Real-ime Sochasic Evacuaion Models for Decision Suppor in Acual Emergencies ARTURO CUESTA, DANIEL ALVEAR, ORLANDO ABREU and DELFÍN SILIÓ Transpors and echnology projecs and processes Universiy of Canabria Los casros s/n CP 39005, Sanander, Spain ABSTRACT This paper inroduces and proposes he use of evacuaion models for decision suppor during acual emergencies. Two examples are presened: EvacTrain 2.0 and EvacTunnel. The proposed models are essenially sochasic, quick and easy o use and can generae and process resuls of several simulaions wihin a few seconds. The main oupu parameer is he percenile (0.90, 0.95 or 0.99 h) of oal evacuaion imes. They also provide oher saisical characerisics and addiional oupus. Boh models have been compared wih oher validaed evacuaion models. Resuls sugges ha he proposed models provide consisen and reliable resuls. The general findings described in his paper sugges ha i is possible o develop efficien evacuaion models for supporing emergency decisions in real-ime. KEYWORDS: modelling, real-ime, Mone Carlo mehods, fire safey managemen, human behaviour INTRODUCTION In mos fire incidens a successful evacuaion and rescue can mean he difference beween life and deah. Therefore, here is a need o predic hese processes. Currenly, evacuaion calculaions are becoming a par of fire safey science. In some cases, hand calculaions are performed, and in ohers, modelling and simulaion are being used [1]. In he lieraure, here are some evacuaion models reviews. According o Gwynne [2], evacuaion models fall ino wo caegories: hose which only consider human movemen and hose which aemp o link movemen wih behaviour. The firs caegory of evacuaion models do no ake ino accoun he psychological aribues of people involved in he emergency. The second caegory of evacuaion models are more realisic and ake ino accoun he individual behaviour (personal reacion imes, exi preference, ec.). According o Tabares [3], he evacuaion models are classified ino hree groups: macroscopic approach, microscopic approach and effec-based simulaion approach. In he macroscopic approach, he occupans are modelled as a homogeneous populaion. In he microscopic approach, occupans are modelled as heerogeneous populaion (agen based, cellular auomaa, ec.). In he effec-based approach, he aciviy based models and hose models ha incorporae social scienific processes are included. These ools have grown more sophisicaed over he ime and have been mainly used for performance-based assessmens and/or forensic analysis. However, here is an expansion of applicaion opporuniies for he evacuaion modelling. Gwynne and Kuligowski presen a classificaion of he differen applicaion modes: Naïve, Operaional, Predicive, Engineered, Ineracive and Real-ime [4]. These applicaion modes require differen levels of daa and user experise. In he real-ime mode, he user can acquire feedback from he model during an acual emergency. This requires inpus from he siuaion o he model which should run significanly faser han real-ime. Some evacuaion models can run in realime [5-7]. Bu, o he auhors knowledge, hese models have no been developed specifically for decision suppor in case of emergency. This paper presens wo evacuaion models developed for decision suppor during acual emergencies. One of he main problems in developing real-ime evacuaion models is ha hey are likely o be less sophisicaed and produce limied informaion due o ime consrains. The challenge is o obain equilibrium beween run imes and providing enough deail in he model o allow sufficien accuracy. To address his, Mone Carlo mehods can be used varying he key random parameers in order o capure all he possible siuaions in which a given scenario migh be evacuaed. This means generaing represenaive and significan samples of oupus variables. The real-ime applicaions require processing he oupus quickly enough and he informaion provided need o be easy o inerpre and wih a high confidence level. 1063
