1 Introduction

Evacuation modelling is fast becoming a key instrument in fire safety. Several evacuation models have evolved briskly over the past decades while their capabilities [1,2,3,4,5], scrutiny [6,7,8,9,10] and validation [11,12,13,14,15] have become central issues in the literature. Moreover, there exists a continuous demand to extend the use and application of these tools [16, 17]. Therefore, empirical data and validation evidence for new scenarios are still of crucial importance.

A considerable amount of literature has been published on marine evacuations. Several studies provide datasets [18,19,20,21], models [22,23,24,25,26,27,28] and/or validated simulation tools [29,30,31] to estimate the time required for passengers and crew to get assembly locations and abandon passenger ships. Moreover, the International Maritime Organization (IMO) provides a guideline to this attempt [32]. However, the evacuation of workers onboard vessels in dockyards is a new challenging scenario. Ship repairing, and shipbuilding are inherently hazardous operations including “hot works” (e.g., welding, oxygen cutting) that require additional measures to prevent fires from heat, sparks, molten metal or direct contact with flame. These activities increase the range of emergency scenarios, comprising fires and spread of combustion products throughout the ship, leading to a complex evacuation process of workers. During the repair period a number of the ship systems may be out of operation. The crew may thus be exposed to unexpected and unfamiliar emergency situations. The provision of warnings (detections and the alarms) may vary between worksites and some workers may not immediately hear the alarm (or other means such as steam whistles) due to noise and/or ear protectors. Also, verbal warnings from fire wardens may not be effective due to language barriers. Furthermore, some workers may spend time shutting down critical operations (e.g., turning gas supply off, closing valves, etc.) before evacuating producing delays in the evacuation response. Vessels in dockyards are complex environments with enclosed or confined spaces (e.g., cargo holds, storage tanks, clearances, engine crankcases, casings, etc.), changing layouts due to working operations, temporary evacuation means (e.g., scaffolds, cat ladders, and manholes) not present in other evacuation scenarios, and often limited number and capacity of evacuation routes and exits (e.g., gangways). These factors are critical as they can have a major influence on the time required to perform a successful evacuation ensuring that all workers are able to reach a safe spot.

These problems have been identified by us based on our own knowledge and experiences. Note that this is a new scenario for evacuation modelling analysis and no previous references for this were found in the literature. To deal with these challenging conditions and to ensure a safe evacuation during shipbuilding and ship repair operations, decision-makers (e.g., competent authorities, employers, shipowners) often use general requirements, provisions and static evacuation plans [33,34,35]. Also some similar evacuation challenges were described for building construction sites [36,37,38].

Modelling and simulation can help to gain additional knowledge regarding the evacuation process and, therefore, improve life safety for these sites. One major issue concerns the appropriate evacuation model selection. There are three possible options: (1) select and use an existing evacuation model, (2) modify/adapt an existing evacuation model to represent the new scenario, or (3) develop a model specifically designed to this purpose [16]. We found examples of these options in the literature regarding building construction sites; relatively new (or not fully explored) evacuation scenarios with similar characteristics to vessels in shipyards. The first study uses existing models i.e. Pathfinder in combination with PyroSim and FDS (Fire Dynamics Simulator) to determine the RSET (Required Safe Escape Time) and the ASET (Available Safe Escape Time) from a prefabricated building construction site in case of fire [36]. This study employs deterministic inputs for evacuation and lacks model validation. The second study presents datasets on workers' evacuation performance and adapts a current evacuation model (BuildingEXODUS) to represent temporary vertical devices and travel speeds on different surfaces [37]. In this study, model validation was conducted. The third study proposes a conceptual agent-based simulation that specializes in high-rise building construction projects [38]. The proposed framework provides a profile of estimated evacuation times for the dynamically changing environment throughout the construction project. However, as reported by the authors further development efforts are needed including a prototype and performance evaluation. Consequently, no validation was conducted in this study. Whilst some research has been carried out on construction sites, to our knowledge, no single study exists which concentrates on the evacuation simulation of workers onboard vessels in shipyards.

