ARTICLE IN PRESS

Building and Environment 41 (2006) 1881–1891

www.elsevier.com/locate/buildenv

Quasi-adaptive fuzzy heating control of solar buildings

M.M. Goudaa, S. Danaherb, C.P. Underwoodc,
a
Faculty of Industrial Education, Cairo, Egypt
b
School of Engineering, University of Northumbria, Newcastle upon Tyne, NE1 8ST, UK
c
School of Built Environment & Sustainable Cities Research Institute, University of Northumbria, Newcastle upon Tyne, NE1 8ST, UK
Received 4 September 2003; received in revised form 2 June 2005; accepted 1 July 2005

Abstract

Significant progress has been made on maximising passive solar heat gains to building spaces in winter. Control of the space
heating in these applications is complicated due to the lagging influence of the useful solar heat gain coupled with the wide range of
construction materials and heating system choices. Additionally, and in common with most building control applications, there is a
need to develop control solutions that permit simple and transparent set-up and commissioning procedures. This paper addresses
the development and testing of a quasi-adaptive fuzzy logic control method that addresses these issues. The controller is developed
in two steps. A feed-forward neural network is used to predict the internal air temperature, in which a singular value decomposition
(SVD) algorithm is used to remove the highly correlated data from the inputs of the neural network to reduce the network structure.
The fuzzy controller is then designed to have two inputs: the first input being the error between the set-point temperature and the
internal air temperature and the second the predicted future internal air temperature. The controller was implemented in real-time
using a test cell with controlled ventilation and a modulating electric heating system. Results, compared with validated simulations
of conventionally controlled heating, confirm that the proposed controller achieves superior tracking and reduced overheating when
compared with the conventional method of control.
r 2005 Published by Elsevier Ltd.

Keywords: Solar building; Space heating; Fuzzy logic control; Neural network; Adaptive control; PI control; Real-time control

1. Introduction capacity. Additionally, there is a requirement for these

systems to give responsive and stable performance
In response to demands for buildings with increasing characteristics under control in situations where the
levels of energy efficiency as well as reduced environ- mix of space and system dynamics can vary consider-
mental impact, significant progress has been made on ably. Finally, space heating control systems are tradi-
the development of building designs which maximise tionally difficult to commission because of the widely
passive solar heating in winter whilst reducing envelope varying operating characteristics to which the plant is
heat loss. This is achieved through the choice of glazing required to respond as well as interaction with other
surface area, orientation and material, together with a linked control systems. This situation is rarely rectified
judicious balance between insulation and thermal post-commissioning due to the poor maintenance
capacity when selecting the opaque materials of management that many buildings suffer from.
construction. Thus there is a need for a new generation of space
As a consequence, the space heating systems to be heating system controllers that can cater for the
found in such buildings are now required to operate for following:
increased periods of time at low, or very low, output

Stable response at light load (traditionally an operat-
Corresponding author. ing region which tends to be especially non-linear).

0360-1323/$ - see front matter r 2005 Published by Elsevier Ltd.

doi:10.1016/j.buildenv.2005.07.008
 ARTICLE IN PRESS
1882 M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891

Good set-point tracking with the ability to anticipate parameters of the FLC. Egilegor et al. [11] used an
the lagging influence of passive solar heat gain. ANN to adapt a daily comfort offset parameter as a

Easy and transparent set-up and commissioning with control target for a FLC applied to the control of
minimal post-commissioning intervention. heating and cooling in a dwelling.
Of the adaptive fuzzy controllers developed for the
Thus the aim of the work reported here is to develop a control of energy and environmental comfort in build-
new controller for space heating in passive solar ings, most have a limited adaptation horizon and none
buildings which is responsive to the lagging effects of of them account for the substantially lagging influences
solar energy inputs, whilst offering good robustness and of the major microclimate variables. The same applies to
tracking properties and minimal commissioning. the earlier model-based adaptive controller develop-
ments. This problem forms the basis of the work
reported here.
2. Previous work

