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

Abstract

The invention discloses a data-driven building multi-room fire horizontal spreading whole process prediction method, which comprises the following steps: determining a plane structure physical model of a fire spreading scene, and dividing a spreading area; a sample set is constructed, a C-I-WOA-BP neural network based on the chaos adaptive whale algorithm is constructed, network parameter configuration is optimized, and the ignition time of a room in an area closest to a fire source room is predicted; adding the ignition time of the ignited area room into a sample set as input of a neural network, and predicting the ignition time of the next area room; and a spreading path is obtained according to the time sequence of the fire spreading from the on-fire room to other rooms. According to the method, the accuracy and efficiency of fire prediction are improved, the requirements for a real building fire horizontal spreading scene in actual structure fire resistance research and fire rescue are met, and a safer and more effective fire emergency response scheme can be designed.

Description

Technical Field

[0001] The present invention relates to the field of building fire protection and structural fire resistance, and in particular to a data-driven method for predicting the entire process of horizontal spread of a fire in multiple rooms of a building. Background Art

[0002] Nowadays, research on fire resistance of building structures is gaining more and more attention. In most studies on fire resistance of structures, it is very important to provide a realistic and accurate multi-room fire spread scenario. Currently, most studies are based on assumed spread paths and spread speeds, and the impact of the randomness of fire influencing factors on the fire spread path and the fire time of each room has not been comprehensively considered.

[0003] There is a significant nonlinear relationship between the spread path of multi-room fire and its influencing factors, and it is impossible to accurately express the relationship through a simple formula. With the development of machine learning technology, back propagation (BP) neural network has been widely used in the field of civil engineering in disaster risk prediction and management, material performance prediction, structural health monitoring, etc. Artificial neural network (ANN) represented by BP neural network only needs to give input parameters to output the desired prediction results through the "black box" in the model. This technology has a high prediction accuracy in many regression prediction problems. However, the application of BP neural network in the prediction of horizontal spread path and spread time of multi-room fire in buildings has not yet been found.

[0004] Most existing studies focus on the spread of fire in large spaces and the spread of smoke between shafts and corridors. These studies focus on the smoke characteristic parameters that affect the safe evacuation of personnel. There are fewer studies related to the more common multi-room fire spread. Such studies use a large number of assumptions, such as fixed fire spread time, initial fire conditions, etc., and do not consider key factors such as the temperature characteristics of the fire-affected room. Therefore, the research conclusions have great limitations. In addition, most fire spread studies are based only on probabilistic models such as component trees or Bayesian networks. The application potential of machine learning models (BP neural networks, etc.) in fire spread paths and fire time in each room has not been fully developed, and there is a lack of optimization algorithms suitable for the stated goals. Summary of the invention

[0005] The technical problem to be solved by the present invention is: the present invention aims to provide a data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building, 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 under the horizontal spread of multi-room fire in an office building.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The data-driven prediction method for multi-room fire spread time and path includes the following steps:

[0008] Step 1: Determine the plane structure physical model of the fire spread scenario based on the boundary ventilation conditions;

[0009] Step 2: Divide the spreading area according to the relative consistency of the spreading direction and the fire ignition time in the area;

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

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

[0012] Step 5: According to the time series of the fire spreading from the room where the fire started to other rooms, the spreading path is obtained.

[0013] Furthermore, the method for constructing a sample set includes:

[0014] Several groups of random parameters are generated as sample data of multi-room fire spread through statistics and Latin hypercube sampling;

[0015] The random parameters include room size, fire source room location, fire load density, opening factor and heat release rate HRR;

[0016] Conduct multi-room fire spread simulation to generate the same number of working conditions as sample data and the corresponding fire time of each room;

[0017] Methods for determining the sample set capacity are:

[0018] (a) The sample set capacity is calculated based on the matching relationship between the sample set capacity N and the training error ε. The formula is as follows:

[0019] n ω =l(m+1)+m(n+1),

[0020] (b) The sample set capacity is calculated based on the sample set capacity N being a multiple of the total number of network connection weights. The formula is as follows:

