1

Locating fire stations in Belgium: An integrated GIS approach.

Philippe CHEVALIER1,3, THOMAS Isabelle2,3,4, GERAETS David5, GOETGHEBEUR Els6, JANSSENS Olivier5, PEETERS Dominique2,3, PLASTRIA Frank7

1 Louvain School of Management, Université catholique de Louvain, Belgium. 2 Department of Geography, Université catholique de Louvain, Belgium. 3 Center for Operations Research and Econometrics, Université catholique de Louvain, Belgium. 4 FRS, Fund for Scientific Research, Belgium. 5 Experian Business Strategies SA, Louvain-la-Neuve, Belgium

6 Department of Applied Mathematics and Computer Science and Center for Statistics, Ghent University, Gent, Belgium

7 MOSI, Vrije Universiteit Brussels, Belgium

Corresponding author Philippe Chevalier chevalier@poms.ucl.ac.be Unité de gestion logistique et modélisation , IAG-Louvain School of Management Place des Doyens, 1 B-1348 Louvain-la-Neuve Belgium Tél: +32 10 47 83 96 Fax: +32 10 47 83 24 Keywords: Location-allocation - GIS - Fire-stations – Belgium JEL: C61, R53

2

Locating fire stations in Belgium: An integrated GIS approach.

Abstract.- This paper shows the potential of a decision support system developed for Belgium by a consortium of universities and a private firm, in the framework of a public call of the Ministry of the Interior. The system is designed to let evidence based optimal solutions enter a pragmatic decision process leading to an effective management of fire-stations locations and allocations. The originality of this project consists in including a risk modelling approach developed at a national scale. This analysis involves a multiscale GIS-system which includes a thorough representation of the physical, human and economic spatial realities, a risk modelling approach, an adequate optimal location and allocation model taking into account the queuing as well as a staffing problems. It concludes with an operational interactive tool for defining locations, allocations, staffing, response times, cost/efficiency trade-off, etc. conceived in an assessment as well as a prospective context: it has indeed numerous functionalities including rapid modification of the modelling conditions to allow for immediate scenario analysis, multi-scale analysis, as well as prospective analysis. Keywords: Location-allocation - GIS - Fire-stations – Belgium JEL: C61, R53

-

3

Introduction

Emergency services play a key role in any country. Their operations are high profile since the public is acutely aware of its own vulnerability and possibly vital reliance on the service one day. Service men and women are rightly admired, and even more so as their interventions may come at a risk to their own lives. More generally, these services come at a considerable cost, for a small country like Belgium (10M inhabitants, 30 000 sq. km) resources involve almost 20 000 people (among which about 12 000 volunteers) and an annual budget of about 400 million euros. The efficient utilization of these resources forms a big responsibility and can have an important impact on the quality of the service provided.

In most countries, the patrimony of fire stations has historically grown, one station at a time, as new needs and means for prevention and protection emerged. A call from the Belgian Ministry, offered the rare opportunity to a consortium of scientists and a private developer to evaluate cost-efficiency of the existing global infrastructure and study alternative locations and operational constraints expected to provide an optimal cost-benefit ratio for the service. To evaluate performance of past and future infrastructures, we analyzed data on emergency calls and observed response times over time and space, merged this with census data on the socio-economic and geo-physical environment, and acknowledged the administrative constraints under which the current service providers operate. Based on this input, we evaluate the current system a set of norms that define an acceptable operational response and propose new, improved locations and allocations of resources. In this paper we describe the challenges encountered, and real-time `quick’ solutions used for the development of a new dynamic and interactive decision support system. We highlight the practical potential of the developed system, share the lessons learned and point to further work and possible improvements.

This paper deals with the development of an optimal location and allocation strategy for fire stations in Belgium, one of the most important strategic decisions for the efficiency of the emergency services. It describes the setup and main results of a research project financed by the Ministry of the Interior (Home Office) aimed at developing an operational and interactive decision support tool related to the spatial management of disaster and emergency services: location of brigades, design of service areas, staffing and equipment (Janssens et al., 2006). The fundamental question underlying this research is: given a time norm (e.g. 8 minutes) within which 90% of the fires should be reached, what is the least cost solution that can achieve this? The main decision variables are the locations and the staffing of the fire stations. In other words, the proposed approach is intended to guide retrospective assessment as well as prospective planning. It was designed to answer questions like: is the present spatial organization efficient and equitable? Is it reasonable to consider a reorganization in Belgium? How much would it cost? How can the steps of this reorganization be planned sequentially? However, our project limits itself to the development of an operational tool rather that proposing feasible solutions. This last task belongs to the Public Authority who has to calibrate realistically the model.

