FAST APPROACH TO ARTIFICIAL NEURAL-NETWORK TRAINING AND ITS APPLICATION TO ECONOMIC LOAD DISPATCH

Artificial Neural Networks (ANN) are gaining popularity in various fie lds of engineering including electrical power systems due to their hig h computational rates and robustness. One of the ANN models extensivel y used for power system applications is the multilayer perceptron mode l based on back propagation algorithm. However,its training requires l arge number of input-output data sets which increases with system size and may become prohibitively large and time extensive, Moreover, the back propagation algorithm offers slow convergence with random initial weights. This paper presents a new approach to minimize the number of training patterns for ANN by using variable slope of the sigmoidal fu nction for different test cases. In addition, the paper suggests the u se of new functions for generating initial weights for training. The A NN models so developed have been tested to solve economic load dispatc h (E.L.D.) problem on IEEE-14 bus test system and 89-bus Indian system . The proposed approach provides tremendous saving in the training tim e of ANN and provides fast and accurate results of E.L.D.