[0021] N=k×(l·m+m·n),

[0022] Where k is the empirical multiplier factor based on the total number of connection weights, 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 training set, validation set and 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] Firstly, the neural network framework structure is determined to be a BP neural network model; the whale optimization algorithm is used to optimize the weights and biases of the BP neural network;

[0026] Using chaotic mapping to initialize the population of the whale optimization algorithm;

[0027] In the learning process, an adaptive weight factor I is introduced to adjust the network weight coefficient, that is, the parameter of the network topology structure, and the network weight is automatically adjusted according to the current value and the target value;

[0028] The input layer parameters of the CI-WOA-BP neural network, including room span, depth, floor height, and maximum heat release rate (HRR), are determined by the Spearman correlation coefficient and variance inflation coefficient. max and HRR duration, a total of 5 variables; the fire time of the room is the output layer variable;

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

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

[0031] A plane frame structure with a horizontal 4×4 span is constructed, in which the fire spread area is a 3×3 span, including 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 in sequence in the 3×3 span fire spread area.

[0032] Furthermore, according to the locations of the two fire source rooms, the nine rooms in the fire spread area were divided into four areas along the diagonal direction;

[0033] When the fire source is located in the corner room 1, the area is divided into: Area 1, including room 2 and room 4; Area 2, including room 3, room 5 and room 7; Area 3, including room 6 and room 8; Area 4, including room 9;

[0034] When the fire source is set in the middle room 9, the areas are divided into: Area 1, including room 6 and room 8; Area 2, including room 3, room 5 and room 7; Area 3, including room 2 and room 4; Area 4, including room 1.

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

[0036] The fire time of the burned room is added to the neural network input layer parameters, including:

[0037] When predicting the fire time of the room in area 2, the fire time of the room in area 1 is added to the neural network input layer parameters; when predicting the fire time of the room in area 3, the fire time of the rooms in areas 1 and 2 is added to the neural network input layer parameters; when predicting the fire time of the room in area 4, the fire time of the rooms from area 1 to area 3 is added to the neural network input layer parameters;

[0038] When the input layer samples gradually increase and the adjacent rooms are not burning, there is no ignition time data, and the null value data is deleted. The samples added in this type of input layer are minority class samples. SMOTE data enhancement is used to synthesize new samples by interpolating between minority class samples in the original data set. The formula is:

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

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

[0041] Furthermore, the parameters of the model, including activation function, learning rate, training accuracy, number of iterations and training algorithm, are trained and verified through trial algorithms to determine the optimal configuration of network parameters.

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

[0043] First, the present invention takes into account a variety of factors that affect the spread of fire. Based on this, the specific random scenarios are determined based on survey statistical data and Latin hypercube sampling, and a total of several groups of sampling results under different fire random scenarios are obtained. On this basis, FDS is used for simulation, so it can accurately reflect the characteristics of real multi-room fire scenes; second, the two representative fire source room locations of corners and centers are comprehensively considered, which is suitable for a variety of multi-room fire spread scenarios; third, based on the chaos adaptive whale algorithm (CI-WOA-BP), the fire spread path and the ignition time of each room are accurately predicted through known fire source parameters; fourth, the BP neural network and the chaos adaptive whale algorithm (CI-WOA) are combined at the same time to effectively improve the prediction accuracy, enhance the generalization ability of the model in the face of different fire scenarios, and can also accurately describe the dynamic process of fire spread and infer the spread path, which is of great value for establishing an efficient early warning system for fire, deepening the understanding of the fire spread mechanism and formulating fire emergency plans.

[0044] The present invention improves the accuracy and efficiency of fire prediction, and meets the needs of actual structural fire resistance research and fire rescue for real building fire horizontal spread scenarios. The present invention provides technical support for establishing an efficient early fire warning system, deepening the understanding of the fire spread mechanism, and formulating fire emergency plans. The method of the present invention will help building designers and safety engineers better understand and predict the spread behavior of fire in building complexes, so as to design safer and more effective fire emergency response plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is the overall flow chart of the method of the present invention.

[0046] Figure 2 It is a diagram of the area of ​​fire damage to inner and outer columns when the fire spread area is 3×3 span.

[0047] Figure 3 It is the temperature of the measuring point at the same position in each fire room under the two working conditions of 4×4 span and 5×5 span.