Early developments in adaptive control applicable to

energy plant in buildings were due to Jota and Dexter 3. Test site
[1], who developed general minimum-variance and
general predictive controllers for heating and cooling The vehicle for controller development and testing
coils in air handling plant. Problems with robustness of was a 3.5 m  3 m  2.3 m (approx.) Test Cell located at
these early controllers led to work on ‘‘jacketing’’ [2] and Cranfield University in Bedfordshire (Fig. 1). The site
receding-horizon controllers [3]. In spite of these early co-ordinates are latitude 52.071 (north), longitude 0.631
efforts, a generally applicable robust control method for (west) and altitude 100 m above sea level. A single glazed
building energy plant that could be applied without the wall section of the Test Cell (as can be seen in Fig. 1)
need for a sophisticated model of the plant remained faces due south, the window panel measuring
elusive and progress was generally overtaken by new 2 m  1.2 m (high). The walls and roof consist of a thin
approaches using fuzzy and neuro-fuzzy systems. proprietary cladding panel on the outside with 50 mm
The using of fuzzy logic methods to develop styrofoam insulation and an air space forming inner
controllers applicable to the inherently non-linear and layers, finished internally with plasterboard. The Test
multi-modal plants found in buildings has received Cell floor is 900 mm above the ground and constructed
widespread attention [4–6]. Early progress tended to similarly but, in addition, with a 36 mm dense concrete
concentrate on ‘‘static’’ fuzzy logic controllers (FLCs) slab with a vinyl tile finish to the interior. This provides
(i.e. those with a fixed rulebase and inference mechan- the main contribution to the thermal storage capacity of
ism) with one or two inputs (usually the feed-back the space.
controlled variable and its rate-of-change) and thus The Test Cell was equipped with a 2 kW (nominal
these developments lacked adaptiveness to changes in capacity) oil-filled electric convector-radiator mounted
plant operating mode or changes in boundary condi- beneath the window panel. A pulse-width-modulation
tions. power control system was developed for the control
Some attention has been given to mechanisms for of power supply to the convector-radiator as shown in
making fuzzy controllers adapt to their operation Fig. 2. The power supply was adjusted to saturate at
domains. Haissig and Woessner, [7], introduced an 1.1 kW—established as the design heating load for the
adaptation mechanism based on updating the location Test Cell under normal UK design conditions.
of the output membership function of a domestic hot
water boiler FLC. Kolokotsa et al. [8], describe the
development of a model-following FLC based on scaling
fuzzy inputs and outputs to respond to an idealised
second-order model of a building space in order to
achieve control over a variety of internal comfort
conditions. The use of artificial neural networks (ANNs)
as a means of adapting the parameters of a fuzzy
controller has also been considered a promising way of
addressing the adaptiveness shortcomings of static
FLCs. ANNs have found applications in areas ranging
from pattern recognition to feed-back control [9]. So
et al. [10] reported on the development of a self-learning
FLC for a building air handling plant which used an
ANN to monitor plant conditions and update the Fig. 1. Test Cell (viewed from approximately south).
 ARTICLE IN PRESS
M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891 1883

