This paper presents a single-instruction-multiple-data (SIMD) based implementation of the transient stability simulation on the Graphics Processing Unit (GPU). Two programming models to implement the standard method of the transient stability simulation are proposed and implemented on a single GPU. In the first model the CPU is responsible for part of the simulation, while the onerous computations were offloaded to the GPU, creating a hybrid GPU-CPU simulator. In the second model, the GPU performs all the computations, while the CPU simply monitors the flow of the simulation. The accuracy of the proposed methods are validated using the PSS/E software for several large test systems. A substantial increase in speed was observed for the GPU-based simulations.

SIMD-Based Large-Scale Transient Stability Simulation on the Graphics Processing Unit

This paper proposes large-scale transient stability simulation based on the massively parallel architecture of multiple graphics processing units (GPUs). A robust and efficient instantaneous relaxation (IR)-based parallel processing technique which features implicit integration, full Newton iteration, and sparse LU-based linear solver is used to run the multiple GPUs simultaneously. This implementation highlights the combination of coarse-grained algorithm-level parallelism with fine-grained data-parallelism of the GPUs to accelerate large-scale transient stability simulation. Multithreaded parallel programming makes the entire implementation highly transparent, scalable, and efficient. Several large test systems are used for the simulation with a maximum size of 9,984 buses and 2,560 synchronous generators all modeled in detail resulting in matrices that are larger than 20, 000 × 20, 000.

Large-Scale Transient Stability Simulation of Electrical Power Systems on Parallel GPUs

A new algorithm for optimal tearing of networks is presented. The algorithm is based on the simulated annealing principle and is able to divide a power system network model into a number of subnetworks in order to optimise the use of parallel computer systems for network analysis. Computational examples are given which illustrate the properties of the new algorithm. The results obtained are compared with those obtained by the iterative improvement method, and it is shown that the annealing algorithm finds significantly better solutions.

Optimal network tearing using simulated annealing

Power system online dynamic security assessment (DSA) is a challenging computing problem. A key problem in DSA is the analysis of a large number of dynamic stability contingencies every 10-20 minutes using online data. In order to speed up the transient stability analysis, parallel processing has been applied and several results can be found in the literature. In this paper, the authors present a distributed approach for real-time transient stability analysis. Distributed computing is economically attractive providing the processing power of supercomputing at a lower cost. Several distributed software environments like the parallel virtual machine (PVM) allow an effective use of heterogeneous clusters of workstations. Both functional and domain decomposition of the transient stability problem were tested under PVM on a homogeneous cluster of eight DEC ALPHA and on an IBM SP2 machine. Functional decomposition has been obtained by the Shifted-Picard algorithm, whereas domain decomposition has been obtained concurrently running different contingencies on different nodes of the cluster, using the very dishonest Newton 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 power network comprising 2583 buses and 511 generators.

A distributed computing approach for real-time transient stability analysis

Parallel processing is a promising technology for the speedup of the dynamic simulations required in power system transient stability analysis. In this paper, three methods for dynamic simulation on parallel computers are described and compared. The methods are based on the concepts of spatial and/or time parallelization. In all of them, sets of linear algebraic equations are solved using different versions of conjugate gradient methods which have been successfully applied in other scientific and engineering applications. The algorithms presented in the paper were tested in a commercially available parallel computer using an actual large power system model. The results obtained in the tests showed a considerable reduction in computation time.

Conjugate gradient methods for power system dynamic simulation on parallel computers

A network of computers arranged in a modified ring structure has been designed for solving in parallel the equations describing dynamic disturbances on large power systems. The operation of this network when solving a 1732-bus, 398-generator system using an algorithm based on the Bonneville Power Administration's Transient Stability Program has been simulated on a CDC 6600. The purpose of the simulation was to determine how efficiently the algorithm can be executed in parallel on this multiprocessor configuration. The critical part of the algorithm is the forward and back substitution passes through the LDU decomposed I = YE equations. Parallel execution times for various combinations of processors and the corresponding gains over a conventional serial solution are given.

Simulation of a Multiprocessor Network for Power System Problems

A network of computers arranged in a modified ring structure has been designed for solving in parallel the equations describing dynamic disturbances on large power systems. The operation of this network when solving a 1723-bus, 398-generator system using an algorithm based on the Bonneville Power Administration's Transient Stability Program has been simulated on a CDC 6600. The purpose of the simulation was to determine how efficiently the algorithm can be executed in parallel on this multiprocessor configuration. The critical part of this algorithm is the forward and back substitution passes through the LDU decomposed I = Y E equations. A scheduling algorithm that assigns tasks to the processors taking into account the precedence relationships necessary in the solution, the communication requirements, and the loadings on the individual processors has been developed to minimize the computing time for this part of the solution. This method gives better results than were obtained with other methods tried earlier.

S. C. Kang

Papers

Simulation of a Multiprocessor Network for Power System Problems

Design of multiprocessor structures for simulation of power-system dynamics. Final report