0 evaluări0% au considerat acest document util (0 voturi)
17 vizualizări5 pagini
OVNI's aim is to represent full-size power system networks for real-time and online simulation. OVNI implements the MATE concept ("Multi-Area Tearing Equivalent") for network partitioning. To prevent accumulated phase drifts that can drive longtime simulations unstable, all parts of the network are solved simultaneously and decoupling through [?]t delays are not allowed.
OVNI's aim is to represent full-size power system networks for real-time and online simulation. OVNI implements the MATE concept ("Multi-Area Tearing Equivalent") for network partitioning. To prevent accumulated phase drifts that can drive longtime simulations unstable, all parts of the network are solved simultaneously and decoupling through [?]t delays are not allowed.
OVNI's aim is to represent full-size power system networks for real-time and online simulation. OVNI implements the MATE concept ("Multi-Area Tearing Equivalent") for network partitioning. To prevent accumulated phase drifts that can drive longtime simulations unstable, all parts of the network are solved simultaneously and decoupling through [?]t delays are not allowed.
AbstractThis paper presents the solution approach and imple-
mentation strategy followed in the development of the OVNI (Ob- ject Virtual Network Integrator) power system simulator. OVNIs aim is to represent full-size power system networks for real-time and online simulation. To cope with the dimensions of the prob- lem, partitioning techniques are required, both in the system so- lution and in the hardware implementation. OVNI implements the MATE concept (Multi-Area Tearing Equivalent) for net- work partitioning. In addition to the considerable solution speed- ups that can be achieved, the MATE concept is particularly suited for software implementation using object-oriented techniques and for hardware implementation using parallel processing schemes. Keywords: Real-time power system simulators, network tear- ing techniques, objected-oriented software design. I. INTRODUCTION The Electromagnetic Transients Program EMTP [1] is the standard tool worldwide for power system transients simula- tion. Even though the EMTP maintains its position as a pow- erful and effective tool, the availability of high computational power at consumer market prices (e.g., Pentium II worksta- tions) has been raising the ante of what can be achieved in power system simulation. In particular, real-time performance of about 50-100 s per solution step for testing of protective relaying equipment has now been possible for over five years [2], [3], [4]. The OVNI simulator (Object Virtual Network Integrator) currently under development at UBC is aimed at permitting continuous direct interaction between the user and the simu- lated system (computer) and between the simulated system and external equipment (e.g., an actual relay or controller). OVNI performs a continuous-time simulation of the power system behaviour and allows for events to be applied (by the user or through feedback from the relay or controller) at any point in the simulation. The requirement of continuous simulation imposes very demanding conditions on the numerical behaviour of the solu- tion algorithms. To prevent accumulated phase drifts that can drive longtime simulations unstable, all parts of the network are solved simultaneously and decoupling through t delays are not allowed. OVNI: An Object Approach to Real-Time Power System Simulators OVNI achieves the proposed design objectives using off-the- shelf Pentium-based personal workstations. Simultaneous ef- ficient solution of large networks is achieved using diakoptic techniques [5] to partition the network along boundaries of maximum solution efficiency. The partitioning technique de- veloped for this purpose is called MATE for Multi-Area Tear- ing Equivalent. MATE extends the concept of subdividing a network along transmission line links (widely employed in real- time simulator implementations, e.g., [2], [3], [4]) to the gen- eral case of an arbitrary partitioning across linear or nonlinear branches connecting subsystems. From a software design point of view, partitioning into sub- systems (object-based modelling) is ideally mapped into ob- ject-oriented programming, which has a number of advantages in the software design of large systems. From a hardware point of view, the solution of weakly coupled subsystems can be eas- ily parallelized for simultaneous execution in multiple proces- sors. The hardware implementation adopted for OVNI is based on off-the-shelf Pentium II personal workstations with com- mercially available real-time operating system and I/O boards. II. REAL-TIME SOLUTIONS Our real-time simulation