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

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

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

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

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

[0053] Fig. 9 It is the improvement effect of CI-WOA optimization on the prediction performance of BP neural network.

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

[0055] Fig.11 It is the comparison between the predicted value and the actual value of the fire time of each room when the fire source is in room1.

[0056] Fig.12 It is the comparison between the predicted value and the actual value of the fire time of each room when the fire source is in room 9. DETAILED DESCRIPTION

[0057] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0058] The data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building according to the present invention has the following process: Figure 1 As shown in the figure, based on the chaotic adaptive whale algorithm to optimize the BP neural network (CI-WOA-BP), the specific implementation process is as follows:

[0059] Step 1: Determine the physical model.

[0060] The present invention aims to provide a real fire spread scenario for structural anti-collapse research. According to European regulations, when the damaged area of ​​the outer column after fire reaches 15% of the total floor area, or the damaged area of ​​the inner column reaches 30%, the building structure is considered to have a risk of continuous collapse. With reference to the physical model in the structural anti-collapse research, the embodiment of the present invention constructs a 5×5 plane frame model, in which the fire spread area is a 3×3 cross-area, which meets the collapse judgment criteria of the above-mentioned European regulations.

[0061] Figure 2 It is a diagram of the area of ​​fire damage to inner and outer columns when the fire spread area is 3×3 spans as described in the present invention. Among them, boundary ventilation conditions are only provided for the room on fire. When simulating the fire spreading to the 3×3 span area, the area of ​​damage to the outer columns caused by the fire reaches more than 15% of the floor area, and the area of ​​damage to the inner columns caused by the fire reaches more than 30% of the floor area, which meets the collapse judgment standard of the above-mentioned European regulations. Therefore, it is preliminarily determined that the physical model of multi-room fire spread is a horizontal frame structure with 5×5 spans, in which the fire spread area is 3×3 spans (9 rooms), and other rooms provide boundary ventilation conditions corresponding to 5×5 spans for the rooms on fire.

[0062] Considering the computational efficiency of the Fire Dynamics Simulator (FDS), horizontal 4×4 span and 5×5 span frame structures were modeled to provide boundary ventilation conditions for the 3×3 span fire room. The burner power, combustible ignition point, and opening size were set the same. The two working conditions of 4×4 span and 5×5 span were simulated respectively, and the temperatures of the measuring points at the same position in each fire room under the two working conditions were compared. Figure 3 , the temperature curves of the measuring points at the same position under the two working conditions are basically the same, indicating that the temperature rise curves of each room in the fire spread area under the two working conditions are similar. It can be seen that the boundary ventilation conditions corresponding to the modeling of 4×4 spans can replace the 5×5 spans. Therefore, this embodiment simplifies the physical model as follows Figure 4 The plan frame structure shown is a 4×4 span structure with a fire spread area of ​​3×3 spans, for a total of 9 rooms.

[0063] like Figure 5As shown in the figure, two fire source room locations are considered, namely room1 and room9. In a 4×4 span structure, room1 represents the corner room of the structure, which has better ventilation conditions (connected to the outside world by windows) and can slow down the spread of fire; room9 represents the middle room of the structure, which has poor ventilation conditions. The two settings of fire spread are intended to contain different spread speeds.

[0064] Step 2: Divide the fire spread area into four areas.

[0065] Considering the complexity of the multi-room fire spread path, the present invention divides the 9 rooms in the fire spread area into 4 areas along the diagonal partition according to the relative consistency of the spread direction and the fire time in the area and the distance from the room of the fire source. Figure 5 As shown in the figure, when the fire source is in room 1 (corner room), the area is divided into: area 1 (room 2, room 4), area 2 (room 3, room 5, room 7), area 3 (room 6, room 8), area 4 (room 9); in contrast, when the fire source is set in room 9 (middle room), the order of area division is adjusted to: area 1 (room 6, room 8), area 2 (room 3, room 5, room 7), area 3 (room 2, room 4), area 4 (room 1). This division facilitates the systematic analysis of the fire time of each room under different fire source locations.

