This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

Available online at www.sciencedirect.com

Journal of Process Control xxx (2008) xxx–xxx www.elsevier.com/locate/jprocont

**Applying the unscented Kalman ﬁlter for nonlinear state estimation
**

Rambabu Kandepu a, Bjarne Foss a,*, Lars Imsland b

a

Department of Engineering Cybernetics, Norwegian University of Science and Technology, Trondheim 7491, Norway b SINTEF ICT, Trondheim 7465, Norway Received 9 July 2007; accepted 4 November 2007

Abstract Based on presentation of the principles of the EKF and UKF for state estimation, we discuss the diﬀerences of the two approaches. Four rather diﬀerent simulation cases are considered to compare the performance. A simple procedure to include state constraints in the UKF is proposed and tested. The overall impression is that the performance of the UKF is better than the EKF in terms of robustness and speed of convergence. The computational load in applying the UKF is comparable to the EKF. Ó 2007 Elsevier Ltd. All rights reserved.

Keywords: Nonlinear state estimation; Kalman ﬁlter; Constraint handling

1. Introduction In the process industries one of the main goals is to make the end product at the lowest possible cost while satisfying product quality constraints. State estimation often play an important role in accomplishing this goal in process control and performance monitoring applications. There are many uncertainties to deal with in process control; model uncertainties, measurement uncertainties and uncertainties in terms of diﬀerent noise sources acting on the system. In this kind of environment, representing the model state by an (approximated) probability distribution function (pdf) has distinct advantages. State estimation is a means to propagate the pdf of the system states over time in some optimal way. It is most common to use the Gaussian pdf to represent the model state, process and measurement noises. The Gaussian pdf can be characterized by its mean and covariance. The Kalman ﬁlter (KF) propagates the mean and covariance of the pdf of the model state in an optimal (minimum mean square error) way in case of linear dynamic systems [2,11].

*

Corresponding author. Tel.: +47 73594476; fax: +47 73594599. E-mail address: Bjarne.Foss@itk.ntnu.no (B. Foss).

All practical systems posses some degree of nonlinearity. Depending on the type of process and the operating region of the process, some processes can be approximated with a linear model and the KF can be used for state estimation. In some cases the linear approximation may not be accurate enough, and state estimator designs using nonlinear process models are necessary. The most common way of applying the KF to a nonlinear system is in the form of the extended Kalman ﬁlter (EKF). In the EKF, the pdf is propagated through a linear approximation of the system around the operating point at each time instant. In doing so, the EKF needs the Jacobian matrices which may be diﬃcult to obtain for higher order systems, especially in the case of time-critical applications. Further, the linear approximation of the system at a given time instant may introduce errors in the state which may lead the state to diverge over time. In other words, the linear approximation may not be appropriate for some systems. In order to overcome the drawbacks of the EKF, other nonlinear state estimators have been developed such as the unscented Kalman ﬁlter (UKF) [1], the ensemble Kalman ﬁlter (EnKF)[10] and high order EKFs. The EnKF is especially designed for large scale systems, for instance, oceanographic models and reservoir models [10]. The UKF seems to be a promising alternative for process control applications [15,16,20].

0959-1524/$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.jprocont.2007.11.004

Please cite this article in press as: R. Kandepu et al., Applying the unscented Kalman ﬁlter for nonlinear state estimation, J. Process Contr. (2008), doi:10.1016/j.jprocont.2007.11.004

u k À1 Þ . uk Þ . and the KF theory is applied to get the mean and covariance of y. how to calculate the mean ðy UKF Þ and covariance ðP UKF Þ y of y. p g. uk Þjv¼v : (b) Compute predicted state mean and covariance (time update) b xÀ x k À 1 . EKF principle and algorithm To illustrate the principle behind the EKF. Four examples are studied in Section 4 to compare the performances of the UKF and EKF. . Discussion on the diﬀerences between the UKF and EKF is presented in Section 5 and conclusions are drawn in Section 6. . The present paper focuses on using the UKF for nonlinear state estimation in process systems and the performance is evaluated in comparison with the EKF.11. Then the following steps are involved in approximating the mean and covariance: Propagate each sigma point through the nonlinear function.004 . The EKF algorithm for this system is presented below: Initialization at k = 0: b x 0 ¼ E ½x0 . uk Þjx¼b . p ¼ 2 n þ 1. Algorithm.ARTICLE IN PRESS 2 R. At the end of the section. v k À1 . . 1: (1) Prediction step. The question is how to compute the pdf of y given the pdf of x? For example. . The paper proposes a simple method to incorporate state constraints in the UKF. Section 2 describes the principles and algorithms of EKF and UKF. n 2 Rnn the observation noise.1016/j. how to calculate the mean ðy Þ and covariance (Ry) of y? If g is a linear function and the pdf of x is a Gaussian distribution. . The question is how the UKF compute pdf of y given the pdf of x. by using linearization.jprocont. the KF is optimal up to the ﬁrst two moments in the class of linear estimators [9]. Kandepu et al. P x0 ¼ E½ðx0 À b P v ¼ E½ðv À vÞðv À vÞ . v. In the case of a nonlinear function (g(x)). 2. consider the following example. the nonlinear function is linearized around the current value of x.. nk . The EKF and UKF algorithms for nonlinear state estimation We present the principles and algorithms of the EKF and UKF.2007. Section 3 introduces a method to incorporate the state constraints in the UKF state estimation and compares it with the EKF. x 0 Þðx0 À b x 0 Þ . The nonlinear functions f and h are need not necessarily be continuous. P xk ¼ ðI À K k H xk ÞP À xk : 2. 2. in other words. y k ¼ hð x k . (2008). . In other words. Applying the unscented Kalman ﬁlter for nonlinear state estimation. UKF principle and algorithm Consider now the same example as in the previous section. J. Even if the pdf is not Gaussian. xÀ k xÀ Dn ¼ rn hðb k . 2. uk Þ . ð2Þ ð3Þ where ($g) is the Jacobian of g(x) at x. P n ¼ E½ðn À nÞðn À nÞ : For k = 1. u the input and (similar to the random samples of a speciﬁc distribution function in Monte Carlo simulations) with each point being associated with a weight w(i). n. P EKF y ¼ ðrgÞP x ðrgÞ . i 2 f1.. v. x k À1 T T T Gv ¼ rv f ðb x kÀ1 . k ¼ f ðb T T PÀ xk ¼ F xk P xk F xk þ Gv P v Gv : (2) Correction step. uk Þjn¼n : (b) Update estimates with latest observation (measurement update) T À T T Kk ¼ P À xk H xk ðH xk P xk H xk þ Dn P n Dn Þ .1. Kandepu et al. Process Contr. ð4Þ ð5Þ where x 2 Rnx is the system state. n. then Kalman ﬁlter (KF) is optimal in propagating the pdf. nÞ. À1 b xÀ xÀ xk ¼ b k þ K k ½y k À hðb k . v . given the mean ðxÞ y and covariance (Px) of the pdf of x are calculated as follows: y EKF ¼ gðxÞ. Let x 2 Rn be a random vector and y ¼ g ðx Þ ð1Þ be a nonlinear function. y ðiÞ ¼ gðxðiÞ Þ Please cite this article in press as: R. diﬀerent characteristics of the EKF and UKF are compared. Let a general nonlinear system be represented by the following standard discrete time equations: x k ¼ f ð x k À1 . (a) Compute the process model Jacobians: F xk ¼ rx f ðx. ukÀ1 Þjx¼b . / Journal of Process Control xxx (2008) xxx–xxx The UKF propagates the pdf in a simple and eﬀective way and it is accurate up to second order in estimating mean and covariance [1]. (a) Compute observation model Jacobians: H xk ¼ rx hðx. T y the noisy observation of the system. The KF is extended to the class of nonlinear systems termed EKF. . g : Rn ! Rm .2. the mean ðy EKF Þ and covariance ðP EKF Þ of y. in the case of being Gaussian? Consider a set of points xðiÞ . in the case of being Gaussian. These sample points are termed as sigma points. v 2 Rnv the process noise. doi:10.