For insance, given a scenario, he main oupu parameer may be he percenile (90, 95 or 99 h) of oal evacuaion imes. The evacuaion models presened in his paper operae in he manner described above. EVACTRAIN 2.0 Moivaion Fire incidens inside passenger rains can consiue a significan risk o life. The rain crew are responsible for passenger safey during an on-board fire emergency. The firs prioriy is o direc passengers away from fire and inform he rain operaions cenre abou he siuaion. The second prioriy is o deermine when and how perform he evacuaion. For insance, wheher o reach an appropriae place for he evacuaion (i.e., he closes saion, plaform) or sop he rain as soon as possible and perform evacuaion o he racks. Furhermore, he crew needs o know he number of passengers on-board, he dangers presen inside and ouside he rain, a safe area where passengers should be moved and which doors should be opened. Therefore, i is no easy o make he correc decisions. The use of compuer modelling analyses in real-ime could improve such decisions under a variey of emergency condiions. Overview of he model EvacTrain 2.0 is an objec-oriened evacuaion model developed by GIDAI Group. The purpose of he model is o simulae differen evacuaion sraegies in rains. The rain spaces are represened by a coarse nework. Each node represens a passenger coach wih exis. Due o he fac ha EvacTrain 2.0 is a model ailored o he inended rain, daa relaed o he rain characerisics are included by defaul in he model. The model considers he following basic scenarios: 1. Emergency evacuaion o plaform (unnel). 2. Emergency evacuaion o rack level. 3. Evacuaion o plaform. 4. Evacuaion o rack level. 5. Evacuaion o oher rain. Evacuaion scenarios 3-5 are no considered as emergency siuaions. Therefore, hey are excluded from he presen descripion. I is assumed ha passengers are ready o sar evacuaion once he rain sops. However, he model includes he preparaion ime. This is a random variable defined as he ime elapsed from he sop of he rain unil passengers sar evacuaion. The values vary according o he emergency evacuaion scenario. In evacuaion o plaform, his variable is he ime spen in opening he rain doors. In evacuaion o rack level, his variable represens he ime spen o se up he porable ladders or ramps. EvacTrain 2.0 focuses on he simulaion of he flow hrough he available exis. The model considers he flow as a random variable individually assigned o each passenger from normal disribuions. The values of he disribuions vary according o he differen exi condiions (o plaform, direcly o rack and hrough emergency ladders). The values used by defaul in he model are obained from [8-10]. I should be noed ha he model is flexible and allows he user o modify hese daa. Some exis may be unavailable in cases of evacuaion. The user has he opion o block he exis (or o se up he available exis for evacuaion) in order o reproduce he real siuaion or o explore he poenial oucomes of an evacuaion sraegy. In hese cases, he model simulaes he relocaion process providing a realisic disribuion of he number of passengers a each exi. Then he criical exi is defined as he exi used by he maximum number of passengers. The model performs 250 ieraions by defaul. The evacuaion ime of i-h ieraion is given by: ( k) ( k) e = p + cri i = 1, n (1) i i i ier Where: k - The evacuaion scenario (emergency evacuaions o plaform or o rack level); (k ) - Preparaion ime for he k evacuaion scenario and i-h ieraion; p i 1064
(k ) - Criical exi ime: cri i m pas ( k ) 1 cri = i ( k ) j= 1 f j (2) Where: m pas - Number of passengers ha use he criical exi; (k ) f - Random flow assigned o j-h passenger according o he evacuaion condiions (k). j Due o is inended use, he presened model is quick and easy o se-up. The inpu parameers consis of: Fracion of he full load: value beween 0 and 1 known by he rain crew. Type of inciden: 1) fire, 2) collision, 3) derailmen, 4) echnical failure. Evacuaion desinaion: 1) plaform, 2) rack level or 3) oher rain (only for ransfers). Number of available exis. EvacTrain 2.0 generaes and process resuls in a few seconds. The oupus produced by he model are displayed a he screen and hey can be saved in x files as well. The model saisically reas he sample of oal evacuaion imes and fis i o a known disribuion (if possible). Oherwise, densiy esimaions are given. The main oupu parameer is a percenile of egress imes (0.90, 0.95 and 0.99 h). I also provides oher saisical characerisics: mean, variance, maximum and minimum values. Addiional oupus include he number of available exis, he criical exi and he number of passengers ha use i. Comparison wih STEPS In his secion we describe a deailed comparison beween EvacTrain 2.0 and STEPS model [11]. As