Edinburgh Explorer Towards the development of a probabilistic approach to informal settlement fire spread using ignition modelling and spatial metrics

: Large conflagrations of informal settlements occur regularly leaving thousands of people 15 homeless daily and taking tens of thousands of lives annually. Over the past few years a large 16 amount of data has been collected from a number of full-scale informal settlement fire experiments. 17 This paper uses that data with a semi-probabilistic fire model previously proposed by the authors, 18 to illustrate the potential applications of the fire spread method proposed. The current model is 19 benchmarked against a 20 dwelling full-scale informal settlement fire experiment, and the effects of: 20 a) the ignition criteria; b) wind direction; and c) wind speeds, on the predicted fire spread rates are 21 investigated through the use of a parametric study. Colour maps of the fire spread rates and patterns 22 are then used to visually interpret the effects of different types of fire scenarios and fire breaks. 23 Finally, the fire spread capability within B-RISK is used to derive a linear equation for the potential 24 fire spread rate as a function of the settlement spatial metrics (e.g. density and distance to nearest 25 neighbour). To further illustrate the potential application of this work, the fire spread rate equation 26 is then applied across the whole of Cape Town, South Africa to show the 10 informal settlement 27 areas most at ‘risk’ of large conflagrations.


32
Informal settlements, also known as shantytowns or slums, are settlements that are typically not 33 formally planned and consist of makeshift structures built on land that has not been designated for 34 residential use. These structures, more commonly known as shanties, shacks or informal settlement 35 dwellings (ISDs), are typically built from materials that are immediately available in the inhabitants' 36 surroundings, many of which are combustible. Informal settlements are extremely vulnerable to large 37 conflagrations as a result of these combustible structures coupled with the close proximity at which 38 these dwellings are built and prevailing weather conditions.

39
In South Africa alone there are more than 5000 ISD fires per annum, and the number of fires are 40 increasing annually [1]. According to the World Health Organization (WHO), fires cause 41 approximately 180,000 deaths globally per annum, with the majority of those deaths and associated

With permission from Ryan Heydenrych
The study of informal settlement fires is a relatively new research field. Previous research has 48 set out to better understand ISD enclosure fire dynamics (individual scale) and informal settlement 49 fire dynamics (macro scale). A number of large-scale ISD experiments have been conducted [5][6][7][8][9], 50 ranging from single dwellings to 20 dwellings in a single burn. In previous work, simulations using 51 Fire Dynamics Simulator (FDS) have been undertaken to demonstrate the software's ability to predict 52 the fire behaviour of single dwelling fires [7]. However, these comprehensive simulations took weeks 53 to run on the High Performance Computer of Stellenbosch University, which made it impractical to 54 run scenarios consisting of multiple dwellings. Cicione et al. [6] proposed some simplifications that 55 were incorporated into those FDS simulations, which significantly reduced the computational time 56 needed to run the multiple dwelling cases. However, the simplified simulations were found to be 57 extremely sensitive to input parameters and, although the simplifications reduced the computational 58 time neededrequirements, the time needed to simulate entire settlement scenarios would still be 59 impractical.

60
As an alternative, Cicione et al. [10] have developed a preliminary semi-probabilistic model of 61 informal settlement fire spread using B-RISK (a two-zone fire modelling software tool). The aim was 62 to take the first step towards developing a tool that could assist authorities of countries with large 63 informal settlements to provide predictive capabilities that can help in identifying high risk areas or 64 quantify the magnitude of an incident to which municipalities may need to respond. The semi-Fire 2020, 3, x FOR PEER REVIEW 3 of 32 not only be randomly selected based on floor area (as done by Cicione et al. [10]), but also based on 69 the cladding/lining material (as discussed in this paper) and their expected heat release rates.

70
Using spatial analysis with Geographic Information Systems (GIS), the layout of informal 71 settlements and the spatial arrangement of individual dwellings relative to each other (referred to as 72 spatial metrics) have been postulated to be indicative of fire spread risk. Identified fire spread risk 73 spatial metrics can then be applied to settlements so that those most at risk of fire spread can be 74 identified. For example, Gibson et al. [11] used burn areas identified from satellite imagery to 75 empirically obtain spatial metric values of settlements from their dwellings within the burn areas.

