A Decentralized Observer for Ship Power System Applications: Implementation and Experimental Validation

440

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 2, FEBRUARY 2010

A Decentralized Observer for Ship Power System Applications: Implementation and Experimental Validation Andrea Benigni, Student Member, IEEE, Gabriele D’Antona, Senior Member, IEEE, Ugo Ghisla, Antonello Monti, Senior Member, IEEE, and Ferdinanda Ponci

Abstract—In the last few years, the growing complexity of the electrical power networks, mainly due to the increased use of electronic converters together with the requirements of a higher level of reliability and security, pushed the development of new techniques for the state estimation of the power systems. In this paper, the authors focus their attention on the implementation and experimental validation of a decentralized observer for the state estimation in an electric ship, whose power network is characterized by fast dynamics and by the presence of many electronic devices. The proposed solution implements a decentralized information filter (DIF). Index Terms—Decentralized information filter (DIF), decentralized observer, state estimation. Fig. 1.

I/O of a state estimator for an electrical power system.

I. I NTRODUCTION

S

TATE estimators have a critical role in power systems by allowing the implementation of all the necessary controls for the safe and reliable operation of the network. Traditionally, a state estimator for electrical power systems has been conceived as an element of a central control system receiving as input measured data gathered from the network and stored in a central database (see Fig. 1). On the basis of the state estimate, decisions concerning power system management are taken in the central control system [1], [2]. A typical state estimator must be capable of dealing with the following: 1) uncertainties on telemetered data; 2) uncertainties on power system parameters; 3) bad or missing data due to network transients and metering/communication failure; 4) errors in the network topology due to wrong switch status information. Manuscript received June 30, 2008; revised April 22, 2009. First published September 1, 2009; current version published January 7, 2010. This work was supported in part by the U.S. Office of Naval Research under Grants N00014-02-1-0623 and N00014-07-1-0603. The Associate Editor coordinating the review process for this paper was Dr. Y. Rolain. A. Benigni, A. Monti, and F. Ponci are with the E.ON Energy Research Center, RWTH Aachen University, 52056 Aachen, Germany (e-mail: [email protected]; [email protected]). G. D’Antona is with the Department of Energy, Politecnico di Milano, 20133 Milano, Italy. U. Ghisla is with the Department of Electrical Engineering, University of South Carolina, Columbia, SC 29208 USA. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TIM.2009.2024695

To enhance the performance and reliability of the processing effort required from a control system, a decentralized approach to the problem can be taken into consideration. This perspective, leading to the concept of a decentralized control system (versus the central control system), brings together new paradigms that are particularly related to decentralized decision and estimation functions ancillary to the control systems [3], [4], [6]. In this paper, we will focus our attention on the decentralized state estimation functionality, with reference to applications concerning ships and aircraft, characterized by well-characterized load dynamics. Furthermore, a novelty of the approach proposed here concerns the formalization of the estimation problem. Classical power system estimators support their robustness versus bad data and uncertainties on redundancies: The number of measurement data is far larger than the number of state variables, and the estimation is based on the minimization of some norm of a residual. Usually, the state estimation of this system is based on a static model [1], [2]. Because of this, the estimator can be used only if the system is considered quasi-static. As it will be clearer in the next paragraph, in the approach proposed here, the estimation is based not only on the measurement data but also on a dynamic model of the power grid (either numerical or analytical), based on the predictability of many electrical loads. This model allows the relaxation of the redundancy condition on the number of measurement data, increasing the reliability of the estimation task against failure of some metering and/or data communication channels. Furthermore, due to the introduction of electronic power converters in the last few years, the network can no longer be

0018-9456/$26.00 © 2009 IEEE

BENIGNI et al.: DECENTRALIZED OBSERVER FOR SHIP POWER SYSTEM APPLICATIONS

assumed to be quasi-static, and the implementation of a dynamic model has become necessary. Moreover, the use of converters has introduced a significant increase in the complexity of the network. Given this complexity, a centralized system becomes difficult to deploy and manage. These problems have required the development of new solutions, based on distributed [3], [4] and decentralized systems [5], [6]. In this paper, we present a first implementation of a decentralized information filter (DIF) [8] for the state estimation on board an electric ship, whose power network is characterized by fast dynamics and by the presence of many electronic devices. The proposed solution involves the use of several decentralized estimators distributed along the network, each of which makes an estimate for the subnetwork and communicates with all the others. In this type of approach, the estimation is based not only on the measurements but also on a dynamic model of the network. The main characteristics of this solution are given as follows: 1) increased computational efficiency, due to the distribution of the computational burden among the estimators; 2) increased reliability of the estimator, due to the distribution of the resources; 3) scalability of the estimation system. In the following, the results obtained from a physical implementation of this system are also shown. The experimental activity has been designed to particularly analyze the following problems: 1) implementation of real-time software; 2) target synchronization; 3) communication. Preliminary experimental results were introduced in [9]. Here, the results are extended and used to support a theoretical analysis of the timing of the procedure. This analysis is supposed to provide references for the design of a realistic application, where the number of state variables can significantly be bigger than that in the laboratory setup. Tradeoffs between the computational cost and the communication cost are discussed.

