Tunnel smoke spread real-time prediction method based on physical priori and time sequence network

Abstract

The application discloses a tunnel smoke spread real-time prediction method based on physical priori and timing network, comprising the following steps: acquiring multi-source timing data and preprocessing, and constructing multi-working condition data set; constructing a mixed physical priori containing smoke spread length priori and smoke layer height field shape priori; adopting intrinsic orthogonal decomposition to reduce the dimension of the smoke layer height field, and extracting spatial main mode and mode coefficient; constructing a multi-task prediction model based on a causal timing convolution network, and synchronously predicting the smoke backflow length and the mode coefficient; using a mixed loss function fusing data fitting loss, physical priori loss and smoothing constraint, and combining a progressive loading mechanism to train the model; using the trained model to output the prediction result in real time, and reconstructing the smoke layer spatial distribution. The application fuses physical priori and causal timing network, realizes real-time and high-precision prediction of the tunnel fire smoke backflow length and the smoke layer spatial distribution, and has the advantages of strong physical rationality, high generalization ability and stable prediction.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of civil engineering disaster prevention and mitigation, tunnel fire protection engineering, and artificial intelligence time-series prediction. In particular, it relates to a real-time prediction method for tunnel smoke spread based on physical priors and time-series networks. This method is applied to intelligent tunnel fire monitoring systems, smoke control auxiliary decision-making systems, and fire safety assessment platforms to achieve rapid and high-precision prediction of the backflow length of smoke and the spatial distribution of smoke layers in tunnel fires. Background Technology

[0002] Tunnels are typical narrow, confined spaces where high-temperature, toxic smoke spreads rapidly in the event of a fire. When longitudinal ventilation is insufficient to suppress the upstream spread of smoke, a smoke backflow occurs, directly threatening the evacuation of personnel and firefighting efforts upstream of the fire source. Therefore, real-time and accurate prediction of the length of the smoke backflow and the spatial distribution of the smoke layer is a core prerequisite for tunnel fire prevention and emergency decision-making.

[0003] Current methods for predicting tunnel flue gas backflow can be mainly divided into three categories:

[0004] The first category is numerical simulation methods based on computational fluid dynamics (CFD) (such as FDS). This method can obtain high-precision results of the spatiotemporal distribution of flue gas through fine mesh generation and solving physical equations. However, a single numerical simulation often takes several hours or even days, which cannot meet the real-time prediction requirements in fire emergency scenarios.

[0005] The second category consists of simplified models based on theoretical and empirical formulas. These models establish a correlation between smoke backflow length, fire source power, and ventilation velocity through fire dynamics derivation and experimental fitting. While this type of method is computationally fast, it can only estimate macroscopic lengths and cannot describe the spatial distribution of the smoke layer. Furthermore, its prediction accuracy significantly decreases under complex ventilation and variable-power fire source conditions.

[0006] The third category is data-driven prediction methods based on deep learning. Recent studies have employed recurrent neural networks (LSTM / GRU) for time-series prediction of smoke backflow length, but these methods suffer from three core shortcomings: First, purely data-driven models lack constraints from the physical laws of fire, easily leading to predictions that do not conform to the physical characteristics of smoke diffusion, and exhibiting poor generalization in unseen conditions. Second, existing models mostly focus on single-point prediction of smoke backflow length, failing to simultaneously predict the spatial distribution of smoke layer thickness, thus hindering the fulfillment of full-space situational awareness requirements. Third, traditional recurrent neural networks suffer from insufficient temporal causality, vanishing gradients in long-sequence predictions, and difficulty in handling prediction scenarios with insufficient early-stage fire data, resulting in poor prediction stability across all time periods.

[0007] To address the aforementioned issues, this invention proposes a real-time prediction method for tunnel smoke spread based on physical priors and temporal networks. This method aims to balance prediction speed, physical plausibility, and spatial distribution perception capabilities, providing reliable technical support for tunnel fire emergency response. Summary of the Invention

[0008] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a real-time prediction method for tunnel smoke propagation based on physical prior knowledge and temporal networks. By integrating physical prior knowledge with causal temporal convolutional networks, a tunnel smoke backflow prediction method is constructed that combines physical rationality, spatial awareness, and real-time prediction capabilities.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] This invention provides a real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks, comprising the following steps:

[0011] S1. Acquire multi-source time-series data in tunnel fire scenarios, perform data preprocessing, and construct a multi-condition dataset;

[0012] S2. Based on the multi-source time-series data, construct a hybrid physical prior, which includes a prior for flue gas propagation length and a prior for the shape of the smoke layer height field;

[0013] S3. The intrinsic orthogonal decomposition is used to reduce the dimensionality of the smoke layer height field, extract the spatial principal mode and the corresponding temporal mode coefficients, and transform the high-dimensional spatial field prediction into low-dimensional mode coefficient prediction.

[0014] S4. Construct a multi-task prediction model based on a causal temporal convolutional network, wherein the multi-task prediction model is used to simultaneously predict the flue gas backflow length and the temporal modal coefficients;

[0015] S5. Using the dataset, and based on a hybrid loss function that combines data fitting loss, physical prior loss, and smoothing constraints, combined with a progressive ramp-in loading mechanism, the multi-task prediction model is trained.

[0016] S6. Using the trained multi-task prediction model, process the real-time input time-series data, synchronously output the predicted value of flue gas counterflow length and the predicted value of time-series modal coefficient, and reconstruct the real-time spatial distribution of the smoke layer based on the predicted value of the time-series modal coefficient and the spatial principal mode.

