Subway carriage fire temperature prediction method

Abstract

The invention discloses a subway carriage fire hazard temperature prediction method comprising the following steps: carrying out fire hazard working condition simulation on a subway carriage to obtain fire hazard experiment data, and preprocessing the fire hazard experiment data to obtain a preprocessed data set; constructing an initial temperature prediction model based on the multi-layer feed-forward network and a genetic algorithm; training the initial temperature prediction model based on the preprocessed data set to obtain a temperature prediction model; and carrying out compartment top temperature prediction on the subway compartment through the temperature prediction model. According to the method, a temperature database is constructed based on a subway carriage fire numerical model, and proper hyper-parameter values are determined. A GA-BPNN algorithm model composed of an input layer, a hidden layer and an output layer is adopted to predict ceiling temperature distribution of subway train carriage fire, and operation management of a tunnel subway system is expected to be further improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of car temperature prediction, and particularly relates to a subway car fire temperature prediction method. BACKGROUND

[0002] Due to the complexity of the tunnel subway car fire mechanism, the traditional physical model is difficult to be quickly applied in the case of fire, and can only provide simple theoretical reference. In recent years, machine learning has achieved breakthrough development, and by using known data for machine learning, a decision scheme is quickly derived, which can be applied to various industries including fire fighting. Machine learning application is a self-evolution process, and the more data obtained, the more accurate it is. As a branch of machine learning, neural network can automatically extract the required information from the input raw data through multiple adjustments and optimizations, and calculate the accurate results according to the requirements. Neural network is widely concerned in the law prediction field due to its unique identification, memory and learning functions, and the back propagation neural network (BPNN) and its improved algorithm are more studied. BPNN is a powerful algorithm that can realize nonlinear mapping, complex pattern recognition and prediction. Neural network is composed of a series of nodes called neurons connected by internal network, and the data features are transmitted to the network through the input layer, forming a linear combination of multiple features and interacting with the weights in the network. Then, the neurons perform nonlinear changes on the linear combination, so that the neural network model can represent multiple features in a complex way. In order to enhance the representation ability of input data features, neural network uses multi-layer network structure. BPNN can store and learn a large number of mapping relationships without explaining the mapping relationship equation between input and output. BPNN corrects network weights and thresholds through back propagation training to find the minimum network error. However, in the back propagation training process, the network is easily trapped in local minimum value, and cannot well realize car temperature prediction. SUMMARY

[0003] To solve the above technical problems, the present application provides a subway car fire temperature prediction method to solve the problems existing in the prior art.

[0004] To achieve the above purpose, the present application provides a subway car fire temperature prediction method, comprising the following steps:

[0005] Through fire condition simulation of the subway car, fire experiment data is obtained, and the fire experiment data is preprocessed to obtain a preprocessing data set;

[0006] An initial temperature prediction model is constructed based on a multi-layer feedforward network and a genetic algorithm;

[0007] The initial temperature prediction model is trained based on the preprocessing data set to obtain a temperature prediction model;

[0008] The roof temperature of the subway car is predicted by the temperature prediction model.

[0009] Preferably, the fire experiment data set includes heat release rate, longitudinal distance, time and longitudinal roof temperature distribution.

[0010] Preferably, the method for obtaining the preprocessed data set comprises:

[0011] Delete the values containing missing values and outliers in the fire experiment data to obtain preprocessed data;

[0012] The preprocessed data is divided into a training set, a test set and a validation set;

[0013] The training set is used for the initial temperature prediction model to learn by fitting the data samples; the test set is used to evaluate the performance of the trained model; and the validation set is used to correct the hyperparameters of the model and further verify the performance of the model.

[0014] Preferably, the activation function expression in the hidden layer and the output layer of the initial temperature prediction model is:

[0015]

[0016] f2(β i )=β i

[0017] In the formula, α i represents the output of the i-th neuron in the hidden layer, β i represents the output of the i-th neuron in the output layer, f1(α i ) is the activation function of the hidden layer, which is used to convert the input α i of the hidden layer neuron to the output, f2(β i ) is the activation function of the output layer, which is used to convert the input β i of the output layer neuron to the output.

