Data-driven building multi-room fire horizontal spread whole process prediction method

Abstract

The application discloses a data-driven building multi-room fire horizontal spread whole process prediction method, comprising the following steps: determining a plane structure physical model of a fire spread scene, and dividing a spread area; constructing a sample set, constructing a C-I-WOA-BP neural network based on a chaotic adaptive whale optimization algorithm, optimizing network parameter configuration, and predicting the ignition time of the room in the area closest to the fire source room; adding the ignition time of the room in the already ignited area into the sample set as the input of the neural network, and predicting the ignition time of the room in the next area; and obtaining a spread path according to the time sequence of the spread of the fire from the room where the fire breaks out to other rooms. The application improves the accuracy and efficiency of fire prediction, meets the demand for real building fire horizontal spread scenes in actual structure fire resistance research and fire rescue, and is helpful for designing a safer and more effective fire emergency response scheme.

Description

Technical Field

[0001] This invention relates to the field of building fire protection and structural fire resistance, specifically to a data-driven method for predicting the entire process of horizontal spread of multi-room fires in buildings. Background Technology

[0002] Research on fire resistance in building structures is receiving increasing attention. In most structural fire resistance studies, providing a realistic and accurate multi-room fire spread scenario is crucial. Current research largely relies on assumed spread paths and speeds, and has not yet comprehensively considered the impact of the randomness of fire-related factors on fire spread paths and the ignition time of each room.

[0003] The spread path of multi-compartment fires exhibits a significant nonlinear relationship with its influencing factors, making it impossible to accurately express this relationship using simple formulas. With the development of machine learning technology, back propagation (BP) neural networks have been widely applied in civil engineering for disaster risk prediction and management, material performance prediction, and structural health monitoring. Artificial neural networks (ANNs), represented by BP neural networks, only require given input parameters to output the desired prediction results through the "black box" within the model, demonstrating high prediction accuracy in many regression prediction problems. However, no applications of BP neural networks have yet been found in predicting the horizontal spread path and spread time of multi-compartment fires in buildings.

[0004] Existing research largely focuses on the fire spread patterns in large spaces and the propagation of smoke in shafts and corridors, emphasizing smoke characteristics that affect safe evacuation. Research on the more common multi-room fire spread is scarce. Such studies rely heavily on assumptions, such as fixed fire spread times and initial fire conditions, without considering crucial factors like the temperature characteristics of the affected rooms, thus limiting their scope. Furthermore, most fire spread studies are based solely on probabilistic models like tree structures or Bayesian networks. The potential of machine learning models (such as backpropagation neural networks) for understanding fire spread paths and ignition times in individual rooms remains largely untapped, lacking suitable optimization algorithms. Summary of the Invention

[0005] The technical problem to be solved by this invention is: This invention aims to provide a data-driven method for predicting the entire process of horizontal spread of multi-room fires in buildings, based on a BP neural network and its optimization algorithm (chaotic adaptive whale algorithm CI-WOA), to solve the problem of predicting the ignition time and spread path of each room in the case of horizontal spread of multi-room fires in office buildings.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A data-driven method for predicting the spread time and path of multi-room fires includes the following steps:

[0008] Step 1: Based on the boundary ventilation conditions, determine the physical model of the planar structure of the fire spread scenario;

[0009] Step 2: Based on the direction of spread and the relative consistency of ignition time within the area, divide the spread area;

[0010] Step 3: Construct a sample set and divide it into training, validation and test sets; construct a CI-WOA-BP neural network; optimize the network parameter configuration; and predict the ignition time of the room in the area closest to the fire source.

[0011] Step 4: Add the ignition time of the rooms in the already-burned area to the sample set as input to the neural network to predict the ignition time of the rooms in the next area;

[0012] Step 5: Based on the time sequence of the fire spreading from the room where it started to other rooms, obtain the spread path.

[0013] Furthermore, methods for constructing sample sets include:

[0014] Several sets of random parameters were generated using statistical methods and Latin hypercube sampling as sample data for the spread of multi-room fires.

[0015] The random parameters include room size, location of the fire source room, fire load density, opening factor, and heat release rate (HRR).

