A challenge in the power system area is the implementation of realtime transient stability analysis. Time domain simulations of power system dynamics have been developed for over three decades and are routinely used for off-line planning and operations planning studies. The computational requirements for simulating power systems dynamics, even for few seconds after fault, represent the main impediment to the use of conventional transient stability programs in real-time. Parallel processing has been extensively used to speed up transient stability analysis and several results can be found in the literature. In recent years, distributed environments built out of pools of networked workstations have been widely used for solving computationally intensive problems that once belonged exclusively to the domain of supercomputers. The success of "cluster computing" is essentially due to the increasing power of workstations, based on fast RISC (Reduced Instruction Set Computer) microprocessors. Moreover, the development of high speed networks, capable to sustain up to Obitis, drastically narrowed the bandwidth and latency gap between an interconnection network in a parallel machine and a communication network in a distributed system. Distributed computing is also economically attractive with respect to traditional parallel computers, allowing to potentially provide at lower cost the processing power of a supercomputer. Heterogeneity and portability are primary goals of the developed distributed systems and programming environments. Heterogeneity allows to connect different machines from different vendors in a single virtual parallel computing environment. These important features can be achieved using programming environments recently developed, such as the Parallel Virtual Machine (PVM) from Oak Ridge National Laboratory and Emory University. The feasibility of real-time transient stability has been investigated exploiting both domain and functional decompositions, on a homogeneous cluster of eight Digital ALPHA and on an IBM SP2 machine. To test the functional decomposition, the Shifted-Picard algorithm has been implemented under PYM, whereas a scaled domain decomposition has been tested running multiple contingencies on different nodes of cluster systems, using the Very Dishonest Newton (VDHN) algorithm, which is the fastest sequential algorithm. In order to assess the performance of these approaches, time domain simulations adopting detailed modeling for synchronous machines have been carried out on a realistic sized network comprising 2583 buses and 511 generators. The distributed Shifted Picard algorithm is characterized by low performance. At first, this is surprising, since previous experiments on parallel environments showed interesting speedups. In order to understand the reason for this behavior, we have conducted a further investigation on the distributed SP algorithm. A communication model has been developed to take into account the main issues involved when a distributed environment is used. The tests carried out have shown that this model must account for the contention of the shared physical medium. Moreover, the SP algorithm was loosely synchronized, to guarantee a deterministic convergence behavior. Thus, the communication pattern resulted to be regular and symmetric, i.e. at a fixed time all the processors attempted to use the shared network. Although this property is effective for parallel processing solutions, our experience has shown that in distributed computing a more chaotic communication pattern would have produced better results, being at any given time the network resources exploited by at most one processing element on average. However, we were forced to this choice because asynchronous algorithms, due to their unforeseeable behavior, are not acceptable for real-time applications. The scaled domain decomposition has been implemented on the same test bed, using on each node the VDHN algorithm. In this case, the size of the problem (i.e the number of contingencies) scales with the available processor number. Overheads due to PVM and disk I/O have been considered in performance evaluation. As shown in Figure 1, domain decomposition and VDHN algorithm has provided the best results: the overall efficiency never drops below 89.16 percent with respect to the linear one on the SP2 machine and 84.88 percent on the ALPHA cluster. This efficiency loss is mainly due to the time spent in program loading, load imbalance, software overhead, multitasking of Unix operating system and communication operations. Although we overlapped whenever possible computation and communication, the main source of this loss remains the protocol support of the Unix I/O subsystem. Exploiting this approach, real-time contingency analysis can be achieved solving several tens of transient stability analyses on an 8 node IBM 5P2 and DEC-ALPHA cluster. (Figure Presented) Figure 1. Speedup of the distributed VDHN vs. number of processors.
A distributed computing approach for real-time transient stability analysis
Bochicchio Mario Alessandro;
1997-01-01
Abstract
A challenge in the power system area is the implementation of realtime transient stability analysis. Time domain simulations of power system dynamics have been developed for over three decades and are routinely used for off-line planning and operations planning studies. The computational requirements for simulating power systems dynamics, even for few seconds after fault, represent the main impediment to the use of conventional transient stability programs in real-time. Parallel processing has been extensively used to speed up transient stability analysis and several results can be found in the literature. In recent years, distributed environments built out of pools of networked workstations have been widely used for solving computationally intensive problems that once belonged exclusively to the domain of supercomputers. The success of "cluster computing" is essentially due to the increasing power of workstations, based on fast RISC (Reduced Instruction Set Computer) microprocessors. Moreover, the development of high speed networks, capable to sustain up to Obitis, drastically narrowed the bandwidth and latency gap between an interconnection network in a parallel machine and a communication network in a distributed system. Distributed computing is also economically attractive with respect to traditional parallel computers, allowing to potentially provide at lower cost the processing power of a supercomputer. Heterogeneity and portability are primary goals of the developed distributed systems and programming environments. Heterogeneity allows to connect different machines from different vendors in a single virtual parallel computing environment. These important features can be achieved using programming environments recently developed, such as the Parallel Virtual Machine (PVM) from Oak Ridge National Laboratory and Emory University. The feasibility of real-time transient stability has been investigated exploiting both domain and functional decompositions, on a homogeneous cluster of eight Digital ALPHA and on an IBM SP2 machine. To test the functional decomposition, the Shifted-Picard algorithm has been implemented under PYM, whereas a scaled domain decomposition has been tested running multiple contingencies on different nodes of cluster systems, using the Very Dishonest Newton (VDHN) algorithm, which is the fastest sequential algorithm. In order to assess the performance of these approaches, time domain simulations adopting detailed modeling for synchronous machines have been carried out on a realistic sized network comprising 2583 buses and 511 generators. The distributed Shifted Picard algorithm is characterized by low performance. At first, this is surprising, since previous experiments on parallel environments showed interesting speedups. In order to understand the reason for this behavior, we have conducted a further investigation on the distributed SP algorithm. A communication model has been developed to take into account the main issues involved when a distributed environment is used. The tests carried out have shown that this model must account for the contention of the shared physical medium. Moreover, the SP algorithm was loosely synchronized, to guarantee a deterministic convergence behavior. Thus, the communication pattern resulted to be regular and symmetric, i.e. at a fixed time all the processors attempted to use the shared network. Although this property is effective for parallel processing solutions, our experience has shown that in distributed computing a more chaotic communication pattern would have produced better results, being at any given time the network resources exploited by at most one processing element on average. However, we were forced to this choice because asynchronous algorithms, due to their unforeseeable behavior, are not acceptable for real-time applications. The scaled domain decomposition has been implemented on the same test bed, using on each node the VDHN algorithm. In this case, the size of the problem (i.e the number of contingencies) scales with the available processor number. Overheads due to PVM and disk I/O have been considered in performance evaluation. As shown in Figure 1, domain decomposition and VDHN algorithm has provided the best results: the overall efficiency never drops below 89.16 percent with respect to the linear one on the SP2 machine and 84.88 percent on the ALPHA cluster. This efficiency loss is mainly due to the time spent in program loading, load imbalance, software overhead, multitasking of Unix operating system and communication operations. Although we overlapped whenever possible computation and communication, the main source of this loss remains the protocol support of the Unix I/O subsystem. Exploiting this approach, real-time contingency analysis can be achieved solving several tens of transient stability analyses on an 8 node IBM 5P2 and DEC-ALPHA cluster. (Figure Presented) Figure 1. Speedup of the distributed VDHN vs. number of processors.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