[0018] Preferably, the method for training and predicting the initial temperature prediction model comprises:

[0019] Normalizing the detection data to obtain normalized data;

[0020] Determine the optimal number of hidden layer nodes;

[0021] Construct a genetic algorithm by initializing a population;

[0022] Calculate the fitness function of each population in the genetic algorithm;

[0023] Select the chromosome with higher fitness by using the geometric programming sorting selection algorithm;

[0024] The arithmetic crossover function is called to mix and fuse the parent individual gene information, and a diversified offspring individual is obtained;

[0025] The offspring individual is subjected to a mutation operation based on a non-uniform mutation algorithm, and a new generation of chromosomes is obtained;

[0026] It is checked whether the fitness of the new generation of chromosomes meets a convergence threshold value, and if the fitness is less than a set threshold value, the new generation of chromosomes is decoded to obtain initial weights and biases of the network;

[0027] The initial weight coefficients of all data points in the normalized data are set to the optimal initial weight values determined by the genetic algorithm;

[0028] The maximum number of iterations, the minimum error of the training target, and the learning rate are set;

[0029] The maximum number of iterations under the initial condition is 1, the training data is started, and the trained data is saved in the current external iteration number, and the current prediction error is calculated;

[0030] The maximum number of iterations is iterated to obtain the prediction error of the current external iteration number;

[0031] The weight coefficients of all data points in the training set are updated based on the multi-layer feedforward network to obtain a temperature prediction model;

[0032] The roof temperature at any longitudinal position in the subway car is predicted based on the temperature prediction model.

[0033] Preferably, the expression of the normalization processing is:

[0034]

[0035] In the formula, X min represents the minimum data, X max represents the maximum data.

[0036] Preferably, the expression for determining the optimal number of hidden layer nodes is:

[0037]

[0038] In the formula, m represents the number of input layer nodes, n represents the number of output layer nodes, and a is an integer with a value range of 1-10.

[0039] Preferably, the expression for selecting a chromosome with a higher fitness is:

[0040] prob(i)=(1-p) * p^i

[0041] In the formula, prob(i) is the probability of the i th individual being selected, and p is a constant between 0 and 1.

[0042] Preferably, the expression of the arithmetic crossover function is:

[0043] childGene = w * parent1Gene + (1-w) * parent2Gene

[0044] In the formula, parent1Gene and parent2Gene represent the gene values of two parent individuals at a specific gene locus, and w is a random weight.

[0045] Preferably, the mutation rate reduction function of the non-uniform mutation algorithm is:

[0046] MR = IMR * (1-G / MG) ^ bet

[0047] In the formula, IMR is the initial mutation rate, G is the iteration number of the current iteration, MG is the total iteration number, and beta is the parameter of mutation rate reduction.

[0048] Compared with the prior art, the present application has the following advantages and technical effects:

[0049] The application discloses a subway car fire temperature prediction method, comprising the following steps: obtaining fire experiment data by simulating a fire working condition of a subway car, preprocessing the fire experiment data to obtain a preprocessing data set; constructing an initial temperature prediction model based on a multilayer feedforward network and a genetic algorithm; training the initial temperature prediction model based on the preprocessing data set to obtain a temperature prediction model; and predicting a car roof temperature of the subway car by using the temperature prediction model. The application constructs a temperature database based on a subway car fire numerical model and determines appropriate hyperparameter values. A GA-BPNN algorithm model composed of a genetic algorithm and a back propagation neural network is used to predict a subway train car fire roof temperature distribution, which is expected to further improve the operation management of a tunnel subway system. BRIEF DESCRIPTION OF DRAWINGS

[0050] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application and of the related description are used to explain the application and are not intended to limit the application. In the drawings:

[0051] Figure 1 A prediction model structure diagram of an embodiment of the application;

[0052] Figure 2 A car temperature fitness change curve diagram of an embodiment of the application;

[0053] Figure 3This is a state diagram of the neural network compartment temperature training according to an embodiment of the present invention;

[0054] Figure 4 A comparison chart of the predicted and actual cabin temperature values ​​according to an embodiment of the present invention;