441

tion of a quasi-static ac condition is becoming incorrect. For example, the growing role of dc transmission is calling for approaches that are able to support ac and dc estimation at the same time [16], [17]. HVDC Light drastically changed the approach to dc transmission. A lot of the limitations of the traditional HDVC are solved by HVDC Light: A higher level of controllability and flexible active/reactive power control are among the main advantages given by this new technology [18]–[20]. These technical advantages, together with economical and social advantages, suggest that, in the future, dc transmission will be more and more used [21], [22]. Under the condition of an integrated ac–dc transmission system, the use of PMUs becomes inadequate. As result, a technique such as the one proposed in this paper seems to be a reasonable solution not only for ship power systems but also for future terrestrial power systems. III. T HEORETICAL B ASIS OF THE DIF For the DIF, each observer uses a dynamic model of the system, as well as for the decentralized Kalman filter. The forecasts of each observer are corrected based on locally performed measurements and on information exchanged with other areas. Consider a discrete-time linear power system model described in the standard linear form xi (k) = Ai (k)xi (k − 1) + Bi (k)ei (k − 1) + wi (k − 1) where xi (k) is the state (independent voltages and currents in reactive elements of the grid) at time k, ei (k) is the driving input vector, and wi (k) is the model noise input modeled as a Gaussian independent identically distributed (i.i.d.) random process with zero mean and E[wi (k)wi (j)] = Qi (k)δ(k − j). The index i refers to the ith partition (zone) of the considered power system. The power system is observed by measuring branch voltages and currents at area i modeled according to the following observation equation, derived on the basis of Ohm’s and Kirchhoff’s laws: zi (k) = Ci xi (k) + Di ei (k) + vi (k)

II. S TATE OF THE A RT AND C OMPARISON State estimation is a critical element of every power system. The most recent developments in this area have been related to the application of synchronized phasor measurement units (PMUs) [10]. These developments are mostly related to a significant number of installations that have been made since 1980s, which now makes the PMU a mature technique [11]–[14]. At the same time, PMUs have some limitations: The main limitation is the assumption to be in a quasi-steady state with very slow variations of voltage and frequency. Even if a lot of problems such as frequency variations and fault detection were widely analyzed in literature, this kind of approach does not seem adaptable for the estimation of power systems with fast dynamics [15]. These considerations make the solutions based on PMUs unpractical for our application. It should also be mentioned that, for terrestrial power systems, the assump-

where zi (k) is the observation vector at time k, and vi (k) is the associated noise modeled as a Gaussian i.i.d. random sequence with E[v(k)v(j)] = Ri (k)δ(k − j). As outlined in [8] and [9], given the state vector xi (k) and the covariance matrix of its estimate PA i (k), the DIF requires the introduction of two new variables, which are denoted by the information matrix Yi (k) and the information state yi (k), i.e.,  −1 Yi (k) = PA (1a) i (k)  A −1 · xi (k). (1b) yi (k) = Pi (k) The implementation of the DIF is based on the following three fundamental steps, as schematically shown in Fig. 2. A. Local Prediction In this phase, the information matrix and state are predicted by the power system model from their estimates YiA and yiA at

442

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 2, FEBRUARY 2010

Fig. 3.

Fig. 2. Block diagram of a DIF.

the previous time step, i.e.,  −1  −1 YiF (k) = Ai · YiA (k − 1) · ATi + Qi (k − 1) (2) yiF (k) = Li (k) · yiA (k − 1) + YiF (k) · Bi · ei (k − 1) (3) (4) Li (k) = YiF (k) · Ci · YiA (k − 1)−1 . It is important to underline that the variations of Ai (k) and Bi (k) are due to variations of the breaker status vector, which is composed of all the logic signals that allow one to reconstruct the topology of the power system. B. Communication At this stage, the observer of each area i communicates with the others, exchanging the matrix Ii and the vector ii obtained using the following formulas: ii (k) = CTi · Ri (k)−1 · [zi (k) − Di · ei (k)] Ii (k) = CTi · Ri (k)−1 · Ci .

