Natural Ventilation and Indoor Air Quality Practice for Sustainable Buildings a Design Appr

Natural Ventilation and Indoor Air Quality Practice for Sustainable Buildings: a Design Approach Mario De Grassi 1 and Berardo Naticchia 2

1Dept. of Architecture and Town Planning, University of Ancona

Via delle Brecce Bianche, 60131 Ancona, Italy

Phone +39 071 2204584 Fax +39 071 2204582

E-mail: Degrassi@Idau.Unian.It

2IDAU, Facoltà di Ingegneria,

Via delle Brecce Bianche, 60131 Ancona, Italy

Phone +39 071 2204584 Fax +39 071 2204582

E-mail: Naticchia@Idau.Unian.It

ABSTRACT: An essential role of architecture is to provide built environments that sustain occupants’ safety, health, physiological comfort, and productivity. Because of complexity of modelling for environmental quality design, its importance has often been overlooked in the quest for energy and environmental conservation. About the air quality issue, the benefits of suitable natural ventilation go beyond the need for oxygen and the energy saving in cooling phases. Continuous recirculation of interior air exposes people to concentrated levels of bacteria and chemicals within the building and forced ventilation may often produce incompatible noise level with activity and well-being of occupants.

Green building practices offer an opportunity to create environmentally sound and resource efficient b uildings by using an integrated approach to design. Integration means to face architectural and environmental aspects of design contemporarily, without disjoining them in a hierarchical sequence of resolving. This requires designer ability of understanding and handling complex physical knowledge in order to correctly address preliminary design decisions rather than commissioning the performance assessment when the design features are truly defined. This paper describes VENTPad, a design environment we are developing to help students and architects in understanding and therefore planning suitable internal fluid-dynamic behaviour of buildings (especially focused on wide internal spaces, like concert halls and theatres, which behaviour is more manifold, thus pedagogically more interesting, than residential buildings).

A novel contribution of the system approach, compared with conventional numerical simulation techniques, is the ability to generate explanations about the building physical behaviour, diagnosing contributes of each relevant design decision to the system performance.

KEYWORDS: Preliminary Sustainable Design, Qualitative Modelling, Education

1. INTRODUCTION

Buildings have diverse effects on the environment during their entire life cycles. Although the tangible impacts are visible only after construction begins, decisions made on the drawing board have long-term environmental consequences. To achieve environmental sustainability in the building sector, architects must be educated about environmental issues during their professional training. Faculties have to foster environmental awareness, introduce students to environmental ethics, and developing their skills and knowledge base in sustainable design. In spite of the urgent need, teaching materials specifically designed for sustainable architecture have been virtually non-existent.

While there is a universal consensus on the importance of environmental education in architecture, the questions of what, when, and how to teach specific subjects related to environmental sustainability cannot be easily answered. One reason for this is that architecture covers a vast number of disciplines ranging from art to science; determining the level and extent of environmental education within design, technology, history, theory, practice, and environmental behaviour is a formidable task.

At present, in the absence of a clear pedagogical framework, sustainable design is being presented as an ethical issue rather than science. While a change of lifestyles and attitudes toward the local and global environments is important, the development of scientific approach that addresses the implementation of environmental design goals is urgent.

The unique way to this aim, is that of giving students the scientific skills and modelling bases to seek and find sustainable design solutions rather than giving them a set of typological solutions.

About this issue, while many energy conservation materials and quantitative analysis methods have been developed since the 1970s’, resources for addressing larger environmental design techniques are greatly lacking. To date, advances made in methods to predict and measure building airflows have truly revolutionised the fields of building ventilation and air quality research in the past two decades, so that many simulation techniques for natural ventilation and

indoor

air quality prediction are well established. Anyway a variety of problems arise in synthesising and applying design principles in the context of ‘green buildings’ design in view of the total environmental performance.

This difficulty stems largely from the complexity of the fluid-dynamic behaviour, because of its evident non-linearity and instability. This often forces to place one’s trust in computational or experimental numerical results, without a chance of explaining the causal framework of calculated performance or forecasting the stability of observed physical behaviour.