The main objectives of this paper are to investigate the feasibility of simulating and predicting the evacuation of workers aboard vessels in dockyards using an existing model, and to validate the accuracy of the model by applying a validation protocol to the observed evacuation of 150 workers during an unannounced evacuation drill of a Ro-Pax ferry during the repair period in a dry dock. The expected impacts of this study are threefold. Firstly, it aims to provide a better understanding and quantification of workers’ evacuation performance during shipbuilding and repair operations on vessels in dockyards. Secondly, it aims to assist others in developing, validating, and applying evacuation models for these types of scenarios. Thirdly, it aims to demonstrate that flexibility is an important factor for the applicability of evacuation modelling to new types of scenarios and evacuation conditions.

The paper has been divided into four parts, apart from this introductory section. Section 2 (Data) presents the scenario and geometries, population, data collection methodology and provides the observed datasets. Section 3 (Model) includes information on the model selection and configuration, as well as how behavioral uncertainty was addressed for running the simulations. Section 4 (Validation) conducts a detailed analysis of the evacuation process and the results of the validation protocol which can help to improve the accuracy of evacuation models for vessels in dockyards. Section 5 (Discussion). This section interprets the results and discusses the implications of the key findings. Section 6 (Conclusion) summarizes the main findings of the study and provides recommendations for future research.

2 Data Collection and Case Study Description

This section presents the scenario and geometries, the population, the data collection methodology and the derived datasets. The information provided is intended to assist interested parties in model development, validation and application when dealing with the evacuation process of vessels in dockyards.

2.1 The Scenario

A Ro-Pax ferry in a dry dock was considered in this study. The vessel has ten decks: engine rooms on decks 1 and 2, garages on decks 3–5, passengers’ accommodation (cabins, cinemas, restaurants) on decks 6–9 and the bridge on deck 10. The heights between decks are: 3.1 m (deck 1–2), 3.3 m (deck 2–3), 6 m (deck 3–5), 3.5 m (deck 6–7 and deck 7–8) and 2.8 m (deck 8–9 and deck 9–10). The vessel has seven regular dogleg staircases (Figure 1). Stair A (nominal width = 1 m; horizontal run = 1.92 m) extends from deck 3 to deck 9, Stair B (nominal width = 0.9 m; horizontal run = 1.92 m) extends from deck 3 to deck 10; Stair C (nominal width = 0.9 m; horizontal run = 1.70 m) extends from deck 2 to deck 10; Stair D (nominal width = 0.9 m; horizontal run = 1.40 m) extends from deck 2 to deck 6; Stairs E, one on the starboard side and one on the portside (nominal width = 0.9 m; horizontal run = 1.92 m), extend from deck 6 to deck 8 and Stair F is a double stair (nominal width = 1.5 m; horizontal run = 1.78 m each part) that extends from deck 6 to deck 9. Working activities (Figure 1) concentrated on three main working sites: (1) engine rooms (decks 1 and 2), (2) the casing/tank (decks 7–10) and (3) the new pump room (decks 2 and 3 on the portside). Figure 2 shows the evacuation network schematization of the ship during the working operations. Temporary means of egress included vertical connections narrower and/or steeper than regular ones: a scaffold z stair from the engine rooms (on deck 2) to deck 3 (Figure 3a), a scaffold z stair (L-shape) near the casing/tank connecting deck 8 and deck 7 (Figure 3b), an external scaffold with vertical ladders leading directly from the pump room to the dock floor (Figure 3c) and two ladders in the pump room (Figure 3d). As any other construction site, the connections between paths were constantly changing making it more difficult for the workforce to way find. From the Figure 4 it is possible to see the three different evacuation conditions registered in the pump room during working operations. During phase 1, there were three exits available, however, the bottom and upper decks were not connected. To exit the pump room, workers on the bottom deck had to use one of the two ladders. Workers on the upper deck could exit via a cat ladder or by the cutting open a section of the ship's hull, known as a “cesarean,” and using scaffolding to reach the dock floor. In phase 2, the bottom and upper decks were connected allowing workers to access any of the three available exits. During the evacuation drill the pump room was in phase 3 with two possible exits.