(5a) (5b)

zi (k) represents the vector of measurements at zone i and is related to the state by the measurement model zi (k) = Ci (k)xi (k) + Di (k)ei (k).

(6)

C. Assimilation In this phase, each observer assimilates the data obtained from other areas with the data calculated during the locally made prediction and measurements, to obtain the following new values of YiA and yiA :  YiA (k) = YiF (k) + Ij (k) (7a) j=i

yiA (k) = yiF (k) +



ij (k).

(7b)

j=i

To implement the information filter on the grid of a ship, the Ai and Bi matrices must be obtained by describing the network using the state-space model. A criterion for the definition of the subgrids of the system must be identified, together with the selection of the quantities y to be measured for each zone.

Electric grid used for benchmarking.

The relation among the measurements and the state and input variables is expressed by means of the Ci and Di matrices. The topology discussed in this paper is a slightly simplified version of what has previously been proposed in [8]. This network is a reasonable notional power system for an electric ship application. The power system is composed of a threephase distribution line and three zones, each composed of an ac–dc converter and an RC load, in which R is variable (Fig. 3). The two inductors L1 and L2 are three-phase inductors that simulate the inductances of the ac lines of the ship. The three-phase line represents the distribution system of the ship, and each zone models the structure of an area of the ship. Each zone is connected to the main ac grid by an ac–dc converter. The main simplifications of our approach concern the realization of only three areas (a real ship will contain many more zones) and the choice of RC loads, as shown in Fig. 4 (the loads on a ship are obviously expected to be more complex). The main idea is to introduce some significant equivalent dynamics to be analyzed with the DIF. The modular structure of the chosen network has many similarities with the grid on board a ship and facilitates the implementation of a decentralized estimator. The energy buffer created by the capacitors of each converter decouples the dc area from the ac area. This decoupling represents a natural separation among the three zones. The process of partitioning a complex power system is discussed in [23]–[26]. Even if these converters are beneficial in their creation of a natural separation between zones, they represent an interesting challenge from the modeling standpoint. For example, for the implementation of the information filter, a linear model of the power system needs to be formulated in terms of state variables. A linear time-variant model approach is adopted for the converters. The model of the converter (Fig. 5) is based on voltage- and current-controlled sources in the Park domain [27]. For the aforementioned reason, a linear model of the Darlington transistor, which is used as a breaker, is also required. We chose to model the Darlington as a variable resistance, which assumes two values. 1) R = 0.5 Ω when the Darlington is on (conduction). 2) R = 10 kΩ when the Darlington is off (block). The approximation obtained with this solution introduces an error in the model. This error has been considered as the most relevant error in the model, and no other errors have been considered in the calculation of Q. The matrix Q is given in the equation shown at the bottom of the next page.

BENIGNI et al.: DECENTRALIZED OBSERVER FOR SHIP POWER SYSTEM APPLICATIONS

Fig. 4.

443

Schematic of a zone: ac–dc converter and variable load.

A. Hardware

Fig. 5.

Modeling of the ac–dc converter using the Park transformation.

The values in this matrix have experimentally been calculated by evaluating the variation of the resistance of the Darlington in the off condition. As result, the network is composed of only linear components, and the linear state-space formulation (4) and (6) can be adopted. The model of the converter is time variant because its variables are dependent on which diodes are in conduction at a given time. The time variance is reflected in the coefficient K. The choice of the correct value for K is performed by determining which diodes are conducting at every given time. This evaluation is required for the hardware implementation of the design of a custom hardware to keep the simulation synchronized with the network operation. IV. I MPLEMENTATION OF THE E XPERIMENTAL T EST B ENCH State estimators have a critical role in power systems by allowing the implementation of all the necessary controls for the safe and reliable operation of the network.