[0017] Preferably, step S1 specifically includes:

[0018] S11. Obtain four types of core data under different fire source conditions in the tunnel through fire dynamics simulation or on-site monitoring, including: distributed temperature sensor DEVC time series data, heat release rate HRR time series data, smoke backflow length calibration data, and spatial distribution data of longitudinal smoke layer thickness in the tunnel.

[0019] S12. For the dual-header format of unit row and true header row of DEVC time series data and HRR time series data of distributed temperature sensor, the header structure is automatically identified and unified; missing values ​​in the data are filled with linear interpolation, and outliers that exceed the physical reasonable range are replaced with the average of the previous and next frames.

[0020] S13. Extract features from the time series data of the distributed temperature sensor DEVC to generate basic features including temperature statistical features, thermal mass features, coarse localization features of the flue gas front and thermal centroid features. Then, perform first-order difference and exponential moving average (EMA) processing on the basic features to generate time series derived features.

[0021] S14. Perform Z-score standardization on the basic features and time-series derived features, and independently standardize the flue gas counterflow length calibration data and the spatial distribution data of tunnel longitudinal smoke layer thickness to ensure the consistency of training data distribution; finally, divide the training set and validation set according to time order, with the validation set accounting for 20%~30%.

[0022] Preferably, in step S13, the method for feature extraction from the time-series data of the distributed temperature sensor (DEVC) includes the following steps:

[0023] S131. Calculate the temperature rise d at each measuring point based on the preset environmental reference temperature. T The temperature distribution characteristics of the entire tunnel are statistically generated, and the temperature distribution characteristics include at least the mean, specified high quantile, maximum value, and number of supercritical temperature measurement points.

[0024] S132. Based on the temperature rise data, calculate the thermal mass characteristics and thermal energy characteristics of the entire tunnel section, the upstream area of ​​the fire source, and the downstream area of ​​the fire source, and further calculate the heat ratio characteristics of the upstream and downstream areas; wherein, the thermal mass characteristics are obtained by summing the positive temperature rise values ​​of each measuring point, and the thermal energy characteristics are obtained by summing the square values ​​of the positive temperature rise values ​​of each measuring point;

[0025] S133. Based on a preset critical temperature rise threshold, determine the coarse positioning length L of the flue gas front upstream of the fire source. up,coarse and downstream flue gas propagation length L down,coarse Based on the temperature rise weighted calculation method, the location of the thermal centroid upstream of the fire source is determined, and the corresponding flue gas length L is generated using the thermal centroid method. up,cm and its downstream corresponding characteristics;

[0026] S134. For all the basic features generated in the above steps, generate corresponding exponential moving average features using multiple preset time constants τ, and calculate the first-order time difference of each basic feature to generate rate of change features.

[0027] S135. The basic features, exponential moving average features, and rate of change features are concatenated to form a complete time-series feature matrix, which serves as the input feature part of the dataset.

[0028] Preferably, in step S2, the step of constructing the hybrid physical prior is as follows:

[0029] S21. Based on the aforementioned multi-source time-series data, the coarse location length of the flue gas front and the flue gas length obtained by thermal center-of-mass method are fused using a linear weighted fusion method to generate a time-varying prior flue gas propagation length L. phy (t), the calculation formula is:

[0030] (1);

[0031] In the formula, w coarse For the weight of the thick front, w cm The weight of the thermal centroid is w. coarse +w cm =1;

[0032] S22. Based on the empirical law of longitudinal attenuation of smoke in tunnel fires, with the fire source location as the center and the a priori smoke propagation length L... phy (t) is used as the upstream attenuation characteristic length to construct the theoretical distribution field H of the smoke layer thickness, which exhibits an exponential attenuation distribution along the tunnel longitudinal direction. phy (x,t) serves as the prior for the shape of the smoke layer height field;

[0033] S23. Prioritize the shape of the smoke layer height field H phy Projecting (x,t) onto the spatial principal modes extracted by the intrinsic orthogonal decomposition generates the corresponding prior values ​​C of the modal coefficients. phy,k (t), serving as the physical constraint for smoke layer prediction, where k is the mode number.

[0034] Preferably, step S3 specifically includes:

[0035] S31. The collected spatial distribution data of longitudinal smoke layer thickness in the tunnel throughout the entire time period is organized into a two-dimensional matrix H according to the time series, with dimensions (T, N). x ), where T is the total number of time steps, N x This represents the number of longitudinal sampling points in the tunnel.

[0036] S32. Decenter the two-dimensional matrix H and calculate its mean vector x in the spatial dimension. mean The decentralized matrix H is obtained.c ;

[0037] S33. For the decentralized matrix H c The Singular Value Decomposition (SVD) is performed using the following formula:

[0038] (2);

[0039] In the formula, U is the left singular matrix, S is the singular value diagonal matrix, and V T The transpose of the right singular matrix; the first K principal modes form the mode matrix with dimensions (K, N). x K represents the preset number of main modes, which can be 3 to 5.

[0040] S34. Based on the principal mode matrix, calculate the modal coefficient matrix for the entire time period, and transform the original high-dimensional smoke layer height field into K-dimensional time-series modal coefficients to complete the dimensionality reduction process.

[0041] Preferably, step S4 specifically includes:

[0042] S41. Input the multi-task prediction model input layer as a fixed-length sliding window temporal feature matrix with dimensions (W, F), where W is the sliding window time length and F is the input feature dimension;

[0043] S42. Construct a causal convolutional basic module. The causal convolutional basic module adopts the CausalConv1d structure. The convolutional output is guaranteed to be consistent with the input temporal length through pre-padding, and there is no leakage of future information, thus ensuring temporal causality.