Consequently, we presently find ourselves armed with a veritable arsenal of tools to evaluate the thermal comfort, air quality and energy conservation efficacy of existing and proposed building ventilation systems. Yet, ironically, we have yet to develop tools to directly answer simple design questions relating to building ventilation: How wide should windows be opened in a given building for wind-driven cross ventilation on a moderate summer day? How should I configure the roof to mitigate air recirculation which obstacles the dilution of internally generated pollutants and exposes people to concentrated levels of bacteria and chemicals within the building?

Incidentally, the role that spaces shape and openings geometry play in setting a local as well as global sound fluid-dynamic behaviour is quite central, so that architectural and environmental design decisions can be never disjoined to study them in a strictly hierarchical sequence.

Then the challenging opportunity facing researchers today is that of synthesising design techniques to handle and explain the complexity concealed in numerical simulation results and experimental data. A method is needed, to examine building environmental response from an engineering point of view rather than a physicist one that is allowing the arrangement of design features by assuming sustainability criteria as the basis for comprehensive evaluation of the environmental building performance.

Such a tool could support the development of students’ and architects’ abilities to explore, assess, and pursue various alternatives for sustainable design with special attention to the close integration between architectural and scientific-technical issues.

The aim we propose requires a qualitative understanding and representation of how the studied system behaves, by deriving from raw and unstructured numerical data and from a ground knowledge of general physical laws, a framework of explicit relationships among a selected set of relevant design variables.

2. THE QUALITATIVE PHYSICAL MODELLING APPROACH

About this aim, computer applications in design have pursued two main development directions: analytical modelling and information technology. The former line has produced a large number of tools for reality simulation (i.e. finite element models), the latter is producing an equally large amount of advances in conceptual design support (i.e. artificial intelligence tools). Nevertheless we can trace rare interactions between computation models related to those different approaches. This lack of integration is the main reason of the difficulty of analytical methods application to the preliminary stage of design, where logical and quantitative reasoning are closely related in a process that we often call ‘qualitative evaluation’.

In this paper, after a brief survey about the current state of qualitative physical modelling applied to design, we propose a general approach of building natural ventilation modelling by means of Bayesian networks.

We are employing this technique to develop VENTPad, a tutoring and coaching system to support natural ventilation modelling of buildings in the preliminary stage of design.

This tool explores the possibility of modelling the causal mechanism that operate in real systems in order to allow a number of integrated logical and quantitative inference about the fluid-dynamic behaviour of buildings.

It represents an innovative connection tool between logical and analytical modelling in preliminary design aiding, able to help students or unskilled architects, both to guide them through the analysis process of numerical data (i.e. obtained with sophisticate Computational Fluid Dynamics software) or experimental data (i.e. obtained with laboratory test models) and to suggest improvements to the design.

VENTPad relies on a probabilistic causal representation, to qualitatively express the knowledge of fluid-dynamics needed to explain the ventilation behaviour of a space.

We view VENTPad as part of a virtual laboratory, a conceptual CAD environment consisting of facilities for assembling, analysing and testing design ideas. By working in this software environment, students can ‘build’ their designs and try out improving them without expense or danger. In simpler domains some commercial software exists that can be viewed as virtual laboratories (e.g. Electronics Workbench) but a novel contribution of VENTPad, compared with these tools, is the ability to generate explanations about the system response, diagnosing contribute of each relevant design parameter or boundary condition to the system behaviour. For educational applications, explanation generation is vital, to help students see what aspects of a situation are important and to tie what they are observing back to fundamental principles.

This aim requires a qualitative understanding and representation of how the studied system behaves, by deriving from raw and unstructured numerical data and from ground knowledge of general physical laws, a framework of explicit relationships among a selected set of variables, which describes the specific system behaviour.

This idea constitutes the original motivations of the research area called ‘qualitative physics’ whose main aim is the development of intelligent tutoring systems and learning environments for physical domains and complex systems. This paper demonstrates how a synergistic combination of qualitative physics and other AI techniques can be used to create

an intelligent learning environment for students learning to analyse and design natural ventilation in buildings. Pedagogically this problem is important because natural ventilation involves the integration of complex physical, thermodynamic and fluid-dynamic knowledge, an area normally closed to architects.

