Dept. of Electrical Engineering Jadavpur University, Kolkata – 700 032 India P.Rajamani* S.Chakravorti K.Karthikeyan J. Electrical Systems x-x (xxx): x-xx Regular paper A comparative study on the performance of modern function estimation tools for estimation of non-linear functions in high voltage system To safeguard the electric power system, protecting the overhead transmission line from lightning stroke is an important task. For evaluating the lightning performance of transmission line several algorithms were used. Choosing a proper learning algorithm to train the network is very important for better estimation. Gradient based function estimation tools like Radial Basis Function (RBF) and back propagation Artificial Neural Network (ANN) are mainly used for this purpose. In this paper an attempt is made to compare the performance of modern function estimation tools, viz. Particle Swarm optimization (PSO) aided Wavelet neural Network (WNN), gradient based WNN, and back propagation ANN for estimation of such non-linear functions. The simulated input, output data of lightning impulse voltage, non-sinusoidal current and chopped impulse voltage were used for training and testing data of developed network for estimation. Results are presented for modern function estimation tools for learning these non-linear functions and showed that the PSO aided network estimates function more accurately than gradient based algorithms. The PSO aided WNN method can be used by electric power utilities as a useful tool for the design of electric power systems, alternative to the conventional analytical method. Keywords: PSO aided WNN, WNN, function estimation, impulse voltage, non-sinusoidal current I. INTRODUCTION Developing models from observed data, is a fundamental need in many fields, viz. signal processing, forecasting, statistical data analysis and classification. This is frequently referred as function estimation and it involves estimating the underlying relationship from a given finite input-output data set. Estimation of general functions by non-linear networks such as multilayer perceptron and radial basis function networks are important tools for system modelling and identification. For more than a decade, neural networks have received considerable attention due to its self-learning, massive parallelism, and ability to self-adapt. It finds many applications particularly in the field of signal processing. The feed forward multi-layer neural networks, along with the back-propagation (BP) training algorithm, have been widely used for function estimation since they provide a generic, non-linear, black box for function representation. Besides this, the wavelet decomposition, being a powerful tool for signal processing, has also emerged as a new tool for function estimation. Because of the similarity between wavelet decomposition and one-hidden-layer neural networks, the idea