Neural Networks 22 (2009) 833–841 Contents lists available at ScienceDirect Neural Networks journal homepage: www.elsevier.com/locate/neunet 2009 Special Issue Comparison of a spiking neural network and an MLP for robust identification of generator dynamics in a multimachine power system Cameron Johnson ∗ , Ganesh Kumar Venayagamoorthy, Pinaki Mitra Real-Time Power and Intelligence Systems Laboratory, Missouri University of Science and Technology, Rolla, MO 65401, USA article info Article history: Received 7 May 2009 Received in revised form 7 June 2009 Accepted 25 June 2009 Keywords: MLP Multimachine power system Neuroidentification Spiking neural network abstract The application of a spiking neural network (SNN) and a multi-layer perceptron (MLP) for online identification of generator dynamics in a multimachine power system are compared in this paper. An integrate-and-fire model of an SNN which communicates information via the inter-spike interval is applied. The neural network identifiers are used to predict the speed and terminal voltage deviations one time-step ahead of generators in a multimachine power system. The SNN is developed in two steps: (i) neuron centers determined by offline k-means clustering and (ii) output weights obtained by online training. The sensitivity of the SNN to the neuron centers determined in the first step is evaluated on generators of different ratings and parameters. Performances of the SNN and MLP are compared to evaluate robustness on the identification of generator dynamics under small and large disturbances, and to illustrate that SNNs are capable of learning nonlinear dynamics of complex systems. © 2009 Elsevier Ltd. All rights reserved. 1. Introduction Online identification of generator speed and terminal voltage characteristics is very much essential for fast, intelligent and adap- tive control in today’s power system (Singh & Venayagamoorthy, 2002). Classical controllers for generators are generally designed based on linearized models obtained around some nominal operat- ing point. But, in a real world power system, the environment con- tinuously changes and a generator’s dynamics also change since it is an integrated part of the power system (Park, Venayagamoorthy, & Harley, 2005). In these situations, the performance of classical controllers such as an automatic voltage regulator (AVR) generally degrades, and intelligent AVR designs are called for. A Neural Net- work (NN) is a very effective tool for designing these types of in- telligent controllers. In order to take the correct control action in a dynamically changing environment, an NN based controller needs a neuroidentifier, which provides an estimation of the speed and terminal voltage characteristics of a generator from past values of speed and terminal voltage. The method of neuroidentification is also very effective for wide area monitoring and control (Venayag- amoorthy, 2007) and finding dynamic equivalents of large power systems (Azmy, Erlich, & Sowa, 2004; Stankovic, Sarik, & Milosevic, 2003). ∗ Corresponding author. E-mail addresses: cameron.e.johnson@gmail.com (C. Johnson), gkumar@ieee.org (G.K. Venayagamoorthy), pm33d@mst.edu (P. Mitra). So far, different types of neural network architectures and their performances have been studied for the purpose of neuroidentifi- cation (Azmy et al., 2004; Park et al., 2005; Singh & Venayagamoor- thy, 2002; Venayagamoorthy, 2007). This includes Multilayer Perceptrons (MLPs), Radial Basis Functions (RBFs), Recurrent Neu- ral Networks (RNNs), and Echo-State Networks (ESNs). But with the advancement of neuroscience it has become clear that none of these networks actually represents the structure and function of biological neurons (Mishra, Yadav, Ray, & Kalra, 2007). In a traditional neural network, a complete set of inputs is fed into the network, each neuron is allowed to activate once, and a single set of outputs is produced in a single time slice. There is no master clock; the network operates in time steps defined by these samples of inputs and their resulting outputs. The biological neuron (and any network based thereon), however, operates in continuous time. Inputs come in as a string of voltage spikes called ‘‘action potentials’’. A traditional artificial neuron uses weighted multipliers and simple summation to generate a net input value for an activation function, whose output is the neuron’s output. Inputs and outputs are real-valued numbers. A biological neuron receives the action potentials, which drive up the voltage on its main body’s membrane. The voltage on the main body (called the ‘‘soma’’) decays quickly, but if enough spikes arrive (usually from multiple neurons) in a short enough period of time, the biological neuron fires (Mishra et al., 2007). The artificial neuron designed to more closely model those found in biological systems is known as a spiking neuron, and a network based upon this type of neuron is referred to as Spiking Neural Network (SNN). First proposed in the Hodgekins/Huxley model in 1959 (Mishra et al., 2007), an artificial spiking neuron 0893-6080/$ – see front matter © 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.neunet.2009.06.033