4. Test Cell modelling and control benchmarking

A thirteenth-order lumped parameter room model

was derived for the Test Cell to form a vehicle for
preliminary controller development. The model was
built up by treating each opaque construction element
(oriented-wall, floor, etc.) as a second-order function
with five parameters (three resistances and two capaci-
tances). The room air volume was treated as a single
Fig. 2. Convector-radiator power control system. thermal capacitance whilst the window panel was
assumed to have negligible thermal capacity and was
treated as a steady-state conduction path. It was
assumed that the direct component of solar radiation
Two integrated-circuit-based temperature sensors entering via the window panel was absorbed by the floor
with 4–20 mA outputs and a measurement uncertainty element whilst the diffuse component was assumed to be
of 70.1 K were used for internal and external air absorbed uniformly by all opaque room surfaces.
temperature measurement. The internal sensor was Fig. 3 gives an equivalent analogue circuit representa-
located in the middle of the Test Cell and the external tion of the Test Cell thermal model and a full
sensor was shielded from solar radiation. Since the data description of the room modelling procedure including
acquisition card inputs were voltages, a high-precision the method used to calculate the five parameters for
resistor (500 O) was connected in series with each each construction element can be found in [12]. The
temperature sensor, and a 2200 mf capacitor was overall resistance and capacitance values for the Test
connected in parallel with this resistor to act as a low Cell model are listed in Table 1.
pass filter with a cut-off frequency of 0.57 Hz. The The Test Cell space heating system, consisting of an
ventilation rate into the Test Cell was fixed at a constant electric oil-filled convector-radiator, can be modelled
rate (equivalent to 0.5 air changes per hour) as measured using the following (BS EN 442-3, 1997) [13]:
by a differential pressure manometer across an orifice dT r
plate in the air supply duct. ¼ ½u  p  kðT r  T i Þn =C om (1)
dt
Two Kip and Zonen type LI-200SA pyranometers
were used to measure the vertical solar radiation in which T r is the radiator surface temperature, T i is the
incident on the window (these can be viewed to the Test Cell air temperature, u is the power controller
right of the Test Cell in Fig. 1). The pyranometers were signal u 2 0; 1, p is the maximum power emitted from
based on a silicon photovoltaic detector mounted in a the radiator (kW), C om is the combined thermal capacity
fully cosine-corrected miniature head. The first of the of the oil and the radiator casing material, k and n are
pyranometers measured the total vertical component of constants. n may be taken to be 1.3 for a convector-
solar irradiance whilst the second, fitted with a shading radiator (BS EN 442-3, 1997) and because p and T i are
ring, was used to measure the diffuse component. The known at design boundary conditions, the constant k
current outputs of both were calibrated under natural can be calculated from Eq. (1) at steady-state conditions
daylight conditions in unit of Watts per square meter; (i.e. dT r =dt ¼ 0).
each pyranometer was found to give 10 mV m2 W1 and As a means of benchmarking the adaptive controller, a
a high-precision amplifier was used to amplify this signal conventional controller was first defined and evaluated.
to 10 mV m2 W1. Power to the convector-radiator was
measured using a current transformer in the AC power
supply.
In terms of control signalling, a dSPACE DS1102
data acquisition card was used for analogue (control)
inputs and the single analogue output required to
‘‘position’’ the power supply to the convector-radiator.
The DS1102 is a single board system, which is
specifically designed for development of high-
speed multi-variable digital controllers and real-time
applications. It is based on a third-generation floating-
point digital signal processor (DSP) providing fast
instruction cycle time. The DSP contained 128 K words
of memory (fast enough to allow for a zero wait state
operation). Fig. 3. Model realisation of the Test Cell.
 ARTICLE IN PRESS
1884 M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891

Table 1 Fig. 4 using a sample of local weather data. Though the

Overall thermal resistances and capacitances for the Test Cell
tracking performance of the control system is good
Resistance Resistance Capacitance Capacitance during periods of low solar irradiance, evidence of
Ref. (Fig. 3) value (K W1) Ref. (Fig. 3) value (J K1) afternoon overheating is clear during periods of
significant solar irradiance. The control system is unable
R1 1.035  102 C1 8.939  103
to adapt to the useful passive heating effect resulting
R2 5.171  102 C2 5.066  104
R3 4.140  102 C3 1.286  105 from the solar input with the consequence that both
R4 1.178  101 C4 2.270  104 comfort and energy use are both compromised. A
R5 9.424  102 C5 1.083  105 further difficulty lies in the establishment of controller
R6 2.356  102 C6 1.915  104 parameters—the fixed settings used above are difficult to
R7 1.396  101 C7 7.272  105
R8 1.11  101 C8 1.283  105
obtain during practical space heating commissioning
R9 2.793  102 C9 1.696  105 and are, in any case, unlikely to be satisfactory across
R10 8.909  102 C10 2.993  104 the essentially broad operating conditions experienced
R11 7.128  102 C11 1.085  105 by the plant during a typical heating season. Thus an
R12 1.781  102 C12 1.915  104 adaptive controller with simple commissioning and
R13 8.928  102 C13 2.969  104
good robustness is needed in this application. The first
R14 7.142  102
R15 1.786  102 of these objectives is realised through the development
R16 1.396  101 of a forecasting model. The remaining two objectives are
R17 1.117  101 realised through the application of a fuzzy logic
R18 2.793  102 controller design.
R19 1.083