Figure 1
figure 1

Regular staircases. Gangways and the main working sites onboard

Full size image
Figure 2
figure 2

Evacuation network schematization of the ship

Full size image
Figure 3
figure 3

Temporary means of egress

Full size image
Figure 4
figure 4

Changing evacuation conditions during working operations in the pump room

Full size image

2.2 The Data Collection Process

The method adopted was to collect data in both working conditions and during evacuation. Table 1 displays the dates, the number of video cameras used, the method employed, the amount of information collected and recording procedure (stationary and moving). In the former, cameras were positioned at selected points to measure conditions passing (e.g., regular stairs, temporary stairs, and gangways) while reducing their impact on working operations and the evacuation process. In the later, researchers wore cameras on their helmets during visits and during the evacuation drill to capture additional information such as subjective experiences, changes in the geometry, environmental conditions (e.g., noise, visibility), people performances, population distribution, and record interviews.

Table 1 Data Collection Summary
Full size table

2.3 The Population

The population mainly consisted of males (95.8%) between 18 and 65 years (approx. 30% < 30 years; 45% 30–50 years; 25% > 50 years). Table 2 provides the estimated number and distribution of population during the four days of data collection.

Table 2 Estimated Occupation Onboard
Full size table

2.4 The Datasets

The data collected consisted of (1) presentation times of the main working sites (pump room, engine rooms and casing/tank) observed during the evacuation drill, (2) vertical speeds on temporary scaffold stairs and ladders collected during normal conditions and evacuation and (3) the evacuee arrival times at the gangways during the evacuation drill for a subsequent simulation analysis. These datasets can be used in conjunction with further datasets to enhance understanding as well as part of model development and/or validation processes. The data and the geometry that support the findings of this study are available from the corresponding author, A.C, upon reasonable request.

2.4.1 Presentation Times

The warning system of the ship was used by combining sound alarm and voice messages during the evacuation drill. Because of the position of the cameras (i.e. outside the working sites), it was only possible to capture the interval between the time at which a warning was given and the time at which each worker left a given working site of origin (pump room, engine rooms, casing/tank) towards the gangways to abandon the vessel (Table 3). The concept is taken from PD7974-6:2019 [39]. The exact frame at which each worker arrived the reference point was registered and noted into a spreadsheet. It is argued that this does not significantly influence our understanding of the likely pre-evacuation times given the provided granularity and the fact that workers covered short distances (< 25 m) from their starting locations to the measuring points. The exact time (frame) at which each worker evacuation time for each worker was defined as the frame when the worker’s trailing leg leaves the gangway. Note that this variable can be affected by several factors [40]. For instance, a considerable proportion of workers (85.7%) began travel within 3 min while the last few workers took longer (> 5 min), resulting in the lognormal nature of the distributions for the pump room and the engine rooms. This was not the case for the casing/tank where workers started to move within 3 min. Furthermore, the presentation time for crew members was longer than those observed for workers revealing that they were engaged in previous tasks before start evacuation in a group. We did not have cameras inside the engine rooms. However, we presume that crew members took the opportunity to simulate their own drill procedures like communicating with the bridge, simulating operations at the engine control room (e.g., stopping the main engine and securing the propellers, closing watertight doors) before starting evacuation.