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

[0067] (1) Sample set determination

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

[0069] This embodiment generates 354 sets of random parameter combinations through statistics and Latin hypercube sampling. Considering the importance of the fire source room location, two common fire source room locations (corner room and middle room) are focused on, and 177 sets of multi-room fire spread sample data are obtained for each fire source. These samples cover the fire spread under different fire source locations, room sizes, room opening factors, combustible types, and heat release rates (HRR).

[0070] The multi-room fire spread simulation was performed through FDS, and the same number of working conditions and the corresponding fire time of each room were generated. The input layer of the following CI-WOA-BP neural network was determined by the Spearman correlation coefficient and the variance inflation factor (VIF), including the span, depth, floor height, and maximum heat release rate HRR. max and HRR duration, a total of 5 parameters.

[0071] There are two estimation methods for determining the sample set capacity:

[0072] (a) According to the matching relationship between the sample set capacity N and the training error ε, we get:

[0073]

[0074] n ω =l(m+1)+m(n+1)

[0075] (b) The sample set capacity N should be a certain multiple of the total number of network connection weights, and we get:

[0076] N = k × (l·m+m·n)

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

[0078] According to the empirical formula and the specific parameters of the model structure of the present invention, the sample set capacity N∈(90,890) is calculated. The present invention considers two different initial fire source room designs: the fire source is in the corner room and the fire source is in the middle room. Each fire source room has 177 sets of sample data. Given that 177 is within the calculated sample capacity range and based on the consideration of the balance between model requirements and actual data volume, the sample set capacity N is finally determined to be 177.

[0079] (2) Sample set division.

[0080] The division of the sample set includes the ratio of training set, validation set and test set, as well as the selection of specific division methods. This embodiment comprehensively considers the division ratios of 8:1:1, 7:1.5:1.5 and 6:2:2 and the allocation methods of random division, sequential division and bootstrap. After 5 rounds of training and verification, based on the average prediction error, 8:1:1 is determined to be the optimal division ratio, and the sequential division method is selected to allocate samples in the original order.

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

[0082] This example compares four different data normalization methods: the maximum value method, the adjusted maximum value method, linear normalization, and Min-Max normalization. By performing five trials on each of these methods, and using the mean of the root mean square error as the evaluation index, and following the criterion of minimizing the error, linear normalization is determined to be the optimal data normalization 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 chaos mapping algorithm.

[0085] The Whale Optimization Algorithm (WOA) is inspired by the unique hunting behavior of humpback whales (bubble net fishing technology). The algorithm mainly searches the solution space by simulating three behaviors of whales: surrounding prey, spiral approach to prey, and random search. In the surrounding prey phase, the whale determines the location of the prey (optimal solution) and moves towards it; in the spiral approach phase, the whale approaches the prey in a spiral form; the random search phase is used to explore other areas in the solution space in case the global optimal solution is outside the current area.

[0086] The present invention uses the whale optimization algorithm (WOA) to optimize the weights and biases of the BP neural network. In order to achieve a balance between global search 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 coefficient, i.e., the parameters of the network topology structure, and automatically adjust the network weight according to the current value and the target value. In addition, the algorithm population is initialized in combination with the Cubic map in chaos theory, aiming to enhance the algorithm capability, avoid falling into the local optimal solution, and improve the application in complex optimization problems and the convergence speed of the algorithm.

[0087] Combining the above methods, the present invention proposes a chaotic adaptive whale algorithm (CI-WOA) to optimize the BP neural network. By organically combining the traditional whale algorithm (WOA), adaptive mechanism (I) and chaotic mapping (C), it not only enhances the performance of WOA, but also optimizes the parameter adjustment of the BP neural network, showing its application potential in complex system optimization problems.

[0088] Among them, the whale optimization algorithm imitates the whale hunting process, which includes 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] Where 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 is used to prevent the whale from approaching the target by direct movement, but by randomly rotating toward the target, that is, spiral approach. 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, that is, the search direction of the optimization algorithm, and updates the position in real time according to the optimal distance. At this time, if the coefficient A is greater than 1, it means that the current food is not the best (the current optimization direction is not the best), and optimization needs to be made in other directions; a decreases linearly from 2 to 0, r is a random number in the range of 0 to 1, and when |A|<1, the algorithm focuses on optimizing the current node, so as a decreases, the algorithm gradually changes from global search to local search.