[0016] Multi-room fire spread simulation was conducted to generate an equal number of working conditions as the sample data and the corresponding ignition time of each room;

[0017] Methods for determining sample set size include:

[0018] (a) The sample set size is calculated based on the matching relationship between the sample set size N and the training error ε, using the following formula:

[0019] n ω =l(m+1)+m(n+1),

[0020] (b) The sample set size is calculated based on the fact that the sample set size N is a multiple of the total number of network connection weights, using the following formula:

[0021] N = k × (l·m + m·n),

[0022] In the formula, k is an empirical multiplier based on the total number of connection weights, and l, m, and n are the number of nodes in the input layer, hidden layer, and output layer, respectively.

[0023] The sample set is divided into a training set, a validation set, and a test set.

[0024] Furthermore, a three-layer CI-WOA-BP neural network model, i.e., a single hidden layer framework, is constructed as follows:

[0025] First, the neural network framework structure is determined to be a BP neural network model; then, the whale optimization algorithm is used to optimize the weights and biases of the BP neural network.

[0026] The population for the whale optimization algorithm is initialized using chaotic mapping;

[0027] During the learning process, an adaptive weighting factor I is introduced to adjust the network weight coefficients, which are parameters of the network topology. The network weights are automatically adjusted based on the current value and the target value.

[0028] The input layer parameters of the CI-WOA-BP neural network, including room width, depth, floor height, and maximum heat release rate (HRR), were determined using Spearman correlation coefficient and variance inflation coefficient. max In addition to the HRR duration, there are 5 variables in total; the room's fire time is an output layer variable;

[0029] The network is trained using the training set, and the model accuracy is calculated using the validation set after training.

[0030] Furthermore, the following physical model is constructed:

[0031] Construct a horizontal 4×4 span planar frame structure, with a fire spread area of ​​3×3 spans, containing 9 rooms from Room 1 to Room 9. Room 1 represents the corner room of the structure, and Room 9 represents the middle room of the structure. Rooms 1 to 9 are arranged sequentially in the 3×3 span fire spread area.

[0032] Furthermore, based on the location of the two types of fire sources in the rooms, the nine rooms in the fire spread area were divided into four zones along a diagonal partition;

[0033] When the fire source is located in corner room 1, the area is divided as follows: Area 1, which includes rooms 2 and 4; Area 2, which includes rooms 3, 5 and 7; Area 3, which includes rooms 6 and 8; Area 4, which includes room 9.

[0034] When the fire source is located in the middle room 9, the area is divided into: Area 1, which includes rooms 6 and 8; Area 2, which includes rooms 3, 5 and 7; Area 3, which includes rooms 2 and 4; and Area 4, which includes room 1.

[0035] Furthermore, Region 1 is the region closest to the room with the fire source. During network training, the ignition time of the room in Region 1 under different fire source locations is used as the output layer variable.

[0036] The ignition time of the already-burning room is added to the parameters of the neural network input layer, including:

[0037] When predicting the fire time of room 2, the fire time of room 1 is added to the neural network input layer parameters; when predicting the fire time of room 3, the fire times of rooms 1 and 2 are added to the neural network input layer parameters; when predicting the fire time of room 4, the fire times of rooms 1 to 3 are added to the neural network input layer parameters.

[0038] When the number of input layer samples gradually increases while adjacent rooms are not burning, there is no ignition time data, so null values ​​are deleted. These newly added samples in the input layer are minority class samples. SMOTE data augmentation is used, which synthesizes new samples by interpolation between minority class samples in the original dataset. The formula is:

[0039] Y = X + rand(0,1) * (X i -X),

[0040] In the formula, Y is the synthesized new sample, and X is... i Let X be the i-th nearest neighbor sample, and rand(0,1) represents a random number between 0 and 1.

[0041] Furthermore, the optimal configuration of network parameters is determined by training and validating the model parameters, including activation function, learning rate, training accuracy, number of iterations, and training algorithm, through trial and error.