⎡

0.0225 ⎢ 0 ⎢ ⎢ 0 ⎢ ⎢ 0 ⎢ Q=⎢ 0 ⎢ ⎢ 0 ⎢ ⎢ 0 ⎣ 0 0

0 0.0225 0 0 0 0 0 0 0

0 0 0.0225 0 0 0 0 0 0

0 0 0 0.0225 0 0 0 0 0

A power system replica of the model in Fig. 3 has been built. Each zone is made of an ac–dc converter connected to a variable load, as previously shown in Fig. 4. The system has been equipped with a distributed data acquisition system with local elaboration capabilities. The hardware used for the experiment can be divided into two blocks: 1) the power section equipped with sensor and measurement processing and 2) the intelligent module for the software implementation (Fig. 6). The measurement system returns both the measurement data vector, which is composed of the measurements of voltage and current on the dc side of the converter, and the breaker status vector, which is composed of logic signals. These signals reveal the status of the load (a logic signal is associated with the control of each Darlington), the synchronization signal, and the transition to zero of the phase chosen as a reference for the system. The components of the hardware are given as follows: 1) uncontrolled ac–dc converter; 2) variable loads; 3) Variac, which models the line inductors L1 and L2. The characteristics of the ac–dc converters, the variable loads, and the line inductors are reported in Tables I–III, respectively. (Refer to Figs. 3 and 4 for the name of the components.) Two transducers are used for the measurement on the dc side, namely, LA 55-P [28] for the current and LV 20-P [29] for the voltage. The signals of the transducers are conditioned and filtered at 250 Hz. The whole system has been designed for a bandwidth of 100 Hz. The synchronization signal used to determine K is obtained by using a voltage transformer and a

0 0 0 0 0.0225 0 0 0 0

0 0 0 0 0 0.0225 0 0 0

0 0 0 0 0 0 0.0225 0 0

0 0 0 0 0 0 0 0.0225 0

⎤ 0 0 ⎥ ⎥ 0 ⎥ ⎥ 0 ⎥ ⎥ 0 ⎥ ⎥ 0 ⎥ ⎥ 0 ⎥ ⎦ 0 0.0225

444

Fig. 6. units.

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 2, FEBRUARY 2010

Experiment hardware. (a) One zone of the power section: converter and load. (b) Control section. (c) Power section overview: three zones and two Variac TABLE I CHARACTERISTICS OF AC–DC CONVERTERS

comparator. The signal is then filtered at 60 Hz to remove any spikes due to switching. The phase error introduced by the filter and the transformer is tabulated for frequencies between 58 and 62 Hz and is corrected in the software. The uncertainty of the transducers has been considered as the biggest contribution to the measurement noise; thus, only this

contribution has been considered. The measurement noise has been considered to be uncorrelated. As explained in Section II, the measurement noise has been modeled as a Gaussian i.i.d. random sequence with   E v(k)v(j)T = δkj R(k).

BENIGNI et al.: DECENTRALIZED OBSERVER FOR SHIP POWER SYSTEM APPLICATIONS

445

TABLE II CHARACTERISTICS OF THE VARIABLE LOADS

TABLE III CHARACTERISTICS OF THE LINE INDUCTORS

The standard deviation of the measured voltage is 0.12 V, and the standard deviation of the measured current is 0.07 A. The matrix R becomes ⎡

⎤ 0.0144 0 0 0 0 0 0.0049 0 0 0 0 ⎥ ⎢ 0 ⎢ ⎥ 0 0.0144 0 0 0 ⎥ ⎢ 0 R=⎢ ⎥. 0 0 0.0049 0 0 ⎥ ⎢ 0 ⎣ ⎦ 0 0 0 0 0.0144 0 0 0 0 0 0 0.0049 The main specifications for the elaboration and acquisition systems of each area are summarized in the following: A sampling frequency of 200 Hz with a resolution of 12 bits has been defined for the analog-to-digital conversion. For communication purposes, an Ethernet interface has been selected. We instead chose to use standard Ethernet communication to be aligned with the current trends in the automation of power systems. This choice, in effect, enables a future implementation of the IEC61850 standard, which is based on the Ethernet physical layer [30], [31]. These requirements were fixed, also having in mind that the equipment is already available from our laboratory. The acquisition and elaboration systems of each area are organized in the list the follows. 1) Zone 1: National Instrument CompactRio-9002: a) digital I/O: Ni9401; b) analog I/O: Ni9215.

2) Zone 2: National Instrument CompactRio-9002: a) digital I/O: Ni9401; b) analog I/O: Ni9201. 3) Zone 3: National Instrument PXI 8176: a) analog and digital I/O: Ni PXI 7831R.