[0044] S43. Construct the CAUSAL-TCN feature extraction backbone, which includes an input projection layer and stacked CAUSAL-TCN residual blocks; the input projection layer maps the input feature dimension to a preset hidden layer dimension; each CAUSAL-TCN residual block includes group normalization, ReLU activation function, causal convolutional layer and Dropout layer, the causal convolutional layer adopts dilated convolution structure to expand the temporal receptive field, and the residual blocks are connected by residuals to ensure gradient propagation stability;

[0045] S44. Construct a dual-branch multi-task prediction head, including a flue gas counterflow length prediction head and a smoke layer modal coefficient prediction head; the CAUSAL-TCN feature extracts the last-moment features of the backbone output, and inputs them into the flue gas counterflow length prediction head and the smoke layer modal coefficient prediction head respectively; the flue gas counterflow length prediction head and the smoke layer modal coefficient prediction head each contain a bottleneck layer and a fully connected layer, the bottleneck layer is used to transform the input features and use the ReLU activation function, the fully connected layer is a linear output layer without activation function, and outputs a one-dimensional flue gas counterflow length prediction value and a K-dimensional smoke layer modal coefficient prediction value respectively, where K is the preset number of main modes.

[0046] Preferably, step S5 specifically includes:

[0047] S51. Construct a data fitting master loss function, which includes a flue gas counterflow length regression loss term and a smoke layer modal coefficient regression loss term. Both terms are calculated using the SmoothL1 loss function, and the smoke layer modal coefficient regression loss term is weighted using a first weighting coefficient to obtain the weighted master loss value. The formula for calculating the master loss value is:

[0048] (3);

[0049] In the formula, L len L is the regression loss term for flue gas counterflow length. coef λ is the regression loss term for the smoke layer modal coefficients. coef Weights for modal coefficient loss;

[0050] S52. Construct a physical prior loss function, which includes a length prior loss term and a modal coefficient prior loss term, and the calculation formula is as follows:

[0051] (4);

[0052] In the formula, L len,phy The SmoothL1 loss is the difference between the predicted flue gas backflow length and the prior flue gas propagation length. coef,phy The SmoothL1 loss between the predicted and prior values ​​of the time-series modal coefficients is λ. prior,len , λ prior,coef These are the weighting coefficients for the corresponding losses;

[0053] S53. Construct a temporal smoothing constraint loss function. This function penalizes the temporal volatility of the prediction results by calculating the mean of the squares of the differences between the predicted flue gas counterflow lengths at adjacent times, thereby suppressing high-frequency jumps and ensuring the temporal continuity of the prediction results. The calculation formula is as follows:

[0054] (5);

[0055] In the formula, Lˊ(t) is the predicted value of the flue gas counterflow length at time t, and T is the total time series length;

[0056] S54. A progressive ramp-in loading mechanism is adopted, and a linear growth time interval for the physical prior loss is set. Within the preset initial training time interval, the weight coefficients corresponding to the physical prior loss function are linearly increased from zero to the preset target value to avoid the strong pulling effect of physical prior when there is insufficient data in the early stage of the fire.

[0057] S55. The total loss function is constructed by weighted summing of the data fitting main loss function, the physical prior loss function, and the temporal smoothing constraint loss function; the calculation formula is as follows:

[0058] (6);

[0059] In the formula, λ smooth To smooth out the constraint loss weights, values ​​are set between 0.0001 and 0.001.

[0060] S56. An optimizer is used to train the multi-task prediction model end-to-end, and a learning rate decay strategy based on validation set loss and an early stopping mechanism are combined to complete the model training and save the optimal model weights.

[0061] Preferably, step S6 specifically includes:

[0062] S61. Load the pre-trained model weights, feature standardization parameters, spatial principal modes extracted by intrinsic orthogonal decomposition, and their corresponding spatial mean vectors;

[0063] S62. Perform time-by-time sliding window prediction for the preset prediction period. For early prediction times where the data length before the current time is less than the preset sliding window length of the model, fill the window by copying the earliest time data and padding it to the left, so as to achieve continuous prediction throughout the entire time period.

[0064] S63. Perform denormalization on the standardized predicted values ​​output by the model to restore the predicted values ​​of flue gas counterflow length and time-series modal coefficients with physical meaning;

[0065] S64. Based on the predicted values ​​of the temporal modal coefficients after denormalization, the spatial principal modes, and the spatial mean vector, the spatial distribution field of smoke layer thickness at each location in the longitudinal direction of the tunnel is reconstructed;

[0066] S65. Output the prediction results, which include at least the time series data of flue gas counterflow length and the time series data of the reconstructed spatial distribution of smoke layer thickness, and generate a comparative visualization result for evaluating the prediction effect.

[0067] Beneficial effects:

[0068] 1. Existing pure data-driven models rely solely on data fitting, which can easily produce predictions that do not conform to physical laws under unseen fire conditions (such as different combinations of fire source power and ventilation speed). This method constructs a complete physical feature system and introduces hybrid physical priors (smoke spread length prior and smoke layer height field shape prior) as soft constraints for model training. This forces the model to follow the physical laws of smoke spread while fitting the data, which greatly improves the model's generalization ability and the reliability of prediction results under complex conditions.

[0069] 2. Traditional empirical formulas or simple deep learning models can only estimate the macroscopic length of the smoke backflow, failing to provide specific smoke layer thickness distributions upstream and downstream of the fire source. This makes it difficult to meet the spatial situational awareness needs of fire rescue operations. This method employs intrinsic orthogonal decomposition to reduce the dimensionality of the high-dimensional smoke layer spatial field, transforming complex spatial field prediction into the prediction of a few time-series modal coefficients. Through the constructed multi-task prediction model, the smoke backflow length and modal coefficients are output simultaneously, and finally, the spatial field is reconstructed. The model not only outputs the smoke backflow length (macroscopic parameter) but also reconstructs the spatial distribution of smoke layer thickness at various longitudinal locations within the tunnel (microscopic field) in real time, providing more comprehensive decision-making information for emergency command.