The methodological approach of qualitative physics is based on capturing the tacit knowledge engineers use to organise and control knowledge gained through formal training. The initial motivation for qualitative physics was to set up and guide the solution of textbook motion problems (de Kleer 1975). Since then, research has mainly focused on purely qualitative reasoning (Bobrow 1984), and significant progress has been made. I believe the time is right to begin exploring the integration of qualitative and quantitative reasoning again. In particular, the long-range goal of my research is to develop a system which can automatically perform engineering analyses of design problems in a human-like way. This paper describes a first step towards that goal.

Studies of natural ventilation problem solving have tended to focus on quantitative reasoning. We begin instead with the view that qualitative models are the starting point for the accumulation and use of more sophisticated, quantitative models. This view is widely held in the mental models literature (Gentner and Stevens 1983), and widely but less formally in the engineering community.

In problem solving, the analysis begins by constructing a qualitative understanding of the situation. This initial understanding provides the framework for further analyses, such as deriving and solving sets of equations. Developing a correct qualitative understanding of the problem is essential to solving complex problems. Qualitative simulation is used to verify that questions make sense by ensuring that the behaviour mentioned could actually occur.

We have tested these ideas through implementation in a program called VENTPad, which solves simple natural ventilation design problems typically addressed at the preliminary design stage.

Section 3 of this document, describes the pedagogical problems that motivated the design of VENTPad, including a brief overview of nature of fluid-dynamic issues in design. Section 4 outlines the causal modelling approach of building behaviour that we use to integrate predictive with diagnostic support in guiding the preliminary design and its successive improvement or correction. How VENTPad represents the causal framework, which operates in the airflow behaviour of a simple application, is the subject of Section 5, with Section 6 outlining our plans for future work.

3. COMPLEXITY OF NATURAL VENTILATION PROBLEM IN DESIGN

A variety of problems arise when teaching students how to design and analyse the natural ventilation design principles.

The main problem stems largely from the complexity of the fluid-dynamic behaviour, because of its evident non-linearity and instability which often forces to place one’s trust in computational techniques without critics or possibility of explaining numerical or experimental results.

Advances made in methods to predict and measure building airflows have truly revolutionised the fields of building ventilation and air quality research and practice in the past two decades. Tracer gas techniques have been extended and refined to allow more accurate, better-characterised, and more complete multizone measurements of airflows within buildings. Varieties of rigorously defined ventilation effectiveness metrics have grown out of these advances and have placed ventilation system evaluation on a solid objective basis.

Macroscopic methods of airflow analysis have been generalised to allow integrated modelling of wind-driven, buoyancy-driven, and mechanically-forced airflow in multizone building systems of arbitrary complexity. The global predictive capability of macroscopic simulation methods have been complimented by a constellation of microscopic methods of analysis, together placed under the more familiar rubric of Computational Fluid Dynamics (CFD), that allow investigation of the details of airflow around buildings and within single and, at this point, simply and well-connected collections of rooms.

Consequently, we presently find ourselves armed with a veritable arsenal of tools to evaluate the thermal comfort, air quality and energy conservation efficacy of existing and proposed building ventilation systems. Yet, ironically, we have yet to develop tools to directly answer simple design questions relating to building ventilation: How wide should windows be opened in a given building for wind-driven cross ventilation on a moderate summer day? How should a ventilating monitor and building windows be configured to mitigate internal and solar gains on the same summer day? What size fan is needed to assist stack-driven airflow through the monitor on a more extreme summer day?

A typical numerical output of computational fluid-dynamic software is showed in figure 1, where the natural ventilation air velocity field computed for a concert hall design is graphically rendered. As we can readily realise, the sophisticated data that appear to give all needed information of interest, become intractable when for example we ask ourselves what could we do in order to reduce the re-circulation of air due to the vortex that obstruct the extraction of exhaust air above the stage.

There’s no tool at moment to support this design problem.

Figure 1: Air movement patterns within a concert hall, obtained with numerical simulation.

Because of this lack only a trial and error approach is available to building, improving or correcting design from a natural ventilation point of view. This fact leads students or unskilled engineers to avoid exploring multiple design alternatives and to avoid carrying out trade-off studies; moreover they tend to get bogged down in carrying out routine calculations often spending time merely in solving data input problems.

VENTPad was designed specifically to help students learn natural ventilation by providing an intelligent learning environment that handles qualitative routine calculations, facilitates sensitivity analyses, helps students keep track of modelling assumptions, and detects physically effectiveness of designs.

