Parallel Implementation Of Power System Power Flow Algorithm Based On Spark Platform

With the continuous growth of electricity demands and constant expansion of the interconnection scale of power grid, the operation and control of power system are facing many serious challenges. It not only handles larger amounts of power system data, but also puts forward higher real-time requirements. Although using traditional high-performance computers, it spends a lot of time to solve high order power system problems with multiple equations. Traditional serial mode of power flow calculation cannot satisfy the requirements of online analysis and real-time control of large-scale power networks in processing scale or calculation speed. Therefore, a kind of feasible parallel computing scheme that supports fast solutions to complex problems is proposed. The accuracy of power flow calculation is ensured, and the increases of computing time consumption caused by the increasing of the problem scale of power flow are avoided.As a popular big data processing platform at present, Spark has widely been used in industrial fields especially in the Internet field, but research and application in power system is still in an exploratory stage. It is an attempt to use Spark for power flow calculation and analysis. Considering the optimization of node numbers and factor tables, a serial flow program based on Newton-Raphson Algorithm is programmed. Considering the language support and the characteristics of flow program, the program with Python is rewritten and then is made conform to the API Spark specification. By running the program on Spark, the parallel processing of power flow problems is realized. The results show the superiority of parallel computation in large scale problems, the calculation speed is improved indeed, and the larger the node scale, the greater the speedup ratio. Both serial algorithm and parallel algorithm based on block coordinate descent are programmed. Taking a decomposition optimization problem for example, the results show that the serial algorithm is better on the whole, parallel algorithms may require more iterations than the serial algorithm, but its convergence can be enhanced by reducing the value of the penalty parameter c, parallel algorithm is more suitable for independent distributed calculation of each control center.