In france the main factors of differences are : underestimation of energy consumption during conception because our standards use a low indoor temperature (19°C) and more often a steady outdoor temperature. In reality measured indoor temperatures are around 21 °C. Ventilation is also too low so it is done by window openings. Of course users behaviour play a major role, but the defaults in building during the construction phase and also poor maintenance and regulation of the HVAC systems.

Mainly - assumptions made by those who made simulation model...

More serously - depends on what tipe of building is modeled and on which energy performance level it is.

For example - low energy building in cold climate are highly influenced by internal gains and user behavior on temperature set point. But there are lots of on going research on that.

While energy consumption is highly dependent on transmitance loss (at low energy buildings on cold climate) degradation or moisture damage on insulation materials and decrease of thermal transmitance in high efficiency glazing (there are lots of runors that gas filling dissapears after 5 years) could couse discrepancies on results.

If there are interest - I can suggest some energy audit based comparison study on discrepancies between calculated and measured energy consumption in cold climate.

I support Mr. Betz idea about accuracy of models.

Actualy - there are posibility to get idealy similar results for yearly results but for hourly results there will be "huge" discrepancies. While these diferences are positive and negative, error compensation could lead to ideal match on annual results.

Maybe it is quite god to have 10% missmatch? Or better way are to use uncertainty analysis? It looks like ideal solutions for some unknown, non-stable and untrendy influencing parameters...

Human activity and other unforeseeable variables (dynamic variables).

Do not mix the accuracy of the model (i.e the series of algorithms that handle input data to predict the energy use of the building) and the accuracy of input data.

Several studies have shown that, when input data are acurate enough, (i.e. those measured for the modelled building), the predicted energy use is very close to the measured one for most models available and used since some time. These models have been the object of a so-called validation procedure, checking that for a set of measured (i.e actual) input data, they give the same output as the measured one. If this procedure is completely and carefully followed, the model should have no error at all.

The main factors for the discrepancy can be sorted in two families:

1) the errors made by the model user himself. In the frame of the COMIS project, we asked several persons to compute, using COMIS, the airflow rates in two buildings. One of these was very accurately described in very details. All input data was given, the users had only to provide these data into the programme. It was found that several peiople dide some input mistakes, leadfing of course to erroneous results.

2) The input data are never all known accurately, The user has to estimate their value or to pick them in tables that give average or common values, but certainly not those that exactly correspond to the modelled building. The building modelled this way is an "average" or "commonly agreed" building, but not the building that uses the measured amount of energy.

I will end with two examples: Measurements of energy use of series of identical homes in several countries have shown that the energy uses is distributed in a bell-shaped distribution, spread between 50% and 150% of the mean value. This is generally attributed to the inhabitant behaviour. When models are used an average inhabitnat is used as input, and it is very difficult to get input data corredponding to teh behaviuour of a given person.

If you run a few hundreds of time a model with the inpuit data of a given building, changing at eeach run all input vasriables at random but remaining within a reasonable range (e.g. dimensions ± 0.1%, U-values ± 10%, airflow rates ± 20

%, etc.) you get hundreds of results that are also distributed within a pretty large range. The lowest the energy use of the buildign, the largest the range is. This test is a so-called MonteCarlo sensitivity analysis. Other techniques for the sensitivity analysis can give the effect of the variation of each variables or combination of variables on the result.

A model is 'only' a model and all those used in designing buildings to the specifications of the building regulations use 'standard' input parameters. The reason for this is that you want to compare buildings from the point of view of their substance in the same way you would compare vehicle fuel consumption on a 'rolling road'. Everyone knows that actual fuel consumption will be higher and strongly influenced by driver behaviour. And the same is true for buildings ... as was stated before, the (heating) energy consumption in (seemingly?) identical buildings can vary by +/- 50%.

There are four main factors for this

(1) occupants behaviour: is all floorspace evenly heated, how many people in the household work/are at school during the daytime, how many children etc.

(2) weather: has to be corrected for by using heating degree days

(3) correct depiction of the building envelope in the model: many details (e.g. corner three dimensional heat flow) are only covered by average values and factors and this does not necessarily represent an actual building situation

(4) sufficiently detailed modelling of the heating and warm water system and the internal loads: mostly these are only covered by average factors, again.

Finally: you always have to be careful about the first heating season: due to the water stored in the building materials and the drying process EVERY building needs about 20 to 25% more heating energy in the first heating period. This is even true for wooden buildings (although you might not expect this).

I agree with all the answers above, since comparing simulated and measured building energy consumption on an hourly basis is a very tough task, and has multiple answers. We do this work on a routine basis as part of validating our designs. While Fred, Karolis, Roulet, Robert, Deleted (?!) et al. have covered most driving factors behind the discrepancies, I would like to bring out two aspects of uncertainty: aleatory and epistemic in modeling and measuring.

Measuring errors, and accuracy levels of instruments needs to be considered. Thus, a 5% error could be due to instrumentation itself.

If measurement is done at AHU level: either on air or water side: in addition to above, what we measure is heat extraction rate and what is modeled is cooling load.

If measuring at chiller level: either on water side (evaporator, condenser), air side (condenser), or electrical side, what we measure is generation. Losses in transmission (chilled water drop in supply+return could easily be > 0.5 deg.C)  could easily be in excess of 10% (if delta T =5 deg. C) and we equate generation without considering losses (it is very difficult to measure these losses at each AHU) with modeled cooling load.

To counter these, we have found some algorithms to be useful in removing these effects. Further, we first calibrate the model for what we have termed as BEL (building envelope load) with measured values. Once these two steps are applied, the driving factors, as desired by you, will show up (with their ranking) upon performing local Sensitivity Analysis (SA). Operating parameters besides weather parameters (generally temperature) show up as having higher sensitivity towards cooling load. SA suggests not to attach importance to magnitude of the Sensitivity Index but to the ranking alone. I generally prefer to obtain several Sensitivity Coefficients (maybe 3) and average their ranking (not their coefficients or indices).

Request please add / comment. Thanks.