J.k =k À1 ¼ f ðXi.k =k À1 .jprocont.k À1 . . . Applying the unscented Kalman ﬁlter for nonlinear state estimation.11. The parameter a.k=kÀ1 ¼ hðXx i.kÀ1 .4 0 P v 0 5 0 0 Pn where Pv and Pn are the process and observation noise covariance matrices. Kandepu et al. wð c ¼ N þk 1 iÞ ði Þ . 2 3 xk 6 7 xa ð6Þ .ARTICLE IN PRESS R. N .8. RwðiÞ ¼ 1 and the covariance is computed by the weighted outer product of the transformed points. . Kandepu et al. An augmented state at time instant k. 2 N . 1. N ¼ nx þ n v þ nn ð7Þ where a and j are tuning parameters. Algorithm. wð m ¼ N þk k 0Þ þ ð1 À a2 þ bÞ.004 . . . 0 6 a 6 1.kÀ1 5. 1. i ¼ 0. 1: (1) Calculate 2N + 1 sigma-points based on the present state covariance: 8 xa i ¼ 0.k =k À1 À b k ÞðYi.. Xk À1 . This parameter (b) can be used to incorporate knowledge of the higher order moments of the distribution. . process noise and measurement noise. k ¼ a2 ðN þ jÞ À N . pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ c ¼ N þ k.The ith sigma point (augmented) is the ith column of the sigma point matrix. . For a Gaussian prior the optimal choice is b = 2 [8]. v and n. (3) Measurement-update equations: Transform the sigma points through the measurement-update function. . i ¼ 1 . the augmented state covariance matrix is built from the covariance matrices of x. Please cite this article in press as: R.b ð9Þ x k À1 þ c S i . n Yi. . b xÀ k ¼ PÀ xk ¼ 2N X i¼0 2N X i¼0 x x iÞ wð xÀ xÀ c ðXi. 2 3 Px 0 0 6 7 ð8Þ P a . 2 x 3 Xi. respectively.b a a Xi. i ¼ 1. i ¼ N þ 1 . k 0Þ . 2N : ð10Þ Calculate the apriori state estimate and apriori covariance. Xi.k =k À1 À b k Þ : T iÞ x ðwð m Xi. .k =k À1 . ð11Þ ð12Þ Similarly. (2) Time-update equations: Transform the sigma points through the state-update function. . < . > k À1 . i ¼ 0. for background theory.kÀ1 6 Xv 7 a Xi. x v Xx i. ¼ P UKF y p X i¼0 wðiÞ ðy ðiÞ À y Þðy ðiÞ À y ÞT : Both the sigma points and the weights are computed deterministically through a set of conditions given in [1].k =k À1 À b k ÞðXi. . 2. v and n refer to a partition conformal to the dimensions of the state. y UKF ¼ p X i¼0 wðiÞ y ðiÞ . to guarantee the semi-positive deﬁniteness of the covariance matrix. The UKF algorithm is presented below. .1016/j. . . uk Þ. a good default choice is j = 0. Initialization at k = 0: b x 0 ¼ E½x0 . wð m ¼ wc ¼ 2ð N þ kÞ where b is a non-negative weighting parameter introduced to affect the weighting of the zeroth sigmapoint for the calculation of the covariance.k À1 .9].k =k À1 Þ. controls the size of the sigma-point distribution and it should ideally be a small number [8]. Xn i.k =k À1 À b k Þ : T iÞ wð m Yi. . vk 5 4 k nk is deﬁned. > : a . . The augmented state dimension is. Let the system be represented by (4) and (5). . iÞ ðiÞ The weights wð m and wc are deﬁned as. refer to [1.kÀ1 ¼ 4 i. doi:10. b yÀ k ¼ P yk ¼ 2N X i¼0 2N X i¼0 iÞ wð yÀ yÀ c ðYi.2007. We must choose j P 0. . (2008).. 2N ð13Þ and the mean and covariance of the measurement vector is calculated. . . 2N . uk À1 Þ. . . P x0 ¼ E½ðx0 À b x 0 Þðx0 À b x 0 ÞT T a a b x 0 ¼ E½x ¼ E½ b x0 0 0 2 3 Px 0 0 6 7 T a a Pv 0 5 Pa xa xa 0 ¼ E ½ðx0 À b 0 Þðx0 À b 0Þ ¼ 4 0 0 0 Pn For k = 1. i ¼ 0.b x k À1 À c S i .k À1 where the superscripts x. pﬃﬃﬃﬃﬃﬃﬃﬃﬃ S ¼ Pa k À1 : In (9) c is a scaling parameter [8]. where Si is the ith column of the matrix. / Journal of Process Control xxx (2008) xxx–xxx 3 the mean is approximated by the weighted average of the transformed points. Process Contr. i ¼ 0.