[0055] Figure 5 This is a regression fitting diagram of the cabin temperature training set and the validation set according to an embodiment of the present invention;

[0056] Figure 6 This is a tunnel temperature adaptability change curve diagram of an embodiment of the present invention;

[0057] Figure 7 This is a state diagram of the neural network tunnel temperature training according to an embodiment of the present invention;

[0058] Figure 8 A comparison chart of tunnel temperature prediction values ​​and actual values ​​according to an embodiment of the present invention;

[0059] Figure 9 1 is a regression fitting diagram of the tunnel temperature training set and the validation set according to an embodiment of the present invention;

[0060] Figure 10 A comparison chart of cabin temperature prediction errors between the GA-BPNN model and the BPNN model according to an embodiment of the present invention;

[0061] Figure 11 This is a comparison chart of tunnel temperature prediction errors of the GA-BPNN model and the BPNN model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0063] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0064] Example 1

[0065] like Figure 1 As shown, this embodiment provides a subway car fire temperature prediction method, which is characterized by comprising the following steps:

[0066] Fire test data is obtained by simulating a fire condition in a subway car, and the fire test data is preprocessed to obtain a preprocessed data set;

[0067] An initial temperature prediction model (GA-BPNN) is constructed based on a multi-layer feedforward network and a genetic algorithm;

[0068] The initial temperature prediction model is trained based on the preprocessed data set to obtain a temperature prediction model;

[0069] The temperature prediction model is used to predict the roof temperature of a subway car.

[0070] The specific implementation is as follows:

[0071] The GA-BPNN model needs to learn the variation law of the input data, and generates a neural network prediction model by training the data set to realize the tunnel subway train temperature prediction function. The training data set includes two categories, independent variables and dependent variables. First, the main related physical parameters should be determined to describe the longitudinal roof temperature distribution of the tunnel fire. The longitudinal roof temperature distribution is related to the heat release rate Q, the longitudinal distance X, the air density ρa, the heat capacity of air cp, the ambient temperature Ta, the gravitational acceleration g, the tunnel height H, and the ventilation speed V, as shown in formula (1). These physical parameters can be determined as the main parameters for establishing the theoretical model.

[0072] f(ΔT(X), Q, x, ρ α , c p , T a , g, H, V) = 0 (1)

[0073] The selected 9 physical parameters belong to different factors of the tunnel fire scene, including ventilation conditions and tunnel size. However, considering the simplicity of calculation, the main influencing factors should be emphasized. Based on the dimension analysis method and prior knowledge, the independent variables are selected as the fire power Q, the longitudinal distance X, and the time t, and the dependent variable is the longitudinal roof temperature distribution In the process of fire numerical simulation, the tunnel and car interior temperature data are obtained through the set temperature sensor and temperature slice. Pyrosim simulates 4 groups of subway car fire conditions under different fire heat release rates, and the simulation time of each group of conditions is 300 seconds. After the fire condition simulation is completed, the roof temperature data is collected by the thermocouple to construct the data set, and part of the data is shown in Table 1. Each group of conditions has 59738 groups of data, and 4 groups of conditions have 238952 groups of data. Each row of data in the original data set represents the roof temperature parameter of the tunnel subway system at that time.

[0074] Table 1

[0075]

[0076] With the change of working conditions, the fluctuation range of temperature data is large, and the change difference between each measuring point is large, which may cause the problems of long training time of neural network, poor fitting effect, and unsatisfactory prediction effect. In order to avoid the influence of sample set on neural network, the simulated fire characteristic basic data is preprocessed to delete the values containing missing values and outliers. In this embodiment, all the preprocessed data is randomly divided into 3 parts: 75% training set, 10% test set and 15% validation set. The training set is used for model fitting data sample, and the gradient descent of training error is carried out in the training process. The test set is used to evaluate the performance of the trained model, but should not be used as the basis for hyperparameter determination and feature selection. The validation set is a special data set, which is mainly used to correct the hyperparameters of the model and further verify the performance of the model.

[0077] Further, the hyperparameters are determined as follows:

[0078] The main hyperparameters of neural network include the number of neurons, learning rate and training iteration number, etc. The setting of hyperparameters will affect the accuracy of model prediction results. Therefore, the hyperparameters should be optimized during model training to ensure that the prediction results meet the expectations.