3.1 FLUID-DYNAMIC MODELS FOR NATURAL VENTILATION IN DESIGN

A building system may be considered to be continuum within which the state variables of temperature T, pressure p, air velocity v, and concentration of species ‘i’ C i vary in space, x, y, z, and in time, t. The variation of these state variables is governed by fundamental mass, momentum and energy conservation principles, bound by environmental and thermal-mechanical-chemical boundary conditions, that allow prediction of the spatial and temporal variation of these state variables (see Awbi 1991 for an overview).

Broadly speaking, two numerical approaches are commonly used for this prediction, namely microscopic and macroscopic analysis.

Microscopic analysis, based typically on finite difference or finite element techniques, approximates the continuously defined state variables by a finite set of spatially discrete but temporally continuous state variables defined at or associated with discrete (mesh) points ‘j’ within the continuum. Microscopic methods of analysis provide the means to predict comfort variables and, importantly, their spatial variation within rooms (air dry bulb temperature and velocity are directly predicted while mean radiant temperature and RH distributions may be easily computed at each of the room air mesh points from computed surface temperatures and vapour-phase water concentrations respectively). As a result, microscopic analytical evaluation of comfort in rooms has become one of the primary applications of computational fluid-dynamics (see, for example Awbi and Gan 1994, Gan 1996). In spite of the direct utility of the microscopic approach to comfort prediction, several limitations must be noted, because of its expensiveness (in terms of data input and computation time often longer than a day), the special expertise needed to implement it and to evaluate the results. For this reason it remains a research tool and is very seldom applied in practice.

Macroscopic analysis, based on idealising the building system as a collection of one or more control volumes (a space whose behaviour is well known) linked by discrete heat or mass transport paths, also approximates the continuously defined state variables by a finite set of spatially discrete but temporally continuous state variables but now the discrete state variables are associated with either the control volumes or discrete transport paths (windows and doors). Macroscopic methods can provide an economic and accessible means to predict simple measures of thermal comfort within rooms (e.g., spatially averaged room air dry bulb temperature, mean radiant temperature, air velocity, and relative humidity). While they can not provide the spatial detail offered by microscopic analysis (frequently missing local phenomena which may considerably affect comfort as air re-circulation which reduce the diffusion of fresh air in specific zones), macroscopic methods can be readily applied to whole building systems and configured to allow an integrated consideration of interacting building systems (e.g. heat transfer in the building fabric and envelope, HVAC systems, lighting systems, and natural ventilation systems).

As in the microscopic case, however, these methods have been formulated to support only a trial and error approach to building design, nevertheless macroscopic analysis allows to link the response of the system directly to key design parameters. For this reason we use a macroscopic model to illustrate our approach of assembling qualitative models; nevertheless microscopic modelling or experimental data could also be used. In this case a more expensive analysis is required in order to extract the causal relationships between relevant state variables and key design parameters.

This makes it possible the integration of both approaches peculiarity; on the one hand the ability of detecting local characteristic of air motion within buildings but on the other hand the possibility of linking these characters directly with key design parameters.

Thus as an example, consider a building system idealised as a collection of zones linked by discrete airflow paths and conductive heat transfer paths.

Macroscopic discrete state variables of pressure and temperature will be associated to each of the zones (i.e. the pressure associated with a specific elevation within the zone identified in the figure 2 and the temperature associated with the spatial mean air temperature within the zone). Similarly, an outdoor ambient reference node will be associated with the ambient pressure and temperature. Surface temperature variables will be associated with the surface of each of the several conductive heat transfer paths within the building system and finally, the mass flow rate of air through each of the several discrete airflow paths will be identified.

,T i ) l

m &) (P 0

Figure 2: Macroscopic discrete representation of a multizone building ventilation model.

With these variables defined, one may apply mass and energy conservation principles to form systems of equations governing heat transfer and airflow in the building system (see Walton 1989 for details).

The model in usual case is constituted of rather complex coupled systems of non-linear equations. For example the individual pressure-flow relations for the discrete paths are generally non-linear but nevertheless depend on key design parameters of the flow path (e.g. size of window opening, speed of a ventilation fan, or height of a monitor window).