[0042] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0043] First, this invention considers various factors affecting fire spread. Based on this, it determines specific random scenarios using survey statistics and Latin hypercube sampling, obtaining several sets of sampling results under different random fire scenarios. Based on this, it uses FDS for simulation, thus accurately reflecting the characteristics of real multi-room fire scenarios. Second, it comprehensively considers two representative fire source room locations: corners and the center, making it applicable to various multi-room fire spread scenarios. Third, based on the Chaotic Adaptive Whale Algorithm (CI-WOA-BP), it accurately predicts the fire spread path and the ignition time of each room using known fire source parameters. Fourth, by combining BP neural networks and the Chaotic Adaptive Whale Algorithm (CI-WOA), it effectively improves prediction accuracy, enhances the model's generalization ability in different fire scenarios, accurately depicts the dynamic process of fire spread, and infers the spread path. This is of great value for establishing an efficient early fire warning system, deepening the understanding of fire spread mechanisms, and formulating fire emergency plans.

[0044] This invention improves the accuracy and efficiency of fire prediction, meeting the needs of practical structural fire resistance research and fire rescue operations for real-world building fire horizontal spread scenarios. This invention provides technical support for establishing efficient early fire warning systems, deepening the understanding of fire spread mechanisms, and developing fire emergency plans. The method of this invention will help building designers and safety engineers better understand and predict the propagation behavior of fires in building complexes, thereby designing safer and more effective fire emergency response plans. Attached Figure Description

[0045] Figure 1 This is the overall flowchart of the method of the present invention.

[0046] Figure 2 This is a diagram showing the area of ​​fire damage to the inner and outer columns across a 3×3 span.

[0047] Figure 3 These are the temperatures measured at the same locations in each of the fire-affected rooms under both 4×4 and 5×5 span conditions.

[0048] Figure 4 This is a simplified diagram of the physical model.

[0049] Figure 5 This is a diagram showing the division of the fire spread area.

[0050] Figure 6 This is a schematic diagram of a neural network architecture.

[0051] Figure 7 This is a schematic diagram of the Smote oversampling principle.

[0052] Figure 8 It is the distribution of prediction errors of the optimized BP neural network on the test set.

[0053] Figure 9 This demonstrates the improvement in prediction performance of the BP neural network achieved through CI-WOA optimization.

[0054] Figure 10 It represents the prediction error distribution of the two optimization models, CI-WOA-BP and SSA-BP, on the test set.

[0055] Figure 11 This is a comparison between the predicted and actual ignition times of each room when the fire source is in room1.

[0056] Figure 12 This is a comparison between the predicted and actual ignition times of each room when the fire source is in room 9. Detailed Implementation

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0058] The data-driven method for predicting the entire process of horizontal spread of multi-room fires in buildings, as described in this invention, has the following flow: Figure 1 As shown, the specific implementation process of optimizing the BP neural network based on the chaotic adaptive whale algorithm (CI-WOA-BP) is as follows:

[0059] Step 1: Determine the physical model.

[0060] This invention aims to provide a realistic fire spread scenario for structural collapse resistance research. According to European regulations, a building structure is considered to be at risk of progressive collapse when the damaged area of ​​an external column reaches 15% of the total floor area after fire, or when the damaged area of ​​an internal column reaches 30%. Referring to physical models used in structural collapse resistance research, this embodiment of the invention constructs a 5×5 planar frame model, where the fire spread area is a 3×3 span, meeting the aforementioned collapse criteria of the European regulations.

[0061] Figure 2 This invention describes the fire damage area of ​​the inner and outer columns when the fire spread zone is a 3×3 span. It only provides boundary ventilation conditions for the room where the fire originates. When the simulated fire spreads to the 3×3 span area, the fire-damaged area of ​​the outer columns reaches more than 15% of the floor area, and the fire-damaged area of ​​the inner columns reaches more than 30% of the floor area, meeting the collapse criteria of the aforementioned European standards. Therefore, the preliminary physical model for multi-room fire spread is determined to be a 5×5 span horizontal frame structure, where the fire spread zone is a 3×3 span (9 rooms), and the other rooms provide boundary ventilation conditions corresponding to the 5×5 span for the room where the fire originates.