[0079] Iteration number determination:

[0080] In the training process of GA-BPNN model, the iteration number is a very important hyperparameter, which determines the number of times the learning algorithm traverses the entire training data set. Each iteration means that each data may correct the internal model parameters. However, the more iterations do not necessarily mean better model performance, and more importantly, the accuracy of the model should be considered. In this embodiment, 6 different iteration numbers are set: 250, 500, 1000, 1500, 2000, 3000, to determine the best iteration number. Under the premise of keeping other parameters unchanged, the iteration number in the model is adjusted, and the root mean square error (RMSE) and the determination coefficient R2 of the corresponding test set are calculated, as shown in Table 2.

[0081] Table 2

[0082]

[0083] From Table 2, it can be seen that with the increase of the number of iterations, the RMSE of the test set first decreases from 5.4030 to 5.1854. Subsequently, with the increase of the number of iterations, the RMSE value of the test set gradually increases from 5.1854 to 7.2837. The R2 of the test set presents an opposite trend to the RMSE. When the number of iterations is 250, R2 is 0.9717; when the number of iterations increases to 500, R2 increases to 0.9733. Subsequently, with the increase of the number of iterations, R2 gradually decreases. In combination, the number of iterations is finally selected in this embodiment when the RMSE value is the smallest and the R2 value is closest to 1. Therefore, the number of iterations of the GA-BPNN model is determined to be 500.

[0084] Learning rate determination:

[0085] In the training process of the neural network, the adjustment amount of the weight is called the step or "learning rate". The learning rate is an adjustable hyperparameter in the training of the neural network, and is usually between 0.01 and 0.001. When the learning rate is properly set, the model can find an approximate function of the optimal weight within a limited training period and given resource constraints. Generally speaking, a larger learning rate can make the model learn faster, but it may not get a completely converged weight set. On the contrary, a smaller learning rate can make the model find a more optimal or even global optimal weight configuration, but this usually takes a lot of training time. In this embodiment, a suitable learning rate value of 0.001 is determined through multiple experiments.

[0086] Activation function determination:

[0087] In the hidden layer, the information carried by the hidden neuron ai is transmitted to bi by the activation function f1. Then, the activation function f2 further processes bi and transmits it to the output layer. The activation function plays a very important role in enhancing the learning ability, understanding ability and representation ability of the network. The activation functions in the hidden layer f1 and the output layer f2 used in this embodiment are shown in equations (2) and (3) respectively.

[0088]

[0089] f2(β i )=β i (3)

[0090] In the formula, α i represents the output of the i-th neuron in the hidden layer, β i represents the output of the i-th neuron in the output layer, f1(α i ) is the activation function of the hidden layer, which is used to convert the input α i of the hidden layer neuron to the output, and f2(β i) is an activation function of the output layer, used to convert the input β of the output layer neuron to output. i

[0091] Further, the model algorithm is determined as follows:

[0092] The embodiment develops a back propagation neural network algorithm based on genetic algorithm optimization, which predicts the ceiling temperature distribution by establishing a nonlinear mapping relationship between the longitudinal distance X, time t and heat release rate Q, and the GA-BPNN model algorithm structure is shown as follows: Figure 1 The GA-BPNN algorithm steps are as follows:

[0093] Collect N thermocouple data under different heat release rates Q from longitudinal distance X1 to XN as time t changes: Normalize the collected data, convert the input and output data to [-1, 1], and the normalization processing formula is shown as formula (4), and the prediction data set is established after induction and arrangement. Xmin and Xmax are the minimum and maximum data values respectively.

[0094]

[0095] Determine the optimal number of hidden layer nodes H, the value range of hidden layer nodes can be calculated by formula (5), m represents the number of input layer nodes, n represents the number of output layer nodes, and a takes an integer value between 1 and 10. According to formula 3.7, 10 selectable hidden layer nodes are obtained. Set these 10 hidden layer nodes for GA-BPNN in turn and compare the root mean square error RMSE of the test set, and finally obtain the corresponding hidden layer node with the minimum error, that is, determine the optimal hidden layer node.