Table 3 Presentation Times (in s) for the Main Working Sites
Full size table

2.4.2 Vertical Speeds on Temporary Scaffold Stairs and Ladders

The speeds were established by measuring the time taken by each worker to cover the distance between two fixed reference points (i.e. start point and end point of the stair or ladder). The vertical speed of each evacuee (vi) was calculated according to:

$$v_{i} = \frac{d}{{\left( {F_{Bi} - F_{Ai} } \right)*1/Fr}}$$
(1)

where d is the distance between the reference start point and end point (measured directly in the scenario), FAi the frame when the individual’s body crosses the reference starting point and FBi the frame when the individual’s body crosses the end point and Fr is the frame rate (30 frames/s). The video images were analyzed using video processing software. This software allows the footage to be manually advanced frame by frame. The software also provides the frame rate. Note that the data was collected in normal conditions and during evacuation. Therefore, the measured speeds may differ. We conducted Kolmogorov–Smirnov two-sample test to examine whether normal and evacuation samples are drawn from the same distribution. There were no significant differences: cat ladder up (D = 0.46, p = 0.12), cat ladder down (D = 0.28, p = 0.18); scaffold L-shape stair down (D = 0.30, p = 0.28); scaffold straight stair up (D = 0.25, p = 0.08). These results were expected since the vertical speeds are mainly constrained by the geometry of these temporary devices (e.g. width, angle, steps). Consequently, we decided to merge the datasets to provide a more robust representation of vertical speeds. Table 4 provides the observed data. Overall, the vertical speeds do not differ significantly from that which are normally distributed (p > 0.05).

Table 4 Vertical Speeds (in m/s) on Temporary Means of Egress
Full size table

On average workers were slower on the ladder than on cat ladders. A possible explanation is that the measured ladder was a single portable ladder and located in a low visibility area, so workers were cautious when using it. Interestingly, there were no significant differences between descendent and ascendent speeds for ladders t(33) = 0.03, p = 0.97 and cat ladders t(280) = 1.47, p = 0.04. The same movement needed, maintaining a three-point contact both when climbing up and down, may explain this.

2.4.3 Evacuation Curves

The exit arrival time of each evacuee during the unannounced evacuation drill was measured from the video recordings as well. The evacuation time for each worker was defined as the frame when the worker’s trailing leg leaves the gangway. Figures 5 and 6 show the arrival time curves for the gangway D3 and the gangway D5 respectively. The differences in the two curves can be explained by the occupant distributions, the pre-evacuation time, the nature of the means of egress and the evacuation routes used. In total 41 workers took 5 min 10 s to abandon the vessel through the gangway D5. These workers were initially located on the casing/tank (n = 20) on decks 7 and 8 and the surrounding areas on the stern between decks 6 and 10 (n = 21). The gangway D3 was used by 109 workers mainly from the engine rooms (n = 43) and the pump room (n = 31). The rest of the workers that used this gangway (n = 35) came from deck 3 and upper decks on the bow. The total evacuation time through D5 was 12 min 57 s. This information was used for further comparison considering multiple simulations and representative runs for model validation described in the next section.

Figure 5
figure 5

Evacuation curve through gangway D5

Full size image
Figure 6
figure 6

Evacuation curve through gangway D3

Full size image

3 Model Implementation

In this section we first present the model selection, model configuration and how the behavioural uncertainty was addressed.

3.1 Model Selection

There are many computer models that can simulate the emergency egress from different scenarios [5, 8, 9]. To our knowledge, none have been used and/or validated for the evacuation of vessels in shipyards. As mentioned, there are three options to deal with a new evacuation scenario: (1) use an existing model, (2) adapt/modify an existing model and (3) develop a new specific model. We chose the first option and selected the STEPS model (v 5.1) for this study. It is a well-known and widely used agent-based model developed by Mott Macdonald company. This model has been applied to many studies addressing different evacuation scenarios e.g. trains [41], road tunnels [42], hospitals [43], schools [44], buildings [11], and validated against real-life observations [11, 41, 43]. According to Kuligowski [8] this is a partial behavioural model where agents have their own characteristics implicitly represented by random pre-evacuation times and walking speeds and coefficients of patience and familiarity. The model accepts CAD Drawings, and the geometry is represented by a fine network of squared cells (0.5 × 0.5 m by default) between which the occupants move towards the exits through a potential map where each grid cell has a value. Results provided by the model include textual (qualitative and quantitative), two-dimensional output and three-dimensional interface. We selected this model for several reasons. The first reason is availability as we have a perpetual license of this software. The second reason is representativeness because this model is a well-known and widely used evacuation model [7,8,9]. The third reason is flexibility as the model allows the user to customize inputs (e.g., travel speeds and the capacities of the means of egress as well as the interpersonal distances), especially useful to reproduce the movement of occupants on non-conventional means of egress (e.g., temporary scaffold stairs and ladders). The fourth reason is the movement method used (fine network) that enables agents to easily move through narrow spaces. Finally, the model has also batch running allowing the easy execution of multiple simulations.

