Looney, Carl. “Pattern Recognition using Neural Networks-theory and Algorithms for Engineers and Scientists.” University of Nevada, Oxford University Press, 1997Beale, Hagan Demuth. “Neural Network Design”, International Thomson, 1996Haykin, Simon. “Neural Networks: A comprehensive foundation.”McMaster University, McMillan, 1994Schalkoff, Robert. “Artificial Neural Networks” Clemson University, McGraw-Hill, 1997.King,Roger and Novosel Damir. “Using artificial neural networks for load shedding to alleviate overloading in lines.” IEEE Transactions of Power Delivery, Vol 9, no 1, January, 1994.Software: Matlab simulink, Simpower Systems demonstrations.Neural Network Control of Power Systems.Patrick Avoke, Student Member, IEEE (Calvin College)Abstract: Like most other real world dynamical systems, power systems are non-linear hence require a convenient method of controlling the activities of the system. The approach to this problemoften involves linearization of the system and then the application ofvarious methods of linear systems controls to manage the system. Needless to say, the efficacy of the linearization step would determine how effective a selected control method would have on the chosen system. With the emergence of neural networks design, modern methods of controlling nonlinear system have been more accurate and convenient for the engineer to work with. In effect, it is possible to “train” neural networks to monitor a system for any irregularities or disturbances and initiate a process to restore “normal” operational conditions within the system based on forecasted results.I. Introduction.Selecting a control measure is often influenced by economic factors, speed of system, and state of the system as well as its sensitivity to other controls systems. Typical emergency conditions in a power installation involve overloading in the power lines. The primary measures for relieving overloaded lines are phase shifting, load shedding, tie line scheduling, generation shifting and controlled power system generation. Load shedding as a fix for overloaded lines in the long term has a correlation with overload levels, implementation of controlled separation and re-establishing power balance. Some adverse effects of uncontrolled load shedding include an increase in the system voltage, over-shedding as well as some undesired increases in line flow.Adibi and Thorne were one of the many sources of proposed controls solution for large power systems. They proposed a real-time control scheme for load-shedding in underground transmission networks. This brilliant scheme used approximate calculations to accelerate the solution time.Despite the cleverness of this system, it was observe that large interconnected power systems were very difficult to incorporate in any such schemes. A big part of the failure of the system to adequately address the standing problem was the lack of computer or communication support at the local control levels at the time. With the overwhelming preponderance of computer technology today, many more sophisticated control measures have hitherto been developed and tested successfully as a remedy.II. Artificial Neural Network Controls (ANN).Artificial neural networks were first developed in the early nineteen forties when a neurophysiologist, Warren McCulloch and a mathematician, Walter Pitts, wrote a paper on how neutrons might work by modeling a simple electrical circuit to describe the process. The idea with this model was to investigate the activity of neurons in the thinking process. In modern times, questions about incorporating neural networks to drive state of the art power systems grids have precipitated growing interest in ways to simulate and control the power system.Figure 1: General role of Neural Network.The artificial neural network as defined by Schalkoff(Schalkoff, 2), is a network composed of a number of interconnected units with each unit having input or output characteristics that implements a local computation or function. Typical neural networks operate in parallel nodes whose function is determined by the network structure, the connection strengths and the function in each node. Neural networks have the unit ability to “learn”. In other words, the human does not necessarily have to be able to explain the “problem” to the system. Designing neural network solutionsfor systems often starts with a series of questions regarding the system such as: “What sort of problem does one seek to solve?”, “Can the network be trained to solve the problem?” and “What would be the best network structure to solve the problem?”(Schalkoff, 11). Once these questions are addressed, parameters for designing the system can be define to include the network structure, training procedure, testing and input/output parameters of the ship.Neural nets can be conveniently described as black-box computational methods for addressing basic Stimuli-Response processes (S-R). On each side of the black-box (ANN) is a known set of inputs corresponding to their respective output set hence any distortion in the input of the system would employ algorithms and codes within the black-box to produce a unique output for that stimulus. It is throughthis process that the “new” output is added to the already existing set of standard neural network responses for known stimuli. It is important to note that the standard S-R pairs encoded into the artificial neural network ought to represent the stable states of the system during normal operation. The approach to “learning” by ANN’s could take the form of deterministic methods like back-propagation and Hebbian approaches or could involve the stochastic approach such as genetic algorithms or simulated annealing. 1Figure 2: A Multi-layer PerceptronFigure 3: Showing structure of recurrent neural network.III. Problems with Neural Nets.Neural networks work quite well with predicting outcomes of non-linear systems in the event of a fault but would obviously need an “initial standard” called the trainingset to compare any fault signals to. This standard would basically indicate to the system whether parameters coming through fall within normal operational condition. It is almost safe to assume therefore that the efficacy of a neural network within a power installation is premised on the quality of the initial training set. The concern with neural nets in this respect is that composing training
View Full Document
Unlocking...