Emergency material transportation network design method

Abstract

The invention discloses an emergency material transportation network design method, and relates to the technical field of emergency material transportation. Comprising the steps of obtaining a historical disaster data set; constructing a hybrid prediction model CNN-MAFP, and training the hybrid prediction model CNN-MAFP through the historical disaster data set; inputting real-time disaster data of a to-be-detected area into the trained hybrid prediction model CNN-MAFP to obtain a post-disaster emergency demand and road segment integrity degree calculation function of the to-be-detected area; and planning a cross-regional network for transporting emergency materials according to the post-disaster emergency demand of the to-be-measured region and the road segment integrity degree calculation function. According to the method, the distribution or change rule of the disaster data does not need to be assumed in advance, and higher flexibility and adaptability are achieved when complex and changeable disaster data are processed.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency material transportation, and in particular to a method for designing an emergency material transportation network. Background Art

[0002] With the increase in population density, the acceleration of urbanization and the frequent extreme climate change, humans have become more vulnerable to natural disasters. In particular, the time, location and scale of sudden disasters such as earthquakes, mountain torrents and mudslides are often difficult to accurately predict, resulting in the inability to accurately estimate post-disaster needs. As an important part of the pre-disaster preparation stage, emergency network design has the characteristics of being oriented to future needs. Therefore, when designing emergency networks, the disaster-causing factors and the current and future socio-economic characteristics of the region should be comprehensively considered. Researching a post-disaster emergency demand forecasting method that conforms to future realities and combines regional characteristics can effectively guide pre-disaster deployment, thereby improving emergency efficiency and reducing costs.

[0003] At present, the design of emergency networks mainly adopts mathematical regression model analysis. This method constructs a mathematical regression model for disaster prediction by using historical data such as disaster risk index, number of damaged disaster-bearing bodies, and intensity of disaster-causing factors collected from news reports and other channels. The prediction results of the regression model are applied to the construction of the emergency network to ensure that emergency responses can be made in a timely and effective manner when disasters occur.

[0004] The defects of the above-mentioned existing technology are: in the process of constructing a mathematical regression model, it is necessary to make prior assumptions about the distribution or change pattern of disaster data; however, disaster data is complex and uncertain, and will change with time and space. The prior assumptions based on historical data cannot fully and accurately describe the true characteristics of the data, making the prediction results of the regression model inaccurate, which in turn leads to insufficient accuracy of the emergency material transportation network. Summary of the invention

[0005] Based on this, it is necessary to provide an emergency material transportation network design method to address the above technical issues.

[0006] An embodiment of the present invention provides a method for designing an emergency material transportation network, comprising:

[0007] Obtain historical datasets of sudden disasters;

[0008] A feature extraction module is constructed by two stacked one-dimensional convolutional layers (1D CNN). The end of the feature extraction module is connected to a multi-head attention mechanism consisting of multiple independent attention heads, and a feature projection module is added to the end of the multi-head attention mechanism to construct a hybrid prediction model CNN-MAFP.

[0009] The hybrid prediction model CNN-MAFP is trained using the historical dataset of sudden disasters;

[0010] The real-time disaster data of the area to be tested is input into the trained hybrid prediction model CNN-MAFP. The feature extraction module is used to extract features from the real-time disaster data to obtain local features that characterize the temporal pattern and spatial distribution of the disaster. The attention weights of the local features are obtained through the multi-head attention mechanism, and the attention weights are concatenated and weighted summed to model the global dependency between local features, thereby obtaining high-dimensional features that reflect the development of the disaster and the scope of impact. The feature projection module is used to reduce the high-dimensional features to the target output space, and the post-disaster emergency demand and road section integrity measurement function after the disaster in the area to be tested are obtained.

[0011] According to the post-disaster emergency demand and road section integrity measurement function, a cross-regional network for transporting emergency supplies is determined; wherein the cross-regional network includes disaster points, alternative facility points, emergency warehouse points and connecting sections between nodes in different regions.