experience ([3], [6], [7]) has mo- tivated the rethinking of a number of solution issues with re- spect to traditional EMTP techniques [8]. One of these issues is the considerable solution speed-ups that can be achieved by precalculating the inverse matrices ([G -1 ]) of possible topo- logical configurations. This approach works well in the con- text of simulators for relay testing like RTNS [3] because for the type of networks involved in these applications (e.g., 30 nodes and 30 branches), there is a reasonably small number of [G -1 ] matrices that need to be precalculated and prestored to cover all possible switching operations. The approach of precalculating [G -1 ] matrices, however, reaches its limits of practicality in situations like the model- ling of HVDC converter bridges. In the work presented in [6] the objective was to model a 4-bridge, 24-valve HVDC con- verter in real time. In this case, precalculation would have required the storage of 2 24 (16 million) inverse matrices. Us- ing network partitioning techniques, however, allowed us to J. R. Mart, L. R. Linares, J. Calvio, H. W. Dommel, The University of British Columbia, Vancouver, Canada, and J. Lin, Electric Power Research Institute, Beijing, China Proceedings of the 1998 International Conference on Power System Technology, Powercon'98, Beijing, China, August 18-21, 1998 This paper is provided on a personal basis only for the private non-commercial use of the requester for educational or research purposes and is subjected to the publisher's copyright restrictions. 2 achieve real-time performance on a Pentium 200 MHz com- puter with the precalculation of only 4 x 2 6 (256) matrices. Some of the benchmarks obtained for this model are presented later in this paper. Another area where precalculation of [G -1 ] matrices can encounter serious drawbacks is in the context of a full-size power system simulator, like OVNI [7]. In this case, precalcu- lation of [G -1 ] matrices for all possible operational conditions in the network is totally impractical. III. MATE DIAKOPTICS Grouping components into clusters joined by links (Figs. 1, 3, and 4) is the technique adopted in OVNI for efficient real- time solutions. The MATE (Multi-Area Tearing Equivalent) solution frame- work subdivides the system into kindred (of akin proper- ties) subnetworks ([A], [B] in Fig. 1) which are independent of each other except for the links subsystem ([]in Fig. 1) 1 . The structure of Fig. 1 is a common one in power systems where power transmission proceeds mostly radially from re- moved generation plants to large consumption centres. As a result, the power network topology shows a concentration of equipment at plants and substations, with only a few links (transmission lines) connecting the subsystems. In the context of real-time EMTP-based simulators, great advantage is taken ([2], [3], [4]) of the fact that the travelling time in transmission lines can completely decouple the con- nected subsystems. In this (trivial) case, subsystem [] in Fig. 1 is zero (non existent). In a more general context, as for exam- ple, in transient stability studies using phasor solutions (or for short-line sections in EMTP simulations) transmission lines do not totally decouple the subsystems, but they can still con- stitute a few links subsystem along which to separate larger groupings of load and generation. In current research on the work reported in [9], savings in the order of 2 to 5 times over conventional techniques have been achieved using the MATE approach in online security index calculations. In the algorithmic implementation of MATE discussed later in this paper, it is more efficient to use a two-level segmenta- tion scheme (Figs. 3 and 4) to distinguish between transmis- sion line links (complete decoupling) and MATE links (de- coupling through the links subsystem). The main advantage of the MATE diakoptic structure of Fig. 1 in the context of real-time simulation is that the solu- tion of the full network can be formulated in terms of inverse matrices [A -1 ], [B -1 ], and [ -1 ]. These matrices represent smaller subsystems that, in many instances, either remain constant or have a reasonable number of possible states that can be precal- culated before the time-step loop. The MATE formulation is particularly advantageous to rep- resent switching operations in branches connecting subsystems. In traditional EMTP modelling, when an ideal switch closes, the two nodes connected by the switch collapse into a single node. Similarly, when the switch opens two nodes are created out of one. These topological changes slow down the solution process in two fronts: i) There is a change in node subscripts, and ii) The entire system needs to be retriangularized. Suppose in MATE