76
Settlements with similar spatial metric values were then identified within a broader environment and 77 were postulated to be at a high risk of fire spread. This approach relies on threshold values (75 th 78 percentile values of spatial metrics found in the burn areas) to identify either settlements which are 79 at higher risk of fire spread or those which are not. This binary approach is simplistic, where in reality 80 all settlements are at some risk of fire spread and thus a more nuanced, fire science-based approach, 81 is needed.

82
It is with this backdrop that this paper seeks to:  result of radiation from either one or more burning items or from the hot gas layer within the 104 enclosure. This section gives a brief review of the radiation and ignition submodels employed in B-

105
RISK but for more information regarding the model, the reader should refer to the user guide and 106 technical manual [13]. The radiation heat transfer method employed by B-RISK has been studied in-

107
Fire 2020, 3, x FOR PEER REVIEW 4 of 32 depth and has been found to be a suitable method for a variety of cases. Sazegara et al. [14] 108 benchmarked the single item ignition prediction capability of B-RISK using results from the furniture 109 calorimeter against room-size experiments. The method has also showed promise in other fields e.g.

112
In this work, the item-to-item submodel of B-RISK is used to simulate fire spread between ISDs, 113 which is a novel application for which the software was never originally designed for. To simulate 114 spread between ISDs in B-RISK, the dwellings are simplified to items (as in ref.
[10]) and treated as 115 being 'outside', with the settlement being simplified to a 'room' that is fully open (i.e. a room with 5 116 vents the size of the room boundaries to allow all the hot gases to escape to the 'outside'). This 117 effectively removes the 'zone' element from the zone model, but by keeping the radiation and ignition 118 submodels, which is a convenient means of using these submodels rather than recreating them from 119 scratch as a standalone tool. In this casepaper, the same approach is followed. Hence, there will be 120 no hot layer build up and the focus will be on item-to-item ignition (in other words, ISD-to-ISD fire 121 spread

127
where ̇ is the radiant heat flux, measured in kW/m 2 , received by the target item from the flaming 128 burning item; ̇ is the total heat release rate, measured in kW, of the burning item; is the 129 radiative fraction; is the angle between the radial distance (R) and an imaginary line parallel to the 130 floor where R intersects with the target item, as depicted in Figure 3; and R is the radial distance, 131 measured in metres, from the centre of the flaming region of the burning item to the nearest point of 132 the target item. Figure 3 depicts the geometry assumed in this paper and also visually illustrates the 133 variables used in Equation 1. In the B-RISK implementation R will always be the plan view distance 134 so that theta will be zero.

210
Using Equation 6 and the cone calorimeter data, Figure 6 has been constructed where the FTP, n and

211
̇ values for a number of these common lining and cladding materials used in informal settlements 212 have been obtained, and are presented in Table 1.

264
Using the crib model discussed by Babrauskas [27], it was determined that the crib mass loss 265 rate in these dwellings were most likely fuel surface area-controlled. Using the heat of combustion as

268
HRR curve depicted in Figure 9 is obtained. Although the dwellings clad with timber planks will 269 have higher HRR (since the timber planks will contribute to the total fuel load and the total HRR),

270
the initial growth period of timber clad dwellings are assumed to be unaffected by the timber planks 271 (controlled by the cribs) and since the timber planks are thin (12 mm thick) it is assumed that it will 272 burn away rapidly after the planks start burning [6,7]. Hence for simplicity, it was decided to assign 273 the HRR curve depicted in Figure 9 to all dwellings for baseline simulation. However, to investigate 274 the sensitivity of the HRR curve of the timber dwellings, three parametric simulations were run, as 275 discussed below. The HRR values in the curve depicted in Figure 9 were increased by 20%, 50% and 276 100% (i.e. the fuel load contribution of the timber planks have been used to increase the area under 277 the HRR curve [7]), respectively. It was found that when the timber dwellings had HRR values 50% 278 greater than the steel dwellings (Figure 9), the predicted spread rates are closer to the experimental 279 spread rates, as depicted in Figure 10.