explained above, his is a model ailored o he inended rain. The comparison presened here is performed for a high speed rain S 102. This is a rain 200 m lengh wih 11 passenger coaches (+1 lounge) and capaciy for 316 passengers. Two ses of evacuaion scenarios are considered: emergency evacuaion o plaform and emergency evacuaion o rack level. Figure 1 shows he emergency evacuaion scenarios o plaform. Scenario 0 11 10 09 08 07 06 Lounge 05 04 03 02 01 Scenario 3 11 10 09 08 07 06 Lounge 05 04 03 02 01 Scenario 3p 11 10 09 08 07 06 Lounge 05 04 03 02 01 Scenario 22p 11 10 09 08 07 06 Lounge 05 04 03 02 01 Passenger coaches affeced or poenially affeced by he fire Fig. 1. Emergency evacuaion scenarios o plaform. 1065
In emergency evacuaion scenarios o plaform, wo consecuive dynamic processes have o be simulaed: 1) evacuaing passengers o a place of relaive safey along he rain (pre-evacuaion aciviies) and 2) evacuaing from he rain. EvacTrain 2.0 simulaes he relocaion process o provide a realisic disribuion of he number of passengers a each exi. On he oher hand, STEPS allows changing he availabiliy of cerain exis during he course of he simulaion by using exi evens. Using his feaure, he user can open, close or make exis unavailable. When an exi is se o closed, he agens will sill consider he exi when choosing heir arge and form a queue in fron of i. When he exi is unavailable, i is considered o be no longer usable, and nobody moves owards ha exi. Table 1 displays he inpus considered for he comparison of he emergency evacuaion scenarios o plaform. The values are obained from an announced evacuaion drill [8]. Table 1. Inpus for he comparison of emergency evacuaion o plaform. Inpus EvacTrain 2.0 STEPS N of passengers 316 316 Flow hrough he exis (per/s) 0.44±0.20 0.44 Walking speed (m/s) No considered 0.99±0.20 Time o open he doors (s) 35 35 Scenario 1b and Scenario 2b in Figure 2 represen emergencies in which passengers have o evacuae o he rack level by using 1 and 2 emergency ladders respecively. Table 2 displays he inpus considered for he comparison of he emergency evacuaion scenarios o rack level. The emergency ladders are 3 m long and consis of wo separae pars ha have o be assembled. These evacuaion elemens hold wo passengers simulaneously. I is considered an average ime of 3 s spen by each passenger o negoiae he emergency ladder (flow of 0.33 per/s). This value is derived from Volpe Cener egress rials o rack level [12, 13]. The preparaion ime is a random parameer in EvacTrain 2.0. However, i is se as a consan value of 300 s o represen he same condiions in boh models. Scenario 1b 11 10 09 08 07 06 Lounge 05 04 03 02 01 Scenario 2b 11 10 09 08 07 06 Lounge 05 04 03 02 01 Fig. 2. Emergency evacuaion scenarios o rack level rough emergency ladders. Table 2. Inpus for he comparison of evacuaion o rack level. Inpus EvacTrain 2.0 STEPS Occupaion 100% 316 316 Flow hrough he exis (per/s) 0.33±0.10 0.33 Walking speed (m/s) No considered (0.99±0.20)*0.80 Time o insall he emergency ladder (s) 300 300 * Facor for he walking speed on he emergency ladder As menioned previously, in EvacTrain, passengers are ready for evacuaion once he rain sops. The model focuses on he simulaion of exi performance and he flow is represened as a random ime spen by 1066
each passenger o negoiae he exi. Therefore, as Tables 1 and 2 show, walking speed is no included as an inpu variable in he proposed model. Each scenario was run 100 imes. I is assumed ha he passengers are complian wih he rain crew s commands. This assumpion is a basic requiremen for seeing he effecs of he procedures ha are implemened. For he evacuaion o he rack level, no alernaive escape roues, such as oher exi doors where he passengers have o climb higher han 1.1 m, are considered. Figures 3 and 4 show he cumulaive disribuion funcions of oal evacuaion imes. The saisical characerisics are shown in Tables 3 and 4. I is possible o see a wider variabiliy on he sample provided by EvacTrain 2.0 in Scenario 1b (see in Figure 4). This is because of he random flow in he single exi simulaed by he proposed model, no reproduced by STEPS model which uses a consan value. However, here is a good agreemen beween boh models. This is quanified by he Percen Error (PE) of he mean (PEM) and he 95 h percenile (PEP) of oal evacuaion imes. Noe ha he resuls of he