5. A forecasting model
A conventional proportional-plus-integral (PI) controller
was used for this purpose and this was tuned using The stochastic problem of supplying a forward-in-
Ziegler–Nichols rules. This method of control with this time prediction of internal air temperature (the target
method of tuning was adopted as a benchmark because it control variable) is considered in this section. The
is very commonly used in feed-back control of space theoretical model developed in the previous section as a
heating systems in practice [14]. The controller para- means of benchmarking the system under investigation
meters arrived at were a proportional gain (kp) of was found to give good agreement with field monitoring
0.8 K1 and an integral time ðti Þ of 10 min according to results [12]. However, it requires unique and extensive
the Ziegler–Nichols’ prescriptions [15]. It is stressed that parameterisation each time it is used in practice and thus
the intention here was to compare the proposed is not suitable for a practical online implementation of
controller with a conventional PI controller (as a the proposed control system. This type of predictive
practical ‘‘benchmark’’) in order to evaluate both modelling problem has historically been dealt with using
adaptiveness and ease of set-up in the case of the auto-regression techniques but the tendency of ANNs to
proposed controller. It would have been perfectly give superior results with noisy and incomplete data,
feasible to introduce a solar input for the PI controller non-linear problems and better computational perfor-
as well but such a development of the PI controller mance during model implementation has resulted in an
would have made it unsuitable as a practical benchmark. ANN approach being adopted here. In the following, a
Thus our central objective was to develop a fuzzy logic summary of the approach that has been adopted is
heating system controller with responsiveness to passive given. A fuller treatment can be found in [17].
solar heating effects and no need for tuning parameters The problem specification was to provide an estimate
and to benchmark it against a purely conventional of internal air temperature at least 1-h ahead in time and
method of control. This is a development of an earlier to within typical measurement uncertainty. The avail-
work which investigated the performance of a simple able historical time series’ to inform the estimate were
fuzzy logic heating controller with a conventional PI the external air temperature, total solar irradiance,
controller and concluded that, though the fuzzy logic heating system control signal and past values of internal
controller offered better robustness and control variable air temperature. The significance of the minimum 1-h
tracking than the PI controller, a mechanisim for time horizon relates to preliminary modelling of a range
adapting to external disturbances (such as solar heating) of room spaces suggesting that the typical time lag
was desirable (Gouda et al. [16]). The development of between the solar input received by a window and a
this mechanism forms the basis of the present work. significant increase in room air temperature is about 1 h
Preliminary results based on the completed model which fixes the problem control horizon. A further
consisting of the Test Cell space, heating system and justification for this emerges from a simple reasoning
PI controller with closed feed-back path are shown in applied to the model parameters of the Test Cell in that
 ARTICLE IN PRESS
M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891 1885

Fig. 4. Sample simulated results using conventional control. (a) Total solar radiation incident on window; (b) internal air temperature.