The EKF mean is obtained by using (2) giving EKF 110 Xmean UKF 110 100 X mean mean true mean 100 90 80 70 Y Ymean linearization Px=16 Py =2686 Py EKF true EKF mean true 90 80 70 Yukf Y y=g(x)=x2 60 50 40 30 20 10 0 y=g(x)=x2 transformed sigma points P =16 x =2304 60 50 40 30 20 10 0 P true =2686 y UKF Pa =2816 0 5 10 0 5 10 x Fig. T P xk ¼ P À xk À K k P y k K k : 2. J. three sigma points are propagated through the function. and the variance is obtained by performing the linearization as shown in the ﬁgure and using (3). The true mean and variance are 51. Kandepu et al.3.k =k À1 À b k Þ : Y EKF mean ¼ 36:00. Z 1 1 xÀx 2 1 true Y mean ¼ E½gðxÞ ¼ x2 : pﬃﬃﬃﬃﬃﬃ eÀ2ð r Þ dx. Applying the unscented Kalman ﬁlter for nonlinear state estimation. doi:10. b xÀ yÀ xk ¼ b k þ K k ðy k À b k Þ. Later. during the timeupdate step. during the time-update (prediction) step. Y UKF mean ¼ 52:00. 1 by considering the following example. Illustration of principle of EKF and UKF. In case of the UKF. 2pr À1 2 P true ¼ E ½ð g ð x Þ À E ½ g ð x ÞÞ : y pﬃﬃﬃﬃﬃ where r ¼ Rx is the standard deviation. P EKF ¼ 2304: y For the UKF. ð14Þ P UKF ¼ 2816: y The UKF approximates the propagation of the pdf through the nonlinearity more accurately when compared to the EKF as illustrated by the numbers in the example above. An example (state estimation of a reversible reaction) will be considered to illustrate the constraint handling capability of the UKF. respectively.43 and 2686. all the sigma points are propagated through the nonlinear function which makes the UKF a better and more eﬀective nonlinear approximator. x ¼ X mean ¼ 6. The UKF is accurate up to second order moments in the pdf propagation where as the EKF is accurate up to ﬁrst order moment [8].ARTICLE IN PRESS 4 R. the mean is propagated through the nonlinear function. x Please cite this article in press as: R. and the UKF estimate and its covariance are computed from the standard Kalman update equations. Process Contr.11. in other words. y ¼ g ðxÞ ¼ x2 .k =k À1 À b k ÞðYi. and the mean and variance are calculated accordingly. 1. Rx ¼ 16: The ﬁgure illustrates how the mean and variance of x are propagated to obtain mean and variance of y for EKF and UKF. this introduces an error since in general y 6¼ gðxÞ. The true mean and variance are calculated by using the following equations. (2008). The Kalman gain is given by. Kandepu et al. we will see that it is possible to implement state constraints by proper conditioning of the sigma points. / Journal of Process Control xxx (2008) xxx–xxx The cross covariance is calculated according to P xk y k ¼ 2N X i¼0 T x iÞ wð xÀ yÀ c ðXi..004 . In the EKF algorithm.1016/j. x 2 R. Discussion The diﬀerence in the principles of state estimation using UKF and EKF is illustrated based the Fig.2007. The UKF principle is simple and easy to implement as it does not require the calculation of Jacobians at each time step.jprocont. 1 K k ¼ P xk y k P À yk .

estimated concentrations should remain positive.11. Applying the unscented Kalman ﬁlter for nonlinear state estimation. the feasible region by the box constraints can be represented by a rectangle as in Fig. t=k UKF.2007. the constraints can be incorporated in the KF by projecting the unconstrained KF estimates onto the boundary of the feasible region at each time step [18. 2. Kandepu et al. such constraints are due to physical limitations on the states. Typically. (2008). C x-kUKF xkEKF Fig.19]. 2. The numerical optimization at each time step may be a challenge in time-critical applications. xL 6 x 6 xH : We will illustrate the method for x 2 R2 . / Journal of Process Control xxx (2008) xxx–xxx 5 3.004 . C Transformed sigma points xkEKF covariance covariance covariance xkEKF. Please cite this article in press as: R. Assume that the state constraints are represented by box constraints. In this section.ARTICLE IN PRESS R.. doi:10.1016/j. known as ‘clipping’ [14]. Illustration of estimation with state constraints. Process Contr. in which the constraints can be included in the estimation problem in a natural way [17]. In case of a second order system. The ﬁgure shows the illustration of the steps of constraint handling in case of the EKF and UKF from one time step to the Initial set up. In Kalman ﬁlter theory. t=k xkEKF. In MHE. a new and simple method is introduced to handle state constraints in the UKF and it is compared to the standard way of constraint handling in the EKF. t=k-1 covariance xk-1 EKF. the state trajectory is computed taking state constraints into account at the expense of solving a nonlinear programing problem at each time step. for instance. However. Kandepu et al.jprocont. An other way of nonlinear state estimation with constraints is moving horizon estimation (MHE). J. State estimation with constraints Constraints on states to be estimated are important model information that is often not used in state estimation. there is no general way of incorporating these constraints into the estimation problem.