proposed model are considered as he approximae values while he resuls of model of he comparison are considered as he acual values. In he emergency evacuaion o plaform boh he PEM and PEP are lower han 3%. Furhermore, in he emergency evacuaion o rack level scenarios, he PEM and PEP are lower han 1.3 %. Based on hese resuls, i can be argued ha EvacTrain is capable of producing reliable predicions of oal evacuaion imes. Probabiliy Probabiliy Scenario 0 1,00 STEPS 5.0 0,80 EvacTrain2.0 0,60 0,40 0,20 0,00 100 110 120 130 140 150 160 Evacua/on /me [s] Scenario 3p 1,00 0,80 0,60 0,40 STEPS 5.0 0,20 EvacTrain 2.0 0,00 130 140 150 160 170 180 190 Scenario 3 1,00 STEPS 5.0 0,80 EvacTrain 2.0 0,60 0,40 0,20 0,00 170 180 190 200 210 220 230 Evacua/on /me [s] Evacua/on /me [s] Fig. 3. Cumulaive disribuion funcions of he oal evacuaion imes o plaform. Table 3. Disribuions of he oal evacuaion imes o plaform (s). Scenario EvacTrain 2.0 STEPS 5.0 Mean S.D. Range Perc. 95h Mean S.D. Range Perc. 95h Scenario 0 132 5.8 116-151 143 131 3.2 126-142 139 Scenario 3 206 6.2 183-220 217 207 5.4 195-222 214 Scenario 3p 150 6.8 130-171 160 154 5.3 144-171 161 Scenario 22p 190 6.2 169-206 200 194 4.7 186-212 205 Probabiliy Probabiliy Evacua/on /me [s] Scenario 22p 1,00 0,80 0,60 0,40 STEPS 5.0 0,20 EvacTrain 2.0 0,00 170 180 190 200 210 220 230 1067
Probabiliy Scenario 1b 1,00 0,80 0,60 0,40 0,20 EvacTrain 2.0 STEPS 5.0 0,00 880 900 920 940 960 980 1000 Evacua/on /me [s] Evacua/on /me [s] Fig. 4. Cumulaive disribuion funcions of oal evacuaion imes o rack level. Table 4. Disribuions of he oal evacuaion imes o rack level (s). Scenario EvacTrain 2.0 STEPS 5.0 Mean S.D. Range Perc. 95h Mean S.D. Range Perc. 95h Scenario 1b 927 15.3 878-970 950 929 9.5 922-984 939 Scenario 2b 562 14.2 512-597 585 569 16.3 519-606 590 probabiliy 1,00 0,80 0,60 0,40 0,20 0,00 Scenario 2b EvacTrain 2.0 STEPS 5.0 520 540 560 580 600 620 640 EVACTUNNEL Moivaion Road unnels consiue dangerous environmens when a fire occurs. Pas disasers have shown he need for an effecive emergency response and he ragic consequences of incorrec or delayed decision making [14]. The unnel operaor is he firs person o deal wih he emergency. He/she deecs he emergency mainly hrough he Auomaic Inciden Deecion (AID) sysem and/or CCTV. Bu, he informaion may be sparse, incomplee and inaccurae and he/she will be required o make decisions such as closing he unnel, declaring he evacuaion, ec. In many cases, hese decisions are based on fixed proocols and can be made oo lae. The use of predicive ools o suppor he operaor decisions in real-ime could improve road unnel safey. In his sense, he GIDAI Group has developed a Decision Suppor Sysem (DSS) for emergency managemen in road unnels [15]. The DSS analyses he curren siuaion and guides he course of decisions o deal wih he emergency. Furhermore, he sysem provides real-ime esimaion of he severiy of he acciden and he required evacuaion and rescue imes by he evacuaion model inegraed in he sysem: EvacTunnel. Overview of he model EvacTunnel is an objec-oriened evacuaion model developed by GIDAI Group [16]. The purpose of he model is o simulae he evacuaion and rescue processes in road unnels. As Figure 5 shows, he presened model considers wo areas inside he unnel. The Area 1 includes he vehicles and he people direcly involved in he acciden, where i is likely o find injured people who canno evacuae by hemselves (rescue process). The Area 2 includes he vehicles and people rapped inside he unnel no direcly affeced by he acciden. This people can leave he unnel by hemselves (self-evacuaion process). The model calculaes boh scenarios separaely. Area 2 Area 1 Fig. 5. Areas inside he unnel considered in he model. The model allows performing several simulaions (a minimum of 100 runs). In each simulaion, he model regisers he evacuaion ime for all unnel users and considers he ime when he las one leaves he unnel. 1068