the time constant of the internal air temperature with training data set for the heating system control signal,
respect to solar energy received at the floor must be the convector-radiator power regulator was cycled
within the model prediction horizon. Consideration of through a predetermined sequence. All data were
Fig. 3 shows that the time constant relating the internal sampled at 15-min intervals. Because of the need to
air temperature to a step input of solar energy to the supply historic data not only the current values
floor will be: C 8 R11 R12 =ðR11 þ R12 Þ. Inserting the (inputs ðkÞ) were needed but also inputs ðk  1Þ,
appropriate values from Table 1 leads to a time constant inputs ðk  2Þy inputs ðk  iÞ were needed. The value
for this pathway of 1828s. This is comfortably within the of i was increased by trial and error and it was found
1-h time horizon specified for the purpose of this work. that a minimum value of i ¼ 48 was needed to
Typical room air temperature sensors have an uncer- successfully capture the building dynamics. This effec-
tainty of 70.2 K and thus the quality of prediction over tively provided 12 h of historic data to the network, but
actual data will also need to be within this tolerance to led to the unwelcome effect of increasing the number of
be acceptable. effective inputs to 192.
In recognition of the potential problem of overfitting To reduce the training data set to a more computa-
by the ANN, the early stopping method was used in tionally manageable size, the application of singular
which the available training data were divided into three value decomposition (SVD) was investigated. SVD is an
subsets. The first subset was used for training, the efficient numerical technique for the analysis of multi-
second for validation and the third subset was used for variate data. It can be used in most types of multi-
testing. For validation, the error between the evolving variate analysis and can greatly increase computational
network and the validation data subset was monitored efficiency in, amongst other things, non-linear analyses
during the training process. Generally, the resulting such as neural network building. The technique also has
validation error decreased during an initial phase the advantage of removing unimportant information
of training but when the network began to overfit which might, for instance, be related to noise. A brief
the data as evidenced by a rise in the validation subset overview of the method follows (for further details see
error, the training was stopped. The weights and for example Danaher and O’Mongain [18]).
biases corresponding to this minimum validation subset Letting the matrix O (of size m  n which contains
error were returned as the required network solution. both the current values and the past values) be an
The error generated by the third data subset was not observation matrix, the SVD method decomposes this O
used during the training, but was used to test the matrix into three matrices: W, L, and V. Thus the
performance of each network solution. Thus the test singular value decomposition of O is defined as:
subset error was used as a basis to compare different
O¼WLV (2)
network structures.
Meteorological data were used for the external air in which W and V are pseudo-unitary (i.e. are unitary
temperature and solar irradiance. To establish the but contain ‘‘lost points’’ and, as such, are described as
 ARTICLE IN PRESS
1886 M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891

Fig. 5. Forecasting model structure.

pseudo-unitary), containing the first n column eigenvec-

tors of OT O and row eigenvectors OOT , respectively.
The L matrix contains the corresponding eigenvalues, in
decreasing order of significance, on the leading diagonal.
The training inputs were presented as an observation
matrix O of size m  n, containing both the current
values and the past history over 12 h. Thus for four
training variables at a 15-min sampling interval n ¼ 192.
The SVD was calculated using weekly blocks and hence
m ¼ 4  24  7 ¼ 672. The SVD technique was used to
decrease these inputs to 10, this value being found to
retain all the important information in the data.
Various network architectures were investigated, and
a number of different network sizes and learning
parameters tested. The architecture that was ultimately
selected is shown in Fig. 5; a conventional feed-forward Fig. 6. Network training.
multi-layer perceptron consisting of three layers; an
input layer with 10 neurons, seven neurons in a hidden
layer, and a one neuron output layer. consequent solar irradiance profile and any long-term
The ANN was trained by back-propagation using the seasonal weather effects.
Levenberg–Marquardt algorithm. The activation func- Results using this network to predict the internal air
tion used with the input and output neurons was linear, temperature of a low thermal capacity building for 1-h
whereas in the hidden neurons a sigmoid form was used. ahead and 2-h ahead, during a four month heating
The number of hidden neurons was estimated by season period (January–April) are shown in Figs. 7(a)
running the network with different number of hidden and (b). The results were in excellent agreement during
neurons and calculating the mean square error using the totally clear and overcast days. On days of intermediate
training, validation and test data subsets. Based on this, solar irradiance, the results were less good due to the
the number of hidden neurons arrived at was seven as stochastic nature of cloud cover and the consequent
shown in Fig. 6. effect on solar irradiation. Results for 2-h-ahead
The network was designed for training every week- prediction are reasonable but an increase in prediction
end, using a data set consisting of the two previous error is evident and, generally, increases in prediction
weeks of observation data. The training set was the horizon beyond 1-h tended to be less reliable. Most of
second week of data and the validation set was the data the error is located in the range 70.2 K—within what
from the first week and these two weeks of data were might be regarded as a typical measurement uncertainty
changed every weekend. This methodology was chosen for practical commercial temperature sensors used in
in order to capture the slowly varying solar position and this application.
 ARTICLE IN PRESS
M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891 1887