. k À1 i ¼ 0 ..C x Xx i ¼ 0. Applying the unscented Kalman ﬁlter for nonlinear state estimation. its estimate ðb x k Þ and state covariance are selected as shown in the ﬁgure.2 are set to the following values. it diverges and if the initial state is inside the limit cycle. 1 . The system model and the state estimator model are the same unless otherwise speciﬁed. and they are furk ther updated in the measurement-update step.11. The mean and covariance with the constrained sigma points now represents the apriÀ ori UKF estimate ðxUKF Þ and covariance. the true state (xk). (3) Measurement-update equations: Transform the constrained sigma points through the measurementupdate function as in (13) and obtain the UKF estimate by following the same steps given in Section 2. the covariance does not include any constraint information. Kandepu et al. state estimation of a reversible reaction is studied to illustrate the constraint handling capability of the UKF. First. . it ¼ P ðXx i. At t = k. (2008). . The noise is Gaussian with zero mean value.C ðxEKF Þ as shown in the ﬁgure.jprocont. doi:10. . / Journal of Process Control xxx (2008) xxx–xxx x. Hence. . an estimation problem in an induction machine is chosen to evaluate the performances for a nonlinear system. uk À1 Þ. Second. The advantage here is that the new apriori covariance includes information on the constraints. . . .004 . the unconstrained EKF estimate ðb x EKF Þ is outside k the feasible region and is projected to the boundary of the feasible region to get the constrained EKF estimate . 2.b and j in Section 2. . 4. the covariance of the EKF estimate is not changed and thus the constraints have no eﬀect on the covariance. the estimate after the measurement-update (refer (14)) is outside the feasible region.. Finally. it can exhibit both stable and unstable limit cycles [12].k À1 . Alternative linear constraints. depending on the direction of time. the same projection technique can be used. in case.C Xx i . After the propagation of the sigma points from (10). the following assumptions are made: The measurement update frequency of the KF coincides with the system discretization sampling frequency. .1. The constraints information can be incorporated in the UKF algorithm in a simple way during the time-update step. the Van der Pol oscillator is considered to study the behavior of the UKF and also to compare the robustness of the estimator due to model errors with that of the EKF. Xi. a ¼ 1.k À1 Þ i ¼ 0. .k =k À1 ¼ P ðXi. b¼2 and j ¼ 0: 4. where P refers to the projections. 2. This way of the handling constraints in the EKF is termed as ‘clipping’ in literature [14]. Please cite this article in press as: R.1016/j.C v Xx i. at t = k three sigma points which are outside the feasible region are projected onto the boundary (lower right plot in the ﬁgure). Further. J. . .C strained transformed sigma points Xx i. nonlinear system. Simulation studies Four rather diﬀerent examples are considered to compare diﬀerent characteristics of the UKF and the EKF.2007. In case of an unstable limit cycle. It is highly nonlinear and. The proposed algorithm is outlined below: Algorithm (outline). . . Cx 6 d are easily included by projecting the sigma point violating the inequality normally onto the boundary of feasible region. next. if the initial state is outside the limit cycle. The tuning parameters (the initial covariance and process and measurement noise covariances) are chosen to be the same for both the EKF and UKF. 2N : i.k =k À1 Þ Calculate the apriori state estimate and apriori covariance as given in (11) and (12) using the con. . Process noise and measurement noise are applied to the system. 2N . the same projection technique can be extended. In Fig. Extension of the proposed method to higher dimension is straightforward. Van der Pol oscillator The Van der Pol oscillator is considered as it is a widely used example in the literature. a. 2N : Again apply the constrains on the transformed sigma points to obtain the constrained transformed sigma points. It is observed that the new (constrained) covariance obtained at a time step is lower in size compared to the unconstrained covariance. Process Contr. Kandepu et al. In all the examples.k =k À1 ¼ f ðXi. e. At t = k À 1. If UKF estimate violates the constraints. the (unconstrained) transformed sigma points which are outside the feasible region can be projected onto the boundary of the feasible region and continue the further steps. While projecting the EKF k estimate. 1: (1) Calculate 2N + 1 sigma points based on the present state covariance according to (9) and project the sigma points which are outside the feasible region to the boundary to obtain the constrained sigma points. which should make the UKF estimate more eﬃcient (accurate) compared to the EKF estimate. Third. For k = 1.g. (2) Time-update equations: Transform the sigma points through the state-update function. 1. If. .k =k À1 . An example (reversible reaction) is considered in the next section to illustrate the state estimation with constraints with the proposed method. For UKF algorithm. 1. a solid oxide fuel cell (SOFC) combined gas turbine (GT) hybrid system is considered to evaluate the performances in the case of higher order.k À1 .ARTICLE IN PRESS 6 R. the technique can be extended to the sigma points from (9).

with ð16Þ k x ¼ 5. It also includes the corresponding covariances of UKF and EKF estimates at each second. the EKF estimate is not converging to the true state as shown in Fig. the UKF estimates are still able to converge to the true states because of the measurement correction and also due to the increase of the size of the covariance until the estimates converge to the true states. The UKF estimate converges to the true state and stays with it. The UKF is robust to the choices of the P x0 . The choice of P x0 is reasonable here as the initial state estimate is far from the true initial state.1.1. the unstable limit cycle is considered and the initial state is chosen to be just inside the limit cycle and the initial state estimate is considered outside the limit cycle to check the convergence of the UKF and EKF estimates.2007. The initial state estimate ðb x 0 Þ is chosen to be outside the limit cycle. 4 shows the phase portraits of the UKF and EKF estimates for the ﬁrst 5 s. Further it may be observed that the later covariances for the EKF are smaller than for the UKF. doi:10. Compared to Figs. The covariances of the UKF and the EKF decrease a lot between from t = 0 and t = 1 s. Kandepu et al. Fig. Rv ¼ k p I 2 À3 T ð15Þ and Rn ¼ k n I 2 .. and the initial state covariance matrix ðP x0 Þ. Unstable limit cycle The unstable limit cycle (Van der Pol oscillator in reverse time) is represented by the following state diﬀerential equations [12] x1 ¼ À x2 . Please cite this article in press as: R. Estimated states for the Van der Pol oscillator in reverse time using UKF and EKF with an initial state estimate far from the limit cycle: large initial state covariance and small measurement noise. k p ¼ 10À3 and k n ¼ 1: ð17Þ which is just inside the limit cycle. À3 P x0 ¼ k x I 2 . First. The initial state is chosen as x0 ¼ ½ 1:4 0 . The diﬀerence can be attributed to the better nonlinear approximation by the UKF at each time step. In case of stable limit cycle.5 0 —0. Process Contr.ARTICLE IN PRESS R. Applying the unscented Kalman ﬁlter for nonlinear state estimation. l ¼ 0:2: The output vector is deﬁned as y ¼ ½ x1 x2 : T Á Á The true and estimated states using the UKF and EKF are shown in Fig. On the other hand. 3.1. 5 and 6 show the plots and phase portraits corresponding to the choices of a small P x0 and a large Rn. drawn according to 1 x k Þ ¼ 1g: fxjðx À b xk Þ P À xk ðx À b T The system is discretized with a sampling interval of 0. T The covariances may also give an idea of the distribution of the sigma points around the mean at a given time instant in the case of the UKF.5 0 5 10 15 20 25 30 35 40 time (sec) 5 4 3 true UKF EKF x 2 2 1 0 —1 —2 0 5 10 15 20 25 30 35 40 time (sec) Fig. J. Rv and Rn compared to the EKF as is illustrated with the following choices. Even though the small P x0 is a bad choice for the considered b x0. x2 ¼ Àlð1 À x2 1 Þx2 þ x1 .5 1 true UKF EKF x 1 0. 4. 3.5 —1 —1. both the UKF and EKF 2 1.1016/j. 4. (2008).1.004 . k p ¼ 10 and k n ¼ 10 : Figs. 5. Rv and Rn are chosen as b x0 ¼ ½ 0 5 . whereas the EKF estimate could not converge to the true state as the time progresses. 5 and 6. / Journal of Process Control xxx (2008) xxx–xxx 7 converges to zero as time progresses. any non-zero initial state converges to a stable limit cycle. Kandepu et al. Process noise with a covariance of 10À3I2 and measurement noise with a covariance of 10À3I2 is added to the system states and measurements.jprocont. k x ¼ 10À2 .11. These noise characteristics are used throughout Section. respectively.