This is calculaed for Area 1 and Area 2 separaely. The evacuaion ime of each unnel user depends upon he pre-movemen ime, unresriced walking speed and he disance hrough he escape roues: Where: e i d e + v - Evacuaion ime for he i-h person; pm i - Pre-movemen ime for he i-h person; d v mov i mov i - Disance o he exi for he i-h person; - Walking speed for he i-h person. mov i = i pm (3) i mov i For people in Area 1, who canno evacuae by hemselves (assised mobiliy), he pre-movemen ime (!"! ) is calculaed by he following expression: = + + + ( MA) pm no reac mov exam (4) Where: MA - Assised Mobiliy. People who canno evacuae by hemselves; - Delay ime o inform he emergency services; no reac mov exam - Reacion ime of emergency services; - Travelling ime o arrive o he scene ; - Time o examine and prepare he users affeced by he acciden. In Area 2, he pre-movemen ime (!"! ) of he people rapped inside he unnel can be divided ino wo phases: (!) 1) Recogniion Phase (!"!" ).- The ime required o undersand wha has happened.! 2) Response Phase (!"!" ).- The ime spen o leave he vehicle and sar evacuaion movemen. The model implemens he pre-movemen ime by using he crierion of disance from he acciden. Figure (!) 6 shows he Recogniion Phase!"!" as linearly dependen wih he disance respec o he acciden zone. The model calculaes he Recogniion Phase considering he ime needed by he persons nex o he acciden area o reach differen locaions during heir movemen owards he exi wih a speed of 1.55 m/s. This walking speed can be changed by he user of he model. 1069
(1) pmij! d!! + v!"!!"#!!"!! d!! d movij (!) Fig. 6. Linearly dependence of!"!" wih he disance. Therefore he Recogniion Phase of unnel users is: d' I d ( 1) 1 mov pm = ij pm + q vmov q ij (5) Where: pm q - Pre-movemen ime of he firs person o respond near o he acciden (consan value); d - Travelling disance of Area 2; d ' I1 mov ij vmov q - Travelling disance o each unnel user; - Unimpeded walking speed of he firs person o respond near he acciden. (!) The heoreical disribuion for he Response Phase!"!" is derived from an experimen conduced a Universiy of Canabria, Spain. The ime spen by 32 paricipans o leave heir vehicles was measured. The paricipans spen beween 15 s and 120 s. The disribuion has a mean of 67.5 s and a sandard deviaion of 17.5 s. Defaul walking speeds are assigned from a normal disribuion wih a mean of 1.25 m/s and a sandard deviaion of 0.32 m/s. These values are derived from [17]. The real-ime mode requires a direc observaion from he real siuaion (i.e. hrough CCTV) and hen provides his informaion direcly o he evacuaion model. However, during he firs sages of he emergency, here is a high level of uncerainy regarding he number of vehicles and he number and disribuion of occupans. The number and characerisics of people involved in he acciden (Area 1) is prediced by he Incidens Model inegraed in he DSS [15]. In his firs version of he model, he unnel operaor inroduces an esimaion of he number of vehicles in Area 2. Oherwise, his informaion can be obained from he raffic couners in he unnel. The number of people inside he vehicles is a random variable beween a maximum and minimum value ha can be predefined by he user (i.e 1-5 occupans/car, 1-2 occupans/ruck and 20-40 occupans/bus). Therefore, in each simulaion he occupaion load is differen. Taking ino accoun ha all he variables lised above are random variables, heir generaion using Mone Carlo mehods can be represened analyically by using he inverse ransformaion mehod (Smirnov ransform) [18]. Alhough in paricular cases, oher algorihms can be used, such as he Box-Muller for normal and lognormal disribuions, or he numerical inegraion mehod for esimaing he disribuion funcion by a hisogram. 1070
EvacTunnel saisically reas he sample of oal evacuaion imes and fi i o a known disribuion (if possible). Oherwise, densiy esimaion is given using hisogram. The main oupu parameer is a percenile of evacuaion imes (0.90, 0.95 and 0.99). The model also provides oher saisical characerisics: mean, variance, maximum and minimum values. Addiional oupus include he number of people rapped inside he unnel a specific ime and locaion. Comparison wih STEPS, Pahfinder and GridFlow Here we describe he comparison analysis beween EvacTunnel and oher curren evacuaion models: STEPS [11], Pahfinder [19] and GridFlow [20]. The comparison is performed for he self-evacuaion of people rapped inside he unnel. Figure 7 shows he layou of he evacuaion scenario for he simulaions. I consiss of an acciden in he cener of he ube obsrucing he access o he cross passage. The evacuaion is modelled considering he momen in which he vehicles are sopped, queuing behind he vehicles involved in he acciden. I is assumed a oal of 54 vehicles rapped in he unnel: 49 ligh vehicles (cars) and 5 heavy vehicles (rucks). The occupaion load is assumed o be 1 person per heavy vehicle. For ligh vehicles a load facor of 2.32 is considered. Therefore, 119 occupans are considered for he simulaions. Evacuaion flow Emergency exi Enrance 262 m Fig. 7. Layou of he evacuaion scenario considered for simulaions. 