[0062] Considering the computational efficiency of the Fire Dynamics Simulator (FDS), horizontal 4×4 and 5×5 span frame structures were modeled to provide boundary ventilation conditions for a 3×3 span fire room. Burner power, combustible ignition point, and opening size were all set identically. Simulations were performed for both 4×4 and 5×5 span conditions, and the temperatures at the same locations in each fire room were compared under both conditions. (See [reference needed]). Figure 3 The temperature curves at the same location under both operating conditions basically overlap, indicating that the temperature rise curves of each room in the fire spread area are similar under both conditions. Therefore, it can be concluded that a 4×4 span model can replace the boundary ventilation conditions corresponding to a 5×5 span model. Thus, this embodiment simplifies the physical model as follows: Figure 4 The diagram shows a 4×4 span planar frame structure, with a 3×3 span fire spread area, comprising a total of 9 rooms.

[0063] like Figure 5As shown, two fire source room locations were considered, namely room1 and room9. In a 4×4 span structure, room1 represents a corner room of the structure with better ventilation (connected to the outside by a window), which can slow the spread of fire; room9 represents a central room of the structure with poorer ventilation. The two fire spread settings are designed to include different spread rates.

[0064] Step 2: Divide the area where the fire is spreading into four zones.

[0065] Considering the complexity of multi-room fire spread paths, this invention, based on the relatively consistent spread direction and ignition time within the area, and according to the distance from the room containing the fire source, divides the nine rooms in the fire spread area into four zones along a diagonal direction. For example... Figure 5 As shown, when the fire source is located in room 1 (corner room), the zones are divided as follows: Zone 1 (rooms 2 and 4), Zone 2 (rooms 3, 5, and 7), Zone 3 (rooms 6 and 8), and Zone 4 (room 9). Conversely, when the fire source is located in room 9 (middle room), the zone division order is adjusted to: Zone 1 (rooms 6 and 8), Zone 2 (rooms 3, 5, and 7), Zone 3 (rooms 2 and 4), and Zone 4 (room 1). This division facilitates the systematic analysis of the ignition time of each room under different fire source locations.

[0066] Step 3: Predict the time of fire in room 1 of area 1.

[0067] (1) Determining the sample set.

[0068] The random parameters considered to influence fire spread include room size, location of the fire source room, fire load density, opening factor, and fire source heat release rate curve.

[0069] This embodiment generated 354 sets of random parameter combinations through statistical analysis and Latin hypercube sampling. Considering the importance of the location of the fire source room, it focuses on two common fire source room locations (corner rooms and middle rooms). For each fire source, there are 177 sets of multi-room fire spread sample data. These samples cover fire spread under different fire source locations, room sizes, room opening factors, combustible material types, and heat release rates (HRR).

[0070] Multi-room fire spread simulations were performed using FDS, generating a number of operating conditions and their corresponding ignition times for each room, equal to the sample data. The input layer of the CI-WOA-BP neural network, including bay width, depth, floor height, and maximum heat release rate (HRR), was determined using Spearman correlation coefficient and variance expansion coefficient (VIF). max Including HRR duration, there are a total of 5 parameters.

[0071] There are two methods for estimating the size of the sample set:

[0072] (a) Based on the matching relationship between the sample set size N and the training error ε, we obtain:

[0073]

[0074] n ω = l(m+1) + m(n+1)

[0075] (b) The sample set size N should preferably be a certain multiple of the total number of network connection weights, resulting in:

[0076] N = k × (l·m + m·n)

[0077] In the formula, k is an empirical multiplier based on the total number of connection weights, and its value ranges from 5 to 10; l, m, and n are the number of nodes in the input layer, hidden layer, and output layer, respectively.

[0078] Based on empirical formulas and the specific parameters of the model structure of this invention, the sample set size N∈(90,890) was calculated. This invention considers two different initial fire source room designs: one in a corner room and the other in a central room. Each fire source room location has 177 sets of sample data. Given that 177 falls within the calculated sample set size range, and considering the balance between model requirements and actual data volume, the final sample set size N was determined to be 177.

[0079] (2) Sample set partitioning.

[0080] The partitioning of the sample set includes the ratio of training set, validation set, and test set, as well as the selection of a specific partitioning method. This embodiment comprehensively considers partitioning ratios of 8:1:1, 7:1.5:1.5, and 6:2:2, and allocation methods such as random partitioning, sequential partitioning, and bootstrap. Through five rounds of training and validation, based on the average prediction error, 8:1:1 was determined as the optimal partitioning ratio, and the sequential partitioning method was selected to allocate samples in the original order.