The problem is indeed quite complex due to its dimension and the many constraints (see e.g. Tillander, 2004). The service does not only respond to fire, but also to other emergencies as well as to less urgent missions. It thus faces several types of risk (fire, road accident, floods, etc.) that require different types of equipment and human skills. As these risks vary over space (and its characteristics: physical, social, economic, history of urbanisation, transportation, environment, etc.), location planning should rely on a local as well as global risk assessment. For emergencies, the delay between the brigade’s departure and arrival on the scene is a key

4

determinant of operational efficiency. This delay not only depends on the distance to bridge, but typically also on the peculiarities of space (road network, environment, etc.) and time(congestion; meteorological conditions; etc.). A long delay has a large impact on the consequences of the fire/accident in terms victims and damage. The national administration of the service must manage large geographical differences well known to lead to equity-efficiency trade-off considerations, but also to a multi-scale problem (the spatial accuracy should be higher in large cities and lower in rural areas). Moreover, the occurrence of simultaneous calls is not to be ignored. Hence, nation-wide global decisions might have dramatic local consequences. Since fire incidents are the core tasks of the service, we here call « fire stations » the sites to be optimally located. Since these stations must answer to more than fires, we speak about “emergency services”, unless we specifically talk about calls for fires.

The objectives of the approach proposed here are the following: to define clear criteria representing realistically the location problem, to develop an operational and interactive tool, to determine optimal locations, to design optimal service areas, to assign staff and equipment (optimal size of the brigades), to manage intervention time at the scale of one country (Belgium) and consequently, to test the extend to which spatial reorganization improves on the current situation. Hence, this paper fits well in the present literature (see e.g. Goldberg, 2004 or Yang et al., 2006 for a literature survey): there is a severe need for modelling real-world complexities and suggesting sustainable solutions.

This paper starts with an overview of the architecture of the project. Data input and other pre-processing functions are then described, followed by the location-allocation methods. Data output, and results interpretation are illustrated by means of a case study. The paper concludes with a summary of model potentialities, limitations, but also plans for future work.

Architecture of the project

The global objective is here to construct an architecture coupling adequate Data Base Management Systems (DBMS) with Location-Allocation Models (LAM) in order to provide an operational Spatial Decision Support System (SDSS) (Figure 1).

DBMS (Data Base Management System) refers to software that collects, manipulates, queries, and retrieves tabular data. It is a prerequisite to an operational model as data collection is the basis of the application developed in the project. Data are here collected about the service itself: what are the tasks to be provided? What are the service level requirements today and in the future? What are the costs of running the services? What are the standards in terms of functioning and locations? What are sustainable standards in average intervention time? Etc. These are all basic management questions which do not always find clear-cut answers as we are dealing with a public service in a quite multifaceted country (See Section 3).

LAM (Location-Allocation Models): the optimal location of emergency facilities has a long history in management sciences and operations research literature that we do not intend to review here (see e.g Goldberg, 2004 who summarises 115 papers and issues related to theoretical as well as applied aspects of the problem). In fact, operational location-allocation models have to be specifically developed in order to take into account the specificities of the problem and for? defining the location, the service areas, the size of the fire-stations (staff, building, cars, …), the intervention times, queuing aspects for all types of emergency

5

services. This is detailed in section 4. The LAM is coupled with the GIS in order to provide an operational decision tool.

GIS (Geographical Information System) is an information technology used to maintain and analyze geographic data; it organizes data into layers and relates sets to space. It allows conveying relationships and operational trends in a spatial context. The use of GIS for providing spatial decision support systems is nowadays well-known (Longley et al., 2005): it provides most inputs to the LAM. GIS are used for storing and analysing geographically distributed data in a multiscale perspective (data about the road network, the distribution of potential and present risk factors, the characteristics of the geographical space, the environmental risks, the current and potential location sites, etc.). Compared to the existing literature, the originality of this project consists in including a risk modelling approach developed at a national scale without ignoring local specificities.