10.25 m Two ess are considered. In Tes 1, no behaviour is performed in order o check ha he simulaion of movemen is working saisfacorily. In Tes 2 a behaviour comparison is performed. Table 5 displays he evacuaion imes obained in Tes 1. The evacuaion imes are very similar beween he models. The Percen Error (PE) is no higher han 1% beween he proposed model and he models of he comparison. The small differences are found due o he random disribuion of he occupans who are furher from he unnel poral (heir sar posiion). Resuls from Tes 1 show ha basic movemen componens of EvacTunnel work adequaely. Table 5. Resuls of Tes 1 (s): 1 run. Model Evacuaion ime Pahfinder 260 STEPS 257 GridFlow 258 EvacTunnel 262 Tes 2 provides an opporuniy o validae he proposed model agains oher evacuaion models. In his es he scenario has been run 100 imes o capure sochasic variaions in he resuls. The implemenaion of pre-evacuaion imes in GridFlow, STEPS and Pahfinder models has been done using he crierion of disance from he acciden as i is considered by he proposed model. In his way a phased response of he occupans has been considered. In order o implemen his, he ube was divided ino 13 zones (20 m lengh) wih differen populaion groups and pre-movemen ime disribuions. In GridFlow hese zones were implemened by differen spaces conneced by links (inle and oule). In STEPS his was done by using locaions on he plane. In Pahfinder recangular rooms were used. The pre-movemen imes have been assigned using normal disribuion laws. In Zone 1 i was assumed a pre-movemen ime disribuion wih a mean of 170 s and a sandard deviaion of 17.5s. Then, he mean value has been increased by 13 s per zone in order o reproduce he same domino effec applied by he proposed model. The same 1071
unimpeded walking speed disribuion has been assigned for all occupans in all models. This is a normal disribuion wih a mean value of 1.20 m/s and a sandard deviaion of 0.20 m/s. Figure 8 shows he cumulaive disribuion funcions of evacuaion imes and Table 6 displays a comparison of he mean, maximum, minimum and 95 h percenile of oal evacuaion imes obained by he models. The prediced evacuaion imes do no vary significanly among each model and heir curves are very similar. In his case he evacuaion ime was driven by he ineracions beween pre-movemen ime and he ravel disribuions. Table 7 shows he PEM and PEP when EvacTunnel is compared wih he oher models. The PEM is lower han 1.5 %. The maximum PEP is 5.62 % when he proposed model is compared wih Pahfinder. The resuls from Tes 2 show ha EvacTunnel is able o provide as reliable predicions as he models of he comparison. 1,00 0,80 probabiliy 0,60 0,40 0,20 STEPS Pahfinder 0,00 GridFlow EvacTunnel 400 450 500 550 600 650 700 750 Evacua/on /me [s] Fig. 8. Cumulaive disribuion funcions of oal evacuaion imes. Table 6. Resuls of Tes 2 (s): 100 runs. Pahfinder STEPS GridFlow EvacTunnel Mean 497 496 495 491 S.D. 31 50 42 44 Range 429-624 434-742 419-670 434-671 Perc. 95h 554 580 570 587 Table 7. Percen Errors of Tes 2 when EvacTunnel is compared wih oher models. Model PEM (%) PEP (%) Pahfinder 1.22 5.62 STEPS 1.02 1.19 GridFlow 0.81 2.90 DISCUSSION OF THE RESULTS Making decisions is cerainly he mos imporan ask of a safey manager (operaor) and i is ofen a very difficul one. The use of evacuaion models for decision suppor involves he following basic requiremens: 1) providing enough deail in he model o allow sufficien accuracy and 2) fas simulaion imes. 1072