[0012] Optionally, the historical disaster data set includes disaster characteristic data, socio-economic characteristic data and disaster loss characteristic data, and the matrix is ​​expressed as M=(Dis i,m ,Soc i,n ,Los i,k ),in:

[0013]

[0014]

[0015] Among them, disaster data represents the mth disaster characteristic value of disaster site i, social and economic data Represents the nth socioeconomic characteristic value of disaster site i, disaster loss data Represents the kth disaster loss characteristic value of disaster-affected point i.

[0016] Optionally, the post-disaster emergency demand introduces a regional correction factor K p and seasonal coefficient Q p , which specifically include:

[0017] Regional correction factor K p The calculation formula is expressed as:

[0018]

[0019] Among them, x pi represents the weight of the i-th reference indicator, and n represents the total number of indicators;

[0020] By regional correction factor Kp and seasonal coefficient Q p , the demand for emergency supplies is ranked according to the influence of region and season to obtain the post-disaster emergency demand.

[0021] Optionally, the road section integrity calculation function includes the degree of self-damage of the road body after the disaster and the quantified value of the earthquake damage to the roadbed and pavement due to the earthquake intensity, which specifically includes:

[0022]

[0023] Among them, P rij Indicates the degree of road self-damage after the disaster, k i represents the average earthquake damage index, X i represents the quantified value of earthquake damage to roadbed and pavement caused by earthquake intensity, σ represents the dispersion coefficient of earthquake damage, and P bij It represents the impact coefficient of the collapse of buildings along the street on the traffic capacity of the road section.

[0024] Optionally, feature extraction is performed on real-time disaster data through a feature extraction module, specifically including:

[0025] X′=ReLU(convld(X));

[0026] X″=ReLU(convld(X′));

[0027] in, is the input of historical disaster data, and It is the output value after the channel dimension is increased for the feature. represents a real matrix.

[0028] Optionally, the attention weights of local features are obtained through a multi-head attention mechanism, including:

[0029] The weight vector X″=[x1″,x2″,...,x n ″], the query matrix Q, key matrix K and value matrix V are obtained through three linear transformations, and the calculation formula includes:

[0030]

[0031] in, and is the learned weight matrix;

[0032] The query, key, and value matrices are divided into multiple headers and the computation process is expressed as:

[0033] split(Q,K,V)=[head1,…,head n ];

[0034] Among them, head i represents the i-th head of the query matrix Q, key matrix K and value matrix V;

[0035] Determine the attention of each head, for the i-th head:

[0036]

[0037] in, To stabilize the gradient, d k is the dimension of query and key;

[0038] The concatenated polyhedron output is mapped to the desired output dimension by a linear transformation, which is given by:

[0039] MultiHead(Q,K,V)=Concat(head1,…,head n )W 0 ;

[0040] (X″′,W score )=MultiHead(Q,K,V);

[0041] in, represents the characteristic output value, W score represents the attention score matrix, W 0 Represents the output projection matrix, Concat(head1,…,head n ) indicates the splicing of output signals from multiple heads. represents a real matrix.

[0042] Optionally, the high-dimensional features are reduced to the target output space through a feature projection module, specifically including:

[0043] The input features are projected into a higher-dimensional space through the feature projection layer ProjectionLayer, and then the odd and even features are separated and added after dimensionality reduction operations. Finally, a single-dimensional output is obtained through linear transformation. The calculation formula is:

[0044] Z odd =ReLU(W1·x odd +b1);

[0045] Z even =ReLU(W2·x even +b2);

[0046] Z=Z odd +Z even ;

[0047] Among them, x evenIndicates odd features, x odd represents even characteristics, Z even represents the odd features after dimensionality reduction, Z odd represents the even feature after dimensionality reduction operation, W represents the linear transformation coefficient, b represents the linear transformation constant, and ReLU() represents the linear transformation;

[0048] The single-dimensional output of the feature projection layer ProjectionLayer is mapped to the target output dimension through linear transformation for the final loss prediction. The calculation formula is:

[0049]

[0050] The predicted value of the model output Compared with the actual disaster loss y, the model parameters are optimized by minimizing the mean square error MSE loss function, and the calculation formula is:

[0051]

[0052] in, represents the predicted value of the model output, y represents the actual disaster loss, and N is the total number.