SDSS (Spatial Decision Support System): the complexity of the service as well as its organizational and geographical constraints is coupled with the modelling process in order to propose a user-friendly interface where the decision maker can select various combinations of criteria for evaluating current locations or for predicting the impact of changes in location. The charge was entrusted to an interdisciplinary team with expertise in data management, statistics, risk modelling, geography and operations research, found in a consortium of scientists and Experian Business Strategies, a private firm also in charge of the follow-up (3 years). Given severe time pressures on the project and the poorly developed data base systems at the grassroot levels, an achievable balance had to be sought between a sophisticated scientific approach and a feasible, fast and user-friendly solution.

Figure 1 : Architecture of the project.

6

Data input and processing

Study area and scale of analysis

The model is here developed at the scale of a country. Belgium is a small but highly urbanised European country: more than 10 million inhabitants spread over approximately 30 000 sq. km. Density varies from almost 20 inhab./km² in rural regions to more than 20,000 in highly urbanised areas (see Appendix 1). The country covers differences in topography (rather flat in the north, rugged landscape in the south), in urbanisation and in city networks (history) as well as in economic, social, political, cultural and environmental features. . The Belgian city network is dominated by Brussels, which counts more than 1.5 million inhabitants in its extended urban agglomeration; it is centrally located and sprawls out to a large extent into Wallonia and Flanders, across its administrative border. In the hierarchy of cities, Brussels is followed by Antwerpen (500.000 inhabitants), Gent (250.000), Liège and Charleroi (200.000). Underlying the official division, there is a whole range of urbanisation levels, going from heavily populated agglomerations with high densities through rich and less densely populated suburbs and “commuter zones” that are re-oriented from a local economy and lifestyle to urban involvement, all the way to rural structures (Van der Haegen et al., 1996). 57% of the Belgian population is living in a city region that accounts for 26% of the national surface. This high prevalence adds to the importance of levels of urbanisation: urban sprawl, suburbanisation and peri-urbanisation in Belgium. Living environments are hence quite heterogeneous.

Official statistics are gathered at different levels of aggregation, called Nuts levels, where Nuts stands for “Nomenclature of Territorial Units for Statistics” that is defined according to a number of basic principles by the E.U. (Eurostat, 2005). The Ministry of Interior is administrated federally: it concerns the entire country, corresponding to the Nuts0 level of aggregation Belgium is administratively divided into three partially autonomous administrative regions (Nuts1 division), divided into 10 provinces (Nuts2). Both Nuts1 and Nuts2 borders are seldom crossed by Disaster and Emergency Services. Nuts regions smaller that Nuts2 are administrative divisions of the country. While Nuts5 regions (communes) could have been sufficient for studying national fire-stations locations, these are still too coarse for intra-urban locations. Moreover they would have led to unacceptable aggregation errors in the local data and consequently, in the modelling solutions. Hence `statistical wards’ are our basic spatial units (BSU): they correspond to the finest aggregation level for which data are officially available. The country is divided into 19.781 statistical wards (BSUs). These wards are defined by the Belgian federal division of statistics in terms of internal homogeneity according to the latest census data. They vary widely in shape and size leading to well-known aggregation problems (see Goldberg (2004) for a review). In this project each BSU is used both as a demand point as well as a potential location for a fire station. For convenience, it was decided that the centroid of each BSU forms the basis for distance computation; information related to the surface of the BSU is hence summarized in this point. Given the size of a BSU, we computed that the travel time within a BSU is always smaller than 2 minutes. However, for confidentiality reasons only, the maps included in this paper are drawn at the Nuts5 level: results obtained at the wards’ level are aggregated at the communes’ level.