293
A soot yield of 0.015 g/g, CO2 of 1.33 g/g and radiant loss fraction of 0.3 were taken from Table   294 3-4.14 of the SFPE Handbook [20]. The heat of gasification (1.8 kJ/g) was selected from

301
The results of the 20 dwelling experiment and B-RISK simulations are depicted in Figure 10. For

347
Comparing the spread rates of Curtain 1 and Curtain 2, it seems that the CHF has a greater effect on       Figure 14, it can clearly be seen that, although a 3.5 m separation 431 (i.e., the fire break) did reduce the fire spread rate compared to the no fire break case as depicted in 432 Figure 14, the fire was still able to spread between columns 2 and 3 ( Figure 7). However, it should be 433 noted that piloted ignition has been assumed in the ignition submodel, but with a 3.5 m separation 434 between dwellings it is less likely that flame impingement will occur. On the other hand, increasing 435 the fire break from 3.5 m to 4.5 m we simulated clearly see that fire spread now diddoes not occur 436 and that the fire was is contained to only one half of the mock settlement. Running the simulation for 437 different separation distances, the minimum distance at which fire spread did not occur was 4.2 m.

438
Thus, these B-RISK simulations indicate that when the effects of wind and multiple dwellings 439 burning at the same time are accounted for, a separation distance of 3.5 m is not sufficient, but rather 440 a distance of at least 4.2 m is needed. It is however acknowledged that such a large separation distance 441 is not always possible in reality as a result of socio-economic issues and insufficient spatial planning.

442
Additionally, it should be noted that for higher wind speeds and different wind directions this critical 443 distance might change, however these factors could be captured by using simulation tools such as B-

444
RISK. Also, branding was not accounted for in this work, which could also significantly affect the 445 critical separation distance.

446
The colour maps illustrate the first step towards producing risk maps for informal settlements 447 using B-RISK, which may be a useful tool for fire brigades and local municipalities. In an ideal version 448 of the software, the user would be able to import settlement geometry from a GIS file and run limitless 449 iterations, by (1) Figure 16 illustrates an example of a settlement layout with

494
Density and close proximity of dwellings has been stated as a cause for rapid fire spread in 495 informal settlements [25]. By analysing the density of dwellings together with the average distance 496 to nearest neighbours (NN1…NN5), a more nuanced understanding of the settlement layout and its 497 impact on fire spread can be obtained. For example, if a settlement has a low average distance to NN1 498 but high average distance to NN2…NN5, it implies that fire will more likely spread from the ignited 499 dwelling to NN1 in a stepwise manner, and the fire is more likely to spread in only single directions 500 (i.e. since the distance to NN2…NN5 might be far enough for spread to those neighbours not to 501 occur). However, if a low average distance to NN1...NN5 is discovered, the spread will be radial as 502 an ignited dwelling will be able to spread to more neighbours more easily. Through analysing these 503 spatial metrics together with fire spread rates, it will become apparent which of these metrics are the 504 most influential. For example, it may be that the density metric captures the information contained 505 in the average distance to NN1…NN5 in which case for future studies, distance to NN will not be 506 required, streamlining the processing.

507
The importance of edge density has been raised here since dwellings are ignited and spread from 508 their edges. The logic therefore follows that settlements with a high edge density (i.e. many longer, 509 thinner homes) offer more opportunities for fire to spread than settlements with low edge density.

510
The two previous papers by Gibson et al. which investigated this revealed some correlation and 511 therefore the role of this spatial metric is further explored here to determine its importance.

513
Fire 2020, 3, x FOR PEER REVIEW 20 of 32 Table 2. Spatial metrics calculated for the settlement shown in Figure 16. To determine which spatial metrics are the most influential for informal settlement fire spread, 517 the radiation and ignition submodels of B-RISK are used (discussed in Section 3) to predict fire spread 518 rates for a variety of randomly populated 'informal settlement' configurations. From these, the 519 average spread rates (i.e. depending on which dwelling ignited first in the populated scenario) are 520 have been obtained and the spatial metrics of the corresponding settlement scenario are calculated.