[0096]

[0097] Construct the genetic algorithm by initializing the population. Specify the matrix w1, b1, w2 and b2, which symbolize the connection weights and thresholds between layers respectively.

[0098] Calculate the fitness. Calculate the fitness function of each population in the genetic algorithm, where the population size is represented as Z. The fitness function selects the GabpEval function in the Goat toolbox.

[0099] ​​Selection. The NormGeomSelect algorithm is used to select chromosomes with higher fitness. The standard geometric selection algorithm normalizes the fitness values of individuals and selects parent individuals according to the normalized fitness values, as shown in equation (6). Here, prob(i) is the probability of the i-th individual being selected, and p is a constant between 0 and 1, representing the selection pressure. By using the standard geometric selection algorithm, the genetic algorithm can select parent individuals according to their fitness values.

[0100] prob(i) = (1 - p) * p^i (6)

[0101] Crossover. The ArithXover function is called to implement the mixing and fusion of gene information of parent individuals, thereby generating more diverse offspring individuals. For each pair of parent individuals, their corresponding gene positions (or variables) are weighted and averaged according to a random weight, as shown in equation (7). Here, parent1Gene and parent2Gene represent the gene values of two parent individuals at a specific gene position, and w is a random weight, usually taking a value in the range of 0 to 1.

[0102] childGene = w * parent1Gene + (1 - w) * parent2Gene (7)

[0103] Mutation. The offspring individuals obtained by crossover are subjected to mutation operation, using the NonUnifMutation algorithm, which introduces a mutation rate reduction function, so that the mutation rate gradually decreases with the increase of iteration number. The mutation rate reduction function of the NonUnifMutation algorithm is shown in equation (8), where IMR is the initial mutation rate, G is the current iteration number, MG is the total iteration number, and beta is the parameter of mutation rate reduction. After selection, crossover and mutation, the population will obtain a new generation of chromosomes.

[0104] MR = IMR * (1 - G / MG)^bet (8)

[0105] Recalculating the fitness function. Check if the fitness of the chromosome meets the convergence threshold. If the fitness is less than the set threshold, decode the chromosome to obtain the initial weights and biases of the network.

[0106] Set the initial weight coefficients of all data points in the training set to the optimal initial weight determined by the genetic algorithm to reduce the relative error and improve the prediction accuracy.

[0107] Set the maximum number of iterations K to 500, the training target minimum error to 0.000001, and the learning rate L to 0.001.

[0108] K = 1, start training data, and save the trained data in the current external iteration number K, and calculate the current prediction error RMES.

[0109] Let K = K + 1, go back to step 11 again iteration until the termination condition is met; find the prediction error of the current external iteration number.

[0110] Based on the formula in BPNN, update the weight coefficient w(i) of all data points in the training set (i = 1, 2 … N)

[0111] Using the determined GA-BPNN algorithm to train the prediction model, so as to predict the roof temperature at any longitudinal position in the tunnel subway fire

[0112] This embodiment uses MATLAB software to complete the training of GA-BPNN prediction model, and uses the trained prediction model to predict and test the along-the-way temperature of 4 groups of working conditions. In each group of training data, the heat release rate Q, the longitudinal distance X, and the time t are taken as the input, and the corresponding longitudinal roof temperature As output. Before starting training, the data set obtained by FDS simulation is imported into MATLAB, and the data amount of the training set, test set and validation set is respectively set to 75%, 10% and 15% of the total data amount.

[0113] The analysis of the prediction results of the car roof temperature is as follows:

[0114] The iteration number in this car temperature prediction model is 500 times, the learning efficiency is 0.001; the excitation function is selected as Sigmoid function. The maximum genetic generation is 60, the population size is set to 15, the crossover probability and mutation probability are respectively 0.09 and 0.05. The BPNN best weight is obtained by genetic algorithm optimization, wherein the connection weight between the input layer and the hidden layer NET.IW{1,1} is in the form of 3×11 matrix, and the connection weight between the output layer and the hidden layer NET.LW{2,1} is in the form of 11×1 matrix. The connection weight between the input layer and the hidden layer is shown in Table 3, and the connection weight between the output layer and the hidden layer is shown in Table 4.

