., 1994). The algorithm locates the design point with an iterative procedure: »*•! = -^^[^(uk)uk-(uk)fuG(uky (7) 8G(.uk) du1 (8) where: \G(uk) - dG(uk) 3K„ defines the gradient vector of the limit state surface in the standard normal space. Liu & Der Kiureghian (1991) introduced a merit function to improve the convergence properties of the HL-RF algorithm. A simpler form of the merit functions: m(u) = 1 / 2 |M|+C|G(M)| (9) M. A. Bailey et al. 526 where c ( > 0) is a penalty parameter, is described by Zhang & Der Kiureghian (1994), and has been implemented for the FORM analyses conducted in this paper. Inverse reliability - modified FORM Inverse reliability has been investigated in structural engineering applications (Der Kiureghian et al., 1994) and groundwater transport (Schanz & Salhotra, 1992). The procedure described by Schanz & Salhotra (1992) is a modification of the HL-RF algorithm (equation (7)), and was used to locate the target concentration and design point for several transport models. Zhang & Der Kiureghian (1994) have proposed additional modifications to the HL-RF algorithm to permit general inverse reliability investigations. The performance function is modified to include an unknown deterministic parameter 6, such that G(U) = G(U,B). This parameter may be considered to be the target value of the performance function (as implicitly done by Schanz & Salhotra (1992)), or a component of the model function. Inverse reliability identifies 6 such that the reliability index for the performance function equals a known target /3r: e:min{|£/||G(C/,e)} = P7 (10) The minimization problem is defined by the set of equations (Zhang & Der Kiureghian, 1994): (Ha) 0 u+ KG(u,d)\ VG(u,Q) = 0 G{u,Q) = 0 (lib) (He) These equations are then linearized and solved. With the iteration algorithm described by xi+1 = Xj + dt, the search direction di is given by the solution of the linearized equations: "ft v„G(M,) (12) ^ = dGiu^ldO Zhang & Der Kiureghian (1994) use a merit function to control the value of dh to improve convergence of the algorithm. The function is given by: m(u) = V2\u\2 + c\G(u,Q)\ ( 13 ) Implementation of the algorithm requires c > PT\u\/ô, where <5 is the required tolerance in satisfying equation (lie). Both the techniques of Schanz & Salhotra (1992) and Zhang & Der Kiureghian (1994) require initial normalization of variables, although Schanz & Salhotra do not Accuracy of uncertainty estimation using inverse first-order reliability analysis 527 specify that the variables are to be uncorrelated in standard space. The solution for the input vector Zx of Schanz & Salhotra is similar to that used by Zhang & Der Kiureghian for iteration. RESULTS AND DISCUSSION The behaviour of the first-order reliability method and inverse reliability were examined by application to a simple analytical solution of a two-dimensional contaminant transport problem. The predictions were then compared with an equivalent Monte Carlo analysis, and the mean-centred first-order method of Ang & Cornell (1974). The Monte Carlo simulation used 20 000 iterations to ensure accuracy at low probabilities of failure. The solution for transport through porous media was given by Kinzelbach (1986) as: C(x,y,y) _ C, o 1 4TI./C ( „ 1 \ : exp V2KL; w r2R 4aLvt ry ^ 2a L (14) where C is the concentration at point (x,y) and time t, C0 is the concentration of the incoming contaminant, v is the steady-state velocity, aL is the longitudinal dispersivity, aT is the transverse dispersivity, X is the degradation rate of the contaminant, and R is the retardation factor. The function W(al,a2) is an integral term called the Hantushfunction (Hantush, 1956), r is the weighted radial distance defined by r2 = x2 + (aL/aT)y2, and y = (1 + 4aL\R/u)0-5. The performance function was defined for an ideal contaminant (R = 1, X = 0) at x = 10 m, y = 0.5 m, and t — 10 years. The steady-state velocity, longitudinal dispersivity and transverse dispersivity were treated as uncertain variables, with /xv= 1.0 m day"1, p,aL = 4.5 m, and/x„r = 1.125 m respectively. The standard deviations of the parameters were varied during the analyses. As it was desired to examine the effects of increasing uncertainty, all variables were assumed to be independent and lognormally distributed. It is believed important to establish the relative accuracy of FORM and inverse reliability in comparison with the other procedures at high levels of uncertainty, and to identify any problems in the implementation of the methods. The initial investigations focused on the accuracy and stability of the FORM procedure. The standard deviation of all three uncertain parameters was fixed at 10% of their mean value (a coefficient of variation = 0.1 ). A cumulative distribution function (CDF) of the probability of failure of the performance function was established by incrementally varying the target concentration ratio (i.e. C/CQ). For the initial analysis run, both the mean-value approach and FORM were found to provide very good matches with the probability of failure calculated form the more rigorous Monte Carlo procedure. This was expected, as techniques based on first-order linearization of Taylor series expansions have been found to provide reasonable accuracy for low variances (Melching, 1992). Increasing the standard deviation of all values to 50% of the mean or a coefficient of variation of 0.5 (Fig. 1) was found to contribute to significant divergence of the mean-value approach from the Monte Carlo data. However, the FORM data was found to provide good agreement, while requiring significantly reduced computational effort. 