Rv and Rn) as given in (16) and (17).23 and 0. 6. 4. the Kalman gain is quite diﬀerent in the two estimators due to the diﬀerence in the covariances as can be observed in Fig. In Figs.5 2 –2 –1. / Journal of Process Control xxx (2008) xxx–xxx UKF EKF x 3 2 1 0 –1 –2.2007.4.18 and 0. In addition.5 2 2. estimates converge very fast in Figs. The MSEs corresponding to the UKF Please cite this article in press as: R. The corresponding mean square errors (MSE)s of the state estimate errors for the EKF and UKF are 0. until 5 s. J. 5 and 6.5 1 1. Both states of x0 and b x 0 were chosen from a normal distribution with zero mean and standard deviation 0.004 .11. Process Contr.5 –1 –0. 3 and 4. large initial state covariance matrix and small measurment noise and with covariances plotted at each second. even though the measurement noise is higher and the initial covariance is a bad choice.ARTICLE IN PRESS 8 8 7 6 5 4 R. and 0.02. two Monte Carlo simulations with 100 runs are performed varying the initial state. The reason is that the assumed measurement noise is less in Figs. doi:10. In addition the initial state estimate is also varied. Phase portraits of UKF and EKF estimates with an initial state estimate far from limit cycle. respectively. the UKF estimate is able to converge quite fast.09. it is reasonable to assume that the correction step in the UKF also improves convergence as compared to the EKF. Kandepu et al. and hence the feedback in the measurement-update makes the estimates converge quite fast. 3 and 4. To verify these results. and the process and measurement noise realizations.5 6 5 4 3 2 UKF EKF 2 x 1 0 –1 –2 –3 –4 –4 –3 –2 –1 x1 0 1 2 3 Fig. Hence.1016/j. As the UKF gives a better approximation in timeupdate step.5 x1 0 0. the assumed measurement noise is higher meaning that the feedback from the measurement-update is less eﬀective. (2008).. The Monte Carlo simulations diﬀer in the choices of the tuning parameters (P x0 . Applying the unscented Kalman ﬁlter for nonlinear state estimation. Kandepu et al.jprocont.

1016/j. small initial state covariance matrix and large measurment noise covariance. Stable limit cycle – with model error In [13]. Phase portraits of UKF and EKF estimates with an initial state estimate far from limit cycle. it is observed that when there is a signiﬁcant model error the EKF is not able to converge to the true states in case of the Van der Pol equation with a stable limit cycle. (2008). are considerably lower compared to that of EKF. This cycle is considered to evaluate the robustness of the UKF to model errors and it is compared to that of EKF.1. doi:10. Kandepu et al. Kandepu et al. 6. The state and measurement equations are x1 ¼ x2 . until 5 s. small initial state covariance matrix and large measurment noise and with covariances plotted at each second.004 . J. x2 ¼ lð1 À y ¼ ½x1 Á Á ð18Þ x2 : x2 1 Þx2 T À x1 .2007. ð19Þ ð20Þ Please cite this article in press as: R. l ¼ 0: 2. 5. / Journal of Process Control xxx (2008) xxx–xxx 4 2 9 true UKF EKF 1 x 0 –2 –4 0 5 10 15 20 25 30 35 40 time (sec) 4 2 true UKF EKF x 2 0 –2 –4 0 5 10 15 20 25 30 35 40 time (sec) Fig. 4.2.ARTICLE IN PRESS R.. Estimated states for Van der Pol equation using UKF with an initial state estimate far from limit cycle.jprocont. Process Contr.11. Applying the unscented Kalman ﬁlter for nonlinear state estimation. and hence it can be concluded that the UKF gives improved performance compared to the EKF. 6 5 4 3 2 UKF EKF 2 x 1 0 –1 –2 –3 –4 –4 –3 –2 –1 x 0 1 2 3 1 Fig.

11. 7 and 8. T An initial state is chosen as x0 ¼ ½ 0:5 0 . From Figs. k1. doi:10. new EKF-based algorithms are developed for the considered estimation problem as the EKF estimator cannot produce eﬀective results. kp and kn values. 7. The inputs.2007. it is clear that the EKF estimates are sensitive to the model error in l. Á Á Á Á Á The initial estimated state is assumed to be b x 0 ¼ ½ 0:5 P x0 ¼ k x I 5 . Kandepu et al. A initial state estimate ðb x 0 Þ.5. The model is discretized with a sampling interval of 0. The reason for the better performance of the UKF is attributed true UKF EKF 4 2 x 0 –2 0 5 10 15 20 25 30 35 40 45 50 1 time (sec) 3 2 1 true UKF EKF x2 0 –1 –2 –3 0 5 10 15 20 25 30 35 40 45 50 time (sec) Fig. The errors in the estimated states in polar coordinates are shown in Fig. The UKF and EKF are used to estimate the states with the actual initial condition x 0 ¼ ½ 0: 2 À0:6 À0:4 0: 1 0: 3 : T The estimates states using UKF and EKF are compared in Fig. The state space model for a symmetrical three-phase induction machine is [4]. k p ¼ 10 and k n ¼ 10À2 : ð22Þ The state estimation results are shown in Fig. the simulation results show that the UKF performs better than the EKF.1.k6 are parameters depending on the considered drive. Applying the unscented Kalman ﬁlter for nonlinear state estimation. the model parameters and inputs are set to the values given in Table 1.x4 are the components of the stator and the rotor ﬂux. To develop high-performance induction motor drives. 4.1016/j. . / Journal of Process Control xxx (2008) xxx–xxx In the state estimator model the value of l is chosen as 0. x3 ¼ k 3 x 1 þ k 4 x3 þ ðz1 À x5 Þx4 . x1 ¼ k 1 x 1 þ z1 x 2 þ k 2 x 3 þ z2 . . the frequency and the amplitude of the stator voltage are denoted by z1 and z2 respectively. it may be necessary to have rotor ﬂux estimators and rotor time constant identiﬁcation schemes [3]. respectively. J. 9 where the UKF and EKF estimates of states x1. x5 ¼ k 5 ðx1 x 4 À x 2 x 3 Þ þ k 6 z3 . .ARTICLE IN PRESS 10 R. T k x ¼ 1. Process Contr. 7. Rv ¼ k p I 5 and À4 Rn ¼ k n I 2 . with parameters k7 and k8. respectively. y 1 ¼ k 7 x1 þ k 8 x3 . For simulation. Process noise with a covariance of 10À4I5 and measurement noise with a covariance of 10À2I2 is added to the system states and measurements. Kandepu et al.1. 0: 1 0:3 À0:2 4 T : The tuning parameters are selected as. k p ¼ 10À3 and k n ¼ 10À3 : ð21Þ where x1. In [3] and [4]. and x5 is the angular velocity.jprocont. b x0 ¼ ½ 5 À1 . Rv and Rn are chosen as before with the following kx.2. x4 ¼ k 3 x2 À ðz1 À x5 Þx3 þ k 4 x4 ..004 . x3 and x5 are compared with the true states. and initial state covariance matrix ðP x0 Þ. Comparison of state estimates using UKF and EKF with a model error. The system is discretized with sampling interval of 0. . All the state variables are normalized. The output equations are given by. Please cite this article in press as: R. The outputs are the normalized stator currents y1 and y2. y 2 ¼ k 7 x2 þ k 8 x4 . x2 and x3. with k x ¼ 1. x 2 ¼ À z1 x 1 þ k 1 x 2 þ k 2 x 4 . whereas the UKF shows robust performance to the model error. and the load torque is denoted by z3. With the initial predicted state far from the actual initial state. State estimation in an induction machine The UKF and EKF are applied to a highly nonlinear ﬂux and angular velocity estimation problem in an induction machine. 8. (2008).

