Unified probability propagation prediction method and system suitable for wui fire

Abstract

The application discloses a unified probability propagation prediction method and system suitable for WUI fire, and the method comprises the following steps: constructing a unified state space and discretizing; constructing an external driving model of building layer fire propagation; constructing a building layer dynamic Bayesian network with each building as a node, parameterizing a conditional probability table of the building layer dynamic Bayesian network based on a physical driving mechanism; establishing a mapping relationship between an internal state of the building layer and an external observation category; realizing state correction and rolling prediction based on Bayesian updating; when external observation is obtained, the system state is updated based on Bayesian updating in combination with observation information, a posterior state distribution is obtained, the posterior state distribution is taken as an initial state of the next time step prediction, a rolling assimilation closed loop is formed, and finally a multi-state probability risk field of building scale evolution over time is output. The application can realize unified coupling of a landscape, a building and a flying fire, obtain state probability output over time, and can be assimilated and corrected.

Description

Technical Field

[0001] This invention relates to the field of fire risk early warning technology, specifically to a unified probability propagation prediction method and system applicable to WUI fires. Background Technology

[0002] Wildfire disasters have evolved from traditional natural ecological events into complex urban disasters causing massive damage to the built environment, especially in Wildland-Urban Interfaces (WUIs). In WUIs, fires often exhibit a cascading process of "wildfire-community fire-urban fire": on the one hand, fire lines formed by burning vegetation continue to advance; on the other hand, buildings within communities ignite non-locally and at multiple points due to heat radiation, flame contact, and flying embers, leading to widespread structural damage and risks to life and property. WUI fires also have a ripple effect on the operation of critical infrastructure such as power supply, communications, transportation, and healthcare, significantly increasing disaster response and recovery costs. Unlike some sudden disasters, wildfires typically possess observable and updatable characteristics during their development: with the continuous influx of information from remote sensing, drones, ground inspections, and emergency reports, management departments often have a window of opportunity to dynamically adjust emergency response strategies over timescales ranging from hours to days as the fire spreads. Therefore, fire risk forecasting for WUI scenarios not only needs to provide a qualitative judgment on "whether it will burn", but also needs to continuously absorb new observations during the disaster, and correct the risk ranking and spatial distribution in real time, so as to provide actionable quantitative support for resource scheduling, key defense and evacuation decisions.

[0003] Unlike traditional wildfires, WUI fires primarily cause damage in the built environment, and their propagation mechanism involves a multi-mechanism coupling of "landscape-building-flying fire" (e.g., Figure 1In a WUI (World Uncontrolled Inferior Fire) scenario, surface fire spreads along combustible fuel, short-range radiation between buildings ignites the fire, and stray fire is transported long distances by wind and deposited downwind to form new ignition sources. These three factors interact and rapidly alter the risk field. Therefore, in a WUI fire emergency scenario, achieving decision-oriented disaster risk forecasting typically requires a predictive model capable of rapidly generating the fire's impact range and exposure intensity at future time steps, given environmental factors such as wind field, fuel, and terrain. Existing technologies commonly employ approaches including: landscape propagation simulation based on fire spread mechanisms to extrapolate fire advance and combustion intensity; and building risk assessment models oriented towards the built environment to convert fire exposure into building-level ignition or loss risks. To characterize the common long-distance multi-point ignition in WUI scenarios, some methods also introduce statistical or semi-physical models of stray fire transport and deposition to provide the ignition "pressure field" or ignition probability in the downwind area. Overall, by combining information such as landscape propagation, building ignition, and flying fire ignition, it is theoretically possible to generate temporal risk output at the building scale, providing a quantitative basis for emergency responses such as key defense, resource allocation, and evacuation.

[0004] However, existing technologies still have key gaps in predictive output for WUI fire emergency decision-making. First, current methods often model landscape spread, inter-building ignition, and stray fire separately or splice them together in a loosely coupled manner, lacking a framework that can provide a unified probabilistic description of the multi-mechanism propagation process within the same state space. This makes it difficult to consistently transmit the influence between different mechanisms within the model, and the outputs struggle to maintain a consistent semantic caliber. Second, many predictions are more biased towards fire line extent, intensity field, or static loss assessment, failing to directly provide the building-scale and time-evolving state probabilities required for emergency action (e.g., risk ranking and hotspot evolution over multiple future time steps). This limits their ability to support resource allocation, key defense, and dynamic decision-making. Third, in disaster risk forecasting scenarios, to dynamically correct and consistently update prediction results using continuously arriving external information, a model framework with clearly defined states and probabilistically meaningful outputs is needed to support the "updates," enabling them to proceed with unified semantics and scale. Existing technologies are still insufficient in this regard, and therefore there is an urgent need for a prediction method that can achieve integrated representation of multiple mechanisms, provide temporal probability output at the building scale, and provide a feasible basis for subsequent data fusion and updates. Summary of the Invention

[0005] To overcome the shortcomings of the above-mentioned technologies, this invention provides a unified probabilistic propagation prediction method and system applicable to WUI fires. This invention uses buildings within a community as the basic modeling object, defining the discrete states of buildings during the fire process; it uses landscape-level fire evolution and fire transport and deposition as external driving forces for building-level state transitions; it characterizes the local fire propagation process using the propagation relationships between building neighborhoods; it recursively derives the system state at discrete time steps, and uses observational information to consistently correct the prediction results after obtaining external observations. Thus, it achieves dynamic assimilation prediction of the WUI fire propagation process.