[0070] 3. Traditional recurrent neural networks are prone to gradient vanishing when processing long-term time-series data, and some model structures may unintentionally leak future information, undermining the causality of predictions. This method constructs a temporal convolutional network based on causal convolution, ensures that the output at the current moment depends only on information from past moments through pre-padding, and expands the receptive field through dilated convolution, ensuring strict temporal causality in the prediction process. At the same time, it effectively captures long-distance temporal dependencies and can still maintain stable prediction performance when there is little data in the early stage of a fire or when the fire situation changes abruptly.

[0071] 4. Data is scarce in the early stages of a fire, and directly applying physical constraints may mislead the model. Furthermore, a purely data-driven model may produce drastically fluctuating predictions, which is detrimental to engineering applications. This method avoids early-stage forced loading by designing a hybrid loss function and employing a progressive ramp-in loading mechanism (allowing the physical prior loss to grow linearly from zero). Simultaneously, it introduces a temporal smoothing constraint loss to penalize prediction jumps between adjacent time points, ensuring model stability in the early training phase and resulting in a smoother, more continuous prediction curve that better reflects the actual physical process of fire evolution, facilitating engineering interpretation and application. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 A flowchart of a real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks provided in an embodiment of the present invention; Figure 2 This is a structural diagram of the CAUSAL-TCN residual block provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of flue gas propagation length prediction provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the prediction of flue gas spread height provided in an embodiment of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] like Figure 1 As shown, this invention provides a real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks, comprising the following steps:

[0076] S1. Acquire multi-source time-series data under tunnel fire scenarios, perform data preprocessing, and construct a multi-condition dataset; specifically including:

[0077] S11. Obtain four types of core data under different fire source conditions in the tunnel through fire dynamics simulation or on-site monitoring, including: distributed temperature sensor DEVC time series data, heat release rate HRR time series data, smoke backflow length calibration data, and spatial distribution data of longitudinal smoke layer thickness in the tunnel.

[0078] Among them, fire dynamics simulation requires the use of FDS to construct a numerical model of the tunnel. The specific steps for constructing the numerical model of the tunnel are as follows:

[0079] Step 1.1.1: Construct a numerical model of the tunnel using FDS. The tunnel dimensions are 300m long, 9m wide, and 6m high. Three fire source locations are set at distances of 50m, 100m, 150m, 200m, and 250m from the tunnel entrance. The fire source power is set in 8 levels, ranging from 5MW to 40MW in 5MW intervals. The longitudinal ventilation wind speed is set in 8 levels, ranging from 1m / s to 12m / s in 1m / s intervals.

[0080] Step 1.1.2: The fire source is set as a rectangular area with a cross-section of 2m × 5m, 0.5m above the ground. Its power growth follows the t² fire model, and the total simulation time is not less than 600s; the fuel type used is heptane (C7H). 16 The carbon monoxide generation rate in its combustion products is set to 0.006, and the particulate matter production rate is set to 0.015; the tunnel wall material is defined as concrete, and its thermal parameters are set using the default configuration of the FDS software.

[0081] Step 1.1.3: A dynamic mesh generation strategy is adopted. A high-precision mesh of 0.25m×0.25m×0.25m is used in the 50m radius area of ​​the fire source center and the 5m area around the fan. A standard mesh of 0.5m×0.5m×0.5m is used in the rest of the tunnel area. The boundary conditions at both ends of the tunnel are set to "OPEN" to simulate an open atmospheric environment.

[0082] Step 1.1.4: The ventilation system consists of two sets of jet fans, symmetrically arranged at 75m and 225m from the tunnel entrance. Each set contains two fans. The size of a single fan is 3m×1.5m×1.5m. Its top is installed 1m from the tunnel ceiling and is equipped with a 1m diameter air supply duct. The air supply volume flow rate is precisely controlled by the HVAC system built into the FDS to achieve the target wind speed condition.

[0083] Step 1.1.5: Install temperature sensors at 5m intervals along the longitudinal centerline of the tunnel, with a total of 61 temperature sensors and a sampling time interval of 1s; set up a temperature monitoring plane at the longitudinal center of the tunnel, export temperature cloud map data, and the image sampling time interval is 1s; the total simulation time is 600s, and the data sampling time interval is 1s.

[0084] S12. For the dual-header format of unit row and true header row of DEVC time series data and HRR time series data of distributed temperature sensor, the header structure is automatically identified and unified; missing values ​​in the data are filled with linear interpolation, and outliers that exceed the physical reasonable range are replaced with the average of the previous and next frames.

[0085] S13. Extract features from the time series data of the distributed temperature sensor DEVC to generate basic features including temperature statistical features, thermal mass features, coarse localization features of the flue gas front and thermal centroid features. Then, perform first-order difference and exponential moving average (EMA) processing on the basic features to generate time series derived features.

[0086] A method for feature extraction from time-series data of a distributed temperature sensor (DEVC) includes the following steps:

[0087] S131. Based on the preset ambient reference temperature of 20℃, calculate the temperature rise d at each measuring point. T Based on the temperature rise data, statistical characteristics including the mean temperature of the entire tunnel, the 75th quantile, the 90th quantile, the 95th quantile, the 99th quantile, the maximum value, and the number of supercritical temperature measuring points are generated.