[0081] (3) Data standardization (normalization).

[0082] This embodiment compares four different data standardization methods: the extreme value method, the adjusted extreme value method, linear normalization, and Min-Max normalization. By performing five trials on each method and using the root mean square error as the evaluation metric, and following the principle of minimizing error, linear normalization was determined to be the optimal data standardization method.

[0083] (4) Determine the neural network framework structure.

[0084] A BP neural network model was constructed and optimized by combining the Whale Optimization Algorithm (WOA), Adaptive Weight Factor (I), and Chaotic Mapping Algorithm.

[0085] The Whale Optimization Algorithm (WOA) is inspired by the unique hunting behavior of humpback whales (bubble net fishing technique). The algorithm primarily searches the solution space by simulating three whale behaviors: surrounding the prey, spiraling towards the prey, and random search. In the surrounding phase, the whale determines the location of the prey (the optimal solution) and moves in its direction; in the spiraling approach phase, the whale approaches the prey in a spiral pattern; the random search phase is used to explore other regions of the solution space to prevent the global optimum from being located outside the current region.

[0086] This invention employs the Whale Optimization Algorithm (WOA) to optimize the weights and biases of a backpropagation (BP) neural network. To achieve a balance between global and local search capabilities and improve learning efficiency and prediction accuracy, an adaptive weight factor (I) is introduced during the learning process to adjust the network weight coefficients, i.e., parameters of the network topology, automatically adjusting the network weights based on the current and target values. Furthermore, a Cubic map from chaos theory is used to initialize the algorithm population, aiming to enhance the algorithm's capabilities, avoid getting trapped in local optima, and improve its application in complex optimization problems and convergence speed.

[0087] Based on the above methods, this invention proposes a Chaotic Adaptive Whale Algorithm (CI-WOA) to optimize BP neural networks. By organically combining the traditional Whale Algorithm (WOA), the adaptive mechanism (I), and the chaotic mapping (C), it not only enhances the performance of WOA but also optimizes the parameter adjustment of BP neural networks, demonstrating its application potential in complex system optimization problems.

[0088] The whale optimization algorithm mimics the whale hunting process and consists of three steps: encirclement, spiral approach, and random search. The core formula is as follows:

[0089] D=|C*X′(t)-X(t)|

[0090] X(t+1)=X′(t)-A*D

[0091] A = 2a * ra

[0092] C = 2r

[0093] In the formula, X′(t) is the optimal food (optimal solution) for the whale; X(t) is the current position of a whale, so D aims to estimate the straight-line distance between the two; C is a random factor, which makes the whale not approach the target by moving directly, but by randomly rotating towards the target, i.e., approaching in a spiral. Setting C to twice the random value can introduce randomness in the solution process and effectively balance global search and local search; X(t+1) describes the position of the whale at the next moment, i.e. the search direction of the optimization algorithm, and updates the position in real time according to the optimal distance. If the coefficient A is greater than 1, it indicates that the current food is not optimal (the current optimization direction is not optimal), and optimization needs to be moved to other directions; a decreases linearly from 2 to 0, and r is a random number in the interval 0 to 1. When |A|<1, the algorithm focuses on optimizing the current section. Therefore, as a decreases, the algorithm gradually changes from global search to local search.

[0094] Construct a three-layer CI-WOA-BP neural network model, i.e., a single hidden layer framework, such as... Figure 6 As shown. Area 1 is the area closest to the room containing the fire source. Select the room's width, depth, floor height, and maximum heat release rate (HRR). max Five variables, including the fire time and HRR duration, were used as input layer parameters; meanwhile, the ignition time of room 1 in different fire source locations was set as the unique output layer variable.

[0095] (5) The optimal configuration of network parameters is determined by performing five rounds of training and validation on the model parameters, such as weights and biases and hyperparameters, including activation function, learning rate, training accuracy, number of iterations and training algorithm, through trial and error. See Table 1.

[0096] Table 1

[0097]

[0098] (6) Training and testing of the model.

[0099] The selected network structure was systematically trained. After the model training was completed, the error between the training set and the validation set was calculated. The small error difference proved that the model achieved high accuracy in fitting the training data and also showed similar excellent performance on the validation set. No overfitting was observed, which shows that the model has good generalization ability.