[0006] Terminology Explanation:

[0007] 1) DBN: Dynamic Bayesian Network.

[0008] The technical solution adopted by this invention to overcome its technical problems is:

[0009] This invention discloses a unified probability propagation prediction method applicable to WUI fires, comprising the following steps:

[0010] Constructing and Discretizing a Unified State Space: Constructing a unified state space including the state of the building layer, the state of the landscape layer, and the state of the flying fire layer for the WUI fire propagation process, and defining the discrete fire state of each building in the building layer.

[0011] Establish external drivers: Construct an external driving model for fire propagation in building layers. The external driving forces include two parts: fire evolution in the landscape layer and fire transport and deposition.

[0012] Construct and parameterize a dynamic Bayesian network for the building layer: Construct a dynamic Bayesian network for the building layer with each building as a node to represent the temporal propagation relationship of building states; Parameterize the conditional probability table of the dynamic Bayesian network for the building layer based on the physical driving mechanism;

[0013] Construct an observation model: Establish a mapping relationship between the internal state of the building layers and the external observation categories;

[0014] State correction and rolling prediction based on Bayesian update: After obtaining external observations, the system state is updated in Bayesian manner by combining the observation information to obtain the posterior state distribution. The posterior distribution is then used as the initial state for prediction in the next time step, forming a rolling assimilation closed loop. Finally, a multi-state probability risk field that evolves at the building scale over time is output.

[0015] Furthermore, the discrete fire states of each building in the building layer include four types: unburned state, ignition state, fully burning state, and burnout state, which are used to represent the staged evolution characteristics of the building during the fire propagation process.

[0016] Furthermore, the landscape layer fire evolution is based on topography, fuel, wind field and initial fire data to advance the temporal spread process of wildfires outside the community on the landscape layer, and obtain the combustion status, fire line advancement information and related heat release characteristics of the landscape grid at each time step.

[0017] Furthermore, based on the results of the fire evolution in the landscape layer, the flyfire transport and deposition identifies units in the landscape layer and building layer that are in an active combustion state as flyfire source items, and calculates the migration and deposition process of flyfire particles in space by combining at least wind speed, wind direction, transport distance and deposition probability, so as to obtain the amount of flyfire exposure borne by each building at the current time step.

[0018] Furthermore, a dynamic Bayesian network for the building layer is constructed using each building as a node, specifically including:

[0019] A building-layer dynamic Bayesian network is constructed with each building as a node. The set of parent nodes for the current time step of each building's state node includes the state node of the previous time step of the building, the state nodes of neighboring buildings at the current time step, and the fire exposure node that acts on the building at the current time step.

[0020] Furthermore, the conditional probability table of the dynamic Bayesian network of the building layer is parameterized based on the physical driving mechanism, specifically including:

[0021] The conditional probability table of the dynamic Bayesian network of the building layer is parameterized using a physical quantity-driven approach. Specifically, it includes the probability of fire ignition driven by the amount of fire exposure, the probability of radiative ignition driven by the thermal radiation of neighboring buildings, and the state transition probability of internal development after the building is ignited.

[0022] Furthermore, the calculation of the probability of ignition of flying fire driven by the amount of flying fire exposure specifically includes: based on the cumulative amount of flying fire exposure that the building experiences in a single time step as the driving factor, and using a preset ignition probability function, calculating the probability of the building transitioning from an unburned state to an ignited state;

[0023] The calculation of the probability of radiative ignition driven by thermal radiation from neighboring buildings specifically includes: using the cumulative radiative energy or equivalent exposure of the target building to the burning neighboring buildings as the driving force, and using a preset ignition probability function, calculating the probability that the neighboring buildings cause the target building to ignite.

[0024] The probability of state transitions in the internal development of a building after it is ignited specifically includes the probability of the building developing from an ignition state to a fully burning state and the probability of it developing from a fully burning state to a burnout state.

[0025] Furthermore, a mapping relationship is established between the internal state of a building floor and the external observation categories, specifically including:

[0026] The observation operators of the observation model map the discrete fire status of building nodes to external observation categories, so as to realize the connection between the internal status of building floors and external observation information under a unified interface. The external observation categories include two categories: undamaged and damaged.

[0027] Furthermore, after obtaining external observations, the system state is updated using Bayesian methods based on the observation information to obtain the posterior state distribution. This posterior distribution is then used as the initial state for prediction in the next time step, forming a rolling assimilation closed loop. Finally, a multi-state probability risk field that evolves at the building scale over time is output, specifically including:

[0028] After obtaining the external observation information at the current time step, the system state is updated using Bayesian methods based on the external driving model, the building-level dynamic Bayesian network, and the observation model, combined with the external observation information at the current time step. This yields a posterior state distribution that reflects the actual fire situation at the current time step. The updated posterior distribution is then used as the initial state for fire propagation prediction at the next time step and continues to advance, thus forming a rolling assimilation closed loop of "prediction-observation-correction-reprediction". Finally, a multi-state probability risk field that evolves at the building scale over time is output, achieving dynamic assimilation prediction of the WUI fire propagation process.

