The possibility of using neural networks in the field of the energy coefficients correction of a power supply system with uneven load in phases is being studied. This need is justified by the fact, that the calculation of the necessary parameters of the symmetry-compensating device was previously based on the Nelder – Mead search optimization method. Search optimization performing is computationally expensive, takes long computation times, and may calculate anomalous values. The article develops the idea of using technology for predicting the parameters of a symmetry-compensating device, based on neural network modeling using Bayesian regularization. For a given set of initial data, the best selected configuration turned out to be a neural network of two layers, implemented in the MATLAB package using the machine learning tool Neural Network Toolbox. The network input parameters are a set of tuples, consisting of load values in each of the three phases of the power supply system, which are resistive-inductive in nature. There are six input quantities in total (load resistance and inductance values in each of the three phases) and all their values are different, which causes current asymmetry in the network and reactive power. The target matrix is formed from tuples, consisting of three values, which are the parameters of the symmetrical compensating device, calculated by the optimization method, in such a way as to compensate reactive power and to balance currents in the network. The number of data tuples, required to train a neural network was determined empirically. During the experiments, the optimal number of neurons in the neural network was also revealed. The use of the generated neural network to calculate the parameters of the symmetry-compensating device determined approximate solutions is comparable in accuracy to the values, found by optimization methods. With the help of the generated neural system, adequate quasi-solutions for calculating the parameters of the symmetry-compensating device were determined, which, in case of calculation, using the optimization method, led to anomalous values, that didn’t optimize the energy coefficients of the power supply system to the required extent. Also, such neuropredictions protect the system from receiving excessive high parameters of symmetry compensating device, which can be obtained with an optimization approach.
Read full abstract