Without accurate and realistic travel times, location models would have little value; distance is a key element of the modelling procedure. Shortest distances are here computed in travel time between the centroids of the BSUs along the road network. Average speed and traffic conditions are taken into consideration (Data base used: MultiNet from TeleAtlas; Computation: Experian Business Strategies). Average speeds were attributed to each type of

7

road segment, and speed reduction factors have been taken into consideration within urban areas as well as for crossroads. Different speed conditions were used ranging from 15km/h at intersections to 80km/h on highways (see Janssens et al., 2006, p. 13). These speeds were defined by a workgroup of experienced firemen from various fire stations. However, no specific traffic regulation measures or congestion situations were included in the computation. The flexibility of the tool allows it to be updated to incorporate more refined information when the latter becomes available.

Demand: Risk modelling

Risks were divided into two types: recurring risks and sporadic risks. Recurring risks concern events that occur with a frequency such that it can be estimated from historical data. Sporadic risks on the other hand are linked to rare or unlikely events (like a major Tchernobyl-like event) for which it is not possible to determine a probability of occurrence based on direct observation of historical data, but that could have consequences of such magnitude that they have to be taken into account in the planning of emergency services. They correspond for instance to education centres of all sizes and types, health care centres, industrial locations with more that 50 employees, firms with a high environmental risk called Seveso-type, all types of halls such as theatre, movie, etc., underground pipelines, tunnels as well as high rise buildings. Consequently, the estimation of each type of risk is very different. The recurring risks were forecasted by regression methods using historical intervention data, while the sporadic risks were evaluated by gathering exhaustive lists for potential causes of catastrophic events.

In order to estimate recurring risks, a survey was conducted in all 251 existing fire stations on a ten year period (1993-2004). Not all fire stations could provide full data, responses varied widely from no answer at all, to some years, to full report on the ten year period. Also, the quality of the data provided varied widely from completely useless by lack of precision to fully exploitable. Sufficient information could be gathered on about 839 000 incidents including 157 000 fires. The yearly mean amount for Belgium is around 30 000 fires and 150 000 other incidents. A statistical model linked these data to (complete) statistics on risk factors, which allowed - under certain assumptions - to predict risk levels for the entire country. The explanatory variables (characteristics of the population residing, working, shopping, visiting, etc. the commune, as well as land-use characteristics) of the statistical model were mostly drawn from the 2001 national census.

Several models were built in order to estimate recurring demand in a prospective way but also in order to fill in the gaps in our geographical representation (missing data due to the fact that not all fire-stations answered our survey). The emergencies were divided in 5 categories: fires related to a dwelling, fires not related to a dwelling, medical emergencies, non medical emergencies and not urgent tasks. A separate model was built to estimate the local rates of calls for each category of emergency. Specifically, we fitted a loglinear model to estimate yearly numbers of type-specific events per square kilometre in the BSUs. Model input could rely on predictor values from census information available for all BSUs, but outcomes were missing for all BSU’s covered by fire stations that did not submit event data. The models aimed to estimate the spatial variation of events in spite of the missing data, without the ambition of yielding causal explanations. As expected from the literature (see Peeters and Thomas 2001) population density plays is a key determinant of fire hazards. The estimated spatial information is illustrated for one type of risk (fires in dwellings)(Figures 2 and 3). For confidentiality reasons results are here mapped at the level of the communes (Nuts5). Both

8

maps reveal the strong influence of the urbanisation level and the population density (see Appendix 1).

Figure 2: Estimated number of fires in dwellings by commune and by year.

Figure 3: Estimated density of fires in dwellings (number of fires by year and by square kilometre)

(between brackets : number of communes within each class)

The sporadic risks are based on potential causes of catastrophic events, theses causes are mainly linked to places where many people gather (e.g. health care centres, education centres, shopping centres, entertainment centres, large office buildings, …) or to places with infrastructures presenting potential dangers (e.g. high rise buildings, industrial complexes with hazardous materials (a.k.a Seveso-type sites), tunnels, pipelines, …) A database was build containing a total of 9 categories of sporadic risks. The sporadic risks to be taken into account can be controlled at the category level or at the individual site level.

Supply : Intervention rules

Besides the enquiry on the history of the interventions, a survey was also conducted in order to understand the intervention rules and their application. It is clear that a large variety of interpretations was observed and the application of the rules is far from being unique. Two

9

types of information were collected: one type pertains to the selection of the fire station that will send a team, the other to the composition of the allocated team. The project-team collaborated with representatives of the fire-fighters to establish standards for the composition of the response teams needed for each category of emergency (team + equipment). Standards have also to be established for the escalation scheme when reinforcements are needed. These standards further depend on the location of the emergency; they can later be translated into parameters and costs. These definitions are a useful by-product of this project but highly depend upon the decision maker.