[0029] This invention also discloses a unified probability propagation prediction system applicable to WUI fires, comprising modules:

[0030] State Space Definition Module: Used to construct a unified state space including building layer state, landscape layer state and flying fire layer state for WUI fire propagation process, and to define the discrete fire state of each building in the building layer;

[0031] External driving building module: used to construct an external driving model for fire propagation in building layers. The external driving forces include two parts: fire evolution in the landscape layer and fire transport and deposition.

[0032] Dynamic Bayesian Network Construction Module: Used to construct a building-layer dynamic Bayesian network with each building as a node to represent the temporal propagation relationship of building states; parameterizes the conditional probability table of the building-layer dynamic Bayesian network based on a physical driving mechanism;

[0033] Observation model construction module: used to establish the mapping relationship between the internal state of a building layer and the external observation categories;

[0034] State update module: After obtaining external observations, it combines the observation information to perform Bayesian update on the system state, obtains the posterior state distribution, and uses the posterior distribution as the initial state for prediction in the next time step, forming a rolling assimilation closed loop, and finally outputs a multi-state probability risk field that evolves at the building scale over time.

[0035] The advantages of this invention compared to the prior art are:

[0036] 1. Achieve unified coupling: Provide a unified state expression basis for the coupled propagation of the landscape layer, the fire layer, and the building layer, so that subsequent prediction, correction, and risk output can be completed within the same probabilistic framework.

[0037] 2. By constructing a dynamic Bayesian network, the state probability output that evolves over time is obtained: providing a multi-state probability risk field that evolves over time at the building scale, which facilitates uncertainty propagation analysis and decision-making (such as hotspot identification and resource scheduling).

[0038] 3. Assimilation and correction: By injecting external observation information into the dynamic Bayesian network through the observation model to achieve Bayesian updates, open-loop drift can be suppressed and the hotspot capture capability of subsequent prediction windows can be improved. Attached Figure Description

[0039] Figure 1 This is a schematic diagram illustrating the coupling principle of the landscape layer, building layer, and flying fire mechanism within the unified state space described in this invention.

[0040] Figure 2 This is a flowchart of the unified probability propagation prediction method applicable to WUI fires as described in this invention.

[0041] Figure 3 This is a schematic diagram illustrating the building state division as described in this invention.

[0042] Figure 4 This is a schematic diagram of the structure of the dynamic Bayesian network for building layers described in this invention.

[0043] Figure 5 This represents the cross-union ratio of the landscape layer propagation simulation effect described in this invention.

[0044] Figure 6 This represents the cumulative recall rate of the present invention in the unassimilated state.

[0045] Figure 7 This is a comparison chart of the cumulative recall rates of the present invention under unassimilated and assimilated conditions at different time steps.

[0046] Figure 8 This is a comparison chart of incremental recall rates for the present invention under the conditions of no assimilation, only one assimilation, and cumulative assimilation. Detailed Implementation

[0047] To facilitate a better understanding of the present invention by those skilled in the art, exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These are merely exemplary embodiments of the present invention; however, it should be understood that the present invention can be implemented in various forms and is not limited to the embodiments described herein. These embodiments are provided to enable those skilled in the art to gain a clearer and more thorough understanding of the present invention.

[0048] I. Example 1:

[0049] This embodiment discloses a unified probability propagation prediction method applicable to WUI fires, such as... Figure 2 As shown, it includes the following steps:

[0050] S1. Construct a unified state space and discretize it: Construct a unified state space including the state of the building layer, the state of the landscape layer and the state of the flying fire layer for the WUI fire propagation process, and define the discrete fire state of each building in the building layer.

[0051] In this embodiment, a unified state space is constructed for the WUI fire propagation process, including the building layer state, landscape layer state, and flying fire layer state, such as... Figure 1 As shown, it specifically includes:

[0052] at discrete time step Above, the WUI fire propagation process is defined in time step. The unified state space is:

[0053] In the above formula, Indicates the building floor at time step The set of states, Indicates the landscape layer at time step The set of states, Indicates the time step of the flying fire layer The set of states includes: the landscape layer state, which describes the combustion evolution of vegetation or surface fuel outside the community; the fly fire layer state, which describes the fly fire transport and deposition exposure caused by the combined effects of landscape combustion and building combustion; and the building layer state, which describes the fire development status of each building within the community at different time steps.

[0054] In this embodiment, the discrete fire status of each building in the building floor is defined, specifically including:

[0055] For building floors, such as Figure 3 As shown, the state of a single building is divided into four discrete states: Safe, Ignite, Fire, and Burned, to characterize the phased evolution of the building during the fire propagation process. Based on this, the entire fire propagation and update process is discretized into multiple continuous time steps, and in each time step, external driving calculations, building layer state predictions, observation access, and Bayesian updates are performed sequentially, thus forming a unified temporal recursive framework.