In most practical situations, however, it will not be possible to establish this relationship formally as the combined system of equations will be hopelessly complex. Consequently, it will be necessary to establish the relation numerically by systematically varying key design parameters over a range of reasonable values and solving for the system response (i.e. for a given building and ambient and operating conditions). Whether formally or numerically derived, one may establish the relation between system response and the key design parameters for a given design problem and for a given boundary condition vector:

{}(){}{}()

{}{}(){}φφφRH m T f RH f m f T ===,,&& where T , m & and RH are surface temperature, air mass flow rate and relative humidity vectors while φ is the vector of key design parameters.

This relationship could be very complex, because of the non-linearity and the instability of t h e system whose behaviour can drastically change with varying design parameters or boundary conditions.

By combining the system response results developed in terms of the key design parameters, with a comfort metric (i.e. considering for a given zone ‘i’ the dry bulb temperature and air velocity along with the spatial average of the mean

radiant temperature) we may establish the relation between the T , m

&, RH and the design parameters. As we can readily deduce, the mathematical model embodies in general a system of coupled equations, which it is not possible to explicitly solve in terms of design parameters. So we can use this model (as well as a microscopic numerical model) only to predict the behaviour of a well-defined system and nothing we can say about the strategy to be adopted in order to modify that behaviour in a required way.

4. EXPLICIT CAUSAL MODELING OF PHYSICAL BEHAVIOUR

Our efforts was aimed to support the decision making, with particular attention to the preliminary stage of design when the student is involved in complex inferences which integrate prediction and diagnosis in order to guide its trial and error activity. Numerical analysis approaches are directed instead only to predictive analysis while diagnosing numerical data (obtained through simulations or testing physical models) is essential to take corrective actions.

Bayesian Networks (also known as Belief networks or causal diagrams) we have employed in VENTPad, were developed to model distributed processing in reading comprehension, where both semantical expectations and perceptual evidence must be combined to form a coherent interpretation. The ability to co-ordinate bi-directional inferences filled a void in expert systems technology of the early 1980’s, and Bayesian networks have emerged as a general representation scheme for uncertain knowledge. Bayesian networks are directed acyclic graphs in which the nodes represent variables of interest and the links represent informational or causal dependencies among the variables. The strength of a dependency is represented by conditional probabilities that are attached to each cluster of parents-child nodes in the network.

For variables without parents (as the boundary condition variables), the probabilities are unconditional distributions.

With these data, a Bayesian network allows one to calculate the joint distribution over all variables1 from which all probabilistic queries, involved in reasoning, can be answered coherently using probability calculus.

They can be used t o model the causal mechanisms that operate in real systems rather than, as in many other knowledge representation schemes (e.g., rule-based systems and neural networks), the reasoning process. This model is obtained by representing the causal dependencies among the system variables as probabilistic functions (i.e., the probability that variable C assumes the value z when A assumes the value x and B assumes the value y is equal to 0,85, written P(C=z|A=x,B=y)=0.85, means that there’s a high probability that the state (x,y) forces C to assume the value z).

Bayesian networks effectively allow a number of integrated logical and quantitative inferences about the behaviour of physical systems and their application could be an interesting connection tool between logical and analytical procedures in preliminary design aiding.

The inference process based on bayesian networks is described in large body of literature and is best summarised in (Pearl 1988). Anyway I refer to following basic works for key concepts and terminology related to these issues (Pearl 1988, 1996; Shachter 1990; Jensen 1996; Spirtes et al. 1993). Bayesian networks have been applied to problems in medical diagnosis (Heckerman et al. 1992; Spiegelhalter, et al. 1989), map learning (Dean 1990), language understanding (Charniak and Goldman 1989a, 1989b). In architecture design and construction early applications are related to reliability analysis of innovative building products (Naticchia 1999a) and to diagnosis of building failures (Naticchia 1999b).

4.1 CAUSAL MODEL OF NATURAL VENTILATION

As an example, it is useful to think of the ventilation model from a causal point of view, as a network that links the key

m&). The comfort criteria also define a parameters – namely the ‘design space’ - to the physical variables (i.e. T i and

i

causal network along with the variables. The network that we obtain explicitly depicts the causal model of system behaviour where the nodes represent variables of interest (it is useful to think of them as discrete variables, which values represent intervals of the actual domain) and the links represent informational or causal dependencies among the variables.

Figure 3: Network describing the causal structure of the comfort problem.

