In order to obtain long-term predictions based on short-term data, a neural
network model was developed. The model parameters are indoor and outdoor t
emperature difference and energy for heating and internal use. For purposes
of training the neural network model a method for extending the measured d
ata to represent an annual variation is proposed. The method has been appli
ed on six single-family buildings.
Based on access to data from 2 to 5 weeks, the deviation between predicted
and measured dirunal energy demand on an annual basis was about 4% with a c
orrelation of 90-95%. when access to the indoor and outdoor temperature dif
ference was assumed. For models based on access to data from the warmest pe
riods with a very small heating demand, the deviation was about 2-4 times l
arger. (C) 2001 Elsevier Science B.V. All rights reserved.