[0088] S132. Based on the temperature rise data, calculate the thermal mass characteristics and thermal energy characteristics of the entire tunnel section, the upstream area of ​​the fire source, and the downstream area of ​​the fire source, and further calculate the heat ratio characteristics of the upstream and downstream areas; wherein, the thermal mass characteristics are obtained by summing the positive temperature rise values ​​of each measuring point, and the thermal energy characteristics are obtained by summing the square values ​​of the positive temperature rise values ​​of each measuring point;

[0089] S133. Based on a preset critical temperature rise threshold, determine the coarse positioning length L of the flue gas front upstream of the fire source. up,coarse and downstream flue gas propagation length L down,coarse Based on the temperature rise weighted calculation method, the location of the thermal centroid upstream of the fire source is determined, and the corresponding flue gas length L is generated using the thermal centroid method. up,cm and its downstream corresponding characteristics;

[0090] S134. For all the basic features generated in the above steps, set two time constants τ of 5s and 15s to generate the corresponding exponential moving average features, and calculate the first-order time difference of each basic feature to generate the rate of change feature.

[0091] S135. The basic features, exponential moving average features, and rate of change features are concatenated to form a complete time-series feature matrix, which serves as the input feature part of the dataset.

[0092] S14. Perform Z-score standardization on the basic features and time-series derived features, and independently standardize the flue gas counterflow length calibration data and the spatial distribution data of tunnel longitudinal smoke layer thickness to ensure the consistency of training data distribution; finally, divide the training set and validation set according to time order, with the validation set accounting for 20%~30%.

[0093] Step S2. Based on the multi-source time series data, construct a hybrid physical prior, which includes a prior for flue gas propagation length and a prior for the shape of the smoke layer height field;

[0094] The steps to construct a hybrid physics prior are as follows:

[0095] S21. Based on the aforementioned multi-source time-series data, the coarse location length of the flue gas front and the flue gas length obtained by thermal center-of-mass method are fused using a linear weighted fusion method to generate a time-varying prior flue gas propagation length L. phy (t), the calculation formula is:

[0096] (1);

[0097] In the formula, w coarse For the weight of the thick front, w cm The weight of the thermal centroid is w. coarse +w cm =1;

[0098] S22. Based on the empirical law of longitudinal attenuation of smoke in tunnel fires, with the fire source location as the center and the a priori smoke propagation length L... phy (t) is used as the upstream attenuation characteristic length to construct the theoretical distribution field H of the smoke layer thickness, which exhibits an exponential attenuation distribution along the tunnel longitudinal direction. phy (x,t) serves as the prior for the shape of the smoke layer height field;

[0099] S23. Prioritize the shape of the smoke layer height field H phy Projecting (x,t) onto the spatial principal modes extracted by the intrinsic orthogonal decomposition generates the corresponding prior values ​​C of the modal coefficients. phy,k (t), serving as the physical constraint for smoke layer prediction, where k is the mode number.

[0100] S3. The smoke layer height field is reduced in dimensionality using intrinsic orthogonal decomposition to extract the spatial principal modes and their corresponding temporal mode coefficients, transforming high-dimensional spatial field prediction into low-dimensional mode coefficient prediction; specifically including:

[0101] S31. The collected spatial distribution data of longitudinal smoke layer thickness in the tunnel throughout the entire time period is organized into a two-dimensional matrix H according to the time series, with dimensions (T, N). x ), where T is the total number of time steps, N x This represents the number of longitudinal sampling points in the tunnel.

[0102] S32. Decenter the two-dimensional matrix H and calculate its mean vector x in the spatial dimension. mean The decentralized matrix H is obtained. c ;

[0103] S33. For the decentralized matrix H c The Singular Value Decomposition (SVD) is performed using the following formula:

[0104] (2);

[0105] In the formula, U is the left singular matrix, S is the singular value diagonal matrix, and V T The transpose of the right singular matrix; the first K principal modes form the mode matrix with dimensions (K, N). x K represents the preset number of main modes, which can be 3 to 5.

[0106] S34. Based on the principal mode matrix, calculate the modal coefficient matrix for the entire time period, and transform the original high-dimensional smoke layer height field into K-dimensional time-series modal coefficients to complete the dimensionality reduction process.

[0107] S4. Construct a multi-task prediction model based on a causal temporal convolutional network, wherein the multi-task prediction model is used to simultaneously predict the flue gas counterflow length and the temporal modal coefficients; specifically including:

[0108] S41. Input the multi-task prediction model input layer as a fixed-length sliding window temporal feature matrix with dimensions (W, F), where W is the sliding window time length, ranging from 30 to 60, and F is the input feature dimension;

[0109] S42. Construct a causal convolutional basic module. The causal convolutional basic module adopts the CausalConv1d structure. The convolutional output is guaranteed to be consistent with the input temporal length through pre-padding, and there is no leakage of future information, thus ensuring temporal causality.

[0110] S43. Construct the CAUSAL-TCN feature extraction backbone, including an input projection layer and stacked CAUSAL-TCN residual blocks; the input projection layer maps the input feature dimension to a preset hidden layer dimension, with a hidden layer dimension of 64; each CAUSAL-TCN residual block contains two layers of group normalization, ReLU activation function, causal convolutional layer, and Dropout layer (see...). Figure 2 The causal convolutional layer uses a dilated convolution structure to expand the temporal receptive field. The dilation coefficient increases in powers of 2. The number of stacked residual blocks is 4 to 6, and the dropout rate of the dropout layer is 0.1 to 0.3. The residual blocks are connected by residuals to ensure the stability of gradient propagation.