[0094] Construct a three-layer CI-WOA-BP neural network model, that is, a single hidden layer framework, such as Figure 6 As shown. Zone 1 is the area closest to the fire source room. Select the room span, depth, floor height, maximum heat release rate (HRR max ) and HRR duration are used as input layer parameters; at the same time, the fire time of the room in area 1 under different fire source locations is set as the only output layer variable.

[0095] (5) The model parameters, such as weights and biases, and hyperparameters, including activation function, learning rate, training accuracy, number of iterations, and training algorithm, were trained and verified for 5 rounds through trial and error algorithms to determine the optimal configuration of network parameters, see Table 1.

[0096] Table 1

[0097]

[0098] (6) Model training and testing.

[0099] The network structure of the selected configuration is systematically trained. After the model training is completed, the errors of the training set and the validation set are calculated. The small error difference proves that the model has achieved high accuracy in fitting the training data, and also shows similar excellent performance on the validation set without overfitting, which shows that the model has good generalization ability.

[0100] Subsequently, the model was validated on an independent test set to ensure the reliability of its generalization performance. The mean absolute error (MAE) was chosen as the criterion for measuring model performance.

[0101]

[0102] Where N represents the number of test samples.) *is the actual value of the kth sample, is the neural network prediction value of the kth sample.

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

[0104] Taking the fire source in room 1 (corner room) as an example, when the fire spreads to room 3, room 5 and room 7 in area 2, room 2 and room 4 in area 1 are already burning, indicating that the ignition time of the room that has been on fire is a key factor affecting the prediction accuracy of subsequent rooms. In order to improve the prediction accuracy of the ignition time of rooms far away from the fire source, the ignition time of the room that has been on fire is included in the input layer parameters of the neural network.

[0105] Figure 6 The neural network of the present invention is a schematic diagram of the neural network architecture. The neural network of the present invention is a single hidden layer. When the output layer parameter is the fire time of the room in area 1, the input layer parameters are the room span, depth, floor height, maximum heat release rate (HRR max ) and HRR duration; when the output layer parameter is the fire time of the room in area 2, in addition to these five variables, the fire times of room2 and room4 in area 1 are included; when the output layer parameter is the fire time of the room in area 3, in addition to these five variables, the fire times of the rooms in areas 1 and 2 are included; when the output layer parameter is the fire time of the room in area 4, in addition to these five variables, the input layer parameters comprehensively consider the fire times of areas 1 to 3.

[0106] As the number of input layer parameters increases, the number of data sets required also increases. When the input layer gradually increases and the adjacent rooms are not burning, there is no ignition time data, so the data has null values ​​and needs to be deleted. The samples added in this input layer become minority class samples. The SMOTE (Synthetic Minority Over-sampling Technique) data enhancement technology is used to synthesize new samples by interpolating between minority class samples in the original data set. The principle of Smote oversampling is as follows Figure 7 As shown. The core formula of the algorithm is:

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

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

[0109] Subsequently, the fire time of the rooms that have caught fire is included in the neural network input layer parameters. Specifically, when predicting the fire time of room 2 in area 2, the fire time of room 2 and room 4 is included; for the prediction of the fire time of area 3, the fire time of rooms in area 1 and area 2 is added; for the prediction of the fire time of room 9 in area 4, the input layer parameters comprehensively consider the fire time of areas 1 to 3. Based on this, the optimization and retraining of parameters and hyperparameters are carried out in turn to determine the optimal configuration of the model (refer to the steps of predicting the fire time of rooms in area 1).

[0110] The fire time of each room under different initial fire source room locations was accurately predicted. When the fire source was located in room 1 (corner room), the prediction accuracy of the fire occurrence time of all rooms exceeded 95%; when the fire source location was changed to room 9 (middle room), the model prediction accuracy increased to more than 96%. The high accuracy rate proves the model's high sensitivity to changes in the location of the fire source room and 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, and the fire events in the adjacent rooms are gradually predicted by continuously adding the area close to the fire room in the input layer of the neural network, so as to obtain the entire fire spread path. The comprehensive method of the present invention is used to predict the spread path of multi-room fires under different initial fire source room locations. This method is based on a neural network, accurately predicts the ignition time of each room, and obtains the spread path training and optimization network under different room sizes, fire source locations, HRR, and fire source types through simulation, so that it can systematically learn the time series of the fire spreading from the source point to the surrounding rooms, thereby realizing the prediction of the spread path of multi-room fires.

