A block of flats differs from an office tower in ways that bear directly on escape. Its occupants include children, older people and people with long-term impairments. They may be asleep. They are not trained, they are not supervised, and the building usually has no fire wardens. Many residential towers have a single stair. For these reasons the fire strategy in several countries has relied on compartmentation: a fire is expected to be contained in the flat of origin, and other residents are advised to remain where they are. British guidance for residential buildings sets this approach out in detail (BSI, 2015).

The fire at Grenfell Tower in London in June 2017 showed what follows when compartmentation fails and a whole building must be emptied down one stair. The first report of the public inquiry recorded the difficulties faced by residents who could not descend unaided and recommended that personal emergency evacuation plans be prepared for residents of high-rise housing whose ability to leave by themselves is compromised (Moore-Bick, 2019). Planning of that kind needs estimates of how long an evacuation would take and where it would be obstructed, and those estimates come from models.

The purpose of this article is to review how well evacuation models serve that need. It considers the types of model in use, the ways in which reduced mobility is represented in them, the data available to support the inputs, and the means of checking that the results deserve confidence.

This is a narrative review of the published literature on building evacuation, with attention to sources that address tall buildings, dwellings, or occupants with reduced mobility. It draws on reviews of evacuation models, on experimental studies of movement, on official investigations and on standards. The literature was not searched according to a formal protocol, and the selection reflects the judgement that a designer or regulator is best served by the studies on which current practice in fact depends. Sensory and cognitive impairments affect evacuation as surely as mobility impairments do, but they raise different modelling questions and are mentioned only where they interact with movement.

The oldest approach treats a moving crowd as a fluid. Building on the pedestrian studies of Fruin (1971), Pauls (1987) derived relationships between the usable width of a stair and the flow of people down it from observations of evacuation drills in tall office buildings. The resulting hydraulic model, in which speed falls as density rises and flow is limited by the narrowest element, is still the standard hand calculation (Gwynne and Rosenbaum, 2016). It describes a homogeneous population. An occupant who moves at a fraction of the general speed, or who occupies the full width of the stair, has no place in it except as an adjustment to the average.

Computer models removed that restriction by representing individuals. Gwynne et al. (1999) classified the early models by how they represent space, population and behaviour, and the survey by Kuligowski et al. (2010) remains the most complete catalogue of their features. In broad terms three kinds of movement model are in use. Cellular automata divide the floor into a grid and move occupants from cell to cell according to local rules (Burstedde et al., 2001). Continuous models place occupants in open space and compute their paths from the distances between them or from attractive and repulsive forces, as in the social force model (Helbing and Molnár, 1995; Helbing et al., 2000). Several widely used simulation tools combine a continuous movement model of this kind with predicted smoke conditions, so that visibility and toxic exposure can slow or redirect occupants.

Agent-based models add decision-making to movement: each occupant has attributes and rules that govern when to leave, which route to take and whether to help others. For the present subject this is the most relevant development, since impairment affects decisions as well as speed. Table 1 compares the families in terms of what they can represent for an occupant with reduced mobility.

An evacuation is usually analysed as a sequence of phases, shown in Figure 1. Reduced mobility lengthens several of them, and the modelling task differs in each.

A model result is only as credible as the checks behind it. Guidance from the United States National Institute of Standards and Technology sets out a series of verification tests for the basic components of an evacuation model, such as speed, flow at doors, movement on stairs and route choice, together with a discussion of validation against experimental data (Ronchi et al., 2013). An international standard now gives a protocol on the same lines (ISO, 2020). These documents include tests for occupants with different speeds and body sizes, but validation data for mixed-ability populations on stairs are sparse, so a model may pass the verification tests and still be unvalidated for the case that matters here.

Uncertainty therefore has to be handled openly. Because behaviour is variable, a single simulation run is one sample from a distribution, and results should be reported as ranges from repeated runs. The more important uncertainty lies in the inputs. The number of residents who will need assistance, where they live in the building, whether a helper is present at night and how long device preparation takes are each capable of changing the answer more than the choice of model. Sensitivity analysis on these inputs is more informative than refinement of the movement algorithm.

The data on which impaired movement is modelled are the weakest part of the field. The principal speed measurements are more than two decades old and were made on a population whose mobility aids, age profile and body size may differ from those of residents today. Powered wheelchairs and mobility scooters, which cannot be carried by one or two helpers, are barely represented. Measurements of long stair descents by people with limited endurance, and of assistance given by untrained helpers, would add most.

A second gap concerns the residential setting itself. Most assisted-evacuation data come from hospitals, and most stair data from offices. Night-time conditions, single-stair layouts, flats used as places of temporary refuge and the interaction between descending residents and ascending firefighters all need study. A third gap is behavioural: whether residents will wait in a lift lobby or refuge when smoke is present, and how the advice to stay or to leave is understood by people who know that they cannot leave unaided. Models can already represent these choices. What is missing is evidence on which to set the parameters.

Evacuation models have progressed from flow calculations for uniform crowds to simulations in which each occupant has individual capabilities, makes decisions, receives help and may use a device or a lift. The structure needed to analyse the evacuation of mobility-impaired residents from tall housing therefore exists. The evidence base has not kept pace. Movement speeds rest on a small number of studies, device performance on trials with trained staff, and lift and refuge behaviour on very little. Until better data are collected, results for such buildings are best presented as ranges tied to stated assumptions about who needs help and who is available to give it, and those assumptions deserve the same scrutiny as the model itself.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.