PNNL

Abstract

Accurate topology identification in a distribution system is essential for its correct operation. The distribution system can have many configurations based on the position of switches, so a generalized state estimator is popular in existing literature for topology error identification; however, such combinatorial problems are computationally expensive. On the other hand, search-based methods such as heuristics procedures and knowledge-based techniques require evaluation of all possible topologies and hence pose computational challenges for large systems. A machine learning model like the ranking support vector machine used in this paper enhances the operation of such topology enumeration process by ranking the possible topologies in order of their likelihood of being the correct topology in descending order. This ranked topology then allows the generalized state estimator and search-based methods to go through the possible topologies in a probabilistic way increasing the chances that a topology will be identified within the first few choices. The ranking support vector machine is accomplished by leveraging trend vectors assembled from measurements received from sensors on a simulated distribution grid. These trend vectors along with topology labels allow for the training of the model for its use in the ranking prediction. The results show that the method can utilize noisy measurements under changing loads.