Search academic texts

Information × Registration Number 0205U002210, 0103U001339 , R & D reports Title Using Feed-Forward Neural Networks with Noniterative Training for Prediction of Consumption of An Electrical Energy popup.stage_title Head Rashkevych Yuriy Mykhaylovych, Registration Date 08-02-2005 Organization Lviv Polytechnic National University popup.description2 In the final report on a theme " Using Feed-Forward Neural Networks with Noniterative Training for Prediction of Consumption of An Electrical Energy" the factors which have an affect on a consumption level (the type of day, time of day, weather conditions) have been selected . It is proved that an accuracy of prediction depends on length of time series, degree of dependence of consumption from time, preservation of conditions, interval of prediction. The software for build-up of load curves, selection of trends, analysis of residuals and finding of periodicity on the basis of methods of a spectral analysis has been advanced; the presence of cycles of a different origin and duration (daily, week and annual) has been ascertained. It is proved, that length of an interval of a prediction depends on length of an interval of a previous history used for an estimation (should be at 5-10 of time less). The algorithm and program unit of input data preprocessing providing detection and elimination of missing data and improbable valueson the basis of principal component analysis have been designed. Using architecture of feed-forward neural network with nonlinear synapses on a basis of a "a functional on set of tabular functions" paradigm with a noniterative base training principle has been offered for prediction. An algorithm of training and program model for the PC including units of training, operation and visualization with two-dimensional situational feature maps have been created. The technique of usage of the program predicting complex in systems of dispatching control of deliveries of energy has been designed. The results of test researches of neural network program complex on the database of JSC "Lvivoblenergo" have confirmed possibility of short-term prediction of a current consumption with an error up to 10 %. Product Description popup.authors popup.nrat_date 2020-04-02 Close