The differences between indoor air temperature measurements and simulations (residuals) are represented in Figure 9. The indoor air temperature measured has been averaged at 1 h time step for comparison with simulation outputs.

* CeFsiu$

2 1,5 1

0,5 0

-0,5 -1 -1,5

72 172 272 372 472 572

Time (hour)

Figure 9: Residuals time evolution From April 28th 2003 to May 19th 2003.

The mean value of the residuals is 0.37 °C and its standard deviation is 0.45 °C. Residuals time evolution is not stationary. It present trends which follow outdoor temperature trends. The model globally underestimates the indoor air temperature during the warmer periods and overestimates it during the colder periods. It reveals a deficient model behaviour. Besides residues are lower during colder days than during warmer ones, it indicates that the model has performed better for colder periods.

The power spectral density of the residuals (see Figure 10) shows that 96% of the variance is concentrated on the frequency range [1, 1/10 h-1] and present peaks at 1/24 h-1 and 1/12 h-1 frequencies. It reveals that model disagreements with measurements appears mainly in the frequency range in which the building is manly excited, as expected according to [5].