with stoichiometric matrix v ¼ ½ À2 1 and reaction rate r ¼ kC 2 A: The state and measurement vectors are deﬁned as CA x¼ . 4.225 k4 À0. 10.05. the model for this system is x ¼ f ð x Þ ¼ vT r : The system is discretized with sampling interval of 0. both the UKF and EKF estimates become negative. 0 36 0 1 0 À2 1 Rn ¼ 10 : 0 1 Á k ¼ 0:16.5 0 –0. Process Contr.ARTICLE IN PRESS R. Because of the clipping in the Please cite this article in press as: R. this time in the case of a highly nonlinear semi-realistic example. The MSE corresponding to the UKF is substantially lower compared to those of the EKF. However. (meaning negative concentrations) which is not possible physically. From Fig. reversible reaction.5 1 0. Applying the unscented Kalman ﬁlter for nonlinear state estimation. doi:10. Table 1 Model parameters and inputs k1 À0. CB The estimation result for the unconstrained case is shown in Fig. It is assumed that the ideal gas law holds and that the reaction occurs in a well-mixed isothermal batch reactor. J. State constraints are incorporated according to the proposed method in Section 3 for the UKF and standard ‘‘clipping” is used for the EKF.1. The results are shown in Fig. The UKF and EKF are used for state estimation. 36 0 1 0 .004 . Again the UKF performs signiﬁcantly better.178 k3 0. k where Cj denotes the concentration of species j.5 0 5 10 15 20 25 30 35 40 45 50 UKF EKF time (sec) Fig.3.234 k5 À0.13 and 0. Errors in amplitude and phase of the UKF and EKF state estimates.081 k6 4. The result shows that the dynamic performance of the UKF estimates is better compared to that of the EKF. 11 the UKF estimates converge to the true states without violating the constraints. Kandepu et al. with the following setup as used in [14]: T T x0 ¼ ½ 3 1 . during the dynamic response. A Monte Carlo simulation with 100 runs was performed and the MSEs of the EKF and UKF estimates are 0.643 k7 À4.186 k2 0. b x 0 ¼ ½ 0:1 4:5 . from ﬁrst principles. 2 A ! B. respectively.11.jprocont. 11. Then. Rv ¼ 10À6 P x0 ¼ . The EKF performance is very similar to the reported results in [14]. Kandepu et al. State estimation with constraints: a reversible reaction example We will here consider an example to illustrate the constraint handling capability of the UKF compared to that of the EKF. 8. It can be mentioned that the behavior of the EKF is similar to behavior observed in [3].. / Journal of Process Control xxx (2008) xxx–xxx 1 11 error in amplitude UKF EKF 0 –1 –2 0 5 10 15 20 25 30 35 40 45 50 time (sec) error in phase (radians) 1.448 k8 1 z1 1 z2 1 z3 0 to the better nonlinear approximation at each step.5 –1 –1. y ¼ ½ 1 1 x.1016/j. Consider the gas-phase.2007. (2008).

12 shows the phase portraits of the unconstrained and constrained UKF estimates for the ﬁrst 4 s. 12. is an interesting power generation system due to its high eﬃciency in the range of 65–75%. Comparison of estimated states: components of stator ﬂux. Applying the unscented Kalman ﬁlter for nonlinear state estimation.2007. it is clear that it takes longer time for the unconstrained estimate to converge as the corresponding covariances do no include the constraint information. 4. Comparison of estimated states: no constraint handling. natural gas (CH4).4. is partially steam reformed in a pre-reformer before Please cite this article in press as: R. known as SOFC/GT hybrid system. (2008).. Process Contr.5 0 5 10 15 20 25 30 35 40 45 50 time (sec) true UKF EKF x 3 –0. The ﬁgure also includes the corresponding covariances plotted at t = 0. 13. A schematic diagram of the SOFC/GT hybrid system integrated in an autonomous power system is shown in Fig.ARTICLE IN PRESS 12 R.004 . Kandepu et al. 1 and 3 s. The constrained UKF estimate converges faster as the covariances decrease faster.5 0 0.5 –1 0 5 10 15 20 25 30 35 40 45 50 time (sec) 6 4 true UKF EKF 5 x 2 0 0 5 10 15 20 25 30 35 40 45 50 time (sec) Fig.jprocont. 9. The fuel. doi:10. Kandepu et al.5 x 1 0 –0. / Journal of Process Control xxx (2008) xxx–xxx true UKF EKF 0. 10. SOFC/GT hybrid system The solid oxide fuel cell (SOFC) stack integrated in a gas turbine (GT) cycle. CA estimate of EKF did not converge to the true state and the estimate of CB takes much longer time to converge to the true state. From Fig. EKF. which include the constraint information. Fig.1016/j.11. 30 true UKF EKF 20 CA 10 0 0 1 2 3 4 5 6 7 8 9 10 time (sec) 5 0 —5 true UKF EKF CB —10 —15 —20 —25 0 1 2 3 4 5 6 7 8 9 10 time (sec) Fig. J. The results from this example con- ﬁrm that the proposed constraint handling method is promising.

All the models of the system are developed in the modular modeling environment gPROMS [7] as reported in [5]. In [5]. The hot stream from the combustor is expanded using a high pressure turbine (HPT) which drives the compressor. The DC power from SOFC stack is inverted to AC using an inverter. J. Applying the unscented Kalman ﬁlter for nonlinear state estimation.1016/j. doi:10. Kandepu et al.004 . Comparison of state estimates: state constraint handling. control relevant SOFC model is evaluated against a detailed model developed in [6]. a low complexity. / Journal of Process Control xxx (2008) xxx–xxx 3 2 true UKF EKF 13 CA 1 0 —1 0 1 2 3 4 5 6 7 8 9 10 time (sec) 4 true UKF EKF CB 3 2 1 0 1 2 3 4 5 6 7 8 9 10 time (sec) Fig. it enters the SOFC anode.ARTICLE IN PRESS R. 12 UKF unconstrained UKF constrained 10 8 6 x2 4 2 0 —2 —6 —4 —2 0 x1 2 4 6 8 Fig. 12. The inverter and the alternator are connected to the electric load through a bus bar. 11. Air is compressed and preheated in a heat exchanger before entering the SOFC cathode.. Phase portraits of the UKF unconstrained and constrained estimates with covariances.jprocont. A part of anode ﬂue gas is recycled to supply the necessary steam required for the steam reformation in the pre-reformer.2007. Process Contr. (2008). Typically 60–70% of the total power is supplied by the SOFC stack. Kandepu et al.11. The comparisons indicate that the low complexity model is sufﬁcient to approximate the important dynamics of the Please cite this article in press as: R. The HPT ﬂue gas is expanded to atmospheric pressure using low pressure turbine (LPT) which drives an alternator. The remaining part of the anode ﬂue gas is supplied to a combustor where the unused fuel is burnt completely in presence of oxygen coming from the cathode ﬂue gas.

