384
HVAC&R RESEARCH
research requiring weather modeling has focused on the development of predictive algorithms
for heating and cooling loads directly rather than the ambient weather conditions (Ferrano and
Wong 1990; Forrester and Wepfer 1984; MacArthur et al. 1989; Rupanagunta et al. 1995; Seem
and Braun 1991). In particular, Kawashima et al. (1995) compared autoregressive integrated
moving average (ARIMA), exponential weighted moving average (EWMA), linear regressive
(LR), and artificial neural network (ANN) models in prediction of thermal loads over 24 h; the
results indicated that the ANN model produced the lowest prediction errors. Weather models for
use in building design load analysis have also been investigated. Jiang and Hong (1993) used a
stochastic weather model to obtain diurnal changes in the climatic variables and shape factors to
estimate the hourly values. Yoshida and Terai (1992) used an autoregressive moving average
(ARMA) model, which took account of the deterministic annual and diurnal periodicity and the
stochastic variations in the climatic variables, including the ambient air temperature, solar radia-
tion, and absolute humidity. Among the methods investigated for predicting weather conditions
for use in predictive control are stochastic dynamic models (Holst et al. 1987), sinusoidal func-
tions (Athienitis 1988), and shape factors and look-up tables (Chen and Athienitis 1996).
Although much research has focused on weather prediction, or on the result of climatic condi-
tions on building thermal loads, no research has been conducted addressing the prediction of
ambient conditions for use in set-point scheduling control. The requirement and assumption
made of such a prediction algorithm are that the climatic conditions relating to the buildings are
measured hourly and are available to update the model prediction. The model is required only to
predict the ambient conditions for the next 24 h. The model need not address the annual period-
icity or seasonal effect separately, because the forecasting is to be conducted daily, wherein the
parameters of the model are updated to include the effect of the measured data from the previous
24 h. Three adaptive models for the diurnal prediction of ambient temperature and solar radia-
tion are examined here, each candidate model having been selected through an understanding of
the properties of ambient temperature and solar radiation time series.
In this paper, two sets of weather data have been selected to analyze the performance of the
prediction models. The U.K. Chartered Institution of Building Services Engineers (CIBSE)
example weather year, measured in London, U.K., from October 1964 to September 1965, and
the data measured in Garston, U.K., during 1994 were used. The CIBSE example year is consid-
ered suitable for use in predicting average building energy consumption, whereas the data from
1994 are useful for predicting the potential overheating risk in low-energy buildings. Clearly,
these data are for a temperate climate, but one that can be subject to significant stochastic
changes in climate, and therefore is potentially challenging to model.
A weather model is complicated by the fact that the climatic variables are correlated to each
other. Yoshida and Terai (1990-1991) suggested that the model could be simplified if it is
assumed that temperature is only correlated to solar radiation, but that solar radiation is not cor-
related to any other variables.
Figure 1 shows the profiles for the hourly mean and standard deviation of global radiation for
the CIBSE example year and 1994 (diffuse and direct radiation have also been considered, but
are not presented here). It is observed that the properties of the radiation data are time dependent
and thus are a nonstationary stochastic time series. The global radiation has a strong periodicity;
the radiation at night is deterministically zero, reaching the peak in the middle of the day. The
standard deviations indicate a high variation of the hourly global radiation in the middle of day,
2
2
the highest variation being 218 W/m , 68.5% of the highest mean radiation 318 W/m . This
implies a significant change in the hourly radiation throughout an entire year.
Figure 2 shows the profiles for the hourly mean and standard deviation of the ambient temper-
ature for the CIBSE example year and 1994. It is observed that the temperature series are also
time dependent and thus are a nonstationary stochastic time series.