Firsly, a good decision should no be suppored by an oucome alone. Deerminisic models are likely o produce an inaccurae represenaion of he evacuaion process as hey only consider one or a few poenial siuaions. This is due o he uncerainy relaed o he siuaion and he uncerainy relaed o he human behavior during evacuaion process. The use of Mone Carlo mehods permis he represenaion of all possible siuaions and he generaion of samples of oal evacuaion imes. The sochasic simulaions generae more reliable and consisen resuls. Furhermore, he use of disribuions of he oal evacuaion imes provides more powerful crieria for decision making. For insance, he use of perceniles (90, 95 or 99 h). Secondly, real-ime evacuaion models have o be a simplified represenaion of he acual siuaion. Therefore, hey should concenrae in he mos essenial parameers and ineracions and ignores he less essenial ones. The proposed models are based on he idea ha evacuaion calculaions can be performed by addressing a small se of random parameers ha have impac in he oucomes [21]. For he analysis of evacuaion process in passenger rains, we sugges ha he dominan parameers are he ime spen o prepare for evacuaion and he flow hrough he available exis. The firs parameer involves specific evacuaion procedures such insalling he emergency ladder. The second parameer depends on he evacuaion condiions defined by he number of passengers per available exi and he evacuaion desinaion (plaform, rack level). For he analysis of evacuaion process in road unnels, we sugges ha he dominan parameers are he pre-movemen imes, he walking speeds and he ravelling disances of individuals. Clearly, he resuls of he comparison sugges ha he addiional complexiy of he curren evacuaion models may no yield significanly differen resuls han he proposed models. In rain evacuaion scenarios, boh he PEM and he PEP are lower han 3% when EvacTrain is compared wih STEPS model. In road unnel evacuaion scenarios, he PEM is no higher han 1.22 % when EvacTunnel is compared wih oher hree evacuaion models (Pahfinder, STEPS and GridFlow). Furhermore, he PEP is no higher han he 5.62 % when EvacTunnel is compared wih Pahfinder. Therefore, i can be argued ha he presened models provide consisen and reasonable resuls. Tables 8 and 9 show a comparison of he simulaion imes and capabiliies beween he evacuaion models used for his sudy. The informaion provided is based on our experiences and esimaions for 100 runs, once he scenarios have been implemened (geomery, number of occupans, ec.). From he Tables 8 and 9 i is possible o see ha he proposed models can perform several simulaions and process he resuls saisically by hemselves in a few seconds. Table 8. Comparison of simulaion imes and capabiliies for rain evacuaion analyses. Model Bach run? Bach run ime Ploing Saisical (100 runs) daa? processing? EvacTrain 2.0 Yes <5 s Yes Yes STEPS Yes 70-130 s (Evacuaion o plaform) 210-390 s (Evacuaion o rack level) No No Table 9. Comparison of simulaion imes and capabiliies for road unnel evacuaion analyses. Model Bach run? Bach run ime Ploing Saisical (100 runs) daa? processing? EvacTunnel Yes <5 s Yes Yes STEPS Yes 400 s No No Pahfinder No >3600 s No No GridFlow Yes 403 s Yes No 1073
These real-ime evacuaion models are appropriae and accurae in specific siuaions. They can be used o make criical decisions during he firs sages of he emergency. In EvacTrain 2.0 he user can explore differen evacuaion processes and choose and appropriae evacuaion sraegy. For insance, if a fire is deeced in he running rain wheher o sop as soon as possible or reach a plaform for he evacuaion. EvacTunnel enables he user o overcome he uncerainy during he firs sages of he emergency. For insance, he unnel operaor can declare he evacuaion based on he model predicions, as a prevenive sraegy, or inform abou he required imes for evacuaion o emergency services before heir arrival o he scene. CONCLUSIONS From his work i is concluded ha evacuaion calculaions can be used for supporing imely decisions during acual emergencies. I is proposed he use of sochasic models wih he capabiliy o perform several simulaions by changing key random parameers o capure all poenial oucomes. These models should process he resuls by hemselves and provide informaion easy o inerpre for decision making. All his process should be performed wihin a few seconds. Two evacuaion models which operae in he manner describe above have been presened and parially validaed agains oher evacuaion models. Noe ha he evacuaion models presened here can also be used for oher applicaions such as performance-based assessmens and/or risk analysis. Mos of he inpu parameers included in he models are obained from empirical research. However, i should be noed ha he flexibiliy of he models allows he user o change hese values. Therefore, i is recommended as a good pracice o use reliable daa from rials and/or evacuaion drills in he scenarios where he evacuaion models are going o be implemened. The curren versions of he proposed models have limiaions and new challenges o be addressed. Fuure research will include furher validaion agains experimens and evacuaion drills for a se of new possible scenarios. ACKNOWLEDGMENTS The auhors would like o hank o he Spanish Minisry of Economy and Compeiiveness for he EVACTRAIN Projec gran, Ref.: BIA2011-26738, co financed by FEDER funds. 1074