Fig. 7. ANN-predicted internal air temperature and modelling error frequencies. (a) 1-h-ahead prediction; (b) 2-h-ahead prediction; (c) 1-h-ahead
error frequency; (d) 2-h-ahead error frequency.

Fig. 8. Quasi-adaptive fuzzy logic controller.

6. A quasi-adaptive fuzzy logic heating controller the first of two inputs to the fuzzy controller. The second
input was the predicted 1-h-ahead internal air tempera-
The development of a quasi-adaptive fuzzy logic ture enabling the controller to adapt to the lagging
controller is considered in this section. The controller is influence of solar heat gain. In this sense, the terminol-
divided into two main modules; a conventional static ogy quasi-adaptive is used for the controller to distin-
fuzzy controller and feed-forward neural network with guish it from the more familiar adapting-rulebase/
SVD algorithm as shown in Fig. 8. The inputs needed adapting-inference paradigms such as is found in
for this controller are the set-point temperature ex- classical neuro-fuzzy controllers.
pressed by the user, and the current and past values of The fuzzy sets of the error were defined as NB
external air temperature, solar radiation and internal air (negative big), NM (negative medium), NS (negative
temperature. The error between the internal air tem- small), ZE (zero), PS (positive small), PM (positive
perature and the set-point temperature was applied as medium), and PB (positive big), structured after a little
 ARTICLE IN PRESS
1888 M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891

trial-and-error as shown in Fig. 9a. The fuzzy sets of the Table 2

Applied fuzzy rules
predicted temperature were defined as LW (low), ML
(medium low), SL (slightly low), MD (medium), SH Error Predicted internal air temperature
(slightly high), MH (medium high), and HG (high)
(Fig. 9b). Overlapping triangular membership functions LW ML SL MD SH MH HG
were used for input (fuzzification) and output (defuzzi- NB NB NB NB NB NB NB NB
fication) of the fuzzy system. Seven fuzzy sets for the NM NB NB NB NB NB NB NB
output variable were defined as NB (negative big), NM NS ZE NM NM NM NB NB NB
(negative medium), NS (negative small), ZE (zero), PS ZE PS ZE NM NM NM NB NB
PS PM PS ZE NM NM NM NM
(positive small), PM (positive medium), and PB (positive
PM PM PM PS ZE NS NS NS
big) (Fig. 9c). Thus according to the number of the fuzzy PB PB PB PB PM PS PS PS
sets of the inputs and the output, there are 49 fuzzy
rules. Table 2 gives details of the fuzzy rules that were
applied.
Mamdani’s minimum operator method was used for conventional PI controller as described in Section 4.
inference and this was defuzzified to a crisp control This was applied in order to benchmark the perfor-
signal using the centre-of-gravity method. The control mance of the heating system to what would be expected
signal was then applied directly to the power regulator in conventional practice. The second was the quasi-
of the oil-filled convector-radiator. adaptive fuzzy logic controller (QAFLC) as described
above.
The PI controller was implemented as described in
Section 4 using the settings determined from the
7. Controller testing and evaluation preliminary step response test. Saturating the integral
term at the lower and upper limits of the applied control
Two controllers were applied independently to the signal effected a simple anti-wind-up mechanism. The
Test Cell heating system. The first of these was a QAFLC was implemented in real-time as shown in
Fig. 10. Two weeks of external air temperature, solar
radiation, internal air temperature and control signal
generated by the benchmark controller were collected
and stored in a matrix to act as a stack. Using the
principle of a stack, the data matrix had two weeks of
data at each sampling time. The network was trained
and validated using these data after applying the SVD
algorithm to generate a 1-h-ahead prediction of the
internal air temperature. The practical implementation
of both controllers was carried out using a dSPACE
type DS1102 floating point controller board consisting
of a four-channel analogue-to-digital converter (ADC)
for inputs (only three of these were used) and a four-
channel digital-to-analogue converter (DAC) for out-
puts (see Fig. 10), supported with an appropriate real-
time software application.
For the performance of both control systems to be
compared, it would have been necessary to have two
Test Cell configurations operating simultaneously and
identical in every respect apart from the choice of
controller. This was not possible for practical and
economic reasons and, therefore, the following proce-
dure was adopted. An initial test run was carried out
using the real-time PI controller and these results were
compared with results generated by the simulation
model (with the same controller specification) described
in Section 4. With minimal adjustment of model
Fig. 9. Fuzzy membership functions. (a) Temperature control error parameters an excellent agreement between the simu-
input; (b) predicted 1-h-ahead internal air temperature input; (c) lated and experimental results was obtained as shown in
control signal output. Fig. 11. Indeed this was expected based on earlier work
 ARTICLE IN PRESS
M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891 1889