[0056] Step S1 aims to provide a unified state representation basis for the coupled propagation of the landscape layer, the fire layer, and the building layer, so that subsequent predictions, corrections, and risk outputs can be completed within the same probabilistic framework.

[0057] S2. Establish external driving force: Construct an external driving force model for fire propagation in building layers. The external driving force includes two parts: fire evolution in landscape layers and fire transport and deposition.

[0058] Based on the unified state space constructed in step S1, an external driving model for fire propagation in the building layer is further constructed. The external driving force includes two parts: the evolution of the fire field in the landscape layer and the transport and deposition of flying fire.

[0059] The landscape layer fire evolution is based on topographic, fuel, wind field, and initial fire data. It performs progressive calculations on the temporal spread of wildfires outside the community on the landscape layer, obtaining the combustion status, fire line advancement information, and related heat release characteristics of the landscape grid at each time step. This section describes the dynamic background of the approach, intrusion, and persistence of external wildfires towards the community boundary.

[0060] Based on the results of fire evolution in the landscape layer, the flyfire transport and deposition method identifies units in an active combustion state within both the landscape and building layers as flyfire source terms. It calculates the spatial migration and deposition process of flyfire particles by incorporating at least wind speed, wind direction, transport distance, and deposition probability, thereby obtaining the flyfire exposure of each building at the current time step. This flyfire exposure reflects the intensity of flyfire input deposited on the building surface or its adjacent area within a given time step and serves as one of the important external drivers of building layer state transitions.

[0061] Through the two external driving methods mentioned above, step S2 describes the continuous effect of the fire on the building complex from the external landscape of the community, and explicitly introduces the non-local long-range ignition mechanism caused by flying sparks, so that the spread of building fires is no longer limited to the local spread relationship between adjacent buildings.

[0062] This step aims to provide dynamic external fire drivers for building floor state prediction. Its effect is twofold: it reflects both the overall evolution of the threat posed by external fires to the community and characterizes the mechanism by which flying embers ignite buildings over long distances, thereby improving the ability to describe the complex propagation characteristics of WUI fires.

[0063] S3. Construct and parameterize the building layer dynamic Bayesian network: Construct a building layer dynamic Bayesian network with each building as a node to represent the temporal propagation relationship of building states; parameterize the conditional probability table of the building layer dynamic Bayesian network based on the physical driving mechanism.

[0064] After obtaining external driving input, a building-layer dynamic Bayesian network is constructed with each building in the community as a node. For any building, its state at the current time step is not only related to its own state at the previous time step, but also influenced by the combustion state of neighboring buildings and the fire exposure input at the current time step. Therefore, the propagation of building-layer states is jointly determined by its own development law, the fire propagation effect between neighbors, and the ignition effect of external fire. Thus, a building-layer dynamic Bayesian network is constructed with each building as a node, as follows: Figure 4 As shown, the set of parent nodes for the current time step of each building includes the state node of the previous time step of the building, the state nodes of neighboring buildings at the current time step, and the fire exposure amount node that the current time step applies to the building.

[0065] The innovation of this embodiment lies not in simply using the general tool of dynamic Bayesian network, but in that, under a unified state space, the three types of effects—the evolution of the building itself, the propagation of neighboring buildings, and the remote ignition of flying fire—are all incorporated into the set of parent nodes of the same building state node. Furthermore, a physical-driven approach is used to parameterize the conditional probability table, thereby forming a building-scale temporal probability propagation model suitable for WUI fire scenarios.

[0066] Specifically, such as Figure 4 As shown, this embodiment constructs a building-layer dynamic Bayesian network (DBN) with each building as a node. Let the set of buildings be . . Figure 4 middle, Indicates the first The time step The status node of each building This indicates the building At the current time step Updated state nodes after the propagation of the "Flying Fire" effect in the Inner Canon. Indicates the action on the first The fire input node for each building. Figure 4 In the time step, "Time Step 1", "Time Step 2" to "Time Step T" represent the structure of the building layer dynamic Bayesian network after it has been expanded along the time axis. For any building... Define it at time step The discrete fire states are:

[0067]

[0068] In the above formula, Safe, Ignite, Fire, and Burned correspond to the unburned state, ignition state, fully burned state, and burnout state, respectively. Therefore, the conditional dependency of a building layer within a single time step can be written as follows:

[0069]

[0070] In the above formula, Indicates the building in the building's floors State Nodes The set of parent nodes; Representation and architecture A set of neighboring building indexes that have direct propagation influence relationships.

[0071] This architecture makes the output of the building layer inherently probabilistic and can couple with the inputs of the landscape layer and the fire layer under the same semantics, thus laying the architectural foundation for the output of the temporal probabilistic risk field.

[0072] To ensure the interpretability of the DBN output and avoid "purely empirical scoring," this embodiment employs a physical quantity-driven approach to parameterize the DBN's Conditional Probability Table (CPT). To characterize the uncertainty of the ignition threshold and material heterogeneity, a logistic probability response function is introduced, representing the physical driving quantity... The mapping is to the ignition probability, and the functional expression of the ignition probability is as follows:

[0073]

[0074] In the above formula, Indicates the median threshold. This represents the steepness coefficient.