[0111] S44. Construct a dual-branch multi-task prediction head, including a flue gas counterflow length prediction head and a smoke layer modal coefficient prediction head; the CAUSAL-TCN feature extracts the last-moment features of the backbone output, and inputs them into the flue gas counterflow length prediction head and the smoke layer modal coefficient prediction head respectively; the flue gas counterflow length prediction head and the smoke layer modal coefficient prediction head each contain a bottleneck layer and a fully connected layer, the bottleneck layer is used to transform the input features and use the ReLU activation function, the fully connected layer is a linear output layer without an activation function, and outputs a one-dimensional flue gas counterflow length prediction value and a K-dimensional smoke layer modal coefficient prediction value respectively, where K is the preset number of main modes.

[0112] S5. Using the dataset, and based on a hybrid loss function that combines data fitting loss, physical prior loss, and smoothing constraints, combined with a progressive ramp-in loading mechanism, the multi-task prediction model is trained; specifically including:

[0113] S51. Construct a data fitting master loss function, which includes a flue gas counterflow length regression loss term and a smoke layer modal coefficient regression loss term. Both terms are calculated using the SmoothL1 loss function, and the smoke layer modal coefficient regression loss term is weighted using a first weighting coefficient to obtain the weighted master loss value. The formula for calculating the master loss value is:

[0114] (3);

[0115] In the formula, L len L is the regression loss term for flue gas counterflow length. coef λ is the regression loss term for the smoke layer modal coefficients. coef Weights for modal coefficient loss;

[0116] S52. Construct a physical prior loss function, which includes a length prior loss term and a modal coefficient prior loss term, and the calculation formula is as follows:

[0117] (4);

[0118] In the formula, L len,phy The SmoothL1 loss is the difference between the predicted flue gas backflow length and the prior flue gas propagation length. coef,phy The SmoothL1 loss between the predicted and prior values ​​of the time-series modal coefficients is λ. prior,len , λ prior,coef These are the weighting coefficients for the corresponding losses;

[0119] S53. Construct a temporal smoothing constraint loss function. This function penalizes the temporal volatility of the prediction results by calculating the mean of the squares of the differences between the predicted flue gas counterflow lengths at adjacent times, thereby suppressing high-frequency jumps and ensuring the temporal continuity of the prediction results. The calculation formula is as follows:

[0120] (5);

[0121] In the formula, Lˊ(t) is the predicted value of the flue gas counterflow length at time t, and T is the total time series length;

[0122] S54. A progressive ramp-in loading mechanism is adopted, and a linear growth time interval for the physical prior loss is set. Within the preset initial training time interval of 0s to 120s, the weight coefficients corresponding to the physical prior loss function are linearly increased from zero to the preset target value to avoid the strong pulling effect of physical prior when there is insufficient data in the early stage of the fire.

[0123] S55. The total loss function is constructed by weighted summing of the data fitting main loss function, the physical prior loss function, and the temporal smoothing constraint loss function; the calculation formula is as follows:

[0124] (6);

[0125] In the formula, λ smooth To smooth out the constraint loss weights, values ​​are set between 0.0001 and 0.001.

[0126] S56. The multi-task prediction model is trained end-to-end using the AdamW optimizer, with an initial learning rate of 5×10⁻⁻⁻⁴. 4 ~1×10⁻³, combined with a learning rate decay strategy based on validation set loss and an early stopping mechanism, to complete model training and save the optimal model weights.

[0127] S6. Using a trained multi-task prediction model, process the real-time input time-series data, synchronously output predicted values ​​for flue gas counterflow length and time-series modal coefficients, and reconstruct the real-time spatial distribution of the smoke layer based on the predicted time-series modal coefficients and the spatial dominant mode; specifically including:

[0128] S61. Load the pre-trained model weights, feature standardization parameters, spatial principal modes extracted by intrinsic orthogonal decomposition, and their corresponding spatial mean vectors;

[0129] S62. Perform time-by-time sliding window prediction for the preset 0~600s prediction period. For early prediction times where the data length before the current time is less than the preset sliding window length of the model, fill the window by copying the earliest time data and padding it to the left, so as to achieve continuous prediction throughout the entire time period.

[0130] S63. Perform denormalization on the standardized predicted values ​​output by the model to restore the predicted values ​​of flue gas counterflow length and time-series modal coefficients with physical meaning;

[0131] S64. Based on the predicted values ​​of the temporal modal coefficients after denormalization, the spatial principal modes, and the spatial mean vector, the spatial distribution field of smoke layer thickness at each location in the longitudinal direction of the tunnel is reconstructed;

[0132] S65. Output the prediction results, which should include at least the time-series data of the flue gas counterflow length (refer to...). Figure 3 ), reconstructed temporal data of the spatial distribution of smoke layer thickness (refer to Figure 4 ), and generate comparative visualizations for evaluating the predictive performance; The visualization results simultaneously plot time-series curves of the actual values ​​and model predictions. The actual values ​​are FDS simulation calibration data, and the model predictions are the output results of this method. (See attached image.) Figure 4 In the diagram, the horizontal axis represents the fire duration (0-300s), and the vertical axis represents the smoke spread height (m). Three key moments, 30s, 100s, and 600s after the fire started, were selected to compare the actual and predicted values ​​of the smoke spread height. The horizontal axis represents the fire duration, and the vertical axis represents the smoke spread height, which intuitively demonstrates the prediction accuracy and temporal stability of this invention.

[0133] Figure 3The true value in the figure refers to the flue gas spread length data collected through FDS numerical simulation, while the model prediction value is the flue gas spread length output by the multi-task prediction model trained using this invention. Figure 4 The true value refers to the smoke spread height collected through FDS numerical simulation, while the model prediction value refers to the smoke spread height output by the multi-task prediction model trained using this invention.

