Neural Network Training Compared for BACKPROP QUICKPROP and CASCOR in Energy Control Problems

Abstract

Gradient-based neural network learning algorithms, specifically the backpropagation (BACKPROP), quickpropagation (QUICKPROP), and cascade correlation (CASCOR), have been re-implemented, tested, and compared for making predictions of time series for daily energy (gas) consumption. Daily records of energy consumption form a time series with information that can predict energy demand for a few days. Initially, the basic prediction is made from the past series without taking external influences into consideration. Later, other independent predictive data series can be included to improve prediction. Repeated presentation of hundreds of sample patterns systematically trains a network which converges upon connection configurations that predict expected demand. It has been demonstrated that the learning algorithms can consistently train networks to predict, one-day-ahead over the succeeding year, with the average error reduced to below 20%.