[0075] In this embodiment, the conditional probability table of the building layer dynamic Bayesian network is parameterized using a physical quantity-driven approach. Specifically, it includes the probability of fire ignition driven by the amount of fire exposure, the probability of radiative ignition driven by the thermal radiation from neighboring buildings, and the state transition probability of internal development after the building is ignited.

[0076] (1) The calculation of the probability of spark ignition driven by spark exposure includes:

[0077] Based on the cumulative fire exposure of the building within a single time step as the driving force, the probability of the building transitioning from an unburned state to an ignited state (the transition probability of Safe→Ignite) is calculated using a preset ignition probability function (i.e., the function of formula (4)). For example, it can be written as:

[0078]

[0079] In the above formula, This indicates that the building was damaged by flying fire within the current time step. The probability of transitioning from a safe state to an ignited state; Represents architecture At the current time step The cumulative fly-through exposure within the building, specifically the cumulative fly-through mass deposited on the building's surface; The steepness coefficient of the fly spark ignition probability curve is used to control the effect of changes in fly spark exposure on the rate of increase in ignition probability. This represents the characteristic fly-fire exposure threshold corresponding to a 50% probability of ignition by a flying spark, i.e., the median threshold of fly-fire intensity. If differences in building materials, roof types, or fire resistance performance are considered, the parameter... and parameters It can be adjusted according to the building attributes, or obtained through experimental calibration, historical case fitting or experience setting, and the state transition terms caused by the flying fire effect can be filled accordingly.

[0080] (2) The calculation of the probability of radiative ignition driven by thermal radiation from neighboring buildings specifically includes:

[0081] Based on the cumulative radiation energy or equivalent exposure of the target building to the neighboring burning buildings as the driving force, the neighboring buildings are calculated using a preset radiation ignition probability function (i.e., the function of formula (4)). Cause the target building The probability of ignition, for example, can be written as:

[0082]

[0083] In the above formula, Indicates neighboring buildings For the target building The probability of it being ignited by thermal radiation; Indicates the neighboring buildings within the current time step. Acting on the target building The cumulative radiant energy is used to characterize the intensity of the thermal impact of neighboring building fires on the target building; The steepness coefficient of the radiation ignition probability curve is used to control the effect of changes in heat exposure on the rate of increase in ignition probability. This represents the median threshold corresponding to a 50% probability of radiative ignition, i.e., the median threshold at which the target building will ignite under thermal radiation. The total probability of radiative ignition is then obtained in the union form of "ignited at least once" under the conditional independence approximation. :

[0084]

[0085] This allows the neighborhood propagation effect to be incorporated into the CPT in a computable and interpretable manner.

[0086] (3) The probability of the internal state transition after the building is ignited includes the probability of the building developing from the ignition state to the fully burning state and the probability of the building developing from the fully burning state to the burnout state.

[0087] This embodiment describes the internal development process of a building after it has been ignited using state transition probabilities at discrete time steps. Specifically, when a building is in an ignited state, there is a certain probability that within the current time step... It further develops into a fully combusted state; when the building is in a fully combusted state, there is a certain probability that it will further develop into a fully combusted state in subsequent time steps. It further develops into a burnout state. To ensure the completeness of the state evolution, this embodiment sets the burnout state as an absorption state, that is, once the building enters the burnout state, it remains in the burnout state for subsequent time steps and does not transition to other states.

[0088] Through the above processing, the internal development process of a building can be uniformly incorporated into the conditional probability table of a dynamic Bayesian network, thus forming a building state transition mechanism together with fire exposure and neighborhood propagation. In this way, the output of the building-level dynamic Bayesian network is no longer a single deterministic judgment at a certain time step, but rather a multi-state probabilistic risk outcome that evolves over time.

[0089] S4. Construct an observation model: Establish a mapping relationship between the internal state of the building layers and the external observation categories.

[0090] To ensure that external observation information during a fire can be used to influence prediction results, it is necessary to establish a mapping relationship between the internal state of a building floor and the categories of external observations. Specifically, let's assume... The architectural observation set is The corresponding observation model can be expressed as follows:

[0091]

[0092] In the above formula, Represents the state mapping function, Represents architecture At time step The internal state, Indicates that the building At time step The observation categories are obtained by mapping the internal state.

[0093] Since external observations are typically not direct observations of the four states inside a building, it is necessary to map the discrete fire states of building nodes to external observation categories using the observation operators of the observation model. This allows for seamless integration between the internal states of building floors and external observation information through a unified interface. The external observation categories are set to a binary classification based on the actual data format, mapping to two categories: undamaged and damaged.

[0094] Step S4 aims to establish an interface between the internal state of the building layers and the external observation space. Its effect is to provide a unified observation representation basis for disaster-assimilation updates, enabling the unified probability propagation prediction method described in this invention to incorporate actual perceived information rather than solely relying on forward simulation results.