that subnetworks [A] and [B] in Fig. 1 remain topologically constant during the time loop simulation (i.e., only the history sources change from time step to time step), and that all switching operations occur in the link branches (subblock []). Then, whether the link branches are open or closed affects only the links subsystem [ -1 ] and sub- matrices [A -1 ] and [B -1 ] do not need to be retriangularized. The savings achieved in MATE by avoiding full-system redimensioning and retriangularization can be considerable, particularly in the context of real-time simulation, where it is particularly important to keep equal-length solution times at each solution step. Another common situation where MATE can result in con- siderable savings is, for example, when subnetwork [A] is lin- ear and subnetwork [B] is nonlinear. In this case since [A -1 ] does not change, the solution of [A] requires minimum updat- ing work, and the most computational work is concentrated in subsystem [B]. Further savings can be achieved if [B] is mod- elled as a finite nonlinearity (e.g., as piecewise linear), in which case all possible states of [B] can be precalculated and prestored. Other Kindred Subnetworks In addition to the most obvious cases discussed above re- garding topologically constant clusters and clusters connected by a few links, there are a number of other kinship situations that can be advantageously formulated and efficiently solved using MATEs diakoptic approach. Some of the kindred sub- systems that we are currently investigating include: 1. Transmission lines decoupled by travelling time in EMTP solutions (trivial case where full decoupling exists). 1 Detailed papers by the first two authors on MATE and its OOP realization are being submitted in May 1998 to IEEE Transactions on Power Systems. A B
Fig. 1. MATE diakoptics. [A] and [B] are cluster
subsystems. [] are the links. 3 2. Dense subsystems joined to each other through a few links (e.g., by transmission lines in transient stability phasor so- lutions). 3. Subsystems where the [G] matrix remains constant con- nected to subsystems where the [G] matrix changes from time step to time step. 4. Subsystems with linear elements connected to subsystems with nonlinear elements. 5. Subsystems solved using a phasor solution (e.g., a power network during a transient stability study) linked to subsys- tems solved with an EMTP solution (e.g., an HVDC con- verter). 6. Slow subsystems where large solution steps t are adequate, connected to fast subsystems where much smaller ts are required. Here the slow subsystem does not need to be solved at each time step of the fast subsystem solution [10]. In EMTP studies, for example, a power network modelled with a t in the hundreds of microseconds may be connected to an HVDC converter where, for proper firing angle resolu- tion, the t should be in the tens of microseconds. IV. OBJECT ORIENTED IMPLEMENTATION To simulate a network in EMTP solutions, a suitable inte- gration rule is first applied to each element in it [1]. At the end of this discretization process, the original network has been replaced by a DC-resistive network containing equivalent re- sistances, and current sources that depend on past history val- ues of currents and voltages. At every time step of the solution loop, this DC-resistive network needs to be updated. This up- dating may or may not include topological changes, but it will always incur a history sources computation. Applying nodal analysis to this discretized network, a system of equations of the form of Eq. (1) is obtained, [ ][ ] [ ] G v h = (1) where [G] is the matrix of conductances, [h] is the vector of total nodal currents, and [v] is the vector of unknown nodal voltages. Topological changes represent changes in [G], and perhaps in the way history sources are accumulated into [h]. Computationally speaking, topological changes are very ex- pensive (step (c) below, and in Fig. 2). During every integration step, the OVNI (Object Virtual Network Integrator) goes through the steps outlined in Fig. 2: a) Updating all history sources in the system; b) Accumulating all current sources into nodal current vec- tors; c) Reassessing the system conductance matrix [G]; d) Solving for the node voltages; e) Presenting the state of the network for output. 