Heuristic Methods and Machine Learning for Distribution Network Reconfiguration

Abstract

The problem of distribution network reconfiguration (DNR) has been addressed for some fourty years in scientific literature and industry. The aim of reconfiguration is to find the best radial operational configuration in a given electrical state. The main objectives are to reduce active losses and improve the profile. The reconfiguration problem is combinatorial and non-linear, making it practically impossible to solve for real-size electrical networks. To counter this problem, numerous heuristics and metaheuristics have been proposed in the literature.

In this thesis, we explore and compare different strategies of reconfiguration. The first method used is based on a minimum spanning tree algorithm to generate an initial solution. This solution is then refined using two heuristics: local search (LS) and tabu search (TS). Finally, a machine learning model is trained to mimic the behavior of these heuristics. This model uses various features to predict whether or not a line is part of the optimal configuration.

All methods are tested on different scenarios with and without distributed generators from the IEEE 33 and IEEE 69 networks. Results show that combining classical heuristics with learning-based approaches provides a balance between performance and computational cost, especially for applications requiring fast decision-making. This work contributes to ongoing research into the reconfiguration of constantly evolving electrical distribution networks.