High Performance Basic Algorithm Of Integrated Grid Based On GPU-Acceleration

With the continuous development of smart grid,the scale of the power system is becoming higher and higher and the requirement of computing accuracy is stricter,at the same time,there also requires to be fast enough to satisfy the requirement of real-time computation.As a result,calculation of power system is being challenged.At the same time,the general purpose computing of GPU is developing very quickly,and its computing speed is far better than the CPU in the field of parallel computing,so it should be used in the calculation of power system.A set of high performance basic GPU algorithm of power system and its implementation is mainly introduced in this paper.First of all,this paper analyses the some of the common algorithm in power system to find out how it can be parallelization and acceleration.The result shows the key factor is large-scale sparse group of linear equations solving.And then,this paper analyses the parallelization of large-scale sparse group of linear equations solving.There are two main methods to solve large-scale sparse group of linear equations,which is the direct methods and iteration methods.LU decomposition is the most commonly used methods in direct methods,and the main difficulty is to reduce injection numbers when performing decomposition.To solve this problem,node ordering algorithm based on graph partitioning is proposed in this paper.This methods works well in reduce injection numbers and cost little time comparing to traditional algorithm.On the other hand,LU decomposition performs badly in large-scale systems,so iterative method should be considered to improve the speed of calculation.It is generally knows the Krylov subspace method is have a better performance above other iterative methods.Krylov subspace method is a series of collections,the BICG methods is one of it and works well in power flow calculation.The Krylov subspace is fast in speed but low in reliability,if the condition number of coefficient matrix is too big or spectral of coefficient matrix dispersively,the algorithm will not convergence.At this time,a preconditioner should be used to improve the distribution of spectral.A good preconditioner should balance the time consuming and result,and on the other hand,convenient to parallelization.There are a number of general preconditioner but no preconditioner optimize direct to power system.This paper presents a preconditioner suitable to N-L power flow,and it can works on GPU fast and efficiently.After the basic algorithm is developed,how to communicate with other system is still a problem for a calculation package.If offline computation is been used,a large number of manpower and material resources to transmit data,and the time reduced by better algorithm will be vanished.So it is very important to make the calculation package a service online.This paper use Protobuf to serialize data and transmit it with ActiveMQ,and make it an efficient and safe API.After unit testing and integration testing,high performance basic algorithm package of integrated grid based on gpu-acceleration finished.The result of test shows this package can solve power flow calculation of systems with millions nodes.