3.2 Model Configuration

3.2.1 Geometry

The CAD drawings of the decks were imported by the model. In this process grids of squared cells (0.5 × 0.5 m) where occupants can walk were defined and walls and other obstructions automatically became blockages. Then, stairs and exits were created. A default value of 1.365 per/m/s was assumed as the maximum flow for the gangways, regular stairs, and scaffold stairs [45]. Occupant spacing on stairs was not measured in this study. However, for safety reasons, only one person was observed using ladders and cat ladders at a time. This was implemented in the model by setting a minimum interpersonal distance between two agents i.e. occupant spacing equal to the length of the ladder/cat ladder.

3.2.2 Population Characteristics

The horizontal walking speed for virtual workers was assumed as normally distributed (Mean = 1 m/s, SD = 0.2). The model allows the possibility to apply speed factors to different surfaces and vertical connections (e.g., stairs, ramps). Hence the reduction factors for temporary means of egress were applied according to data collected (Table 4) as follows: 0.30 for cat ladders (up-down), 0.23 for ladders, (up), 0.57 for the scaffold straight stair (up), and 0.55 for the scaffold L-shape stair (down). It should be noted that ship stairs connecting deck 2 and deck 1 in the engine rooms are steeper (> 50º angle) than building stairs.

The walking speeds for these stairs were automatically adjusted by the model as follows: k = 0.387/sin q where k is the reduction factor and q is the angle of the slope for a specific stair. The starting locations and pre-evacuation times of workers and crew members were established by observing the video recordings (Table 5). As mentioned, the number of cameras was insufficient to cover all decks and locations. However, it was possible to determine overall starting areas and the associated pre-evacuation times for workers on decks 3–6 based on information partially recorded. For instance, a camera located on deck 3 allowed us to determine a group of 9 workers who were located next to the gangway D3. The pre-evacuation time assumed for this group ranged from 0 to 60 s as they were observed leaving the vessel first. This camera also allowed us to identify another group of 8 workers who came from the stern on deck 3 at around 350 s from the alarm. Hence the starting location of this group was the stern and the pre-evacuation time ranged from 250 to 450 s. The distribution of the rest of workers on deck 3 (n = 19) was assumed randomly (all deck) with pre-evacuation times derived from the PD7974 [39]. The same was assumed for workers initially located on deck 5 (n = 7) and workers from deck 6 (n = 9) (Table 5). Workers had a clear understanding of the layout and exit use was observed to be performed by proximity (Table 5). Consequently, we adopted a deterministic approach in the model, and agents selected the specific route based on proximity as well (i.e., shortest path).

Table 5 Population Distribution, Pre-evacuation Times and Exit Use Implemented for the Simulations
Full size table

Although the complexity of the internal layout of the working sites was represented in the model (e.g., scaffolds, cat ladders, temporary stairs), the precise initial location of the workers in these areas were unknown as well as the starting location of workers on decks (3, 5 and 6) who represented 42% of the population. Furthermore, the collected presentation time distributions were assumed as pre-evacuation times in the model.