SOFC-GT hybrid system integrated in an autonomous power system. In several cases the improvement is substantial. 13. In this section. The results albeit for one example are promising. The state estimation methods have been Table 2 Available measurements No. Process Contr. The UKF shows consistently improved performance as compared to the EKF.ARTICLE IN PRESS 14 R. in addition to providing better performance. The initial state estimate is diﬀerent from the actual initial state and from the simulation results it can be concluded that UKF estimate converges to the true states very fast compared to the EKF estimate. This is quite different from the EKF where ‘clipping’ usually is performed without adjusting the state estimate covariance.2007. Our results indicate that the UKF with the proposed constraint handling may be an interesting alter- Please cite this article in press as: R. EKF and UKF are used to design the state estimator. Applying the unscented Kalman ﬁlter for nonlinear state estimation. and 14 measured outputs which are listed in Table 2. the EKF. / Journal of Process Control xxx (2008) xxx–xxx I dc DC AC I ac ∠0 Vac ∠0 Inverter Ejector Splitter Anode Bus bar XT V t ∠ δt I t ∠φt Fuel Exhaust Air blow-off Mixer Reformer Electrolyte Cathode Combustor SOFC stack Heat exchanger shaft I R ∠φR I RL ∠φRL R I RC ∠φRC RL I IK∠ φIK RC XC I PK∠ φ PK PPK QPK I m∠ φm Pd Qd Compressor Air HPT I g ∠φ g Pm XL LPT Generator Vg ∠0 Fig. A rationale for using the MHE is often attributed to its constraint handling ability.. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Measurement Pre-reformer temperature (K) Shaft speed (rad/s) Heat exchanger hot stream temperature (K) Heat exchanger cold stream temperature (K) SOFC outlet temperature (K) Combustor outlet temperature (K) Fuel mass ﬂow rate (kg/s) Anode recycle ﬂow rate (kg/s) Flow to the combustion chamber (kg/s) Air blow-oﬀ ﬂow rate (kg/s) Air mass ﬂow rate (kg/s) SOFC current (A) SOFC voltage (V) Generator power (kW) compared using the same tuning parameters to make the comparative study as credible as possible.004 . J. In the EKF. both in the convergence rate as well as in the long term state estimation error. and their performances and computational load are compared. The purpose of a state estimator is to design an advanced controller and as well for monitoring and fault diagnosis purposes. Both the state estimators are tuned to obtain good performance. and the simulation results are presented in Figs. A method for constraint handling in UKF is proposed in Section 3. These runs reinforced the hypothesis that the UKF is indeed an interesting alternative to the EKF. the Jacobians are calculated numerically at each time step. A regulatory control layer is designed and it is included in the model. Kandepu et al. 5. From this example. The ﬁgures show selected states for a typical run. An important feature of the proposed constraint handling method is the fact that it adjusts the covariance in addition to the state estimates themselves. doi:10. The computational load of the UKF is almost the same as that of the EKF. Monte Carlo simulations have also been run with diﬀerent initial states.1016/j. Four diﬀerent examples have been tested. Discussion The UKF as a tool for state estimation has been compared to the standard method for nonlinear state estimation. The SOFC/GT hybrid system model has 3 inputs. SOFC and can hence be used for operability and control studies.jprocont.11. (2008). 14 and 15. Kandepu et al. it is clear that the UKF can be used to design state estimators for higher order systems and the UKF performance is again favorable. 18 states. A state estimator of the hybrid system is designed using both the UKF and EKF. The tuning parameters for both the UKF and EKF are the same to ensure a fair comparison. The initial state estimate cannot be chosen too far from the actual state to avoid numerical problems in gPROMS.

The latter implies that the computational load using the UKF is signiﬁcantly lower than what can be expected for the MHE. In this case the computational load for the UKF is comparable to the EKF. 7462 1350. of CO 0 10 20 30 40 50 2 0. J. (2008). There is usually a price to pay for improved performance. In many cases. doi:10. 15.11.1 1350.0166 0 time (sec) Conc.1016/j. Applying the unscented Kalman ﬁlter for nonlinear state estimation.1015 0.98 9..04 0. Please cite this article in press as: R.043 0 10 20 30 40 50 time (sec) time (sec) Fig. particularly for higher order systems. Conclusions This paper shows that the UKF is an interesting option to the EKF because of improved performance. native to MHE.06 1350. of H2 in anode 10.0168 50 0. it is very simple to implement the proposed method to incorporate the constraints in the UKF compared to the MHE.0172 0. 14.0432 0. Kandepu et al.94 0 10 20 30 40 true UKF EKF Conc.ARTICLE IN PRESS R.96 9. Kandepu et al. Process Contr.101 0.04 1350. The improved performance of the UKF compared to the EKF is due to two factors. the increased time-update accuracy and the improved covariance accuracy. prereformer and heat exchanger cold stream temperatures. Further.jprocont.12 true UKF EKF SOFC shaft speed (rad/s) 7460 7458 7456 7454 7452 7450 0 10 20 30 40 50 1350.14 1350. Furthermore. Comparison of state estimation using UKF and EKF: concentration of O2 in cathode and concentrations of H2.004 .0434 0. of O in cathode 10. of H O in anode true UKF EKF time (sec) in anode 0.102 2 Conc.08 1350.0436 true UKF EKF 0.02 0 10 20 30 40 true UKF EKF T 50 time (sec) 913 true UKF EKF 1210 1205 time (sec) 912 THX.02 10 9. Comparison of state estimation using UKF and EKF: Compressor shaft speed and SOFC. the Jacobians for the EKF are computed using ﬁnite diﬀerences. 6. In this case the computational load increases when moving from the EKF to the UKF if the Jacobians are computed analytically. / Journal of Process Control xxx (2008) xxx–xxx x 10 –3 15 Conc. H2O and CO2 in anode of SOFC. This again makes a diﬀerence in terms of diﬀerent Kalman gains in the measurementupdate equation and hence the eﬃciency of the measurement-update step. cold 1200 1195 true UKF EKF Tref 911 910 0 10 20 30 40 50 1190 0 10 20 30 40 50 time (sec) time (sec) Fig. The covariance estimation can be quite diﬀerent for the two ﬁlters as shown in Fig.017 true UKF EKF 10 20 30 40 50 2 0. 6.2007.