B. Software In a physical implementation of the DIF, two main issues arise: 1) role of the communication protocols; 2) synchronization of the execution. Last but not least, the overall system has to be designed according to the requirement of being a hard real-time system. With this goal in mind, all functions were timed, and particular attention was paid to communication. The final structure of the software is reported in Fig. 7. A first step in the execution is the initialization, during which the matrices Y and y are calculated and the ports for communication are opened. When this first phase is completed, the software enters an infinite time loop. The previously described equations in the prediction and measurement assimilation stages are calculated in the local filter and in the global assimilation steps (see Fig. 7). The steps of synchronization, communication, and transmission-to-host are described in the following sections. 1) Synchronization: For synchronization, we used an external shared clock. With this solution, the error between two

446

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 2, FEBRUARY 2010

Fig. 7. Block diagram of the software.

different observers is less than 200 μs. This synchronization technique is easy to implement and produces good results. However, it results in strong limitations for the distribution of the system: This is not a limiting factor for this experimental validation, but it may be a problem in future applications in systems characterized by a long distance between the observers. In these applications, the synchronization technique must be replaced with a technique of synchronization based on Ethernet communication: This issue has already been analyzed and resolved in many publications, such as in [30]. 2) Communication: During the execution of this block, each observer transmits to the others the values of I and i, which are calculated in the local filter, and the values of the breaker status vector, for the elements in each observer’s area. The User Datagram Protocol (UDP) was used as the communication protocol. Although it does not possess the control-of-error capabilities that the Transmission Control Protocol (TCP) has, it was chosen because of its execution speed. The broadcast technique was used for transmission. Using the UDP with broadcast transmission, a complete exchange of information among the three areas can be achieved in 6 ms. 3) Transmission-to-Host: During the execution of this step, each observer transmits the values of Y and y to the host, which are calculated during global assimilation. Although the communication with the host is essential to monitor the operation of each observer, it is not necessary for the operation of the filter. Therefore, the transmission-to-host step is not a critical point in the implementation of the software and does not have to be timed like the communication between one target and another. The transmission-to-host step is executed in the background, and the execution time of the program is reduced by 2 ms. All the software was written using LabVIEW 8.0. V. A NALYSIS OF THE DIF E XECUTION T IME We will now proceed with the analysis of the DIF execution time. The main parts to be analyzed are the communication time and the computational time. Letting Nz be the number of the zones in which the grid is divided, a complete exchange of information among all the zones takes Nz ΔT (in milliseconds), where ΔT is the time required for each transmission. In this implementation, that time is 2 ms. The time required for one transmission depends on the number of data that has to be transmitted. The time of 2 ms is enough only for the transmission of a single data package. The maximum dimension of a package for the UDP is 548 bytes.

Each observer must transmit the following data: 1) I and i; 2) breaker status vector; 3) counter value, which is used by the receiving observer to calculate the value of the coefficient K associated with the conduction status of the ac–dc converter of each zone. For the transmission of the data of the counters, 2 bytes are required for each counter of a subgrid. The number of bytes required for the breaker status vector depends on the number of possible configurations of the subgrid; anyway, the allocation of 2 bytes for it is more than enough. According to (2), (3), (5a), (5b), and (7) describing the DIF algorithm, the highest computational burden is due to the matrix inversions in (2), (3), (5a), and (5b). If Ni is the number of local state variables and Mi is the number of local measurements, the computational cost is O(Ni3 + Mi3 ). Since, in general, the measurement uncertainty is time invariant, the inversion in (5a) and (5b) can be evaluated offline, reducing the online computational cost to O(Ni3 ). For the application discussed in this paper concerning nine state variables and adopting a Pentium IV processor, the computational time of the DIF was about 100 μs. This time was negligible if compared with the communication time (6 ms). To estimate the computational complexity of the DIF for real applications, characterized by a larger number of zones and state variables, it is necessary to develop a rough model for the computational time and the communication time. Being Nz the number of zones partitioning the power system and N the overall number of state variables, the local number of state variables for each local information filter is Ni ∝ N/Nz . The computational time is consequently given by TC = αN 3 /Nz3 = α Ni3 . Instead, the communication time is proportional to the number of communicating subzones; under the hypothesis that this number R is fixed and that the data transmitted, according to (5a) and (5b), are proportional to Ni2 , the estimated transmission time is TT = βN 2 /Nz2 + β(R − 1)N 2 /Nz2 = βRN 2 /Nz2 = β  Ni2 . In the experiment reported here, the number of local variables Ni is 9, and the constant quantities α and β  are 0.14 and 74 μs, respectively.

BENIGNI et al.: DECENTRALIZED OBSERVER FOR SHIP POWER SYSTEM APPLICATIONS

447

Fig. 9. Comparison between the measured voltage and the estimated voltage on capacitor Cb of zone 1.

Fig. 10. Detail of the comparison between the measured voltage and the estimated voltage on capacitor Cb of zone 1. Fig. 8. Expected computational time, transmission time, and cycling time of the implemented DIF versus the number of local state variables.