1 In this paper we avoid focusing on calculation problems related to the resolution of bayesian networks. These issues

are best treated in (Pearle 1988).

To translate the causal network in a Bayesian Causal Network we must express the conditional probabilities which link any variable value with any value combination of its parent nodes.

The solving algorithms provide to spread the effect of an assertion (made by updating the likelihood of a variable), to any variable of the net. In this way the causal framework represented in the net, highlights the system states compatibles with observed data, by increasing the likelihood of specific values of the nodes. As written, this feature allows the investigator to answer a variety of queries, including: abductive queries, such as ‘What is the most plausible explanation for a given set of data?’; and control queries; such as ‘What will happen if we intervene and enlarge the inlet?’. Answers to these queries depend on the causal knowledge embedded in the network. The probabilistic basis of Bayesian networks offers a coherent semantics for co-ordinating top-down and bottom-up inferences, thus bridging information from high-level concepts and low-level percepts. This capability is important for achieving selective attention that is, selecting the most informative data in a specific step of the design. In other words the student is guided not only in diagnosing the reason of a behaviour but also in focusing its attention to weigh the importance of each variable in determining that response or in confirming the diagnosis. The capability of integrating in a unique graphical model, directly accessible to the user, many abstraction levels of the domain is a novel feature which enhance the tutorial role of the expert system, and modify the interaction with the user that directly interact with the graphical knowledge representation envisioning the effect of data he provides directly on the whole knowledge. This allows a sort of sensivity analysis about variables and parameters relevance, which enhances the user skill in perceiving the integrity of design problem normally hidden in other expert systems.

5. APPLICATION

In this section we’ll demonstrate the approach of causal network modelling through a simple case of stack ventilation of a single zone building under steady state conditions of heat transfer.

5.1 STACK VENTILATION OF A SINGLE ZONE BUILDING

Consider a simple single-zone model of a building utilizing a stack ventilation strategy 2.

Figure 4: Single-zone stack-ventilated building model.

A steady wind, characterized by a stagnation pressure p 0, approaches the building from the left as air passes through a window ‘a’ of a cross sectional area A a and exits at a higher opening ‘c’ of a cross sectional area A c , located a distance ?z above the lower opening.

These three variables (A a , A c , ?z ) will be taken as the key design parameters that will be adjusted to achieve thermal comfort. Wind pressure coefficients at the windows C pa and C pc , building conductances ΣUA , internal gains q gain , outdoor air temperature T 0 and heat capacity of air p c

?, are assumed known a priori. Two unknown state variables are associated with this simple idealization, the zone air temperature T i and the zone air pressure p i defined relative to a specific elevation, which will be taken along the horizontal centreline through the lower window.

With the problem thus defined, we can form the heat transfer system equations by demanding conservation of thermal energy:

()()gain a p i c p i q T m c T m c T T UA =?+?∑00??&&

2 This example is adapted from Axley 1997.

We’ll model the airflow through the windows using the familiar orifice equation as it has proven to be a reliable model.

In the absence of wind-driven pressures, both indoor and outdoor air pressures will be assumed to vary hydrostatically in proportion to the indoor and outdoor air densities ρi and ρ0 respectively. Then:

()i a d a p p A C m ?=0?2ρ

and ()()()z g p z g p A C m i i c d c ?????=00?2ρρρ

The airflow system equations may be formed by demanding conservation of airflow, namely by using the equation m a +m c =0. Finally we’ll use the ideal gas law to estimate air densities indoor and out.

In this case the heat transfer and airflow equations are coupled through both the airflow rate and buoyancy terms and the resulting nonlinearity is pathological. Consequently, it was not possible to explicitly express the resulting equations for the indoor air temperature T i in terms of the design variables A a , A c and ?z . Nevertheless, we can numerically solve the equation and plot the indoor air temperature as depicted in figure 5. In addition the inlet air velocity is also plotted as it varies with the inlet window opening A a :

a

a A m v ρ&=

A ( c 2A a

m ) 2

( m 2

,7 m K/W orifice opening discharge coefficient ) 3

2 2 gain pb

Figure 5: Isothermal curves (T i =26°C) and inlet air velocity of the model solution (given the listed variable values) in

terms of the key design parameters.