In summary, the model configuration was affected by uncertainties associated with the stochastic nature of human behaviour (e.g., variability) as well as uncertainties arising from the lack of data and/or specific information regarding the original conditions during the evacuation drill. The input behavioral uncertainties included potential discrepancies in speed, path selection, initial position, and pre-evacuation times. Although we were able to incorporate vertical speed distributions based on our observations (as indicated in Table 4), we had to assume horizontal speeds for the workers. Furthermore, agents selected the shortest paths to the gangways (as outlined in Table 5), but this implicit behavior may not accurately represent current wayfinding behavior. Additionally, we did not have precise information about the initial location of workers, and the pre-evacuation time distributions were derived from either presentation times with corresponding errors or based on arbitrary assumptions. These uncertainties were highly likely to introduce variations in the model predictions and, as a result, impact the level of agreement between the model and the observed drill data.

3.3 Behavioural Uncertainty

The convergence method proposed by Ronchi et al. [46] was applied to assess the impact of behavioural uncertainty on the variability of the results produced. The method uses error estimations and functional analysis to assess the convergence of the evacuation curves produced by the model. We considered five key variables: (1) total evacuation time (TET), the time for the last arrival from the vessel, (2) standard deviation (SD) of total evacuation times: measure of the spread of last arrival times, (3) Euclidean relative distance (ERD): the average difference between arrival curves, (4) Euclidean projection coefficient (EPC): the best possible agreement between arrival curves, and (5) secant cosine (SC): similarity between the curve shape by assessing the gradients produced along arrival curves. We checked the convergence of the measures over 100 runs and calculate the progressive difference between the threshold value and the values of each convergence measure. The acceptance criteria thresholds below 5% would permit the assessment of the required safe egress time with a reasonable accuracy [47]. Here we assumed a threshold 1% (0.001) for 10 consecutive runs also used in a previous study [48]. Figure 7 shows the convergence results for the gangway D3 and the gangway D5 respectively. The model converges at different rates. The SD of total evacuation time is the slowest variable to converge, while the other variables converge relatively quickly. We conclude that the variability of the simulated results (representing scenario unknowns and behavioural uncertainty) is compliant with the chosen acceptance criteria. Overall, the criteria result in a total number of repetitions lower than 58 runs for gangway D3 and 51 runs for gangway D5. Consequently, we considered 60 randomized simulations for the analysis.

Figure 7
figure 7

Convergence of model results for evacuation through the gangways. Total evacuation time (TET), standard deviation (SD) of total evacuation time, euclidean relative distance (ERD), euclidean projection coefficient (EPC), and secant cosine (SC)

Full size image

4 Results: Validation of the Simulation Model

This section presents the results of the proposed validation protocol comprising three steps: (1) graphical validation by visually comparing graphs of prediction and observation, (2) classical hypothesis testing by comparing the predicted and the observed evacuation samples and (3) functional analysis to quantify the agreement between the predicted and the observed evacuation curves.

4.1 Graphical Validation

From Figs. 8 and 9 it is possible to see that the average exit time curves (in sec) produced by each evacuee from 60 repeat simulations with the range of variation between the minimum and maximum are reasonable approximations to the drill data. Comparing the predicted average evacuation curve with the experimental data, we note that the total evacuation time through gangway D3 is underpredicted by 6 s or 0.8% (772 s compared with 777 s), while the time for workers to exit through the gangway D5 is overpredicted by 14 s or 4.3% (324 s compared with 310 s).

Figure 8
figure 8

Predicted evacuation curves (max., min. and average of 60 runs) vs observed evacuation curves in gangway D5

Full size image
Figure 9
figure 9

Predicted evacuation curves (max., min. and average of 60 runs) vs observed evacuation curves in gangway D3

Full size image

4.2 Classical Hypothesis Testing

The second validation procedure is introduced here and uses hypothesis testing to confirm that the model generated reliable samples [49]. Classical statistical testing is a method used to validate a model by comparing the model’s predicted output to actual data. This process considers whether the difference between the prediction and observation is statistically significant. We used to two common statistical tests for model validation: Mann–Whitney–Wilcoxon test (MWW) and Kolmogorov–Smirnov test (K–S) were used. MWW tests the null hypothesis that the two samples (average predicted and observed) come from the same population whereas K–S tests the null hypothesis that the two samples are drawn from the same distribution. Table 6 shows the summary of statistics. The null findings support the second validation criterion i.e. results do not contradict the hypotheses that the two samples (observed and predicted) came from the same population and that they are drawn from the same distribution.