using this general type of model for simulations towards (Fig. 12(a)). The QAFLC is seen to give superior
relatively short time horizons [19]. Thus the QAFLC tracking of the set-point reference temperature. In
was then implemented in real-time and results were particular, the afternoon overheating arising from the
compared with simulations based on the validated solar radiation peaking at noon is evident under
model of PI-controlled heating system using identical conventional PI control but the QAFLC successfully
weather data. reduces this undesirable effect significantly thus improv-
Fig. 12 shows a sample of results from the real-time ing comfort and reducing heating energy costs.
implementation of the QAFLC compared with simu-
lated results of the conventional PI-controlled heating.
The results are based on a cold winter day in February
2002 with high solar irradiance, peaking at around noon 8. Conclusions

The development of a quasi-adaptive fuzzy logic

controller (QAFLC) for space heating control in solar
buildings is described the main aim of which is to reduce
the lagging overheating effect caused by passive solar
heat gain to a room space.
The controller’s adaptive mechanism was effected by
training an ANN to predict the controlled variable up to
1-h ahead and this prediction thus formed one of two
inputs to an otherwise static fuzzy logic controller. To
reduce the computational demands consistent with a
real-time implementation of the controller, singular
value decomposition was used to reduce the training
data set from 192 values to 10.
The QAFLC was implemented in real-time using a
25 m3 (approx.) Test Cell on an exposed site equipped
Fig. 10. Real-time implementation of the QAFLC. with an electric heating system. A model of the Test Cell

Fig. 11. Results: comparison of simulated and real-time PI control. (a) Total solar radiation incident on window; (b) internal air temperature; (c)
convector-radiator output.
 ARTICLE IN PRESS
1890 M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891

Fig. 12. Results: comparison of real-time QAFLC and simulated PI control. (a) Total solar radiation incident on window; (b) internal air
temperature; (c) convector-radiator output.