REFERENCES [1] Kuligowski, E. D., Modeling Human Behavior During Building Fires, NIST Naional Insiue of Sandards and Technology, NIST Technical Noe 1619, USA, December, 2008. [2] Gwynne, S., e al., (1999) A Review of he Mehodologies Used in he Compuer Simulaion of Evacuaion from Building Environmen, Building and Environmen 34, 741-749. [3] Tabares, R. M., (2009) Evacuaion Process Versus Evacuaion Models: Quo Vadimus?, Fire Technology 45, 419-430. [4] Gwynne, S., and Kuligowski, E., Applicaion Modes of Egress Simulaions, Proceedings of he 4 h Inernaional Conference Pedesrian and Evacuaion Dynamics, Springer, 2008, pp.397-409. [5] Kisko, T.M, and Francis, R.L., (1985) Evacne+: a compuer program o deermine opimal evacuaion plans, Fire Safey Journal 9, 211-220. [6] Lin, Y., e al., Agen-Based Simulaion of Evacuaion: An Office Building Case Sudy, Proceedings of he 4 h Inernaional Conference Pedesrian and Evacuaion Dynamics, Springer, 2008, pp.345-357. [7] Yamashia, T., e al. Exhausive esing plan wih high speed evacuaion simulaor, Proceedings of he Inernaional Scienific Conference Emergency Evacuaion of People from Buildings, 2011, pp.357-363. [8] Capoe, J. A., Alvear, D. M., Abreu, O. V., Cuesa, A. and Alonso, V., (2012) A Sochasic Approach for Simulaion Human Behavior during Evacuaion Process in Passenger Trains, Fire Technology 48: 911-925. [9] Capoe, J. A., Alvear, D. M., Abreu, O. V. and Cuesa, A., (2012) Analysis of evacuaion procedures in high speed rains fires, Fire Safey Journal 49:35-46. [10] Norén, A., and Winér, J., Modelling Crowd Evacuaion from Road and Train Tunnels-Daa and design for faser evacuaions, Repor 5127, Deparmen of Fire Safey Engineering Lund Universiy, Sweden, 2003. [11] STEPS Simulaion of Transien and Pedesrian movemens: User Manual, unpublished, available wih egress model from Mo MacDonald. hp://www.momac.com. [12] Markos, S.H, and Pollard, J.K. Passenger Train Sysems: Single-Level Commuer Rail Car Egress Experimens, Prepared by Volpe Cener/USDOT for FRA/USDOT. Final Repor. In FRA repor approval process as of May 2013. [13] Galea, R.E., e al. The Developemen and Validaion of a Rail Car Evacuaion Model, Proceedings of he 13 h Inernaional Fire Science & Engineering Conference INTERFLAM 2013, Inerscience, 20013, pp.1013-1034. [14] Carvel, R., and Marlin, G., A hisory of fire incidens in unnels, The Handbook of Tunnel Fire Safey, Alan Beard and Richard Carvel, UK, 2005, p. 3-37. [15] Alvear, D., Abreu, O., Cuesa, A., and Alonso, V., (2013) Decision suppor sysem for emergency managemen: Road unnels, Tunnelling and Underground Space Technology 34, 13-21. [16] Capoe, A., Alvear, D., Abreu, O., Cuesa, A., and Alonso, V., (2013) A real-ime sochasic evacuaion model for Road unnels, Safey Science 52, 73-80. [17] Boyce, K.E., Shields, T.J, and Silcock, G.W.H., (1999) Toward he Characerizaion of Building Occupancies for Fire Safey Engineering: Prevalence, Type, and Mobiliy of Disabled People, Fire Technology, 35, 1, 35-50. [18] Rubinsein, R.Y., Kroese, D.P., Random Number, Random Variable and Sochasic Process Generaion. Simulaion and Mone Carlo Mehod, 2012, Wiley-Inerscience, pp. 49-80. 1075
[19] Thoron, C., O Konski, R. and Hardeman, B., Inroducing -: An Agen-Based Egress Simulaor, Proceedings of he Fourh Inernaional Symposium on Human Behaviour in Fire, UK, 2009, pp.567-572. [20] Bensilium, M. and Purser. D., GridFlow: an Objec-Oriened Building Evacuaion Model Combining Pre-movemen and Movemen Behaviours for Performance-Based Design, Proceedings of he Sevenh Inernaional Symposium on Fire Safey Science, USA, 2003, pp. 941-952. [21] Purser, D., Dependence of Modelled Evacuaion Times on Key Parameers and Ineracions, Proceedings of he Ninh Inernaional Symposium on Fire Safey Science, Germany, 2003, pp. 353-364. 1076