Study On State Estiamtion Algorithm Of Power System Based On GPU Acceleration

In recent years,the rapid development of UHV AC-DC hybrid large power grids requires the network analysis and application of the power dispatching system to be capable of unified analysis of the whole network.Power system state estimation is the basis of online analysis application.The traditional serial calculational mode of state estimation suffers from excessive communication and calculation burden,which is difficult to meet the real-time requirement of power dispatching system.So it is important to develop new state estimation algorithms for large power grid,with fast computing speed and numerical stability,which also effectively shorten the execution cycles.In this paper,the state estimation algorithm is accelerated in parallel by the graphics processing unit(GPU).The specific research contents are as follows:Firstly,the principle and algorithm of the classical basic weighted least squares method for state estimation are studied.The complex computing tasks involved in the weighted least squares method are analyzed in detail.Secondly,according to the computational model of CPU+GPU heterogeneous platform,the overall process of state estimation parallel algorithm based on GPU acceleration is proposed.The task allocation strategy of CPU and GPU in hybrid architecture,CPU process control,GPU parallelization scheme and data transmission have been optimized.Furthermore,the GPU kernels such as the measurement Jacobian matrix generation,the information matrix solving process and the large sparse matrix multiplication algorithm,the parallel LU decomposition algorithm of the sparse linear equations and the forward-backward substitution of triangular equations are designed in detail.By means of optimization strategies such as memory merging access and fine-grained parallelism mining,the measurement Jacobian matrix of the 9241 node power system generated by GPU only needs 0.45 ms.The kernel for large sparse matrix multiplied by its transposition in the information matrix algorithm can reach the highest acceleration of 3.9 times compared with the CUDA library function.The sparse LU decomposition method optimized by domino technology is designed to solve sparse linear equations.It takes only 59 ms to solve the sparse linear equations on GPU for 9241 nodes power system.Finally,based on the K40 C GPU card,five power system examples of different sizes are used to test the performance of the above GPU acceleration state estimation algorithm.The test results show that the parallel algorithm proposed in this paper control the state estimation time of the 9241 node power system within 1s and achieve an excellent acceleration effect compared to CPU algorithm.