Algorithm selection for power flow management

Abstract

Algorithms are essential for solving many important problems, including in power systems control, where they can allow the connection of new demand and generation whilst deferring or avoiding the need for network reinforcement. However, in many problem domains no algorithm always delivers the best performance for all problems, so better performance can be achieved by using algorithm selection to select the best algorithms for each problem. This work applies algorithm selection to power systems control, with power flow management using generator curtailment examined as a representative power systems control task. The first half of this work focuses on whether potential performance benefits are available if algorithms are selected optimally for each network state. Five power flow management algorithms are implemented, which use diverse approaches such as optimal power flow, constraint satisfaction, power flow sensitivity factors, and linear programming. Four case study power systems – an 11 kV radial distribution system, a 33 kV meshed distribution system, the IEEE 14-bus system, and the IEEE 57-bus system – are used to test the algorithms over a extensive range of network states. None of the algorithms give the most effective performance for every state, in terms of minimising either the number or energy of overloads, whilst minimising curtailment. By optimally selecting algorithms for each state there are potential performance benefits for three of the four case study systems In the second half of this work, algorithm selection systems (selectors) are created in order to exploit and deliver the observed potential performance benefits of per-state algorithm selection. Existing techniques for creating algorithm selectors are adapted and extended for the power flow management application, which includes the development of a training method that allows selectors to consider two objectives simultaneously. The selectors created take measurements of network state as input and use machine learning models to make algorithm selection decisions. The models either directly predict which algorithm is likely to be the most effective, or predict the performance of each algorithm, with the algorithm with the most effective predicted performance then being selected. Both of these approaches are shown to be effective in creating algorithm selectors for power flow management that deliver statistically significant performance benefits. In some cases, the selectors are able to match the optimum performance that could be achieved by selecting between the algorithms.