was first developed to enable simulations of a conven- Acknowledgements

tional proportional-plus-integral (PI)-controlled heating
system to be generated for the purpose of making The financial assistance of the Egyptian government is
comparisons with the real-time implementation of the acknowledged, as is the use of the Energy Monitoring
QAFLC. The simulation model was validated using Company Limited’s Test Cell 2000 at Cranfield Uni-
measured data from the Test Cell and excellent versity.
agreement was obtained.
Results of the real-time implementation of the
QAFLC were compared with simulations of the References
conventionally controlled heating system. The QAFLC
[1] Jota FF, Dexter AL. Self-tuning control of an air handling plant.
was found to give superior tracking of the controlled In: Proceedings of the IEE International Conference Control ’85.
variable set-point and was found to reduce the after- London: Institute of Electrical Engineers; 1985.
noon overheating of the space significantly when [2] Dexter AL, Haves PA. Robust self-tuning predictive controller
compared with conventional control, thus improving for HVAC applications. ASHRAE Transactions 1989;95(2):
431–8.
comfort conditions and reducing heating energy costs. A
[3] Macarthur JW, Woessner MA. Receding horizon control: a
further advantage is that, other than for the need for model-based policy for HVAC applications. ASHRAE Transac-
local weather data measurement, the QAFLC can be tions 1993;99(1):139–48.
implemented with minimal set-up and commissioning, [4] Dexter AL, Trewhella DW. Building control systems: fuzzy rule-
whereas a conventional controller requires the complex based approach to performance assessment. Building Services
process of parameter tuning to be undertaken. Engineering Research and Technology 1990;11(4):115–24.
[5] Huang S, Nelson RM. A PID-law-combining fuzzy controller for
Further work is warranted in three areas. Firstly, the HVAC applications. ASHRAE Transactions 1991;97(2):768–74.
rationalisation of the fuzzy logic controller rulebase [6] So ATP, Chow TT, Chan WL, Tse WL. Fuzzy air handling
which would lead to further improvements in computa- system controller. Building Services Engineering Research and
tional demands for online implementation. Secondly, Technology 1994;15(2):95–105.
the extension of the adaptation mechanism to enable [7] Haissig CM, Woessner M. An adaptive fuzzy algorithm for
domestic hot water temperature control of a combi-boiler.
other sources of heat gain such as lighting and internal International Journal of HVAC & R Research 2000;6(2):117–34.
equipment to be accounted for. Finally, there is a need [8] Kolokotsa D, Tsiaros D, Stavrakakis GS, Kalaitzakis K,
to conduct extensive field trials of the controller. Antonidakis E. Advanced fuzzy logic controllers design and
 ARTICLE IN PRESS
M.M. Gouda et al. / Building and Environment 41 (2006) 1881–1891 1891

evaluation for buildings’ occupants thermal-visual comfort and [14] Underwood CP. HVAC control systems. London: E & FN Spon;
indoor air quality satisfaction. Energy in Buildings 2001;33(6): 1998.
531–43. [15] Ziegler JG, Nichols NB. Optimum settings for automatic
[9] Kalogirou SA. Application of artificial neural networks in energy controllers. Transactions of the ASME 1942;64:759–68.
systems: a review. Energy Conversion and Management [16] Gouda MM, Danaher S, Underwood CP. Thermal comfort based
1999;40(10):1073–87. fuzzy logic controller. Building Services Engineering Research
[10] So ATP, Chan WL, Tse WL. Self-tuning fuzzy air handling and Technology 2001;22(4):237–54.
system controller. Building Services Engineering Research and [17] Gouda MM, Danaher S, Underwood CP. Development of an
Technology 1997;18(2):99–108. artificial neural network predictor for indoor temperature of a
[11] Egilegor B, Uribe JP, Arregi G, Pradilla E, Susperregi L. A fuzzy solar building. In: Proceedings of the IASTED conference on
control adapted by a neural network to maintain a dwelling intelligent systems and control, Tampa, FL, 2001.
within thermal comfort. In: Proceedings of the Building Simula- [18] Danaher S, O’Mongain E. Singular value decomposition in
tion ’97, Prague, vol. II. 1997. p. 87–94. multispectral radiometry. International Journal of Remote Sen-
[12] Gouda MM, Danaher S, Underwood CP. Building thermal model sing 1992;13:1771–7.
reduction using nonlinear constrained optimisation. Building and [19] Gouda MM, Danaher S, Underwood CP. Low-order model for
Environment 2002;37:1255–65. the simulation of a building and its heating system. Building
[13] BS EN 442-3. Specification for radiators and convectors— Services Engineering Research and Technology 2000;21(3):
evaluation of conformity. London: British Standards Institute; 199–208.
1997.