[0134] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks, characterized in that, Includes the following steps: S1. Acquire multi-source time-series data in tunnel fire scenarios, perform data preprocessing, and construct a multi-condition dataset; S2. Based on the multi-source time-series data, construct a hybrid physical prior, which includes a prior for flue gas propagation length and a prior for the shape of the smoke layer height field; S3. The intrinsic orthogonal decomposition is used to reduce the dimensionality of the smoke layer height field, extract the spatial principal mode and the corresponding temporal mode coefficients, and transform the high-dimensional spatial field prediction into low-dimensional mode coefficient prediction. S4. Construct a multi-task prediction model based on a causal temporal convolutional network, wherein the multi-task prediction model is used to simultaneously predict the flue gas backflow length and the temporal modal coefficients; S5. Using the dataset, and based on a hybrid loss function that combines data fitting loss, physical prior loss, and smoothing constraints, combined with a progressive ramp-in loading mechanism, the multi-task prediction model is trained. S6. Using the trained multi-task prediction model, process the real-time input time-series data, synchronously output the predicted value of flue gas counterflow length and the predicted value of time-series modal coefficient, and reconstruct the real-time spatial distribution of the smoke layer based on the predicted value of the time-series modal coefficient and the spatial principal mode.

2. The real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks according to claim 1, characterized in that, Step S1 specifically includes: S11. Obtain four types of core data under different fire source conditions in the tunnel through fire dynamics simulation or on-site monitoring, including: distributed temperature sensor DEVC time series data, heat release rate HRR time series data, smoke backflow length calibration data, and spatial distribution data of longitudinal smoke layer thickness in the tunnel. S12. For the dual-header format of unit row and true header row of DEVC time series data and HRR time series data of distributed temperature sensor, the header structure is automatically identified and unified; missing values ​​in the data are filled with linear interpolation, and outliers that exceed the physical reasonable range are replaced with the average of the previous and next frames. S13. Extract features from the time series data of the distributed temperature sensor DEVC to generate basic features including temperature statistical features, thermal mass features, coarse localization features of the flue gas front and thermal centroid features. Then, perform first-order difference and exponential moving average (EMA) processing on the basic features to generate time series derived features. S14. Perform Z-score standardization on the basic features and time-series derived features, and independently standardize the flue gas counterflow length calibration data and the spatial distribution data of tunnel longitudinal smoke layer thickness to ensure the consistency of training data distribution; finally, divide the training set and validation set according to time order, with the validation set accounting for 20%~30%.

3. The real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks according to claim 2, characterized in that, Step S13, the method for feature extraction of the distributed temperature sensor DEVC time series data, includes the following steps: S131. Calculate the temperature rise d at each measuring point based on the preset environmental reference temperature. T The temperature distribution characteristics of the entire tunnel are statistically generated, and the temperature distribution characteristics include at least the mean, specified high quantile, maximum value, and number of supercritical temperature measurement points. S132. Based on the temperature rise data, calculate the thermal mass characteristics and thermal energy characteristics of the entire tunnel section, the upstream area of ​​the fire source, and the downstream area of ​​the fire source, and further calculate the heat ratio characteristics of the upstream and downstream areas; wherein, the thermal mass characteristics are obtained by summing the positive temperature rise values ​​of each measuring point, and the thermal energy characteristics are obtained by summing the square values ​​of the positive temperature rise values ​​of each measuring point; S133. Based on a preset critical temperature rise threshold, determine the coarse positioning length L of the flue gas front upstream of the fire source. up,coarse and downstream flue gas propagation length L down,coarse Based on the temperature rise weighted calculation method, the location of the thermal centroid upstream of the fire source is determined, and the corresponding flue gas length L is generated using the thermal centroid method. up,cm and its downstream corresponding characteristics; S134. For all the basic features generated in the above steps, generate corresponding exponential moving average features using multiple preset time constants τ, and calculate the first-order time difference of each basic feature to generate rate of change features. S135. The basic features, exponential moving average features, and rate of change features are concatenated to form a complete time-series feature matrix, which serves as the input feature part of the dataset.

4. The real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks according to claim 3, characterized in that, In step S2, the steps for constructing the hybrid physics prior are as follows: S21. Based on the aforementioned multi-source time-series data, the coarse location length of the flue gas front and the flue gas length obtained by thermal center-of-mass method are fused using a linear weighted fusion method to generate a time-varying prior flue gas propagation length L. phy (t), the calculation formula is: （1）； In the formula, w coarse For the weight of the thick front, w cm The weight of the thermal centroid is w. coarse +w cm =1; S22. Based on the empirical law of longitudinal attenuation of smoke in tunnel fires, with the fire source location as the center and the a priori smoke propagation length L... phy (t) is used as the upstream attenuation characteristic length to construct the theoretical distribution field H of the smoke layer thickness, which exhibits an exponential attenuation distribution along the tunnel longitudinal direction. phy (x,t) serves as the prior for the shape of the smoke layer height field; S23. Prioritize the shape of the smoke layer height field H phy Projecting (x,t) onto the spatial principal modes extracted by the intrinsic orthogonal decomposition generates the corresponding prior values ​​C of the modal coefficients. phy,k (t), serving as the physical constraint for smoke layer prediction, where k is the mode number.