Optimization heuristic

The recurrent risk associated with each BSU is represented by 5 rates, each of which corresponds to the 5 categories of emergencies. A 0-1 variable was also defined, indicating the presence inside each BSU of a “sporadic risk” that requires the presence of a fire station to be located within a specific time radius.

The optimal dimensioning problem consists in finding the minimum cost configuration for the emergency services in order to answer a given proportion of emergency calls within a given time threshold (for example answer 90% of calls within 8 minutes), and ensuring that there is a fire station within a given time radius of each sporadic risk (for example there is a fire station within 10 minutes of each location with a sporadic risk). The cost of a configuration depends on the number of fire stations as well as on the number of vehicles and crews present in each station. The time to answer an emergency call includes the time for a vehicle and crew to be ready and the time to drive to the call site; we here only consider the latter.

Given the complexity of the location-allocation problem at the scale of a country, two sub-problems were approached sequentially. The location problem was solved first, and the staffing problem next. For the location problem we assume that a crew is always available in each station, and we seek the minimal number of stations (and their locations) needed to reach the desired service level. This problem is modelled as a Mixed Integer Programming (MIP) optimization problem. In the staffing problem we take the locations as fixed and we optimize the number of crews in each station for the desired service level. A Specific evolution of the hypercube queuing model is used.

The location problem The location problem is modelled as follows:

!i

iyMin ,

s.t. jzyai

jiij!"#

wjzj! P w

j

j

"j

"

! "#$i

iij Sjyb 1

yi ! 0,1{ } "i

z j ! 0,1{ } "j

where: - i = 1,…, n index of the possible locations (BSUs);

10

- yi indicates whether there is a fire station at location (BSU) i; - wj quantifies the expected recurrent yearly event rate at BSU j; - zj indicates whether BSU j is within a distance of a fire station acceptable for recurrent

risk; - aij indicates whether the potential location i for a fire station is within acceptable

distance for the recurrent risk at BSU j; - P is the fraction of the recurrent risk that must be covered by fire stations within

acceptable range; - S is the set of BSUs that contain a source of sporadic risk and must thus be within a

special acceptable range of a fire station; - bij indicates whether the potential location i for a fire station is within acceptable

distance for the sporadic risk site j. Our decision aid tool provides a flexible and user-friendly interface for quick modification of the parameters (e.g. fraction of recurrent risk to cover, list of sporadic risks, list of potential locations for fire stations, specifications of acceptable time-distances for each type of emergency/risk, etc.).

We started by implementing the problem for exact solution with a commercial MIP solver. As the solution times were often quite long (several hours) with the best available commercial solvers, even when reduced to the provincial level (see below), we chose to implement a mean field feedback neural network approach (Ohlsson et al, 2001). This heuristic has a reliably short running time and led to solutions that were almost as good as the optimal ones.

The territory served by an emergency service is nowadays most of the time a commune or, in some cases, an association of communes. A service can nevertheless be called for reinforcing other stations. A first analysis revealed the dramatic impact these boundaries have on performance. It would thus seem advantageous to solve the problem at the national level, ignoring further territory constraints. The national scale was however considered as too large for both administrative and computational reasons. Our calculations showed that by operating within each province (Nuts2) one loses little efficiency for a better control of the service level. This is due to the diversity in Belgium, where some provinces are much more densely populated than others. The ‘optimal’ solution at the national scale will tend to strongly favour more densely populated areas when trying to meet a global service level. At the province scale, it is easier to manage the service level and/or to get a covering of the population and surface. The number of needed fire stations evolves with the percentage of risk covered (Figure 4). By this we mean the percentage of the expected number of calls (according to the estimated occurrence rates of all 5 risk categories) that lie within a given acceptable time-distance of a fire station. The standard time-distance used here was 8 minutes for all categories of recurrent risk.