[0095] S5. Bayesian Update-Based State Correction and Rolling Prediction: After obtaining external observations, the system state is updated using Bayesian methods based on the observation information to obtain the posterior state distribution. The posterior distribution is then used as the initial state for prediction in the next time step, forming a rolling assimilation closed loop. Finally, a multi-state probability risk field that evolves at the building scale over time is output.

[0096] Specifically, when no new external observations are obtained at a certain time step, the system performs prior predictions based on the external driving model and the building-layer dynamic Bayesian network. After obtaining external observations for the current time step, the system updates its state using Bayesian methods based on the external driving model, the building-layer dynamic Bayesian network, and the observation model, combined with the external observation information for the current time step, to obtain a posterior state distribution reflecting the true fire situation at the current time step, as follows:

[0097]

[0098] In the above formula, Indicates as of Cumulative observation information, This represents the prior prediction result obtained by combining historical observation information. This represents the likelihood term corresponding to the observation at the current time step.

[0099] Through the above update process, the system state can be consistently corrected using newly acquired building observation information to obtain a posterior state distribution reflecting the actual fire situation at the current time step. This updated posterior distribution is then used as the initial state for fire propagation prediction in the next time step, continuing the process and forming a rolling assimilation loop of "prediction-observation-correction-reprediction." Ultimately, this outputs a multi-state probability risk field that evolves at the building scale over time, achieving dynamic assimilation prediction of the WUI fire propagation process.

[0100] Step S5 aims to endow the unified probability propagation prediction method of this invention with dynamic update capability during disasters. Its effect is that it not only improves the accuracy of the building state estimation at the current time step, but also enables the transfer of the correction gain from observations to subsequent prediction windows, thereby enhancing the rolling prediction capability and practical value of the unified probability propagation prediction method of this invention.

[0101] To verify the applicability and effectiveness of the unified probability propagation prediction method for WUI fires described in this invention in real WUI fires, a fire that occurred on January 8th of a certain year in a certain country (let's call this fire H) is selected as an example scenario. This fire event H has typical "field-town" cross-media propagation characteristics, with the fire spreading rapidly from the upstream field canyon area into the town (let's call it town P) and then within the building complex. This is suitable for testing the unified representation ability of this invention for landscape layer propagation, building layer propagation, and flying spark ignition, as well as the feasibility of the temporal probability output and discrete observation update at the building scale.

[0102] The study area in this embodiment covers the upstream initial ignition zone and the downstream town of P, with a spatial range of approximately 60 km × 60 km. It is rasterized with a 30 m spatial resolution to represent the combustion state and propagation process at the landscape layer. At the building layer, 11,854 buildings within town P are used as the objects for building propagation and risk output. The simulation period for the fire event H is selected as January 8th of a certain year, from 06:30 to 12:30 local time. The period from 09:30 to 12:30 local time is considered the "urban intrusion phase" after the fire enters town P, used to focus on verifying the ability to predict and prioritize building-level propagation.

[0103] In this embodiment, the landscape layer uses fuel information and wind field as environmental inputs to drive fire spread; the building layer uses building state nodes and their neighborhood interactions and local flyfire exposure as key inputs, and outputs the multi-state probability distribution of each building at each discrete time step. To demonstrate the "assimilation" capability of this invention, this embodiment sets up two comparative scenarios: Scenario A is open-loop prediction without introducing external observation updates; Scenario B introduces building-level observation information for updates at discrete observation time steps, and continues to roll the updated posterior state forward for prediction, to compare the corrective effect of discrete information updates on subsequent prediction windows.

[0104] Regarding the landscape propagation effect, this embodiment uses the fire line obtained by inversion based on NEXRAD (Next Generation Radar) as a control to compare the degree of agreement between the landscape propagation output of this invention and the output of the traditional wildfire simulation FARSITE (Fire Area Simulator). Figure 5 As shown, the results indicate that the method of this invention achieves higher spatial consistency at multiple key time steps, measured by IoU (Intersection over Union): the IoU remains above 0.6 from 08:00 to 10:00 local time. This result demonstrates that the unified state space described in this invention possesses more stable spatial consistency in characterizing event-scale propagation trajectories, providing a more reliable external propagation background for subsequent building-level propagation and risk output.

[0105] In the building layer effect verification, the method of this invention outputs a multi-state probability risk field of buildings at each time step, and can form a high-risk priority list based on this to serve disaster response. To reflect the application semantics of "prioritizing the handling of limited resources", this embodiment uses Top-n recall rate as the core effect indicator: at each prediction time step, buildings are sorted from high to low predicted risk, and the top few (preferably 10% of the total number of buildings in this embodiment) are selected. The proportion of buildings covered by these buildings that subsequently ignite and are destroyed is calculated to obtain the "cumulative recall rate (Top-10%)", which is used to characterize the method's ability to capture "hot spots in the next time window".

[0106] like Figure 6 As shown, under open-loop prediction without introducing external observation information updates, the present invention can still form a usable hotspot identification capability in the early stage of a fire entering a community: in the early critical stage (approximately 08:30–09:30 local time), the cumulative recall rate (Top-10%) is relatively high, with a peak of about 60%, indicating that the method of the present invention can compress most of the subsequent ignition buildings into a smaller high-risk set, providing a basis for rapid deployment in the early stage; however, as the prediction time window is extended, the accumulation of errors in open-loop propagation will cause the risk ranking to gradually drift, and the recall rate will drop significantly in the later stage, dropping to below 15% in the stage close to 11:00 local time, indicating that it is difficult to maintain a stable hotspot capture capability in the long term by relying solely on pure prediction.