[0100] The model was then validated on an independent test set to ensure the reliability of its generalization performance. Mean absolute error (MAE) was chosen as the metric for evaluating model performance.

[0101]

[0102] In the formula, N represents the number of test samples. *For the actual value of the k-th sample, This is the neural network prediction value for the k-th sample.

[0103] Step 4: Predict the time of fire in rooms 2, 3, and 4 of the area.

[0104] Taking a fire originating in room 1 (a corner room) as an example, when the fire spreads to rooms 3, 5, and 7 in area 2, rooms 2 and 4 in area 1 are already burning. This indicates that the ignition time of the already-burning rooms is a key factor affecting the prediction accuracy of subsequent rooms. To improve the prediction accuracy of the ignition time of rooms far from the fire source, the ignition time of the already-burning rooms is incorporated into the parameters of the neural network input layer.

[0105] Figure 6 This is a schematic diagram of the neural network architecture described in this invention. The neural network of this invention has a single hidden layer. When the output layer parameter is the fire time of room 1 in region 1, the input layer parameters are the room's width, depth, floor height, and maximum heat release rate (HRR). max The input layer parameters include five variables: the fire time of room 2 in zone 2, the fire times of room 2 and room 4 in zone 1, in addition to these five variables; when the output layer parameter is the fire time of room 3 in zone 3, the fire times of rooms in both zone 1 and zone 2 are included in addition to these five variables; when the output layer parameter is the fire time of room 4 in zone 4, the input layer parameter takes into account the fire times of zones 1 to 3 in addition to these five variables.

[0106] As the number of input layer parameters increases, the required dataset size also increases. When the input layer gradually increases while adjacent rooms are not burning, there is no ignition time data, resulting in null values ​​that need to be deleted. These newly added samples from the input layer become minority class samples. The SMOTE (Synthetic Minority Over-sampling Technique) data augmentation technique is used to synthesize new samples by interpolating between minority class samples in the original dataset. The principle of SMOTE oversampling is as follows: Figure 7 As shown. The core formula of this algorithm is:

[0107] Y = X + rand(0,1) * (X i -X),

[0108] In the formula, Y is the synthesized new sample, and X is... i Let X be the i-th nearest neighbor sample. `rand(0,1)` represents a randomly selected number between 0 and 1, used to control the range of the synthesized samples. The oversampling factor can be controlled by the number of nearest neighbor samples around the measurement point. This method effectively expands the sample coverage, improving the model's generalization ability and training effect.

[0109] Subsequently, the ignition time of the already-burning rooms was incorporated into the input layer parameters of the neural network. Specifically, when predicting the ignition time of room 2 in region 2, the ignition times of rooms 2 and 4 were included; for the prediction of the ignition time of room 3, the ignition times of rooms 1 and 2 in region 3 were added; and for the prediction of the ignition time of room 9 in region 4, the input layer parameters comprehensively considered the ignition times of regions 1 to 3. Based on this, the parameters and hyperparameters were optimized and retrained sequentially to determine the optimal configuration of the model (refer to the steps for predicting the ignition time of room 1 in region 1).

[0110] Accurate predictions were made for the ignition time of each room under different initial fire source locations. When the fire source was located in room 1 (corner room), the prediction accuracy for the fire occurrence time in all rooms exceeded 95%; when the fire source location changed to room 9 (middle room), the model's prediction accuracy further improved to over 96%. This high accuracy demonstrates the model's high sensitivity to changes in the fire source room location, as well as its excellent adaptability and accuracy.

[0111] Step 5: Predict the spread path.

[0112] Based on steps 3 and 4, the aforementioned neural network model is loaded. By continuously adding elements close to the area of ​​the room where the fire originated to the input layer of the neural network, the ignition events of neighboring rooms are gradually predicted, thereby deriving the entire fire spread path. This invention employs a comprehensive method to predict the spread path of multi-room fires under different initial fire source room locations. This method, based on a neural network, accurately predicts the ignition time of each room. It trains and optimizes the network by simulating spread paths under different room sizes, fire source locations, HRR, and fire source types, enabling it to systematically learn the time series of fire spread from the source point to surrounding rooms, thus achieving the prediction of multi-room fire spread paths.

