Comparative study of blind source separation algorithms for harmonic source identification in power system networks

Abstract

Wide use of non linear loads in electric power systems results in the generation and propagation of current and voltage harmonics in the network which pollutes the entire power system. The harmonics generated by the harmonic sources need to be properly identified and estimated. Such a step will allow the harmonic perpetrators to be penalised and encourages the installation of dynamic filters at suitable locations for the compensation of harmonics. Various techniques are discussed in literature for the identification and estimation of harmonics that exist in power system network; these are mainly classified as single point and multipoint measurements. The methods based on neural networks, Kalman Filter and the like require apriori knowledge of the power system network topology. But, Independent Component Analysis (ICA) is a blind signal processing technique used for harmonic source identification and estimation which is independent of the topology of the power system network. In this research work, four different time structured ICA algorithms are evaluated for harmonic voltage and current estimation. The algorithms employed are Fast ICA (FICA), Efficient Variant Fast ICA (EFICA), Entropy Bound Minimisation (EBM) and Joint Approximate Diagonalisation of Eigen Matrices (JADE) ICA. A four bus system and an IEEE 14 bus system with non linear loads like Static Var Compensator, High Voltage Direct Current Transmission, Adjustable Speed Drives, Pulse Width Modulated Inverter and Electric Arc Furnace are considered to examine the competence of all the four algorithms. For comparison of the effectiveness of these algorithms figure of merit in terms of four error parameters namely maximum absolute error (MAE), absolute mean error (AME), mean square error (MSE) and average absolute percentage error (AAPE) are formulated. The error in harmonic estimation using the four algorithms is inconsistent and is dependent on the probability distribution of the known voltage and current values.