1. Segmentation of the network The solution of Eq. (1) becomes computationally expensive as the number of nodes increases. The OVNI solver reduces the computational burden by subdividing the system into subsys- tems that can be solved as if they were decoupled from the rest of the network. The result of this segmentation process con- verts Eq. (1) into many smaller subsystems of the same form (Figs. 3 and 4), and given that the solution effort per subsys- tem is of the order of the square of the number of equations/ unknowns, the gain in speed can be considerable. The process that the OVNI follows to segment the network into smaller, more manageable sections is divided into two stages. First, it uses topological segmentation provided by the time delay of transmission line links to break the network into completely decoupled blocks. Then, it uses MATE diakoptics to fracture these blocks into even smaller subblocks. The relationship between the networks topological blocks and MATEs subblocks can be seen in Figs. 3 and 4. 2. OVNI s top level classes In OVNIs fragmented scenario, object-oriented design (OOD) is used to render a flexible, reliable, and expandable solution to the simulation problem. At the top level of the OVNIs object-oriented programming (OOP) abstraction is the simulation itself, an object single instantiation, OSI, of the class sim_t. The simulation object drives, under the synchroniza- tion of the clock (an OSI of the class clk_t), the network object (an OSI of the class net_t) to solve itself at every time step. On its part, the network class, net_t, has a container rela- tionship with most other objects in the solution domain; namely: Fig. 2. Steps in simulation cycle per integration step. Update History Sources (Element model' s responsibility) Accumulate Nodal currents start-thi s-step Sol v e f or No d e Vo l t a g e s e n d - t h i s - s t ep r e a d y f o r o u t p u t / i n p u t Re c o mp u t a t i o n o f [ G] ma t r i x 4 it contains all the topological blocks, all MATEs subblocks, all probes (oscilloscopes, voltmeters, sensors, etc.); all network elements; all MATEs links; and all current and voltage sources. While doing this, the OVNI maintains each of those objects very loosely coupled with the rest, which permits easy redefi- nition and expansion or modification of any of those entities. 3. Element-models encapsulation The OVNIs computation/simulation can be separated into two main activities, one that involves the element models ex- clusively, and another that is the sole responsibility of the inte- grator proper. Elements (transformers, inductors, capacitors, HVDC bridges, transmission lines, etc.) are represented in the OVNI as instantiations of some particular class in a hierarchy whose top ancestor is the abstract class elm_t. This clean segmentation of responsibilities, enforced by the encapsulation of OOP, allows the OVNI simulator to profit from models developed by other parties, with the abstract methods interface defined in elm_t as the only constraint. The methods in elm_t maintain the compatibility of interface between the integrator and the model, without burdening the model devel- oper with the need to get into the intricacies of the integrator proper. V. HARDWARE IMPLEMENTATION A major guideline in OVNIs hardware implementation has been to stick to off-the-shelf computer equipment, on the low- price end of the spectrum. In our lab, OVNI runs on Pentium Pro 200 MHz workstations, but the current technology targets are Pentium II 330 MHz machines for the core and Pentium II 200 MHz machines for the interface. Data exchange with the devices under test and the monitor- ing probes is made through a convenient set up of 16-bit ana- log outputs and optically isolated digital inputs. Communication requirements for data exchange among the different OVNI processors impose transmission rates in the or- der of 40 Mbytes/sec, with the communication cards plugged into PCI busses. Given the speed requirements, the chosen hardware archi- tecture favours a master/slave structure instead of a network one. The same higher performance in data transmission needs made us choose a DMA and a radial architecture versus a tree- like architecture. The Pharlap real-time operating system, with full 32-bit addressing capabilities, is used for the hardware platform. However, direct control of the hardware, bypassing the OS, is needed in many critical instances. The processors synchronization adopted is based on a com- mon shared interruption input whose controller is the master Fig. 3. Solution scheme for OVNIs two-level segmentation: topological and MATE. Fig. 4. Illustration of OVNIs two-level segmentation: topological and MATE. topological block MATE sub block link matrix Network's Conductance Matrix topological blocks MATE sub bl ock transmi ssion line link tra n s m is s io n lin e lin k transmissi on line l ink