[0113] This invention uses MATLAB software to establish a BP neural network prediction model optimized by the Chaotic Adaptive Whale Algorithm (CI-WOA). Figure 8The diagram illustrates the prediction error distribution of the optimized BP neural network on the test set. The model's relative prediction error is concentrated in the range of [-0.0783, 0.0399], with the maximum and minimum absolute errors being 0.0783 and 0.0045, respectively, reflecting the relatively stable prediction performance of the optimized model. Notably, the prediction error for most samples is close to 0, further validating the model's accuracy and generalization ability. Compared to the traditional BP neural network, the CI-WOA-BP neural network reduces the average absolute error to 3.76%, compared to 7.08% for the traditional method, demonstrating lower error fluctuation and higher prediction accuracy. Furthermore, no extreme error values ​​were observed, indicating that the model does not exhibit overfitting and possesses stronger generalization ability. The Chaotic Adaptive Whale Algorithm (CI-WOA) significantly enhances the performance of BP neural networks in complex system prediction tasks, demonstrating significant application potential.

[0114] Figure 9 This study demonstrates the improvement in prediction performance of the BP neural network through CI-WOA optimization. After optimization, the CI-WOA-BP neural network model shows a significantly better match between predicted and actual observations than the unoptimized BP network, exhibiting less fluctuation and higher prediction accuracy, indicating superior stability and reliability. Therefore, the Chaotic Adaptive Whale Algorithm (CI-WOA) effectively optimizes the weights and thresholds of the BP neural network, significantly improving the model's prediction accuracy and demonstrating great potential and practical value in handling complex data prediction problems.

[0115] Figure 10 The figure shows the prediction error distribution of two optimized models, CI-WOA-BP and SSA-BP, on the test set. For ease of representation, CI-WOA-BP and SSA-BP are denoted as WOA and SSA, respectively. The histogram provides a visual comparison of the frequency distribution of prediction errors for the two models. The figure shows that the error distribution of the CI-WOA-BP model is more concentrated near the zero error point than that of the SSA-BP model. Furthermore, the CI-WOA-BP error distribution curve has a higher peak near the zero point, indicating that most of its predictions are closer to the true values, demonstrating higher prediction accuracy. In addition, the CI-WOA-BP model's error distribution curve has shorter tails on both sides, meaning it produces large errors less frequently, reflecting better stability and reliability. In contrast, although the SSA-BP model's error distribution is relatively concentrated, its lower peak near the zero point and longer tails indicate that its predictions are less concentrated near the true values ​​and contain more large errors.

[0116] In summary, the Chaotic Adaptive Whale Algorithm (CI-WOA) demonstrates superior prediction performance compared to the Sparrow Algorithm (SSA). Specifically, its prediction error distribution is more concentrated, its mean absolute error is lower, and it exhibits higher stability and reliability across the entire prediction range. These characteristics enable the CI-WOA-BP model to provide more accurate and robust prediction results when facing real-world problems.

[0117] Figure 11 This is a comparison between the predicted and actual ignition times of each room by the CI-WOA-BP neural network when the fire source is in room1. Figure 12 This compares the predicted and actual ignition times of each room using the CI-WOA-BP neural network when the fire source is in room 9. It can be observed that the optimized CI-WOA-BP model significantly improves the accuracy of ignition time prediction and enhances the model's generalization ability, enabling it to more accurately depict the dynamic process of fire spread and infer the spread path.