Table 6 Statistical Testing Results Comparing the Average Representative Sample (60 runs) and the Observed Sample (Two Tailed, α = 0.05)
Full size table

4.3 Functional Analysis

The third validation procedure is the well-known functional analysis [12, 14, 15]. Four variables were considered to quantify the agreement between the model predictions and the observed data: (1) Total Evacuation Time (TET100—the time for the last arrival from the building and TET95—95th percentile of arrival times), (2) Euclidean Relative Distance (ERD—the average difference between observed and simulated data in an arrival curve), (3) Euclidean Projection Coefficient (EPC—the best possible agreement between the observed and simulated data in an arrival curve), and (4) Secant Cosine (SC—similarity between the curve shape by assessing the gradients produced along the simulated and observed arrival curves). A perfect match for the model and the observed curves is represented by the following optimal values i.e. TET = 0.0; ERD = 0.0, EPC = 1.0 and SC = 1.0.

We first selected the performance metrics for the best/worst representative runs (Table 7). As expected, some runs (21%) for the arrival curves in gangway D3 resulted in a great difference in the shape of the curves (SC < 0.60). The evacuation of the last 30 evacuees that used the gangway D3 lasted 420 s (7 min). Therefore, we argue here that simulating this process leads to more variability in the generated shapes of the curves. Then we compared the predicted average arrival times over the repeated calculations and the drill data. Relative differences of total evacuation times (TET) and the functional operators (ERD, EPC and SC) were used as quantitative measures. Results in Table 8 show that the model fulfils the threshold values suggested by Galea et al. [50] (i.e. TET < 0.15; ERD < 0.25; 0.8 ≤ EPC ≤ 1.20; SC > 0.80) indicating a good agreement between the model and the observed data. Therefore, the third validation criterion was satisfied.

Table 7 Best and Worse TET, ERD, EPC and SC Comparing the Observed and the Simulated Evacuation Curves (60 Runs)
Full size table
Table 8 Results of Functional Analysis Comparing the Average Predicted Curves (60 runs) and the Observed Evacuation
Full size table

5 Discussion

Evacuation of workers during shipbuilding and repair operations is challenging. Evacuation modelling can help to increase our knowledge about this process and, therefore, improve workers safety. Empirical data and validation are essential to gain confidence and applicability of such tools to these new scenarios. The present study makes several noteworthy contributions to this effort. The case study involved a Ro-Pax ferry in a dry dock during building and repairing operations. The methodology applied shows the procedure we followed when facing the complexity of the evacuation of vessels in shipyards.

The pre-evacuation time is conceivably the most critical variable for most evacuation scenarios [39]. For practical reasons, our response data focused on the presentation times (i.e. the time when each worker left a given working site onboard from the alarm). The range of the overall distribution [1–334 s] for workers was found to be close to the response time distribution reported in a previous study [0–340 s] for building construction sites [37]. Our observations are in line with the pre-evacuation times suggested in standards [39, 51]. Nevertheless, the period until the last few workers moved was longer (> 5 min) showing the potential impact of isolation and task disengagement in this kind of scenarios. Therefore, care should be taken when using/extrapolating the presented data to another target scenario since the conditions should be like the ones described here. Overall, these observations and precautions warrant further research to increase our understanding on the response of employees onboard to enhance our simulation capabilities. This study has also provided walking speed distributions on temporary means of egress. As expected, vertical speeds (up and down) on ladders, cat ladders and scaffold stairs are slower than those for regular building stairs [52]. The presented data must be interpreted with caution. Rather than fixed structures these devices are often improvised and adapted to diverse gaps, heights, and paths according to specific working operations onboard. Hence the presented data for scaffold stairs might not be generalized. Further research should be done for example to investigate the effect of different scaffold stair slopes on walking speeds. The stated speeds for ladders and cat ladders are more transferable to other scenarios. Interestingly, we found no differences between ascending and descending speeds on these devices. This result is likely related to the same movement needed both when climbing up and down maintaining a three-point contact (two hands and one foot, or two feet and one hand on a ladder) thus holding the bodyweight without taking advantage of gravity when descending. The benchmark datasets for validation of this new scenario were completed by the observed arrival times of 150 workers during an unannounced evacuation drill.