[0113] The present invention adopts MATLAB software to establish a BP neural network prediction model optimized by a chaos adaptive whale algorithm (CI-WOA). Figure 8The prediction error distribution of the optimized BP neural network on the test set is shown in Figure 2. The relative prediction error of the model is concentrated in the interval [-0.0783, 0.0399], where the maximum and minimum absolute errors are 0.0783 and 0.0045, respectively, reflecting that the optimized model has a relatively stable prediction performance. It is worth noting that the prediction error of most samples is close to 0, which further verifies the accuracy and generalization ability of the model. Compared with the traditional BP neural network, the average absolute error of the CI-WOA-BP neural network is reduced to 3.76%, which is 7.08% of the traditional method, showing lower error fluctuation and higher prediction accuracy. In addition, no extreme error values ​​are observed, indicating that the model does not have overfitting and has stronger generalization ability. The Chaos Adaptive Whale Algorithm (CI-WOA) greatly enhances the performance of BP neural networks in complex system prediction tasks and has obvious application potential.

[0114] Fig. 9 The CI-WOA optimization improves the prediction performance of the BP neural network. After optimization, the CI-WOA-BP neural network model predicts values ​​that are significantly better than the unoptimized BP network, showing smaller fluctuations and higher prediction accuracy, indicating better stability and reliability. It can be seen that the Chaos Adaptive Whale Algorithm (CI-WOA) effectively optimizes the weights and thresholds of the BP neural network, significantly improves the prediction accuracy of the model, and has great potential and practical value in dealing with complex data prediction problems.

[0115] Fig.10 The prediction error distribution of the two optimization models, CI-WOA-BP and SSA-BP, on the test set is shown. For the convenience of representation, the CI-WOA-BP and SSA-BP methods are respectively denoted as WOA and SSA in the figure. The frequency distribution of the prediction errors of the two models can be intuitively compared through the histogram. It can be observed from the figure that the error distribution of the CI-WOA-BP model is more concentrated near the position where the error is 0 than that of the SSA-BP model, and the peak of the error distribution curve of CI-WOA-BP near point 0 is higher, indicating that most of its predicted values ​​are closer to the true value, showing higher prediction accuracy; in addition, the error distribution curve of the CI-WOA-BP model has shorter tails on both sides, which means that it has a lower frequency of large errors, reflecting better stability and reliability. In contrast, although the error distribution of the SSA-BP model is relatively concentrated, the peak near point 0 is lower and the tail is longer, indicating that the concentration of its predicted values ​​near the true value is lower and there are more large errors.

[0116] In summary, the Chaotic Adaptive Whale Algorithm (CI-WOA) shows better prediction performance than the Sparrow Algorithm (SSA), which is manifested in that its prediction error distribution is more concentrated, the mean absolute error is lower, and it shows higher stability and reliability in the entire prediction range. These characteristics enable the CI-WOA-BP model to provide more accurate and robust prediction results when facing practical problems.

[0117] Fig.11 It is the comparison between the predicted value and the actual value of the fire time of each room by CI-WOA-BP neural network when the fire source is in room1. Fig.12 The comparison between the predicted value and the true value of the fire time of each room by the CI-WOA-BP neural network when the fire source is in room 9. It can be found that the CI-WOA-BP optimization model greatly improves the prediction accuracy of the fire time and enhances the generalization ability of the model, which can 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 fire in a building, characterized by: The following steps are involved: Step 1: Determine the plane structure physical model of the fire spread scenario based on the boundary ventilation conditions; Step 2: Divide the spreading area according to the relative consistency of the spreading direction and the fire ignition time in the area; Step 3: construct a sample set and divide it into training set, validation set and test set, build a CI-WOA-BP neural network, optimize the network parameter configuration, and predict the fire time of the room closest to the fire source room; Step 4: Add the fire time of the rooms in the burned area to the sample set as the input of the neural network to predict the fire time of the rooms in the next area; Step 5: According to the time series of the fire spreading from the room where the fire started to other rooms, the spreading path is obtained.