[0115] Table 3

[0116]

[0117] Table 4

[0118]

[0119] The iteration fitness decline curve and the neural network training state change are respectively as shown in Figure 2 ,Figure 3 As shown in the figure. During the training process, the fitness value initially decreased slowly. After the 26th round, the fitness value dropped rapidly from 0.01704 to 0.01636. It converged in the 54th round of training, and the fitness value dropped to 0.01625 and remained unchanged. The decrease in the gradient from the 1st round to the 54th round indicates that the model is gradually approaching the optimal solution and the model's predictive ability has improved. The network reached the set threshold (10-6) and completed convergence in the 62nd round of training, with the gradient reduced to 0.00048448. During the network training process, the gradient value initially decreased slowly. From the 43rd to the 45th round, the gradient increased slightly, but overall, the gradient value remained relatively low.

[0120] The GA-BPNN model takes 27.38 seconds to predict the temperature of the subway train compartment and draw the temperature prediction graph, which has a short response time. The error line graph and regression fitting graph of the compartment temperature test set of the GA-BPNN model are as follows: Figure 4 、 Figure 5 As shown in the figure, it can be observed that the predicted data of the train car temperature are all closely close to the zero error line of y=x. Although there is still a deviation between the predicted value and the true value, the error range is within 20°C. The hollow circles in the figure represent the prediction results, and the slash represents the optimal estimate, that is, the prediction result is almost the same as the test true value. The GA-BPNN prediction model can accurately predict the temperature distribution of the tunnel subway car after the fire based on the three input features, and the prediction deviation is small. Under different fire source powers, the maximum error of the car ceiling temperature is 19.10°C, and the average error is 8.17°C.

[0121] To further quantify the predictive performance of the GA-BPNN model, we calculated the performance metrics MAE, RMSE, and R² based on the predicted and true values. The MAE was 8.17, and the RMSE was 9.76. The R² values ​​for the training and test sets of the neural network model were 0.97 and 0.99, respectively. This demonstrates that the neural network-based prediction model can accurately predict the simulation results of the FDS model.

[0122] Similarly, in the tunnel temperature data set obtained in this FDS simulation, the number of training sets was set to 518 and the number of test sets was set to 130. Using the determined prediction model algorithm, the training data was introduced for training to construct a tunnel temperature prediction model. The number of iterations in the model was set to 500, the learning efficiency was set to 0.001, the activation function was selected as the Sigmoid function, the maximum genetic generation was set to 60, the population size was set to 15, the crossover probability and mutation probability were set to 0.09 and 0.05 respectively, and the fitness curve and regression prediction graph are shown in the figure below. Figure 6 、 Figure 7The figure shows that as the number of iterations increases, the fitness curve gradually decreases, and the change is not large after 35 iterations. This shows that the model gradually adapts to the data distribution in the training process and can better generalize to the test data. During the training process, the gradient initially decreases slowly, and at the 21st round, the gradient increases slightly, but still remains in the order of 10-2. It converges in the 60th round of training and the network reaches the set threshold, the gradient is 0.00059522, and the gradient decreases from the 1st to the 60th iteration, indicating that the model is gradually approaching the optimal solution, which means that the prediction ability of the model has improved. In addition, this situation shows that the model has a certain robustness and can maintain stability when facing noise and interference during the training process.

[0123] Figure 8 The error line chart of the GA-BPNN prediction of the tunnel temperature test set under 4 different fire conditions is shown in FIG. 8, Figure 9 is a regression fitting chart of the neural network training set and the test set, and the hollow circles represent the predicted tunnel temperature data values. Figure 8 The predicted scatter values in the figure are almost coincided with the reference line of x=y, indicating that the model training accuracy is high. The overall data in the predicted data has a low probability of error, and individual data prediction has a large error, but the overall error is small. Under different fire power, the maximum error of the tunnel roof temperature is 13.73℃, the average error is 3.98℃, and 94% of the error values are less than 10℃. Therefore, the GA-BPNN model performs well in predicting the smoke temperature in the tunnel fire.