521
In this case, 25 different settlement configurations, consisting of 20 dwellings each (which were the 522 same as the baseline dwellings used in Section 3), were randomly populated (i.e. the location of the 523 dwellings were randomly populated). Each settlement scenario thus had a different dwelling layout 524 configuration resulting in different spatial metrics values, an example of which can be seen in Figure   525 16 and Table 2.

602
Thus, the NN1+3 spatial metrics place a greater role in excluding PFAs from analysis than density. It

603
can be seen in Figure 17 that where the spatial metrics of PFA overlapped with those of the B-RISK 604 scenarios (shown as Reduced PFAs in the figure), the fire spread rate was is highest. This implies that 605 Equation 11 predicts high spread rates (greater than 2500 m 2 /h) more reliably than low spread rates 606 across all PFAs since the equation was developed using scenarios which predict a higher spread rate 607 and those PFAs which likely (but this is yet to be proven) have a lower spread rate were not used in

614
To consider if PFAs with high fire spread rates are in fact affected by large fires in reality, fire 615 spread rates were are obtained for burn areas which were previously mapped from satellite imagery 616 [35]. These burn areas were are assigned fire spread rates by spatially overlaying the Reduced PFAs 617 with the burn areas and assigning the fire spread rate from the PFA to the overlapping burn area.
618 Figure 18 reveals that the burn areas are more likely to be found in PFAs with higher fire spread rates 619 implying that the fire spread rate equation is correct to some degree but since fire spread rates are 620 not known for the mapped fires, this can be considered a qualitative rather than quantitative 621 agreement.

625
The PFAs with the top ten (out of 127 PFAs) fire spread rates across all informal settlements 626 within the City of Cape Town are given in Table 3 and the location of the PFAs is shown in Figure   627 19.

628
629 Figure 19.  dwellings, and radial fire spread [m 2 /h] is assumed, as the fire grows in a larger settlement, the fire 652 spread rate will increase. Furthermore, the shape of a settlement will play a role too, since as radial 653 fire spread is assumed, once the fire front reaches the boundary of a settlment, the fire spread rate 654 will change from 'radial' to linear along the length of the settlement's boundary. In a settlement which 655 has a high perimeter to area ratio, the fire will reach an edge beyond which the fire can no longer   There have been a number of assumptions and simplifications made throughout this paper and 691 these have been highlighted throughout the paper. However, the hope is that the methodologies 692 developed in this paper would ultimately be of use for real settlements as a useful tool for fire fighters 693 and local municipalities. In order to achieve this, it is important that future work refines the 694 methodology by developing more robust methods for the assumptions made. As more data becomes 695 available from informal settlement dwelling experiments and from real fire incidents, the method 696 discussed in this work can be calibrated and updated to account for more variables. Before B-RISK 697 can be used in practice to simulate informal settlement fire spread rates and to determine settlements 698 at risk, the following are some considerations that need to be implemented or investigated in future  effect on the rate of fire spread and can increase the fire spread rate by more than 90%.

726
The paper then takes the next step in developing a tool to identify settlements and areas in 727 settlements most at risk, by post-processing the B-RISK output data to generate colour maps of the 728 linear fire line progression rates and spread patterns. Colour maps of the 20 dwelling experiment and 729 parametric simulations have been created showing that for fire spread not to occur, a critical 730 separation distance of around 4.2 m between dwellings is necessary, based on these simulations and 731 the parameters used. This is larger than the previously proposed separation distance of 3.8 m, because 732 the wind effect and the influence of multiple dwellings burning at the same time were not previously 733 considered, but are accounted for in this work. A next step to this work would be to provide colour 734 maps (risk maps) for large informal settlements to determine which settlements are most at risk and 735 also to identify 'hot spots' within settlements.

736
The use of B-RISK to produce a fire spread rate equation using spatial metrics has been 737 demonstrated. A total of 500 simulations using 25 settlement scenarios were run in B-RISK and 738 average fire spread rates were calculated. Analysis of spatial metrics calculated for each scenario 739 reveal that settlement density and the average distance to the first nearest neighbour plus the distance 740 to the third nearest neighbour are the most influential spatial metric in predicting fire spread rate.