528 M. A. Bailey et al. 0.15 0.2 0.25 0.3 0.35 0.45 TARGET CONCENTRATION RATIO Fig. 1 Comparison of FORM with Monte Carlo (population 20 000) and mean-value first-order methods at coefficient of variation = 0.50. It was expected that further increases in the skewness of the lognormal distribution would contribute to increased divergence of the mean-value first-order technique (Melching, 1992). Cawlfield & Wu (1993) and Jang et al. (1994) have previously considered the effects of increased variability on the behaviour of FORM, and noted divergence when compared with Monte Carlo data. Our study was also concerned with the accuracy of the procedure, but expanded to include Considerations on the effects of uncertainty on the stability on the solution. Further runs were conducted with standard deviations set to 100%, 150% and 200% of the mean values (Figs 2-3). The increased parameter variability was found to 0.15 0.2 0.25 0.3 0.35 TARGET CONCENTRATION RATIO Fig. 2 Comparison of FORM with Monte Carlo (population 20 000) and mean-value first-order methods at coefficient of variation = 1.00. Accuracy of uncertainty estimation using inverse first-order reliability analysis 1 • A 0.9 111 E 0.8 3 0.7- \\ IV, \ V \\ \ 0.5 \ ID < g MC MVFO FORM \Y £ " • o j*. 529 \ 0.3 DC °- 0.2 ^ 0.10 ^ 1 1 1 1 0.0S 0.1 0.15 0.2 1—~- 0.2S i 1 1 0.3 0.35 0.4 mi | 0.45 TARGET CONCENTRATION RATIO Fig. 3 Comparison of FORM with Monte Carlo (population 20 000) and mean-value first-order methods at coefficient of variation = 2.00. ultimately contribute to widespread disagreement in probability estimation. The probability of failure estimated by FORM was found to consistently overestimate that from Monte Carlo simulation for coefficients of variation in excess of 1.0. It was also noted that the ability of FORM to identify an appropriate design point progressively deteriorated, resulting in local (minor) instabilities. Similar behaviour has been observed in other applications of FORM (Melching, 1995). Significantly poorer stability was observed for several simulations conducted with a large standard deviation in a single variable. Further investigation suggested that the choice of initial parameters for the FORM iteration largely contributed to this instability. The inverse reliability procedure was applied and tested in a similar manner to that used for FORM. Simulations were again conducted at increasing levels of standard deviation, and are presented in Fig;T 2, where they are compared against equivalent Monte Carlo analysis. The inverse reliability procedure .is in close agreement with the Monte Carlo results for probabilities of failure corresponding to standard deviations up to 100% of mean values. Beyond this level of uncertainty,, the procedure overestimated the probability of failure associated with a given target level. With standard deviations set at 150% of the magnitude of the mean, significant divergence was observed for probabilities of failure between approximately 0.05 and 0.9. Reasonable agreement occurred at lower probabilities. This agreement deteriorates markedly as the standard deviation is increased further (Fig. 4). This behaviour mirrors that found by Schanz & Salhotra (1992), who found that their procedure calculated accurate estimates of 90th and 95th percentile targets, but poorer estimates of the 50th percentile targets. A feature of the inverse procedure was greater stability in comparison with FORM. Some instability in the algorithm was observed at lower probabilities of failure, as occurred with FORM. However, thèse did not cause excessive inaccuracy in the algorithm, and did not contribute to catastrophic instability. 530 M. A. Bailey et al. PROBABILITY OF FAILURE Fig. 4 Comparison of inverse reliability with Monte Carlo method (population 20 000). CONCLUSIONS The objective of this paper was to provide a perspective into the accuracy and stability of reliability methods applied to hydrological engineering. Emphasis was placed on the ability of the first-order reliability method (FORM) and an inverse reliability procedure to predict probabilities of failure associated with a particular performance function. The analysis confirmed that use of mean-value first-order techniques for uncertainty evaluation is subject to error for applications involving high parameter variances, such as those commonly encountered in practical problems. When compared to the meanvalue procedure, FORM offered greater accuracy up to levels of parameter uncertainty above a coefficient of variation of 1.0. It was also found at these high variances that the robustness of the FORM approach was at times questionable for this particular application, particularly when the search algorithm appeared to identify local minima, leading to numerical instability and ultimately failure of the algorithm. Comparison of the inverse reliability procedure with the results of Monte Carlo simulation was also favourable to coefficients of variation in the order of 1.0. The inverse technique, which is thought to be more applicable to common hydrological