2. The data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building according to claim 1 is characterized in that: Methods for constructing sample sets include: Several groups of random parameters are generated as sample data of multi-room fire spread through statistics and Latin hypercube sampling; The random parameters include room size, fire source room location, fire load density, opening factor and heat release rate HRR; Conduct multi-room fire spread simulation to generate the same number of working conditions as sample data and the corresponding fire time of each room; Methods for determining sample set capacity are: (a) The sample set capacity is calculated based on the matching relationship between the sample set capacity N and the training error ε. The formula is as follows: (b) The sample set capacity is calculated based on the sample set capacity N being a multiple of the total number of network connection weights. The formula is as follows: N=k×(l·m+m·n), Where k is the empirical multiplier factor based on the total number of connection weights, 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 training set, validation set and test set.

3. The data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building according to claim 1 is characterized in that: Construct a three-layer CI-WOA-BP neural network model, that is, a single hidden layer framework, as follows: Firstly, the neural network framework structure is determined to be a BP neural network model; the whale optimization algorithm is used to optimize the weights and biases of the BP neural network; Using chaotic mapping to initialize the population of the whale optimization algorithm; In the learning process, an adaptive weight factor I is introduced to adjust the network weight coefficient, that is, the parameter of the network topology structure, and the network weight is automatically adjusted according to the current value and the target value; The input layer parameters of the CI-WOA-BP neural network, including room span, depth, floor height, and maximum heat release rate (HRR), are determined by the Spearman correlation coefficient and variance inflation coefficient. max and HRR duration, a total of 5 variables; the fire time of the room is the output layer variable; The network is trained using the training set, and after training, the model accuracy is calculated using the validation set.

4. The data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building according to any one of claims 1 to 3, characterized in that: Construct the following physical model: A plane frame structure with a horizontal 4×4 span is constructed, in which the fire spread area is a 3×3 span, including 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 in sequence in the 3×3 span fire spread area.

5. The data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building according to claim 4 is characterized in that: According to the locations of the two fire source rooms, the nine rooms in the fire spread area were divided into four areas along the diagonal direction; When the fire source is located in the corner room 1, the area is divided into: Area 1, including room 2 and room 4; Area 2, including room 3, room 5 and room 7; Area 3, including room 6 and room 8; Area 4, including room 9; When the fire source is set in the middle room 9, the areas are divided into: Area 1, including room 6 and room 8; Area 2, including room 3, room 5 and room 7; Area 3, including room 2 and room 4; Area 4, including room 1.

6. The data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building according to claim 5 is characterized in that: Area 1 is the area closest to the room with the fire source. During the network training process, the fire time of the room in area 1 under different fire source locations is used as the output layer variable; The fire time of the burned room is added to the neural network input layer parameters, including: When predicting the fire time of the room in area 2, the fire time of the room in area 1 is added to the neural network input layer parameters; when predicting the fire time of the room in area 3, the fire time of the rooms in areas 1 and 2 is added to the neural network input layer parameters; when predicting the fire time of the room in area 4, the fire time of the rooms from area 1 to area 3 is added to the neural network input layer parameters; When the input layer samples gradually increase and the adjacent rooms are not burning, there is no ignition time data, and the null value data is deleted. The samples added in this type of input layer are minority class samples. SMOTE data enhancement is used to synthesize new samples by interpolating between minority class samples in the original data set. The formula is: Y=X+rand(0,1)*(X i -X), In the formula, Y is the new sample synthesized, X i is the i-th nearest neighbor sample of sample X, and rand(0,1) represents a random number between 0 and 1.

7. The data-driven method for predicting the entire process of horizontal spread of multi-room fire in a building according to claim 1 is characterized in that: The model parameters, including activation function, learning rate, training accuracy, number of iterations and training algorithm, are trained and verified through trial algorithms to determine the optimal configuration of network parameters.

Images