5. The real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks according to claim 1, characterized in that, Step S3 specifically includes: S31. The collected spatial distribution data of longitudinal smoke layer thickness in the tunnel throughout the entire time period is organized into a two-dimensional matrix H according to the time series, with dimensions (T, N). x ), where T is the total number of time steps, N x This represents the number of longitudinal sampling points in the tunnel. S32. Decenter the two-dimensional matrix H and calculate its mean vector x in the spatial dimension. mean The decentralized matrix H is obtained. c ; S33. For the decentralized matrix H c The Singular Value Decomposition (SVD) is performed using the following formula: （2）； In the formula, U is the left singular matrix, S is the singular value diagonal matrix, and V T The transpose of the right singular matrix; the first K principal modes form the mode matrix with dimensions (K, N). x K represents the preset number of main modes, which can be 3 to 5. S34. Based on the principal mode matrix, calculate the modal coefficient matrix for the entire time period, and transform the original high-dimensional smoke layer height field into K-dimensional time-series modal coefficients to complete the dimensionality reduction process.

6. The real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks according to claim 4, characterized in that, Step S4 specifically includes: S41. Input the multi-task prediction model input layer as a fixed-length sliding window temporal feature matrix with dimensions (W, F), where W is the sliding window time length and F is the input feature dimension; S42. Construct a causal convolutional basic module. The causal convolutional basic module adopts the CausalConv1d structure. The convolutional output is guaranteed to be consistent with the input temporal length through pre-padding, and there is no leakage of future information, thus ensuring temporal causality. S43. Construct the CAUSAL-TCN feature extraction backbone, which includes an input projection layer and stacked CAUSAL-TCN residual blocks; the input projection layer maps the input feature dimension to a preset hidden layer dimension; each CAUSAL-TCN residual block includes group normalization, ReLU activation function, causal convolutional layer and Dropout layer, the causal convolutional layer adopts dilated convolution structure to expand the temporal receptive field, and the residual blocks are connected by residuals to ensure gradient propagation stability; S44. Construct a dual-branch multi-task prediction head, including a flue gas counterflow length prediction head and a smoke layer modal coefficient prediction head; the CAUSAL-TCN feature extracts the last-moment features of the backbone output, and inputs them into the flue gas counterflow length prediction head and the smoke layer modal coefficient prediction head respectively; the flue gas counterflow length prediction head and the smoke layer modal coefficient prediction head each contain a bottleneck layer and a fully connected layer, the bottleneck layer is used to transform the input features and use the ReLU activation function, the fully connected layer is a linear output layer without activation function, and outputs a one-dimensional flue gas counterflow length prediction value and a K-dimensional smoke layer modal coefficient prediction value respectively, where K is the preset number of main modes.

7. The real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks according to claim 4, characterized in that, Step S5 specifically includes: S51. Construct a data fitting master loss function, which includes a flue gas counterflow length regression loss term and a smoke layer modal coefficient regression loss term. Both terms are calculated using the SmoothL1 loss function, and the smoke layer modal coefficient regression loss term is weighted using a first weighting coefficient to obtain the weighted master loss value. The formula for calculating the master loss value is: （3）； In the formula, L len L is the regression loss term for flue gas counterflow length. coef λ is the regression loss term for the smoke layer modal coefficients. coef Weights for modal coefficient loss; S52. Construct a physical prior loss function, which includes a length prior loss term and a modal coefficient prior loss term, and the calculation formula is as follows: （4）； In the formula, L len,phy The SmoothL1 loss is the difference between the predicted flue gas backflow length and the prior flue gas propagation length. coef,phy The SmoothL1 loss between the predicted and prior values ​​of the time-series modal coefficients is λ. prior,len , λ prior,coef These are the weighting coefficients for the corresponding losses; S53. Construct a temporal smoothing constraint loss function. This function penalizes the temporal volatility of the prediction results by calculating the mean of the squares of the differences between the predicted flue gas counterflow lengths at adjacent times, thereby suppressing high-frequency jumps and ensuring the temporal continuity of the prediction results. The calculation formula is as follows: （5）； In the formula, Lˊ(t) is the predicted value of the flue gas counterflow length at time t, and T is the total time series length; S54. A progressive ramp-in loading mechanism is adopted, and a linear growth time interval for the physical prior loss is set. Within the preset initial training time interval, the weight coefficients corresponding to the physical prior loss function are linearly increased from zero to the preset target value to avoid the strong pull effect of physical prior when there is insufficient data in the early stage of the fire. S55. The total loss function is constructed by weighted summing of the main loss function for data fitting, the physical prior loss function, and the temporal smoothing constraint loss function; the calculation formula is as follows: （6）； In the formula, λ smooth To smooth out the constraint loss weights, values ​​are set between 0.0001 and 0.

001. S56. An optimizer is used to train the multi-task prediction model end-to-end, and a learning rate decay strategy based on validation set loss and an early stopping mechanism are combined to complete the model training and save the optimal model weights.

8. The real-time prediction method for tunnel smoke propagation based on physical priors and temporal networks according to claim 5, characterized in that, Step S6 specifically includes: S61. Load the pre-trained model weights, feature standardization parameters, spatial principal modes extracted by intrinsic orthogonal decomposition, and their corresponding spatial mean vectors; S62. Perform time-by-time sliding window prediction for the preset prediction period. For early prediction times where the data length before the current time is less than the preset sliding window length of the model, fill the window by copying the earliest time data and padding it to the left, so as to achieve continuous prediction throughout the entire time period. S63. Perform denormalization on the standardized predicted values ​​output by the model to restore the predicted values ​​of flue gas counterflow length and time-series modal coefficients with physical meaning; S64. Based on the predicted values ​​of the temporal modal coefficients after denormalization, the spatial principal modes, and the spatial mean vector, the spatial distribution field of smoke layer thickness at each location in the longitudinal direction of the tunnel is reconstructed; S65. Output the prediction results, which include at least the time series data of flue gas counterflow length and the time series data of the reconstructed spatial distribution of smoke layer thickness, and generate a comparative visualization result for evaluating the prediction effect.

Images