Empirical neural network studies for multi-port load calculation by the input currents

Abstract

- A linear multi-port is considered a model of a wire communication line with physical quantities sensors or as a power loads supply line. The problems of known methods are shown to determine the multi-port parameters and the calculation of load resistances by specified or measured input currents. In the present work, the “loads‒currents” relationships are approximation tasks of feedforward neural networks. The corresponding input currents are calculated for a particular set of load values, using the multi-port models with one, two, and three loads. This is how the training or input vector (input currents) and the target vector (loads) are composed, the dimension is equal to the amount of input currents or loads, and the size corresponds to the load set. Numerical experiments by the Fit Data package of MATLAB Deep Learning toolbox demonstrate the accuracy of load calculation and capability to generalization. An introduced quantitative index of the quality of training allows us to identify the minimum size of the training vector and the optimal amount of hidden layers’ neurons. The obtained results provide purposeful and fast network training.