GPU-Based Criticality Analysis Applied to Power System State Estimation

Abstract

State Estimation (SE) is one of the main tools in the operation of power systems. Its primary role is to filter statistically small errors and to eliminate spurious measurements. SE requires an adequate number of measurements, varied, and strategically distributed in the power grid. Criticality analysis consists of identifying which combinations of measurements, forming tuples of different cardinalities, are essential for observing the power network as a whole. If a tuple of measurements becomes unavailable and, consequently, unobservable, the tuple is considered critical. The observability condition is verified by the factorization of the residual covariance matrix, which usually is a time-consuming computation task. Also, the search for criticalities is costly, being a combinatorial based problem. This paper proposes a parallel approach of the criticality analysis in the SE realm, through multi-threads execution on CPU and GPU environments. To date, no publication reporting the use of GPU for computing critical elements of the SE process is found in the specialized literature. Numerical results from simulations performed on the IEEE 14- and 30-bus test systems showed speed-ups close to 25 , when compared with parallel CPU architectures .

GPU parallel processing to enable extensive criticality analysis in state estimation

“…This paper proposes a GPU-based high-performance parallel computation platform to enable extCA, which has significantly improved from its early version. 30 Two steps of the parallel algorithm performed on the CPU are identified as a significant hurdle for extCA parallelization since they require data transfer from CPU to GPU (a time-consuming procedure). Thus, these algorithm steps-enumeration (measurement combinations) and update (storage of identified criticalities)-are now adapted and moved to GPU, performed together with the other two algorithm steps.…”

“…The proposed parallel algorithm previously identifies C1s, which are segregated from the dataset to not participate in the combinations of non-critical measurements, thus reducing the problem's search space, as advocated in Reference 7; this reduction may be substantial depending on the number of C1s (refer to the simulation results in Section 6.1). Hence, unlike, 30 with this slight change, C1s do not participate in the residual covariance submatrices defined in Section 3.…”

“…1. Contrasted with the CPU-multithread variant, the outcomes of GPU 30 reflect performance gains solely from migrating Steps 2 and 3 from the CPU to the GPU. On the other hand, the results of GPU [GTX] , which also executes Steps 1 and 4, exhibit superior speed-ups, confirming the benefit of eliminating the data transfer between architectures.…”

“…A challenging solution for managing the extCA problem's heavy computational burden is erecting a parallel computation platform using GPUs. Motivated by this challenge, the authors have advanced their research on extCA to make it viable in enlarged instances, encouraged by the preliminary results obtained in Reference 30.…”

“…When transferring data from the CPU to the GPU, these are copied from the CPU's memory to the GPU's, and vice‐versa on the way back, involving a time‐consuming process representing performance bottlenecks. This paper proposes a GPU‐based high‐performance parallel computation platform to enable extCA, which has significantly improved from its early version 30 . Two steps of the parallel algorithm performed on the CPU are identified as a significant hurdle for extCA parallelization since they require data transfer from CPU to GPU (a time‐consuming procedure).…”