The information collected allows key aspects of the validation process to be configured (e.g., the geometry, the population, the pre-evacuation times, the routes used, and the speeds observed). The next step consists of simulating this new evacuation scenario. The criteria for model selection were based on [8]. A key aspect in the model selection was the flexibility to represent/simulate the particularities of movement on different types of temporary scaffold stairs and ladders such as the flow capacity, the impact of different slope/angles on travel speeds and the interpersonal distance on ladders (i.e., only one person on ladders at a time). These fundamental scenario elements were imposed allowing the simulated initial conditions to be configured to reflect those observed. Yet, there were several uncertainties when configuring the model. For instance, the specific starting location of 42% of the population and their pre-evacuation time were unknown. Hence, we used the convergence method to determine the minimum number of runs for the validation analysis (n = 60).

Overall, the results demonstrated that it is possible to predict the evacuation performance of workers onboard vessels in dockyards with a reasonable level of accuracy as far as the simulation software used is adapted to take into consideration the specific characteristics of these new evacuation scenarios. A validated software can enhance the safety of shipyard workers in several ways (1) establishing the time for workers to leave certain areas and/or abandon the vessel, (2) determining the maximum number of workers in specific areas, (3) designing evacuation systems including temporary evacuation means and/or (4) evaluating the effectiveness of procedural solutions. Such assessments could be performed both during the planning stages and during working operations. In any case, an expert user is needed. It should be noted that validation of an evacuation model against a single data set does not necessarily imply validation for other scenarios or even for the same type of scenarios. The process described in this paper should be conducted when facing a new project especially when the scenario in question has not been analysed/simulated yet as the one addressed here.

This study has limitations. First, it focuses on shipbuilding and repair operations limits its applicability to other types of evacuation scenarios. The findings may not be generalizable to other types of emergency evacuations, such as building fires or natural disasters. Second, the data collected from a single case study, a Ro-Pax ferry in a dry dock, may not be representative of all shipyard evacuation scenarios. The sample size may be sufficient for the specific scenario studied, but it is not necessarily representative of all shipyard evacuation scenarios. Finally, there were many uncertainties in the data used for the simulations, such as the specific starting location of 42% of the population, their pre-evacuation times, and their horizontal walking speeds not collected in this study. These uncertainties may have affected the accuracy of the simulation results and limited the generalizability of the findings to other scenarios. For instance, future research should be systematically devoted to a more efficient data collection and analysis of presentation timing (response of workers) in realistic conditions.

6 Conclusion

This study was undertaken to promote the use of evacuation modelling for safety assessment during shipbuilding and repair operations. New empirical data and model validation has been provided. The presented results suggest that existing agent-based evacuation models, with appropriate inputs and small adaptations, may be applied to these complex scenarios. Further research could explore the generalizability of the evacuation modelling and simulation in different types of evacuation scenarios of the shipyard context. Improving the accuracy of the model by collecting more data of workers evacuation performances while addressing behavioural uncertainties in evacuation modelling predictions, should be pursued. We highlight the importance of using multiple methods for model validation, as each method has its own strengths and weaknesses. Additionally, investigating the effectiveness of different evacuation procedures and systems could inform the development of guidelines or standards for shipyard evacuations and enhance the accuracy and applicability of evacuation modelling in real-world scenarios. We also suggest that similar studies to this may be carried out when dealing with new evacuation scenarios. It is important to have an open-minded approach when evaluating models and simulation, therefore, it's important to critically assess the assumptions, limitations and uncertainties associated with the model, as well as to consider other possible models or methods that might provide additional insights in further research.