A designer could use the numerical solution to guide design decision while maintaining the comfort object (i.e. T i ?26°C, v ?0,5 m/s ). For example, with a stack height of 5 m the designer could achieve the comfort objective with the window opening combination (A a , A c )?(1,5 m 2, 1 m 2).

Unfortunately the number of key design parameter is often more than three, so that the use of plotted graphs becomes practically impossible. In this case only a trial and error approach is available to make design decision.

Consider now the simple causal network of figure 6 where the variables are considered as discrete with listed intervals as feasible values.

A a≡ [0.3÷1, 1÷2, 2÷3];

A c≡ [0.3÷1, 1÷2, 2÷3];

?z ≡ [2.5÷5, 5÷10, 10÷20];

v ≡ [0.2÷0.5, 0.5÷2, 2÷4];

T i≡ [22÷24, 24÷26, 26÷28];

Comfort ≡ [Weak, Reasonable, Good]

Figure 6: Causal model of the single-zone stack-ventilated building.

To specify the probability distribution of the network, one must give the prior p robabilities of all root nodes (nodes with no predecessors) and the conditional probabilities of all non-root nodes given all possible combinations of their direct predecessors. Considering the node v, we have to specify the values:

P(v=0.2÷0.5|A a=0.3÷1,?z=2.5÷5);

P(v=0.5÷2|A a=0.3÷1,?z=2.5÷5);

P(v=2÷4|A a=0.3÷1,?z=2.5÷5);

and so forth with all combinations of ?z and A a.

To express such values, say P(v=0.2÷0.5|A a=0.3÷1,?z=2.5÷5) we had to evaluate the probability that using design parameters included within the intervals (A a=0.3÷1,?z=2.5÷5) the inlet air velocity will be included between the values [0.2÷0.5 m/s]. Many approaches are available to solve and automate this problem, which is typical in the fuzzy set theory3.

Once the Bayesian network has been assembled, it allows effectively investigating the stack ventilation design problem both assessing the consequences of a decision on the comfort variables and going back up the causes of a given behaviour.

For instance we may set the variable T i to a given value and the variable Comfort to the value ‘Good’, getting the most likely design parameter values, which achieve the selected state. Afterwards, by setting one or more design parameters we can address the decision about the remaining design parameters by selecting those values with higher related probability. In this manner the design could achieve a good result in terms of internal comfort also overlooking the investigation about the inlet air velocity, which value is constrained by the ‘Good Comfort’ and the air temperature selections.

Moreover the graph we have assembled can be clearly expanded, introducing further causal dependencies between design parameters (i.e. parameters as T0 that we have fixed in this example, or expressing a design parameter by means of more detailed features) and by adding external variables, which affect the microclimate conditions.

Suppose for instance, we want to introduce the relationship between the direction of airflow through the inlet opening and the pattern of air circulation inside the room.

It is possible to embody in the model the knowledge that the airflow may assume one of the three patterns displayed in figure 7, varying the position of the jalousie window sashes showed in the section ‘d’ of figure.

The different patterns considerably influence comfort and air quality in the different locations of the room, that is the configuration ‘a’ produces a significant recirculation of air showed in figure, while the pattern ‘b’ is characterized by high air velocity in the centre location of the room (because of the narrow airflow path which is limited by the vortex).

By using such a network also an unskilled designer could address detailed decision about these fluid dynamic issues in the preliminary stage of design. For instance, he could explain a computational result that appears like the condition ‘a’, by means of an incorrect air insertion through the inlet (i.e. related to the trees located near the building that modify the airflow direction) and address the design modification by placing sashes of type ‘c’ in front of inlet.

Moreover when the situation ‘b’ is recognised he could improve the ventilation behaviour of building simply placing a low hedge as showed in section ‘c’ of figure.

3 With a complex mathematical model a method to simple compute these values could be the MonteCarlo simulation

technique.

Figure 7: Effects of the design details on the internal airflow pattern.

6. DISCUSSION

The aim of VENTPad is that of demonstrating that qualitative physics has advanced enough to support new applications of AI to educational problems. Bayesian causal networks modelling provides representational tools and techniques that can be used to encode a substantial body of knowledge about civil engineering physics, with both diagnostic and predictive inference capabilities.

Automatically generated explanations enable the user to explore the consequences of his or her assumptions, and figure out what modelling assumptions are needed to make further progress. To date, VENTPad is still under development (using the Hugin shell for Bayesian networks analysis) and has only been tested on simple case study. We will be testing it with undergraduate civil engineering students.