[0053] Optionally, identify a cross-regional network for transporting emergency supplies, including:

[0054] Minimize the cost of site selection and resource pre-positioning before disaster. The formula is:

[0055]

[0056] The formula for minimizing the transportation cost of post-disaster emergency supplies based on N sample data is:

[0057]

[0058] Construct a finite set of uncertain scenarios with N independent predicted actual values Each of the discrete random vectors The probability is equal to Constructing Experience Distribution As an estimate of the true probability distribution The reference distribution of is:

[0059]

[0060] in, and Indicates that the unit mass is concentrated on is the Dirac distribution at , and ∑ represents the Borelσ-algebra.

[0061] Compared with the prior art, the above-mentioned method for designing an emergency material transportation network provided by the embodiment of the present invention has the following beneficial effects:

[0062] The hybrid prediction model CNN-MAFP constructed by the present invention does not need to make any prior assumptions about the distribution or change rules of disaster data, but directly mines the potential rules and characteristics of real-time disaster data to obtain local features; then the attention weights of the local features are spliced ​​and weighted summed to achieve modeling of the global dependency relationship between the local features and obtain high-dimensional features; the high-dimensional features are reduced to the target output space to obtain post-disaster emergency demand and road section integrity measurement functions, which can accurately reflect the demand for emergency materials after the disaster and the degree of damage to the road section, so that the hybrid prediction model CNN-MAFP has higher flexibility and adaptability when processing complex and changeable disaster data, and provides support for rescue departments to formulate plans for cross-regional transportation of emergency materials, thereby improving the accuracy of the emergency material transportation network. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A schematic diagram of a flow chart of a method for designing an emergency material transportation network provided in an embodiment;

[0064] Figure 2 A diagram of a process of collecting historical disaster characteristic data of an emergency material transportation network design method provided in one embodiment;

[0065] Figure 3 A diagram of a random optimization model of an emergency network design for an emergency material transportation network design method provided in an embodiment. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0067] In one embodiment, a method for designing an emergency material transportation network is provided, such as Figure 1 As shown, the method includes:

[0068] Obtain historical datasets of sudden disasters.

[0069] A feature extraction module is constructed by two stacked one-dimensional convolutional layers 1D CNN. The end of the feature extraction module is connected to a multi-head attention mechanism composed of multiple independent attention heads, and a feature projection module is added to the end of the multi-head attention mechanism to construct a hybrid prediction model CNN-MAFP.

[0070] The hybrid prediction model CNN-MAFP is trained using historical datasets of sudden disasters.

[0071] The real-time disaster data of the area to be tested is input into the trained hybrid prediction model CNN-MAFP. The feature extraction module is used to extract the features of the real-time disaster data to obtain local features that characterize the temporal pattern and spatial distribution of the disaster. The attention weights of the local features are obtained through the multi-head attention mechanism, and the attention weights are spliced ​​and weighted summed to model the global dependency between local features, and high-dimensional features that reflect the development of the disaster and the scope of impact are obtained. The high-dimensional features are reduced to the target output space through the feature projection module to obtain the post-disaster emergency demand and road section integrity measurement function after the disaster in the area to be tested.

[0072] According to the post-disaster emergency demand and road section integrity measurement function, the cross-regional network for transporting emergency supplies is determined. The cross-regional network includes disaster points, alternative facility points, emergency warehouse points and connecting sections between nodes in different regions.

[0073] The specific implementation is:

[0074] 1. Construction of disaster data feature set

[0075] Effective deep learning and machine learning prediction results rely on a large number of data sets containing multiple valid features. However, the statistics of historical disaster information in multiple disaster databases do not have accurate statistical values ​​for the demand for various types of materials. Post-disaster emergency needs are generally calculated by counting the losses caused by the disaster. For example, emergency supplies such as water, food, and clothing are calculated by the affected population, and the degree of road damage is calculated by disaster-causing factors. First, based on the regional disaster system theory, a sudden disaster feature set is constructed, and historical sudden disaster data is collected according to the feature set classification to obtain a historical disaster data set. Taking earthquake disasters as an example, the data collection and processing process is as follows: Figure 2 shown.