63–68. T. Applying the unscented Kalman ﬁlter for nonlinear state estimation. J. J. Cramado. Simultaneous registration and fusion of multiple dissimilar sensors for cooperative driving. C. Kalman ﬁltering with state equality constraints. due to the proposed constraint handling method it may also be an interesting alternative to the MHE. Kandepu et al. The ﬁnancial support from The Gas Technology Center.B. Stochastic Processes and Filtering Theory. Khalil. Romanenko. Julier.S. Leung. A critical evaluation of extended kalman ﬁltering and moving horizon estimation. (2008). Van der Merwe. S. State estimation of induction motor using unscented Kalman ﬁlter. 2004. L.1016/j.L. F. Brazil.A. Automatica 34 (9) (1998) 1119–1123. Springer. Mexico. Rawlings. Constrained state estimation for nonlinear discrete-time systems: stability and moving horizon approximations. [19] S. Imsland. 2000. Please cite this article in press as: R. Rawlings. [18] D. in: Round Tables on Non-linear Model Identiﬁcation. [10] G. Process Systems Enterprise Ltd.B. Akin. Process Contr. Modeling and control of a SOFC-GT based autonomous power system. S. Li. Prentice Hall. The computational load for the UKF is comparable to the EKF for the typical case. D. [14] E. Mathematics in Science and Engineering. in: Proceedings of 8th International IFAC Symposium on Dynamics and Control Process Systems. [5] R. Further the ﬁnancial support from the NFR BIGCO2 project is also acknowledged. B.A. IEE Proceedings – Control Theory and Applications 147 (3) (2000) 257–266. TexasWisconsin Modeling and Control Consortium (TWMCC). Stiller. vol. A new EKF-based algorithm for ﬂux estimation in induction machines. 1970. Journal of Power Sources 141 (2005) 227–240. U. B. Aydin Ersak. R.11. B. Acknowledgements The authors are grateful to Professor Biao Huang for discussions on the UKF. Finitevolume modeling and hybrid-cycle performance of planar and tubular solid oxide fuel cells.jprocont. [12] H. Simon. Cancun. [9] B. where the Jacobians are computed numerically. Linear System Theory and Design.H. April 6–7. Thorud. Orguner. References [1] S.V. Unscented ﬁltering and nonlinear estimation. IEEE Transactions on Industrial Electronics 40 (5) (1993) 496–504. Rao. Mayne. Proceedings of the IEEE 92 (2004) 401–422. Ungarala. [2] C. Castro. The proposed method is much simpler to implement compared to the MHE. The Ensemble Kalman Filter. vol. Academic Press.L. Jazwinski. Li.004 . Data Assimilation. Haseltine. Evensen. 2004. 1999.Q. O. [6] C.ARTICLE IN PRESS 16 R. Computer and Chemical Engineering 28 (2004) 347–355. [13] M. IEEE Transactions on Automatic Control 48 (2003) 246–258. Karoliussen. Salvatore. [7] gPROMS.K. Tarchioni. O. Stasi. [15] A. 64. 2002–2003. Foss. [16] B.. Kandepu et al. Sonnemann. Oregon Health and Science University.T.2007. IEEE Transactions on Aerospace and Electronic Systems 38 (2002) 128–136. J. / Journal of Process Control xxx (2008) xxx–xxx [8] R. J. NTNU-SINTEF and Norwegian Research Council is acknowledged.K. [17] C. Technical Report. [4] K. E. L. 2006. Bolland. Berlin. K. The unscented ﬁlter as an alternative to the EKF for nonlinear state estimation: a simulation case study.H. Chia. Ahmed. [3] L. Sigma-Point Kalman Filters for probability inference in dynamic state-space models. PhD Thesis. Wang. IEEE Transactions on Intelligent Transportation Systems 5 (2004) 84–98. [20] W. Reif. H. doi:10. Stiller. Journal of Energy 2006 36 (2007) 406–417. J. Uhlmann. Nonlinear Systems Third Edition. Design of dynamic numerical observers. Thorud. Kandepu. gPROMS introductory user guide. H. Bolland. New York and London. pp. An EKF-based nonlinear observer with a prescribed degree of stability. Oxford University Press.M. Unbehauen. S. Q. Seljbeo. Constrained extended kalman ﬁlter for nonlinear state estimation. Overview of emerging Bayesian approach to nonlinear system identiﬁcation. International Workshop on Solving Industrial Control and Optimization Problems. Huang. New York. 2. IEEE Transactions on Control Applications 2 (2003) 915–919. [11] A. Dolence. 2007. Chen. June 2007. Riyaz.

- HVDC_LCC
- Modeling Unknown Circuit Breakers
- A novel approach for load flow
- 39 New England Power Network
- IguanaTex
- Nonlinear Local Load for PFC
- VSC TRANSMISSION TECHNOLOGIES.pdf
- Venable k Factor
- Bose GridComms Overview Chapter
- Park’s Transformation for Space Vector Control
- 07006792
- 07140832
- DC voltage droop control design for multi-terminal HVDC systems considering AC and DC grid dynamics
- 2012 PES Power
- Advanced Vector Control for Voltage Source Converters Connected to Weak Grids
- PowerFactory Product Manual
- PF_14_0_Tutorial_EN
- Decentralized Control
- Phasor-Only State Estimation
- ABB_US8606416B2.pdf
- Optimization Tools
- gammab
- The Community in Optimization Lecture
- SOP For Distribution
- Hypersim Brochure Lores

Applying the unscented Kalman ﬁlter for nonlinear state estimation

Applying the unscented Kalman ﬁlter for nonlinear state estimation

- System Identification of a Steam Distillation Pilotscale
- 2
- HVDC Power Injection
- Controller Design for Hysteresis Compensation of Piezo Actuators for MicroNano Positioning
- Control RGB Jurnal
- 06701388
- On-load flow control
- Full Text
- Mohanty VLSI Integration 2012Jan SRAM
- Montecarlo
- 05772000.pdf
- Content
- 55
- Model Predictive Heuristic Control
- h-tree.pdf
- Induced Voltage on Pipelines by OHL
- 10.1.1.127
- Mac Martin 96
- MEPowerManagement
- syl 2
- 00067530.pdf
- 05673966[1]
- 02Whole.pdf
- An Experimental Approach for Assessing the Harmonic Impact of Fast Charging Electric Vehicles on the Distribution Systems
- Transitioning to Physics-of-Failure as a Reliability Driver in Power Electronics
- EPSM
- M.tech(Electrical Power Systems)
- Battery Energy Storage Station (BESS)-Based Smoothing Control of Photovoltaic (PV) and Wind Power Generation Fluctuations
- Conference-201304-Chun-Liang Liu-Design and Implementation of a Digitally-controlled LLC Resonant Converter for Battery Charging Applications

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd