Classification of harmonic-producing residential type loads using artificial neural networks

In this thesis, it is aimed to develop artificial neural network (ANN) models for the classification of residential type single-phase nonlinear loads in the system with harmonically distorted source voltage. As residential type loads, six different loads were considered, namely: a full-wave rectifier feeding a resistive load, a half-wave rectifier feeding a resistive load, a dimmer circuit feeding a resistive load, a full-wave rectifier charging a battery, a single-phase asynchronous motor with a permanent capacitor, and a heater. Accordingly, in this thesis, the time-domain model of a system with the considered loads connected to the same bus was first created in the Matlab/SIMULINK environment. In the modelled system, the behaviours of various harmonic and power quality indices of the loads were statistically analysed under randomly generated distorted source voltages. Then, the harmonic and power quality indices measured at the common connection point for possible load combinations under a large number of distorted source voltages were analysed using the Pearson correlation heat map method. Based on the analysis results, 15 features were identified for use in load classification. Thus, a total of 6 different ANN models, each with 15 inputs and 6 outputs, were developed. Note that depending on the number of hidden layers, the models can be divided to single-layer and double-layer models and based on the activation function in the hidden layers, they can be divided to the models with tanh, sigmoid, or ReLU functions. The number of neurons in the hidden layers of the developed models was optimized by considering several performance parameters. Finally, as a result of the comparative evaluations among the optimized ANN models, it was determined that the ANN model with 2 hidden layers, 26 neurons in the first hidden layer and 30 neurons in the second hidden layer, using tanh activation functions in the hidden layers and sigmoid activation functions in the output layer, showed the highest performance with 94.11% training accuracy and 93.56% test accuracy.