[0076] The historical disaster data set includes disaster characteristic data, socio-economic characteristic data and disaster loss characteristic data, and the matrix is ​​expressed as M = (Dis i,m ,Soc i,n ,Los i,k ),in:

[0077]

[0078] Disaster data The mth disaster characteristic value of the disaster site i is represented by the disaster factor value (for example, for earthquakes, including level, intensity, and affected area) and environmental data values ​​(such as season, temperature, and daytime when the disaster occurs), and socioeconomic data It represents the nth socioeconomic characteristic value of the disaster site i, which mainly reflects the vulnerability of disaster sites in different regions, such as local population density, GDP, level and degree of improvement of transportation infrastructure, etc. It represents the kth disaster loss characteristic value of the disaster-affected point i, including the number of casualties and affected people, the rate of house collapse, etc.

[0079] 2. Calculation of post-disaster emergency needs and road damage

[0080] ① Post-disaster emergency demand is generally calculated by counting the losses or impacts caused by the disaster. For example, water and food are calculated by the total number of people affected after the disaster, and medicines are calculated by the total number of injured people. In addition, this method takes into account the emergency urgency, emergency importance and emergency rescue intensity shown by different regions when they are hit by disasters in different seasons, as well as the difference in demand for seasonal materials during sudden disasters, and introduces the regional correction factor K. p and seasonal coefficient Q p .

[0081] Regional correction factor K p The calculation formula is expressed as:

[0082]

[0083] Among them, x pi represents the weight of the i-th reference indicator, and n represents the total number of indicators.

[0084] Introducing the seasonal coefficient Q p For example, when resettling disaster victims in the hot summer, no or only a small amount of cotton-padded clothes are needed; when resettling disaster victims in the cold winter, a large amount of cotton-padded clothes are needed to keep warm.

[0085] However, considering that the temperature is lower at night, the seasonal coefficient for summer is 1, for spring and autumn it is 1.5, and for winter it is 2. However, the effect of the seasonal coefficient on pure water is the opposite, so the seasonal coefficient for summer is 2, for spring and autumn it is 1.5, and for winter it is 1.

[0086] By regional correction factor K p and seasonal coefficient Q p , the demand for emergency supplies is ranked according to the influence of region and season to obtain the post-disaster emergency demand.

[0087] In summary, the calculation methods for the overall demand for emergency supplies in seven major categories in a certain area after an earthquake within the 6-day golden emergency rescue period after an earthquake are listed. At the same time, according to the characteristics of emergency supply demand, many categories of materials are integrated into emergency supply packages. The material settings of each emergency supply package and the calculation method for the overall demand of the disaster-stricken point are shown in Table 1 below:

[0088] Table 1 Information on the categories of emergency material packages required and the method for calculating the overall demand at the disaster site

[0089]

[0090]

[0091] ② The degree of road damage is affected by direct damage caused by disasters or road congestion caused by secondary disasters. For example, the calculation function of the road section integrity includes the degree of self-damage of the road body after the disaster and the quantification value of the earthquake damage to the roadbed and pavement by the earthquake intensity. The calculation formula is:

[0092]

[0093] Among them, P rij Indicates the degree of road self-damage after the disaster, k i Represents the average earthquake damage index, specifically k i =0.2X i , X i represents the quantified value of earthquake damage to roadbed and pavement caused by earthquake intensity, σ is the dispersion coefficient of earthquake damage, and X i The quantitative values ​​and recommended values ​​of σ are shown in Table 2. bij It represents the impact coefficient of the collapse of buildings along the street on the traffic capacity of the road section. This data can be directly expressed from the post-earthquake building collapse rate in the International Disaster Database (EM-DAT).

[0094] At the same time, the transport capacity of emergency materials is affected by the integrity of the road section. In order to facilitate the subsequent model construction, the integrity of the corresponding road section can also be reflected by the degree of road damage, expressed as ζ ij =1-(P rij ×P bij ).