In Fig. 8, the transmission time TT and the computational time TC are shown versus the number of local state variables Ni . In Fig. 8, it is clear that, when the number of local state variables is less than 100, the filter cycle time is determined by the communication time only. VI. E XPERIMENTAL R ESULTS Figs. 9 and 10 represent the comparison between the voltage directly measured on the capacitor Cb (see Fig. 4) of zone 1 and the trend calculated by the software. Fig. 11 represents the comparison between the voltage directly measured on the capacitor Cb of zone 3 and the voltage estimated by the DIF. The estimation error shown in Fig. 10 is mainly due to the oversimplified model used for describing the Darlington transistor; this consideration can be derived from observing the acquired voltage and current on the Darlington shown in Fig. 12, which are far to be linear.

Fig. 11. Comparison between the measured voltage and the estimated voltage on capacitor Cb of zone 3.

As shown in Fig. 12, when the Darlington transistor is in conduction, the voltage and the current on it are constant: In this case, the choice of modeling the Darlington with a constant

448

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 2, FEBRUARY 2010

Fig. 12. Current and voltage on the Darlington transistor.

istic of the hardware and software: The performance of the system can be improved using more sophisticated hardware. The communication and synchronization steps influence the speed of the system, but not its accuracy. A problem that arises from the analysis of this system is about the duration of the communication step: In fact, each communication needs 2 ms. In future applications with multiple observers, it will be necessary to develop an algorithm in which the observers only communicate with those nearest, to limit to overall computational burden derived from the communication protocols. An interesting future development would be the replacement of the UDP with a deterministic communication system: In particular, a good solution would be the use of a real-time protocol derived from the TCP instead of from the UDP. In fact, applications based on the UDP generate a large volume of data and are not sensitive to network congestion; on the other hand, the TCP traffic prevents the network from congestion by means of closed-loop control of the packet loss and round-trip time [33]. With this solution, the performance of the system can strongly be improved, but it has to be checked for the compatibility of the hardware with the transmission protocol: In fact, the performance of the system depends on the control of the delay and jitter in the station nodes [34]. R EFERENCES

Fig. 13. Delay between two steps of the filter.

resistance (0.5 Ω) is justifiable. During the block period, the ratio between the voltage and the current is not a constant value, and the choice of modeling the Darlington with a constant resistance (10 kΩ) introduces an error. This model error is only partially compensated by the measurements. Moreover, even if this implementation is not considered, it is possible to identify a set of measurements to remove the problem. Using what is proposed in [32], it is possible to identify a set of critical measurements that make the system observable without excessively increasing their number. In this particular implementation, we chose to concentrate all the measures on the converters, assuming that we do not have direct access to the loads. Fig. 13 shows the delay between two steps of the filter. VII. C ONCLUSION As shown in Figs. 9–11, estimates are very close to the actual voltage across the capacitor Cb: This result gives a confirmation of the theoretical work described in previous publications. The performance of this system is only conditioned by the execution speed of the filter, which depends on the character-

[1] F. C. Schweppe and E. J. Handschin, “Static state estimation in electric power systems,” Proc. IEEE, vol. 62, no. 7, pp. 972–982, Jul. 1974. [2] A. Monticelli, “Electric power system state estimation,” Proc. IEEE, vol. 88, no. 2, pp. 262–282, Feb. 2000. [3] J. B. Carvalho and F. M. Barbosa, “Parallel and distributed processing in state estimation of power system energy,” in Proc. 9th MELECON, vol. 2, May 18–20, 1998, pp. 969–973. [4] D. M. Falcao, F. F. Wu, and L. Murphy, “Parallel and distributed state estimation,” IEEE Trans. Power Syst., vol. 10, no. 2, pp. 724–730, May 1995. [5] G. D’Antona, A. Monti, F. Ponci, and L. Rocca, “Distributed state estimator for electric power systems in avionic and naval applications,” in Proc. Instrum. Meas. Technol. Conf., Sorrento, Italy, Apr. 2006, pp. 2312–2316. [6] R. Vadigepalli and F. J. Doyle, III, “A distributed state estimation and control algorithm for plantwide processes,” IEEE Trans. Control Syst. Technol., vol. 11, no. 1, pp. 119–127, Jan. 2003. [7] C. Rehtanz, Autonomous Systems and Intelligent Agents in Power System Control and Operation. New York: Springer-Verlag, 2003. [8] G. D’Antona, A. Monti, and F. Ponci, “Decentralized information filter for the state estimation in electrical power system,” in Proc. IMTC, Warsaw, Poland, May 1–3, 2007, pp. 1–5. [9] A. Benigni, U. Ghisla, G. D’Antona, A. Monti, and F. Ponci, “Decentralized observer for electrical power systems: Implementation and experimental validation,” in Proc. IEEE IMTC, May 12–15, 2008, pp. 859–864. [10] A. G. Phadke, J. S. Thorp, and K. J. Karimi, “State estimation with phasor measurements,” IEEE Trans. Power Syst., vol. PWRS-1, no. 1, pp. 233– 238, Feb. 1986. [11] Z. Huang, J. F. Hauer, and K. E. Martin, “Evaluation of PMU dynamic performance in both lab environments and under field operating conditions,” in Proc. IEEE Power Eng. Soc. Gen. Meet., Jun. 24–28, 2007, pp. 1–6. [12] J.-A. Jiang, Y.-H. Lin, J.-Z. Yang, T.-M. Too, and C.-W. Liu, “An adaptive PMU based fault detection/location technique for transmission lines. II. PMU implementation and performance evaluation,” IEEE Trans. Power Del., vol. 15, no. 4, pp. 1136–1146, Oct. 2000. [13] J.-A. Jiang, J.-Z. Yang, Y.-H. Lin, C.-W. Liu, and J.-C. Ma, “An adaptive PMU based fault detection/location technique for transmission lines. I. Theory and algorithms,” IEEE Trans. Power Del., vol. 15, no. 2, pp. 486– 493, Apr. 2000. [14] J. S. Thorp, A. G. Phadke, and K. J. Karimi, “Real time voltage-phasor measurement for static state estimation,” IEEE Trans. Power App. Syst., vol. PAS-104, no. 11, pp. 3098–3106, Nov. 1985.