[0124] In order to further quantify the prediction performance of the GA-BPNN model, the performance indicators MAE, RMSE and R2 are calculated according to the predicted value and the true value. The average absolute error MAE is 3.95, the root mean square error RMSE is 5.63, and the R2 values of the training set and the test set are 0.98 and 0.98 respectively.

[0125] From the perspective of error analysis, due to the complex temperature change mechanism and large temperature fluctuation range caused by the lateral multiple openings of the car, the GA-BPNN model performs better in predicting the smoke temperature in the tunnel car fire than in predicting the car temperature. Overall, the GA-BPNN prediction model has good learning effect on the subway car fire ceiling temperature data, can accurately predict the temperature of each measuring point along the path under 4 fire conditions in the test set, and can obtain satisfactory prediction results. Therefore, the GA-BPNN prediction model can skillfully predict the temperature distribution of the entire tunnel ceiling with high precision.

[0126] The model comparison is also carried out in this embodiment as follows:

[0127] In order to further verify the prediction accuracy of the GA-BPNN algorithm, the commonly used BPNN prediction model is used as a comparison, and a more direct comparison shows the prediction effect of the model. The model is built using the MATLAB toolbox, and in order to control the variables, the same input and output data as the GA-BPNN model are used for simulation. Figure 10 The GA-BPNN model and the BPNN model are compared, Figure 11 The GA-BPNN model and the BPNN model are compared.

[0128] From Figure 10 , 11 It can be seen that for the prediction of the car temperature and the tunnel temperature, the GA-BPNN model is obviously better than the BPNN, and the GA optimization algorithm has a very high prediction accuracy, which can effectively prevent the BPNN model from falling into a local optimum, and greatly reduce the prediction error. From the error analysis, the error range of the GA-BPNN model prediction result is mostly within ±5%, and the error distribution of the GA-BPNN model test value sample is relatively concentrated, with a maximum error of 28.32°C; although the prediction accuracy of the BPNN model is still about 92%, the error distribution is more in the range of -10% to 20%, and the maximum error is more than 70%, the training sample error and the test sample error distribution are relatively dispersed, and the maximum error reaches 107.47°C. The MAE and RMSE of the car temperature of the GA-BPNN prediction model are 10.25 and 15.03, respectively, and the MAE and RMSE of the car temperature of the unoptimized BPNN model are 16.41 and 25.75, respectively, and the MAE and RMSE of the prediction result are reduced by 37.53% and 41.63%, respectively; the MAE and RMSE of the tunnel temperature of the GA-BPNN prediction model are 3.83 and 6.05, respectively, and the MAE and RMSE of the tunnel temperature of the unoptimized BPNN model are 5.36 and 7.67, respectively, and the MSE and RMSE of the prediction result are reduced by 28.54% and 21.12%, respectively, and the determination coefficient R2 is increased by 1.8%. The prediction value obtained based on the GA-BPNN prediction model is very close to the true value, and the test error mean is about 20% lower than that of the BPNN model, and the model accuracy has been significantly improved. From the comparison results, it is found that the prediction result of the GA-BPNN is obviously better than that of the traditional BPNN model. Not only in the prediction of the tunnel ceiling temperature, but also in the prediction of the car fire ceiling temperature, the GA-BPNN optimization algorithm can obtain better prediction results compared with the traditional BPNN algorithm. The error of the prediction result is within the acceptable range, and the prediction method has high accuracy.

[0129] The GA algorithm successfully optimizes the prediction characteristics of the model, and is superior to the traditional model in accuracy and generalization ability. The GA-BPNN model successfully reveals the relationship between the ceiling temperature and the characteristic parameters, and can accurately predict the temperature at different positions in the tunnel subway fire. Therefore, based on the above comparison and analysis, the GA-BPNN method has the advantages of high prediction accuracy, strong generalization ability, good robustness, high practicability, and broad application prospects.