[0107] Furthermore, to demonstrate the "assimilation" effect of the method described in this invention, this embodiment introduces external observation information to perform probability-consistent correction on the risk field at several discrete time steps. Specifically, assimilation is performed at 8:30, 8:30-9:00, and 8:30-9:30 respectively. Results show that observation updates can significantly improve the hotspot capture effect of subsequent prediction windows, such as... Figure 7 As shown: After each update, the risk ranking undergoes a targeted rearrangement, with a more pronounced spatial alignment between the high-risk set and subsequent ignition points, resulting in a continuous increase in cumulative recall (Top-10%). Additionally, as... Figure 8 As shown, compared to only performing one update ( Figure 8 Compared to the yellow bars displayed in the middle, the scrolling multiple updates ( Figure 8 (The blue bars in the image show that the method can more stably suppress error accumulation, allowing the improvement effect to be maintained over a longer subsequent window, thereby enhancing the ability of disaster risk forecasting to support resource scheduling and dynamic redeployment.) The above results demonstrate that the method of this invention can not only output a building-scale temporal probabilistic risk field, but also absorb external information at discrete information update time steps and consistently correct the risk field, thus forming a quantitative gain that can be used for disaster hotspot identification and decision support.

[0108] II. Example 2:

[0109] This embodiment discloses a unified probability propagation prediction system applicable to WUI fires, including modules:

[0110] State Space Definition Module: Used to construct a unified state space including building layer state, landscape layer state and flying fire layer state for WUI fire propagation process, and to define the discrete fire state of each building in the building layer;

[0111] External driving building module: used to construct an external driving model for fire propagation in building layers. The external driving forces include two parts: fire evolution in the landscape layer and fire transport and deposition.

[0112] Dynamic Bayesian Network Construction Module: Used to construct a building-layer dynamic Bayesian network with each building as a node to represent the temporal propagation relationship of building states; parameterizes the conditional probability table of the building-layer dynamic Bayesian network based on a physical driving mechanism;

[0113] Observation model construction module: used to establish the mapping relationship between the internal state of a building layer and the external observation categories;

[0114] State update module: After obtaining external observations, it combines the observation information to perform Bayesian update on the system state, obtains the posterior state distribution, and uses the posterior distribution as the initial state for prediction in the next time step, forming a rolling assimilation closed loop, and finally outputs a multi-state probability risk field that evolves at the building scale over time.

[0115] For the system implementation, since it basically corresponds to the method implementation, the specific implementation process of the functions and roles of each module in the system can be found in the implementation process of the corresponding steps in the method described in Embodiment 1, and will not be repeated here.

[0116] The above description only outlines the basic principles and preferred embodiments of the present invention. Those skilled in the art can make many changes and modifications based on the above description, and these changes and modifications should fall within the protection scope of the present invention.

Claims

1. A unified probability propagation prediction method applicable to WUI fires, characterized in that, Includes the following steps: Constructing and Discretizing a Unified State Space: Constructing a unified state space including the state of the building layer, the state of the landscape layer, and the state of the flying fire layer for the WUI fire propagation process, and defining the discrete fire state of each building in the building layer. Establish external drivers: Construct an external driving model for fire propagation in building layers. The external driving forces include two parts: fire evolution in the landscape layer and fire transport and deposition. Construct and parameterize a dynamic Bayesian network for the building layer: Construct a dynamic Bayesian network for the building layer with each building as a node to represent the temporal propagation relationship of building states; The conditional probability table of the building layer dynamic Bayesian network is parameterized based on the physical driving mechanism. Specifically, the conditional probability table of the building layer dynamic Bayesian network is parameterized using a physical quantity-driven method. Specifically, this includes the probability of fire ignition driven by the fire exposure amount, the probability of radiation ignition driven by the thermal radiation of neighboring buildings, and the state transition probability of internal development after the building is ignited. The calculation of the probability of ignition of flying fire driven by the amount of flying fire exposure specifically includes: based on the cumulative amount of flying fire exposure that the building experiences in a single time step as the driving factor, and using a preset ignition probability function, calculating the probability of the building transitioning from an unburned state to an ignited state. The calculation of the probability of radiative ignition driven by thermal radiation from neighboring buildings specifically includes: using the cumulative radiative energy or equivalent exposure of the target building to the burning neighboring buildings as the driving force, and using a preset ignition probability function, calculating the probability that the neighboring buildings cause the target building to ignite. The probability of state transition in the internal development of a building after it is ignited specifically includes: the probability of the building developing from an ignition state to a fully burning state and the probability of developing from a fully burning state to a burnout state; Construct an observation model: Establish a mapping relationship between the internal state of the building layers and the external observation categories; State correction and rolling prediction based on Bayesian update: After obtaining external observations, the system state is updated using Bayesian methods based on the observation information to obtain the posterior state distribution. This posterior distribution is then used as the initial state for the next time step prediction, forming a rolling assimilation closed loop. Finally, a multi-state probability risk field that evolves at the building scale over time is output, specifically including: After obtaining the external observation information at the current time step, the system state is updated using Bayesian methods based on the external driving model, the building-level dynamic Bayesian network, and the observation model, combined with the external observation information at the current time step. This yields a posterior state distribution that reflects the actual fire situation at the current time step. The updated posterior distribution is then used as the initial state for fire propagation prediction at the next time step and continues to advance, thus forming a rolling assimilation closed loop of "prediction-observation-correction-reprediction". Finally, a multi-state probability risk field that evolves at the building scale over time is output, achieving dynamic assimilation prediction of the WUI fire propagation process.

