LEARNING METHOD FOR NEURAL NETWORK AND ELECTRIC SYSTEM VOLTAGE/REACTIVE POWER CONTROLLER USING THE LEARNING METHOD

摘要

PURPOSE:To promptly complete learning even when a target value is rapidly changed by obtaining the error of the target value and the output value of a neural network and changing the weight of connection between respective neurons in the neural network. CONSTITUTION:By obtaining the error of the target value and the output value of the neural network (ST2), the sum of the square value of the error and the square value of the varied portion of the error is obtained and when the sum is larger than a prescribed value, the sum is varied in a direction to be made smaller by changing the weight of the connection between the respective neurons in the neural network. Since the variation portion of the error is considered in such a manner, whether the variation portions of the error is in the direction to be made smaller or in the direction to be made larger can be recognized as information by changing the weight of the connection between the neurons at the time of changing the weight of the connection between the neurons (ST5, 6) and thus, appropriate correction amount can be obtained and the period of time required for completing the learning can be shortened compared with a conventional method as a result.