Artificial neural network techniques for on-line distribution network reconfiguration based resistive loss reduction

Network reconfiguration for time varying loads is a complicated and non-lin ear optimization problem which can be solved efficiently by Artificial Neur al Networks (ANNs), as the ANNs are capable of learning a tremendous variet y of pattern mapping relationships without having a priori knowledge of a m athematical function. In this paper, four-generalized ANN models are propos ed for on-fine network reconfiguration under varying load conditions. For t he first two models the training sets are generated by considering various load types and fixed load levels to reflect the load variations that are no rmally encountered in practice. Actual loads (P, Q) are used as the input v ectors for the ANN model-I and fixed load levels of various load types are used for the ANN model-II. The training data are generated from the Daily L oad Curves (DLCs) with the same load trends for P and Q in the case of ANN model-III, and different load trends in the case of ANN model-IV. The four ANN models are designed to suit the systems having different complexity and load characteristics, and the one, which is most suitable, can be used to predict the switching status of the dynamic switches in optimizing the netw orks for loss minimization. All the four proposed ANNs are trained using Co njugate Gradient Descent Back-propagation Algorithm, and tested by applying arbitrary input data generated from typical DLCs. The test results of each of the four ANN models are found to be the same as that obtained by offlin e simulation of loss minimization algorithm. A scheme for on-line implement ation of the proposed ANN models is also presented. The proposed ANN techni ques are compared with two other methods in literature and found to give be tter performance, and simple to design and implement.