Techniques for power system simulation using multiple processors
The thesis describes development work which was undertaken to improve the speed of a real-time power system simulator used for the development and testing of control schemes. The solution of large, highly sparse matrices was targeted because this is the most time-consuming part of the current simulator. Major improvements in the speed of the matrix ordering phase of the solution were achieved through the development of a new ordering strategy. This was thoroughly investigated, and is shown to provide important additional improvements compared to standard ordering methods, in reducing path length and minimising potential pipeline stalls. Alterations were made to the remainder of the solution process which provided more flexibility in scheduling calculations. This was used to dramatically ease the run-time generation of efficient code, dedicated to the solution of one matrix structure, and also to reduce memory requirements. A survey of the available microprocessors was performed, which concluded that a special-purpose design could best implement the code generated at run-time, and a design was produced using a microprogrammable floating-point processor, which matched the code produced by the earlier work. A method of splitting the matrix solution onto parallel processors was investigated, and two methods of producing network splits were developed and their results compared. The best results from each method were found to agree well, with a predicted three-fold speed-up for the matrix solution of the C.E.G.B. transmission system from the use of six processors. This gain will increase for the whole simulator. A parallel processing topology of the partitioned network and produce the necessary structures for the remainder of the solution process.