[0130] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting temperature of a subway car fire, characterized in that: The following steps are involved: Fire test data is obtained by simulating a fire condition in a subway car, and the fire test data is preprocessed to obtain a preprocessed data set; Construct an initial temperature prediction model based on a multi-layer feedforward network and genetic algorithm; Training the initial temperature prediction model based on the preprocessed data set to obtain a temperature prediction model; The temperature prediction model is used to predict the roof temperature of the subway car.

2. The subway car fire temperature prediction method according to claim 1, characterized in that: The fire experiment dataset includes heat release rate, longitudinal distance, time and longitudinal roof temperature distribution.

3. The subway car fire temperature prediction method according to claim 1, characterized in that: The method for obtaining the preprocessed data set comprises: Deleting missing values ​​and abnormal values ​​in the fire test data to obtain preprocessed data; Dividing the preprocessed data into a training set, a test set, and a validation set; The training set is used to learn the initial temperature prediction model by fitting data samples; the test set is used to evaluate the performance of the trained model; and the validation set is used to correct the hyperparameters of the model and further verify the model performance.

4. The subway car fire temperature prediction method according to claim 1, characterized in that: The activation function expressions in the hidden layer and output layer of the initial temperature prediction model are: f2(β i )＝β i ； Where, α i represents the output of the i-th neuron in the hidden layer, β i represents the output of the i-th neuron in the output layer, f1(α i ) is the activation function of the hidden layer, which is used to convert the input α of the hidden layer neurons into i Converted to output, f2(β i ) is the activation function of the output layer, which is used to convert the input β of the output layer neurons into i Convert to output.

5. The subway car fire temperature prediction method according to claim 1, characterized in that: The method for training and predicting the initial temperature prediction model includes: Normalizing the detection data to obtain normalized data; Determine the optimal number of hidden layer nodes; Construct a genetic algorithm by initializing the population; Calculate the fitness function of each population in the genetic algorithm; The chromosome with higher fitness is selected using geometric programming sorting selection algorithm; Call the arithmetic crossover function to mix and fuse the genetic information of the parent individuals to obtain diversified offspring individuals; Performing a mutation operation on the offspring individuals based on a non-uniform mutation algorithm to obtain a new generation of chromosomes; Check whether the fitness of the new generation chromosomes meets the convergence threshold. If the fitness is less than the set threshold, decode the new generation chromosomes to obtain the initial weights and biases of the network. The initial weight coefficients of all data points in the normalized data are set to the optimal initial weight values ​​determined by the genetic algorithm; Set the maximum number of iterations, the minimum error of the training target, and the learning rate; Under the initial conditions, the maximum number of iterations is 1, start training the data, save the trained data in the current external iteration number, and calculate the current prediction error; Iterate according to the maximum number of iterations to obtain the prediction error of the current external iteration number; Based on the multi-layer feedforward network, the weight coefficients of all data points in the training set are updated to obtain the temperature prediction model; The roof temperature at any longitudinal position in the subway car is predicted based on the temperature prediction model.

6. The subway car fire temperature prediction method according to claim 5, characterized in that: The expression of the normalization process is: Where, X min Indicates the minimum data, X max Indicates the maximum data.

7. The subway car fire temperature prediction method according to claim 5, characterized in that: The expression for determining the optimal number of hidden layer nodes is: Where m represents the number of input layer nodes, n represents the number of output layer nodes, and a is an integer ranging from 1 to 10.

8. The subway car fire temperature prediction method according to claim 5, characterized in that: The expression for selecting chromosomes with higher fitness is: prob(i)=(1-p) * p^i; Where prob(i) is the probability that the i-th individual is selected, and p is a constant between 0 and 1.

9. The subway car fire temperature prediction method according to claim 5, characterized in that: The expression of the arithmetic crossover function is: childGene=w*parent1Gene+(1-w)*parent2Gene; In the formula, parent1Gene and parent2Gene represent the gene values ​​of the two parent individuals at specific gene positions, and w is a random weight.

10. The subway car fire temperature prediction method according to claim 5, characterized in that: The mutation rate reduction function of the non-uniform mutation algorithm is: MR = IMR * (1-G / MG)^bet; Where IMR is the initial mutation rate, G is the number of generations of the current iteration, MG is the total number of generations of iterations, and bet is the parameter for reducing the mutation rate.

Images