problems than the FORM approach on which it is based, was also found to have greater numerical stability for the example considered. However, this stability cannot be guaranteed for other applications. The results just described were applicable for a relatively simple case of contaminant transport. Requiring further investigation is the behaviour of reliability methods for more complex situations such as those encountered with multi-process (i.e. catchment) models. In such models, inaccuracies in individual processes must be combined to evaluate overall prediction uncertainty. Problems associated with locating the global minima become more difficult as the performance function becomes more complex. In Accuracy of uncertainty estimation using inverse first-order reliability analysis 531 multi-process models, these problems are exacerbated by difficulties in defining performance functions appropriate to the components of the model. The definition of the inverse reliability procedure implies application to a whole system rather than individual processes. Error propagation in multi-process models, and incorporation of techniques such as FORM and inverse reliability, is the subject of ongoing research. REFERENCES Ang, A. H.-S. & Cornell, A. C. (1974) Reliability bases of structural safety and design. J. Struct. Div. ASCE 100(ST9), 1755-1769. Beven, K. (1989) Changing ideas in hydrology — the case of physically-based models. J. Hydrol. 105, 157-172. Binley, A. M., Beven, K. I., Calver, A. & Watts, L. G. (1991) Changing responses in hydrology: assessing the uncertainty in physically based model predictions. Wat. Resour. Res. 27(6), 1253-1261. Cawlfield, J. D. & Wu, M.-C. (1993) Probabilistic sensitivity analysis for one-dimensional reactive transport in porous media. Wat. Resour. Res. 29(3), 661-672. Connell, L. D. (1995) An analysis of perturbation based methods for the treatment of parameter uncertainty in numerical groundwater models. Transport in Porous Media 21, 225-240. Der Kiureghian, A. &Liu,P.-L. (1986) Structural reliability under incompleteprobability information. J. EngngMech.Div. ASCE 117(1), 85-104. DerKiureghian.A., Zhang, Y. &&Li,C.-C. (1994) Inverse reliability problem./. EngngMech. Di v. ASCE 120(5), 11541159. Dettinger.M. D. & Wilson, J. L. (1981) First order analysisof uncertainty in numerical models of groundwater flow. Part 1: Mathematical development. Wat. Resour. Res. 17(1), 149-161. Freeze, R. A. (1975) A stochastic-conceptualanalysis of one-dimensional groundwater flow in nonuniform homogenous media. Wat. Resour. Res. 11(5), 725-741. Guymon, G. L. (1994) Unsaturated Zone Hydrology. Prentice Hall, Englewood Cliffs, New Jersey. Hantush, M. S. (1956) Analysis of data from pumping tests in leaky aquifers. Trans. AGU 37(6), 702-714. Hasofer, A. M. & Lind, N. C. (1974) Exact and invariant second-moment code format. /. Engrs Mech. Div. ACSE 100(EM1), 111-121. Jang, Y.-S., Sitar, N. & Der Kiureghian, A. (1994) Reliability analysis of contaminant transport in saturated porous media. Wat. Resour. Res. 30(8), 2435-2448. Kinzelbach, W. (1986) Groundwater Modelling. An Introduction with Sample Programs in BASIC. Elsevier Science Publishers, Amsterdam. Liu, P.-L. & Der Kiureghian, A. (1991) Optimization algorithms for structural reliability. Structural Safety 9, 161-177. Madsen, H. O., Krenk, S. &Lind, N. C. (1986) Methods ofStructural Safely. Prentice-Hall, Englewood Cliffs, New Jersey. McLaughlin, D. &Wood, E. F. (1988) Distributed parameter approach for evaluating the accuracy of groundwater model predictions. 1. Theory. Wat. Resour. Res. 24(7), 1037-1047. Melching, C. S. (1992) An improved first-order reliability approach for assessing uncertainties in hydrologie modeling. /. Hydrol. 132, 157-177. Melching, C. S. (1995) Reliability estimation. In: Computer Models of Watershed Hydrology (ed. by V. P. Singh), 69-118. Water Resources Publications, Highlands Ranch. Melching, C. S., Yen, B. C. & Wenzel, H. G. (1990) A reliability estimation in modeling watershed runoff with uncertainties. Wat. Resour. Res. 26(10), 2275-2286. Rackwitz, R. & Fiessler, B. (1978) Structural reliability under combined load sequences. Comp. and Struct. 9, 489-494. Schanz, R. W. & Salhotra, A. (1992) Evaluation of the Rackwitz-Fiessler uncertainty analysis method for environmental fate and transport methods. Wat. Resour. Res. 28(4), 1071-1079. Sitar, N., Cawlfield, J. D. & Der Kiureghian, A. (1987) First-order reliability approach to stochastic analysis of subsurface flow and contaminant transport. Wat. Resour. Res. 23(5), 794-804. Tang, D. H. &Pinder, G. F. (1977) Simulation of groundwater flow and mass transport. Adv. Wat. Resour. 1(1), 25-30. Townley, L. R. & Wilson, J. L. (1985) Computationally efficient algorithms for parameter estimation and uncertainty propagation in numerical models of groundwater flow. Wat. Resour. Res. 21(12), 1851-1860. Yen, C.-C. & Guymon, G. L. (1990) An efficient deterministic-probabilisticapproach to modeling regional groundwater flow. 1. Theory. Wat. Resour. Res. 26(7), 1559-1567. Zhang, Y. & Der Kiureghian, A. (1994) Two improved algorithms for reliability analysis. In: Reliability and Optimization of Structural Systems (ed. by R. Rackwitz, G. Augusti & A. Borri) (Proc. Sixth IFIP WG7.5 Working Conf, September 1994), 297-304. Chapman & Hall, London.