[0095] Table 2 Quantitative value X of earthquake damage under different intensities i and recommended values ​​of dispersion coefficient σ

[0096] earthquake intensity Ⅵ Ⅶ Ⅷ Ⅸ Ⅹ Ⅺ <![CDATA[Quantitative value of earthquake damage X i > 1 1 1.05 1.15 1.20 1.3 Recommended value of dispersion coefficient σ 0.40 0.30 0.25 0.20 0.15 0.10

[0097] 3. Constructing the hybrid prediction model CNN-MAFP

[0098] The present invention develops a hybrid prediction model CNN-MAFP by integrating a feature extraction module based on a convolutional neural network, a multi-head self-attention mechanism, and a feature projection module to predict the emergency needs and road integrity of potential disaster-affected points under different disaster scenarios. The hybrid prediction model CNN-MAFP includes the following three core parts:

[0099] ① Feature extraction module based on convolutional neural network

[0100] Convolutional Neural Networks (CNN) were originally designed for image datasets to capture local invariant relationships between different dimensions, but they have also been shown to perform well when processing single-dimensional and multi-dimensional data. CNN models reduce the number of parameters through parameter sharing mechanisms, which not only speeds up the training process but also improves the generalization ability of the model.

[0101] First, we use the ability of one-dimensional convolutional layers (1D CNN) to process sequence data, extract local patterns in historical disaster data through convolutional layers, and use a larger perception field to capture more information. In this case, key feature information cannot be easily omitted. The specific approach is to extract features through two stacked 1D CNNs. The process is as follows:

[0102] X′=ReLU(convld(X));

[0103] X″=ReLU(convld(X′));

[0104] in is the input of historical disaster data, and It is the output value after the channel dimension is increased for the feature. Represents a real matrix, which is a consensus on the use of model symbols.

[0105] ②Multi-head self-attention mechanism

[0106] The multi-head attention mechanism is developed on the basis of the self-attention mechanism. It is a variant of the self-attention mechanism and aims to enhance the expressiveness and generalization ability of the model. It uses multiple independent attention heads to calculate the attention weights separately and concatenates or weighted sums their results to obtain a richer representation. The disaster data collected based on the regional disaster system theory contains a variety of characteristics including the disaster itself, socio-economics, and disaster losses. However, only a small part of the feature data is important or relevant to the prediction results. Therefore, the use of the multi-head self-attention mechanism can usually improve the performance of the prediction method.

[0107] Step 1: Input the weight vector X″=[x1″,x2″,...,x n ″], through three linear transformations, we get the query matrix Q, key matrix K and value matrix V, whose calculation formula includes

[0108]

[0109] in, and is the learned weight matrix.

[0110] Step 2: The query, key, and value matrices are divided into multiple headers, and the calculation process is expressed as:

[0111] split(Q,K,V)=[head1,…,head n ];

[0112] Among them, head i represents the i-th head of the query matrix Q, key matrix K, and value matrix V.

[0113] Step 3: Calculate the attention of each head. For the i-th head:

[0114] MultiHead(Q,K,V)=Concat(head1,…,head n )W 0 ;

[0115] in, represents the stable gradient, d k Represents the dimensionality of the query and key.

[0116] Step 4: Map the concatenated polyhedron output to the desired output dimension through linear transformation.

[0117] (X″′,W score )=MultiHead(Q,K,V);

[0118] in, represents the characteristic output value, W score represents the attention score matrix, W 0 Represents the output projection matrix, Concat(head1,…,head n ) indicates the concatenation of output signals from multiple heads.

[0119] ③Feature projection module

[0120] The feature projection module reduces the high-dimensional features extracted by convolution and attention to the target output space through a multi-layer fully connected network.

[0121] The present invention first projects the input features into a higher-dimensional space through the feature projection layer ProjectionLayer, then separates the odd and even features as follows, performs dimensionality reduction operations on each feature and then adds them together, and finally obtains a single-dimensional output through a linear transformation. The calculation formula is:

[0122] Z odd =ReLU(W1·x odd +b1);

[0123] Z even =ReLU(W2·x even +b2);

[0124] Z=Z odd +Z even ;

[0125] The single-dimensional output of the feature projection layer ProjectionLayer is mapped to the target output dimension through linear transformation for the final loss prediction. The calculation formula is:

[0126]

[0127] The predicted value of the model output Compared with the actual disaster loss y, the model parameters are optimized by minimizing the mean square error (MSE) loss function, and the calculation formula is:

[0128]

[0129] in, represents the predicted value of the model output, y represents the actual disaster loss, and N is the total number.