These results confirmed the Ministry’s presumption that lifting the original intervention rules would have a direct beneficial effect. They also allowed to measure it quite precisely, and showed that relocation/removal/addition of fires-stations might be advantageously considered. They formed a second important set of results of the study.

11

Figure 4: Trade-off between number of fire stations and % of risk covered for a province

Staffing

When the locations of the fire stations are fixed we may optimize the staffing of each station in order to meet the required service level at the minimum cost. In case each BSU could only be reached by a single fire station within the required service time, the staffing problem would consist of a series of independent problems for each station. In practice the topology of the country imposes that in order to cover let us say 90% of the country within a specified time interval, we will have relatively large overlap regions where more than one fire station can intervene within the prescribed time. As the frequency of occurrence of incidents is only relevant for recurrent risks, the staffing part of the model does not take the sporadic risks into account.

This gives rise to a spatial queuing model. For an emergency call, we suppose that as long as the closest fire station has an adequate crew available, it will generate the dispatch. Otherwise, we look for such crew in the second nearest fire station, and so on until an adequate crew is found. We use the hypercube queuing model (Larson and Odoni 1981) to evaluate waiting times of different calls and the service level following the arrival rates of demands and the staffing levels of each fire station. A local search heuristic is then used to optimize the staffing of the different stations in order to meet the global service level.

The notion of service level is slightly modified here. Indeed, since there might be portions of the country that are not reachable within the specified maximum time limits, the optimization model would completely neglect those areas: no matter how many crews are used, the calls from such locations cannot be answered in time. In order to avoid this, our metric is now the fraction of calls answered within the specified time limit in case this is possible for the call location, or answered as fast as technically possible otherwise, when the call location is beyond the specified time limit of any fire station. Essentially the second case means that a crew must be available from the closest fire station.

As mentioned earlier an escalation scheme has been defined to send reinforcements when emergencies require more resources than a single response team. In the decision aid tool, two levels of reinforcement are considered, each level having its own time limit and coverage

12

requirement. The same optimization approach is used for the staffing of reinforcement crews at each level (the reinforcement must be based in one of the fire stations). Data output and results interpretation

The final product takes the form of a flexible software tool…[please explain in general terms the specific use of the tool, before starting the detail]. The user is faced with choices related to the modelling rules and conditions:

(1) the studied area: does he/she consider the location problem for the entire country (Nuts0), or only one region (Nuts1), province (Nuts2) or even to a set of communes corresponding for instance to an urban agglomeration keeping in mind that the model relies on detailed spatial information at the level of the statistical wards (BSUs);

(2) the types of sporadic risks: one can choose which sporadic risks are considered in the modelling process, and the associated time thresholds;

(3) the types of recurring risks: one can restrict the application to a particular kind of recurring risks or consider all of them (5 types), depending on strategic decisions ;

(4) the intervention time thresholds for the recurring risk: the decision maker decides for instance that each BSU should be reached within a 8 minutes intervention time; he/she can then later do the same simulations with 9, 10 or 12 minutes in order to test and measure the impact of such a decision; He can also use a varrying (as a function of the local population density) time

(5) the fraction of risk « covered », i.e. fraction of the expected emergencies according to the estimated rates for all categories, that must lie within the intervention time threshold. This applies only to the recurrent risks; for the sporadic risks all included locations must lie within the corresponding intervention time threshold.

(6) potential locations for a fire-station: the user has the choice to list the BSUs that are /are not potential locations. Some locations may be avoided, otherimposed;

(7) relative weight of the existing fire-stations: the user can weight the inertia of the existing system. This weight depends on the available budget as well as the willingness to change. Moving a fire-station from one location to another is quite expensive; the decision maker can be more or less ready to move.

(8) Service level for staffing: what fraction of expected calls should be serviceable within the special rules used for the staffing calculations.

(9) Time it takes to handle each category of emergency. This is used in the staffing part of the model to compute the unavailability of the crews (“busy-time”).

13

Figure 5: Example of screenshot: when clicking on any BSU, it is highlighted and relevant risk

information gets displayed (one can choose which information is displayed). The red dots are here the locations of sporadic risks.