BENIGNI et al.: DECENTRALIZED OBSERVER FOR SHIP POWER SYSTEM APPLICATIONS

[15] A. G. Phadke and J. S. Thorp, Synchronized Phasor Measurements and Their Applications. Berlin, Germany: Springer-Verlag, 2008. [16] J.-J. Zhang, Q. Liu, C. Rehtan, and S. Rudin, “Investigation of several new technologies for mega city power grid issues,” in Proc. Int. Conf. Power Syst. Technol. PowerCon, Oct. 22–26, 2006, pp. 1–6. [17] G. Asplund, “Application of HVDC Light to power system enhancement,” in Proc. IEEE Power Eng. Soc. Winter Meet., Jan. 23–27, 2000, vol. 4, pp. 2498–2503. [18] M. Jeroense, “HVDC, the next generation of transmission highlights with focus on extruded cable systems,” in Proc. ISEIM, Sep. 7–11, 2008, pp. 10–15. [19] Z. Hu, C. Mao, and J. Lu, “Improvement of transient stability in AC system by HVDC Light,” in Proc. IEEE PES Transmiss. Distrib. Conf. Exhib.: Asia, Pacific, 2005, pp. 1–5. [20] K. Eriksson, “Operational experience of HVDC Light,” in Proc. 7th Int. Conf. AC–DC Power Transmiss. (Conf. Publ. No. 485), Nov. 28–30, 2001, pp. 205–210. [21] Z. R. Zhang, Z. D. Yin, and F. X. Hu, “Research of multi-farms transmission of distributed generation based on HVDC Light,” in Proc. Int. Conf. Power Syst. Technol., PowerCon, Oct. 22–26, 2006, pp. 1–6. [22] Q. Zhong, Y. Zhang, L. Lin, Q. Chen, and Z. Wu, “Study of HVDC Light for its enhancement of AC/DC interconnected transmission systems,” in Proc. IEEE Power Energy Soc. Gen. Meet., Convers. Del. Elect. Energy 21st Century, Jul. 20–24, 2008, pp. 1–6. [23] A. Bremeller, M. N. John, and M. R. Scott, Practical Diakoptics for Electrical Networks. London, U.K.: Chapman & Hall, 1969. [24] P. Zhang, J. R. Marti, and H. W. Dommel, “Network partitioning for realtime power system simulation,” in Proc. IPST, Montreal, QC, Canada, Jun. 19–23, 2005. [25] S. Y. R. Hui and K. K. Fung, “Fast decoupled simulation of large power electronic systems using new two-port companion link models,” IEEE Trans. Power Electron., vol. 12, no. 3, pp. 462–473, May 1997. [26] G. N. Korres, “A partitioned state estimator for external network modeling,” IEEE Trans. Power Syst., vol. 17, no. 3, pp. 834–842, Aug. 2002. [27] G. Superti Furga, “Sulle equazioni di stato dei convertitori statici,” L’Energia Elettrica, vol. LX, no. 5, pp. 187–198, May 1983. [28] Current Transducer LA 55-P, LEM. [Online]. Available: http://www. lem.com [29] Voltage Transducer LV 20-P, LEM. [Online]. Available: http://www. lem.com [30] R. E. Mackiewicz, “Overview of IEC 61850 and benefits,” in Proc. IEEE PES PSCE, Oct. 29, 2006–Nov. 1, 2006, pp. 623–630. [31] L. Hossenlopp, “Engineering perspectives on IEC 61850,” IEEE Power Energy Mag., vol. 5, no. 3, pp. 45–50, May/Jun. 2007. [32] J. B. A. London, N. G. Bretas, and L. F. C. Alberto, “Analysis of measurement set qualitative characteristics for state estimation purposes,” in Proc IEEE Power Eng. Soc. Gen. Meet., Jun. 12–16, 2005, vol. 1, pp. 294–301. [33] N. B. Yoma, J. Hood, and C. Busso, “A real time protocol for the internet based on the least mean square algorithm,” IEEE Trans. Multimedia, vol. 6, no. 1, pp. 174–184, Feb. 2004. [34] T. Skeie, S. Johannessen, and O. Holmeide, “Timeliness of real-time IP communication in switched industrial Ethernet networks,” IEEE Trans. Ind. Informat., vol. 2, no. 1, pp. 25–39, Feb. 2006.