[0130] 4. Two-stage emergency network design stochastic optimization model driven by multi-source data

[0131] The present invention determines a cross-regional network for transporting emergency supplies based on post-disaster emergency needs and a function for measuring the completeness of road sections. The cross-regional network includes disaster points, alternative facility points, emergency warehouse points, and connecting sections between nodes in different regions, and responds to emergency events by sharing emergency warehouses, emergency supplies, and allocating resources across regions.

[0132] Identify a cross-regional network for transporting emergency supplies, including:

[0133] Minimize the cost of site selection and resource pre-positioning before disaster. The formula is:

[0134]

[0135] Minimize the transportation cost of emergency supplies after a disaster. The formula is:

[0136]

[0137] By combining the minimization of pre-disaster site selection and resource pre-positioning costs, the minimization of post-disaster emergency material transportation costs and the reference distribution, a stochastic optimization model for emergency network design is obtained, which is expressed as:

[0138]

[0139] in, represents the fixed cost of construction of emergency facility i, represents the unit cost of materials pre-placed at emergency facility i, y i represents a 0-1 decision variable, h i represents a continuous decision variable, y represents the actual disaster loss, and N represents the size of the data-driven data set.

[0140] In the pre-disaster preparation stage, decision makers need to integrate the results of regional emergency loss forecasts and demand estimates, select certain locations from candidate emergency facility points to build emergency warehouses, and reserve a certain amount of emergency supplies to meet the emergency needs of potential disaster-stricken points after the disaster. In the response stage after the disaster, decision makers select transportation routes to allocate emergency supplies from emergency warehouses to disaster-stricken points based on the needs of the disaster-stricken points and the completeness of the transportation network sections. At the same time, in order to avoid the paralysis of emergency warehouses due to road interruptions or facility damage in the disaster-stricken areas after major disasters, the emergency needs of the disaster-stricken areas are considered to be provided by cross-regional warehouses, thereby enhancing the flexibility and robustness of the emergency network and emergency response.

[0141] However, the emergency network design problem faces challenges such as limited emergency warehouse capacity, uncertain demand, and uncertain post-disaster road network damage information. These challenges may lead to deprivation costs when the needs of the disaster-stricken points cannot be met in time. At the same time, considering that emergency warehouses may be set up at potential disaster points in the pre-disaster stage, once a disaster occurs at this point in the future, it may cause the emergency warehouse to fail. Therefore, when constructing the model, it should be considered that the needs of the disaster-stricken points must be met by emergency warehouses at other points or even cross-regional warehouse transportation.

[0142] The specific implementation process includes:

[0143] Based on the above considerations and requirements, a stochastic optimization model (2ERN-DSO) for emergency network design is constructed. The overall framework of the model is as follows Figure 3 shown.

[0144] The symbols of the model in Table 3 are as follows:

[0145]

[0146]

[0147] ①2ERN-DSO model is expressed as:

[0148]

[0149] The objective function is divided into two parts. The first part is to minimize the cost of site selection and resource provisioning, part 2 is based on minimizing the transportation cost of emergency supplies containing N sample data. The uncertain set in this model Two uncertainties were considered: Represents the uncertain disaster site demand, The uncertainty of the information on the integrity of the road section after the disaster. The true probability distribution of the uncertain parameters assumed by the 2ERN-DSO model is unknown, but the prediction data obtained from the historical samples can be used to construct the fuzzy set of uncertain parameters. The specific method is to consider a finite set with N independent predicted actual values. Each of the discrete random vectors The probability is equal to This allows us to construct an empirical distribution As an estimate of the true probability distribution F ζ The reference distribution of is:

[0150]

[0151] in Indicates that the unit mass is concentrated on is the Dirac distribution at , and ∑ represents the Borelσ-algebra.

[0152] At the same time, the traditional random optimization method only uses one sample or a batch of samples for each update, and the update direction may be affected by noise and fluctuate greatly, and the instability is strong. The data-driven method of the present invention is not limited to limited sample data. Considering that the pre-disaster emergency management decision is greatly affected by the worst historical situation, but only considering the worst situation will lead to overly conservative decision-making plans, resulting in a sharp increase in costs. Therefore, the post-disaster demand value and road damage degree value predicted based on deep learning are used to construct the uncertainty factors in the fuzzy set of the 2ERN-DSO model. The bounded polyhedron support set of is given by:

[0153]

[0154] in d m and m express i∈N p and The upper and lower bound vectors of the support set containing the demand and post-disaster road section damage degree determined by (i,j)∈E, and the value of this vector is provided by the CNN-MAFP prediction program.