topol ogi cal bl ock topol ogi cal bl ock topol ogi cal bl ock topol ogi cal bl ock MATE's Diakopt ic SubBlock MATE's Diakoptic SubBlock MATE's Links 5 processor. Integration step widths t are generated through the timers already incorporated within the workstation hard- ware. VI. SAMPLE CASE: AN HVDC SYSTEM A situation where real-time performance becomes particu- larly difficult is when there are many switching operations in- volved. This is the case, for example, of HVDC converters where the state of the thyristor valves changes at almost every step of the simulation. As indicated earlier, precalculation of all possible inverse matrices (in the order of 16 million for a 24-valve converter) is not practical in this case. As an example of the performance benefits obtainable by OVNIs combination of topological segmentation and MATEs segmentation in an OOP environment, a test case for a 6-valve, 12-valve, and 24-valve HVDC converter was run both under OVNI and under Microtran (UBCs version of the EMTP). Both programs ran on a Pentium-Pro 200 MHz workstation. The timings obtained are shown in Table I. VII. CONCLUSIONS The large computational power available today at low prices has opened a number of opportunities in what can be achieved in systems simulation. At the same time, as the goals become more ambitious (e.g., real-time simulation), new solution ap- proaches are needed to harness this power. OVNI (for Object Virtual Network Integrator) is an ambitious project that aims at attaining real-time solution speeds for large power system networks using desktop computers. To tackle this challenge, a new network solution strategy has been devised. MATE (for Multi-Area Tearing Equivalents) simplifies the network solu- tion problem by splitting the network into manageable pieces. Subsystems are defined which are independent of each other except for a few links. This independence allows the exploita- tion of the solution properties of each subsystem to full advan- tage. For example, subsystems that do not change topologi- cally from time step to time step do not need to be retriangu- larized (as it is currently done in the EMTP solution) just be- cause some of the branches connecting the system to another subsystem change. The MATE solution is a full-system simul- taneous solution and decoupling between subsystems is not approximated by t delays but it is exactly resolved. The con- cept has been implemented in software using OOP design and in hardware using off-the-shelf personal computers. Real-time performance has been obtained in test systems of 30 nodes and 30 branches for relay testing and in the simulation of 24-valve HVDC converters. VIII. ACKNOWLEDGEMENTS The OVNI project is a large project and at present a team of about twenty individuals, including graduate students, visi- tors, and research assistants are participating in different as- pects of it. Funding also proceeds from a number of sources, but particular recognition should be given to the Natural Sci- ences and Engineering Research Council of Canada (NSERC), the British Columbia Hydro Authority (BC Hydro), and Powertech Labs of Surrey, BC, for their continued support. IX. REFERENCES [1] H. W. Dommel, Digital computer solution of electromagnetic transients in single- and multiphase networks, IEEE Trans. on Power Apparatus and Systems, Vol. PAS-88, No. 4, April 1969, pp. 388-399. [2] P.G. McLaren, R. Kuffel, R. Wierckx, J. Giesbrecht, L. Arendt, A real time digital simulator for testing relays, IEEE Trans. on Power Delivery, Vol. 7, No. 1, January 1992, pp. 207-213. [3] J. R. Mart, L. R. Linares, "Real-Time EMTP-based transients simulation," IEEE Trans. on Power Systems, Vol. 9, No. 3, August 1994, pp. 1309-1317. [4] M. Kezunovic, M. Aganagic, V. Skendzic, J. Domaszewicz, J. K. Bladow, D. M. Hamai, S. M. McKenna, Transients com- putation for relay testing in real-time, IEEE Trans. on Power Delivery, Vol. 9, No. 3, July 1994, pp. 1298-1307. [5] G. Kron, Tensor Analysis of Networks, New York: Wiley, 1939. [6] S. Acevedo, L. R. Linares, J. R. Mart, Y. Fujimoto, "Efficient HVDC converter model for real time transients simulation," Accepted for publication in IEEE Trans. on Power Delivery. [7] J. R. Mart, L. R. Linares, R. Rosales, H. W. Dommel, "OVNI: a full-size real-time power system simulator," International Conference on Digital Power System Simulators, ICDS'97, Montreal, Quebec, Canada, May 28-30, 1997. [8] H. W. Dommel, EMTP Theory Book, 2nd Ed., Vancouver: Microtran Power System Analysis Corporation, 1992. [9] M. P. de Arizn, J. R. Mart, "Real-time power system secu- rity and control," IEEE Wescanex Conference Proceedings, Winnipeg, May 22-23, 1997. [10] Linares, L.R. and Mart, J.R., Sub-area latency in a real-time power network simulator, Proceedings IPST95, International Conference on Power System Transients, Lisboa, September 1995, pp. 541-545. Solution times in s/step (Pentium-Pro 200 MHz) Simulator 6-valve 12-valve 24-valve MicroTran 459 983 3120 OVNI 26 45 81 Ratio: 17.7 21.8 38.5 TABLE I BENCHMARKS FOR HVDC CONVERTER MODEL