Figure 6: Example of screenshots: optimal locations of fire stations when 70% of the emergencies lie within 10 minutes of a fire stations. “Sporadic risks” are not considered in the left hand map; the map on the right does impose that all locations with a “sporadic risk” must be within 10 minutes of a fire station. Coloured BSUs (pink) are those that cannot be reached beyond the 10 minutes limit in this

layout.

Conclusion

The analysis tool developed in this project results in an efficient and flexible tool that not only assists decision makers, but also helps to better understand the functioning of the Belgian emergency services. It is important to stress the fact that the quality of the results is highly dependent on the quality of the empirical data. Given the many data sources we, it proved to

14

be a real challenge to collaborate well with all organizations on whose data we rely and for whom, emergency services are often no priority. One additional difficulty for such endeavour is that many countries have decided not to proceed with population censuses anymore, this would make the estimation of spatial variation in risk much more difficult.

One of the key benefits of this project is that it provides a clear and objective basis for discussion between the many stakeholders involved with the emergency services planning. To cite a few of the stakeholders, we can think of the communes, the federal emergency service agency, the firemen unions, … Using this tool, it is much easier to asses the impact of various planning decisions. It is notable that six months after the end of the project, a national reform of the emergency services is already well underway, taking into account several findings of the project, most notably the necessity to pool services for large intervention zones.

This paper discusses Version 1.0 of our computer package. Our model would find all its sense if a permanent platform would be created not only for the updating of the GIS data but also for refining the program itself, following its more intense use. The tool does not output `the ultimate solution’; it forms however an important contribution towards rational, evidence based, informed decisions which will necessarily involve political choices.

Acknowledgements- This paper summarizes a fruitful collaboration of Adhoc Solution (a private company located in Louvain-la-Neuve and now part of the Experian Group), a consortium of scientists from several universities and several disciplines (Chevalier P., Crémer L., Geraets D., Goetghebeur E., Janssens O., Plastria F., Peeters D., Thomas I., Vandaele. N., Voisin O.) as well as the Ministry of the Interior represented by Mme Breyne and Mr. Looze. A special thank to the firemen and their commanders for the fruitful discussions and for the data. They all contributed to the success of this research. References Badri M., Mortagy A., Alsayed A. 1998. A multiobjective model for locating fire stations. European Journal of Operational Research, 110, 243-260.

Eurostat 2005. Regions - Statistical Yearbook 2004, EC Brussels

Goldberg J. 2004. Operations Research Models for the Deployment Emergency Services Vehicles. EMS management Journal, 1:1, 20-39.

Janssens O. , Cremer L., Geraets D., Zwings A., Roumieux V., Voisin O. 2006. Modélisation et quantification d’un réseau optimal de services d’incendie. Rapport Interne Ibz, 121 p.

Larson R.C. Odoni A.R. 1981. Urban Operations Research. Prentice Hall, (http://web/mit/edu/urban_or_book/www/book/ )

Longley P., Goodchild M., Macguire D., Rhind D. 2005. Geographical Information Systems: Principles, Techniques, Management, and Applications Wiley, New York, 2 volumes.

Ohlsson M., Peterson C., Söderberg B. 2001. An efficient Mean Field Approach to the Set Covering Problem. European Journal of Operations Research 133, 583-595.

Peeters D., Thomas I. 2001. Localisation des services publics: de la théorie aux applications. in Sanders L. (ed.), Les modèles en analyse spatiale, collection IGAT, Hermès-Science Publications, Paris, pp. 105-127.

Sauvagnargues-Lesage S., L’Héritier B., Toussardon T. 2001, Implementation of GIS application for French fire-fighters in the Mediterranean area. Computers, Environment and Urban Systems 25, 307-318.

Tillander, K. 2004. Utilisation of statistics to assess fire risks in buildings. Phd Thesis. Helsinki University of Technology. VTT publication 537, 265 p.

15

Vanderhaegen H, Van Hecke E., Juchtmans G. 1996. Les régions urbaines belges en 1991. Etudes Statistiques, 104, 15p

Yang L., Jones B., Yang S. 2006. A fuzzy multi-objective programming for optimization fo fire station locations through genetic algorithms. European Journal of Operational Research (in press)

APPENDIX

Population density in Belgium (Data Source : I.N.S., 1 January 2002)