Our goal is to have VENTPad continuously available in architecture design courses and to enhance it in a commercial version, which features are extended to support the analysis of multizone models (i.e. residential and office buildings) and nearly external environment of building.

7. REFERENCES

[1] Awbi, H.B., (1991) Ventilation of Buildings, E&FN Spon, London.

[2] Axley, J.W., (1997) Macroscopic Formulation and Solution of Ventilation Design Problems, in proceedings of 18th

AIVC Conference, Ventilation and Cooling, Coventry, GB.

[3] Gan, G., (1996) Effect of combined heat and moisture transfer on the predicted thermal environment, Indoor+Built

Environment, Vol 5 (N°3), 170-180.

[4] Awbi, H.B. and Gan, G., (1994) Predicting air flow ande thermal comfort in offices, ASHRAE Journal, February,

17-21.

[5] Bakker P.G., (1991) Bifurcations in flow patterns: some applications of the qualitative theory of differential

equations in fluid dynamics, Kluwer Academic Publishers.

[6] Bobrow, D., (1984) (Ed.) Qualitative reasoning about physical systems, MIT Press, Cambridge.

[7] Charniak, E., and Goldman, R. (1989a) A Semantics for Probabilistic Quantifier-Free First-Order Languages with

Particular Application to Story Understanding. In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, 1074–1079. Menlo Park, CA.

[8] Charniak, E., and Goldman, R. (1989b) Plan Recognition in Stories and in Life. In Proceedings of the Fifth

Workshop on Uncertainty in Artificial Intelligence, 54–60. Mountain View, CA.

[9] Dean, T. (1990) Coping with Uncertainty in a Control System for Navigation and Exploration. In Proceedings of

the Ninth National Conference on Artificial Intelligence, 1010–1015. Menlo Park, CA.

[10] de Kleer, J., (1975) Qualitative and quantitative knowledge in classical mechanics, TR-352, MIT AI Lab,

Cambridge, Mass.

[11] Gentner, D. and Stevens, A., (1983) (Eds.) Mental Models, Erlbaum Associates, Hillsdale, N.J..

[12] Heckerman, D., Horvitz, E., and Nathwani, B. (1992) Toward normative expert systems. Part I: The Pathfinder

project. Methods of Information in Medicine, Vol. 31. pp. 90-105.

[13] Jensen, F.V. (1996) An Introduction to Bayesian Networks. Springer, New York.

[14] Naticchia B. (1999a) Qualitative simulation for assessment of chemical-physical compatibility among building

components, In Proceedings of the 8th International Conference on Durability of Building Materials and Components, Vancouver.

[15] Naticchia B., (1999b) Diagnosing aged housing defects: A bayesian network approach for reliability of

rehabilitation effects, In Proceedings of the XXVII IAHS World Congress on Housing, San Francisco.

[16] Naticchia B., (1999c) Physical knowledge in patterns: Bayesian network models for preliminary design, in

Proceedings of the 17th, eCAADe conference, Liverpool, UK.

[17] Pearl, J. (1988) Probabilistic Reasoning in Intelligence Systems. Morgan Kaufmann, San Mateo, CA.

[18] Pearl, J. (1996) Causation, action, and counterfactuals. In Y. Shoham, editor, Theoretical Aspects o f Rationality

and Knowledge, Proceedings of the Sixth Conference, pp. 51-73. Morgan Kaufmann, San Francisco, CA.

[19] Shachter, R. D. (1990) Special issue on influence diagrams. Networks: An International Journal, Vol. 20 (5).

[20] Spiegelhalter, D.; Franklin, R.; a nd Bull, K. (1989) Assessment Criticism and Improvement of Imprecise Subjective

Probabilities for a Medical Expert System. In Proceedings of the Fifth Workshop on Uncertainty in Artificial Intelligence, 335–342. Mountain View, CA.

[21] Spirtes, P., Glymour, C., and Schienes, R., (1993) Causation, Prediction, and Search. Springer-Verlag, New York.

[22] Walton, G. (1989) Airflow Network Models for Element-Based Buildings Airflow Modelling in ASHRAE

Symposium on Calculation of Interzonal Heat and Mass Transport in Buildings, Vancouver.