2. The method according to claim 1, characterized in that, The discrete fire states of each building in the building layer include four types: unburned state, ignition state, fully burning state, and burnout state, which are used to represent the staged evolution characteristics of the building during the fire propagation process.

3. The method according to claim 1, characterized in that, The landscape layer fire evolution is based on topography, fuel, wind field and initial fire data. It advances the temporal spread process of wildfires outside the community on the landscape layer and obtains the combustion status of the landscape grid, fire line advancement information and related heat release characteristics at each time step.

4. The method according to claim 1 or 3, characterized in that, Based on the results of fire evolution in the landscape layer, the method of flying fire transport and deposition identifies units in the landscape layer and building layer that are in an active combustion state as flying fire source items, and calculates the migration and deposition process of flying fire particles in space by combining at least wind speed, wind direction, transport distance and deposition probability, so as to obtain the amount of flying fire exposure borne by each building at the current time step.

5. The method according to claim 1, characterized in that, A dynamic Bayesian network for building layers is constructed using each building as a node, specifically including: A building-layer dynamic Bayesian network is constructed with each building as a node. The set of parent nodes for the current time step of each building's state node includes the state node of the previous time step of the building, the state nodes of neighboring buildings at the current time step, and the fire exposure node that acts on the building at the current time step.

6. The method according to claim 1, characterized in that, Establishing a mapping relationship between the internal state of a building floor and the categories of external observations specifically includes: The observation operators of the observation model map the discrete fire status of building nodes to external observation categories, so as to realize the connection between the internal status of building floors and external observation information under a unified interface. The external observation categories include two categories: undamaged and damaged.

7. A unified probability propagation prediction system applicable to WUI fires, characterized in that, Includes modules: State Space Definition Module: Used to construct a unified state space including building layer state, landscape layer state and flying fire layer state for WUI fire propagation process, and to define the discrete fire state of each building in the building layer; External driving building module: used to construct an external driving model for fire propagation in building layers. The external driving forces include two parts: fire evolution in the landscape layer and fire transport and deposition. Dynamic Bayesian Network Construction Module: Used to construct a dynamic Bayesian network for each building as a node, to represent the temporal propagation relationship of building states; The conditional probability table of the building layer dynamic Bayesian network is parameterized based on the physical driving mechanism. Specifically, the conditional probability table of the building layer dynamic Bayesian network is parameterized using a physical quantity-driven method. Specifically, this includes the probability of fire ignition driven by the fire exposure amount, the probability of radiation ignition driven by the thermal radiation of neighboring buildings, and the state transition probability of internal development after the building is ignited. The calculation of the probability of ignition of flying fire driven by the amount of flying fire exposure specifically includes: based on the cumulative amount of flying fire exposure that the building experiences in a single time step as the driving factor, and using a preset ignition probability function, calculating the probability of the building transitioning from an unburned state to an ignited state. The calculation of the probability of radiative ignition driven by thermal radiation from neighboring buildings specifically includes: using the cumulative radiative energy or equivalent exposure of the target building to the burning neighboring buildings as the driving force, and using a preset ignition probability function, calculating the probability that the neighboring buildings cause the target building to ignite. The probability of state transition in the internal development of a building after it is ignited specifically includes: the probability of the building developing from an ignition state to a fully burning state and the probability of developing from a fully burning state to a burnout state; Observation model construction module: used to establish the mapping relationship between the internal state of a building layer and the external observation categories; The state update module is used to perform a Bayesian update on the system state after obtaining external observations, combining the observation information to obtain the posterior state distribution. This posterior distribution is then used as the initial state for the next time step prediction, forming a rolling assimilation closed loop. Finally, it outputs a multi-state probability risk field that evolves at the building scale over time, specifically including: After obtaining the external observation information at the current time step, the system state is updated using Bayesian methods based on the external driving model, the building-level dynamic Bayesian network, and the observation model, combined with the external observation information at the current time step. This yields a posterior state distribution that reflects the actual fire situation at the current time step. The updated posterior distribution is then used as the initial state for fire propagation prediction at the next time step and continues to advance, thus forming a rolling assimilation closed loop of "prediction-observation-correction-reprediction". Finally, a multi-state probability risk field that evolves at the building scale over time is output, achieving dynamic assimilation prediction of the WUI fire propagation process.

Images