Daily load curve clustering and prediction by neural model tool box for power systems with non-stochastic load components

摘要

A short-term load curve forecasting method based on neural network models was created by means of a neural network tool box in a two step concept: For selection of appropriate training sets of comparable daily demand patterns typical load profiles for different day-types are classified by Kohonen network. The weather-load-correlation is modelled by a multilayer feed-forward-perceptron. To enlarge the training data base of stochastic load curve samples “uninfluenced” demand profiles are reconstructed by modelling and filtering the effect of deterministic load control. Experiences with real data from an utility are reported.