Claims

1. A data-driven method for predicting the entire process of horizontal spread of multi-room fires in buildings, characterized in that: Includes the following steps: Step 1: Based on the boundary ventilation conditions, determine the physical model of the planar structure of the fire spread scenario. Specifically, construct a horizontal 4×4 span planar frame structure, in which the fire spread area is a 3×3 span, containing 9 rooms from Room 1 to Room 9. Room 1 represents the corner room of the structure, and Room 9 represents the middle room of the structure. Rooms 1 to 9 are arranged sequentially in the 3×3 span fire spread area. Step 2: Based on the direction of fire spread and the relative consistency of ignition time within the area, divide the fire spread area into four zones. Specifically, according to the location of the two types of fire sources in the rooms, divide the nine rooms in the fire spread area into four zones along a diagonal direction. When the fire source is in corner room 1, the zones are divided as follows: Zone 1, including rooms 2 and 4; Zone 2, including rooms 3, 5, and 7; Zone 3, including rooms 6 and 8; Zone 4, including room 9. When the fire source is located in the middle room 9, the area is divided into: Area 1, which includes rooms 6 and 8; Area 2, which includes rooms 3, 5 and 7; Area 3, which includes rooms 2 and 4; and Area 4, which includes room 1. Step 3: Construct a sample set and divide it into training, validation and test sets; construct a CI-WOA-BP neural network; optimize the network parameter configuration; and predict the ignition time of the room in the area closest to the fire source. Step 4: Add the ignition time of the rooms in the already-burned area to the neural network input layer parameters as the input of the neural network to predict the ignition time of the next area. Specifically: Area 1 is the area closest to the fire source room. During the network training process, the ignition time of the rooms in Area 1 under different fire source locations is used as the output layer variable. The ignition time of the already-burning room is added to the parameters of the neural network input layer, including: When predicting the fire time of room 2, the fire time of room 1 is added to the neural network input layer parameters; when predicting the fire time of room 3, the fire times of rooms 1 and 2 are added to the neural network input layer parameters; when predicting the fire time of room 4, the fire times of rooms 1 to 3 are added to the neural network input layer parameters. When the input layer parameters gradually increase while neighboring rooms are not burning, there is no ignition time data, so null values ​​are deleted. The samples added to this input layer are minority class samples. SMOTE data augmentation is used, which synthesizes new samples by interpolation between minority class samples in the original dataset. The formula is: ， In the formula, For the synthesized new sample, For the sample The i-th nearest neighbor sample, This indicates that a number between 0 and 1 will be randomly selected. Step 5: Based on the time sequence of the fire spreading from the room where it started to other rooms, obtain the spread path.

2. The data-driven method for predicting the entire process of horizontal spread of multi-room fires in buildings according to claim 1, characterized in that, Methods for constructing sample sets include: Several sets of random parameters were generated using statistical methods and Latin hypercube sampling as sample data for the spread of multi-room fires. The random parameters include room size, location of the fire source room, fire load density, opening factor, and heat release rate (HRR). Multi-room fire spread simulation was conducted to generate an equal number of working conditions as the sample data and the corresponding ignition time of each room; Methods for determining sample set size include: (a) According to The sample set size N is calculated based on the matching relationship with the training error ε, using the following formula: ， (b) The sample set size N is calculated based on the empirical scaling factor k based on the total number of connection weights, the number of input layer nodes l, the number of hidden layer nodes m, and the number of output layer nodes n, as follows: ， In the formula, k is an empirical multiplier based on the total number of connection weights, and l, m, and n are the number of nodes in the input layer, hidden layer, and output layer, respectively; the sample set is divided into a training set, a validation set, and a test set.

3. The data-driven method for predicting the entire process of horizontal spread of multi-room fires in buildings according to claim 1, characterized in that, The following method is used to construct a three-layer CI-WOA-BP neural network model, i.e., a single hidden layer framework: First, the neural network framework structure is determined to be a BP neural network model; then, the weights and biases of the BP neural network are optimized using the Whale Optimization Algorithm (WOA). The population of the Whale Optimization Algorithm (WOA) is initialized using the chaotic mapping C; During the learning process, an adaptive weighting factor I is introduced to adjust the network weight coefficients, which are parameters of the network topology. The network weights are automatically adjusted based on the current value and the target value. The input layer parameters of the CI-WOA-BP neural network, including room width, depth, floor height, and maximum heat release rate (HRR), were determined using Spearman correlation coefficient and variance inflation coefficient. max In addition to the HRR duration, there are 5 variables in total; the room's fire time is an output layer variable; The network is trained using the training set, and the model accuracy is calculated using the validation set after training.

4. The data-driven method for predicting the entire process of horizontal spread of multi-room fires in buildings according to claim 1, characterized in that, By using trial and error, the model's parameters, including activation function, learning rate, training accuracy, number of iterations, and training algorithm, are trained and validated to determine the optimal configuration of the network parameters.

Images