Andrea Benigni (S’09) was born in Italy in 1983. He received the M.S. degree from the Politecnico di Milano, Milan, Italy, in 2008. He is currently working toward the Ph.D. degree with RWTH Aachen University, Aachen, Germany. Before joining RWTH Aachen University, he was a Visiting Scholar with the University of South Carolina, Columbia. His research activities mainly concern state estimation and real-time simulation.

449

Gabriele D’Antona (S’88–M’06–SM’07) received the M.Sc. degree in electrical engineering and the Ph.D. degree in electrical science from the Politecnico di Milano, Milan, Italy, in 1989 and 1994, respectively. From 1990 to 1996, he was with the Joint European Torus (JET) Laboratory, Abingdon, U.K. From 1996 to 2002, he was with the Department of Electrical Engineering, Politecnico di Milano. From 2002 to 2007, he was an Associate Professor of electrical and electronic measurements with the Department of Electrical Engineering, Politecnico di Milano, where he has been with the Department of Energy since 2008. His current research interests principally concern the development of measurement systems based on probabilistic and variational signal processing techniques for the diagnosis and control of spatially distributed systems. Developments in this field mainly regard the realization of diagnosis and control systems for the magnetic confinement of thermonuclear plasmas and electrical power systems. Dr. D’Antona is a Cochair of the TC-18 IEEE Instrumentation and Measurement Society Technical Committee for “environmental measurements,” a member of the International Measurement Confederation (IMEKO) TC21 Working Group on mathematical tools for measurements, and a member of the Italian Organization for Standardization (UNI) Technical Committee for statistical methods.

Ugo Ghisla received the B.S. degree in electrical engineering and the M.S. degree from the Politecnico di Milano, Milan, Italy, in 2005 and 2008, respectively. He is currently working toward the Ph.D. degree in electrical engineering with the University of South Carolina, Columbia. His research interests are power electronics, measurements, and controls.

Antonello Monti (M’94–SM’02) received the M.S. degree in electrical engineering and the Ph.D. degree from the Politecnico di Milano, Milan, Italy, in 1989 and 1994, respectively. From 1990 to 1994, he was with the Research Laboratory, Ansaldo Industria, Milan, where he was responsible for the design of the digital control of a large power cycloconverter drive. In 1995, he joined the Department of Electrical Engineering, Politecnico di Milano, as an Assistant Professor. From 2000 to 2008, he was an Associate Professor and then a Full Professor with the Department of Electrical Engineering, University of South Carolina, Columbia. He is currently the Director of the Institute for Automation of Complex Power Systems within the E.ON Energy Research Center, RWTH Aachen University, Aachen, Germany.

Ferdinanda Ponci received the Ph.D. degree in electrical engineering from Politecnico di Milano, Milan, Italy, in 2003. In 2003, she joined the Department of Electrical Engineering, University of South Carolina, Columbia, as an Assistant Professor. She is currently with RWTH Aachen University, Aachen, Germany, affiliated with the Institute for Automation of Complex Power Systems of the E.ON Energy Research Center. Her research interests are agent technologies for the control of complex power systems and modeling and simulation of uncertain system for state estimation and control.