[0155] The second stage problem in the 2ERN-DSO model The model is represented as follows:

[0156]

[0157]

[0158] In the first stage of the objective function (1), the model aims to minimize the fixed cost of opening emergency facilities and the purchase cost of pre-positioned relief supplies. ijn represents the second-stage pursuit variable associated with sample n∈N, i.e., the decision variable for the quantity of emergency supplies transported, Represents the predicted realization value of the nth historical observation. For a given first-stage facility location and material pre-positioning decision vector y, h set, minimize the expected second-stage transportation cost under all random scenarios n∈N, including transportation costs, and at the same time, in order to avoid a shortage of pre-positioned resources, add a penalty variable θ 2n In order to avoid insufficient transport capacity of the road section, the penalty variable θ is increased. 1n .

[0159] Constraint (2) The overall transportation relationship can be expressed as:

[0160] Constraint (3) represents the transport capacity constraint of the road segment.

[0161] Constraint (4) indicates that material pre-positioning can only be carried out in the selected emergency facilities.

[0162] Constraint (5) defines the domain of the variable.

[0163] For example, an empirical analysis is conducted on the design of emergency networks and optimization of material transportation in a certain region using the earthquake disaster background. This includes the epicenter location, occurrence time, magnitude, maximum intensity, different intensity ranges, number of casualties, total number of people affected, etc. At the same time, the spatiotemporal data of multiple socioeconomic elements of the panel data in the statistical yearbook of the prefecture-level city where the disaster-stricken area is located are integrated. In order to eliminate the impact of historical nominal prices and real prices involving GDP, per capita wages, etc. in different years in the sample data on the model prediction results, the historical CPI data based on 2023 is corrected to constant prices. At the same time, the regional coefficient, seasonal coefficient, earthquake intensity, and building collapse rate were taken into consideration to estimate the demand for emergency materials caused by various disasters and the degree of integrity of the road sections after the disaster. Finally, a historical earthquake disaster data set was obtained. The data set has a total of 33 eigenvalues, as shown in Table 4. For the inflow / outflow gate control branch, the first 27 eigenvalues ​​including disaster factors, environment, and socio-economics were used as the inflow sequence, and the last 5 eigenvalues ​​(deaths, injuries, homeless people, affected people, and building collapse rate) in the earthquake casualty loss data set were used as the outflow sequence.

[0164] Table 4 Historical earthquake disaster data for a region

[0165]

[0166]

[0167] The prediction capabilities of the proposed hybrid prediction model CNN-MAFP are compared and analyzed with the existing CNN and other prediction methods. The performance metrics include mean absolute percentage error (MAPE), mean absolute error (MAE), root mean square error (RMSE), and determination coefficient (R). 2 Verify the model performance. The four performance indicators are defined as follows:

[0168]

[0169] where y n is the actual value of the nth data. is the predicted value of the nth data, is the average of the true values, and k is the total number of samples. RMSE, MAE, and MAPE values ​​close to 0 indicate that the model has better prediction accuracy. 2 The closer the value is to 1, the better the fit is.

[0170] The proposed hybrid prediction model CNN-MAFP is compared with several other models using historical earthquake disaster datasets, including several popular machine learning models, such as Linear Regression (LR), Support Vector Regression (SVR), Random Forest (RF), Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short Term Memory (LSTM) and Convolutional Long Short Term Memory (CNN-LSTM). To prove the effectiveness of the proposed method in predicting earthquake casualties and losses, all parameters of the comparison model are the same as the hybrid prediction model CNN-MAFP, and the input and output of all models are the same. The prediction performance obtained by the experiment is shown in Table 4. According to the experimental results, for the dataset used, compared with the current single model LR and ensemble learning models such as CNN, LSTM, and CNN-LSTM, the proposed hybrid prediction model CNN-MAFP shows better prediction results.

[0171] Table 5 Prediction effect of hybrid prediction model CNN-MAFP compared with other prediction models

[0172] Model MAPE MAE RMSE R2 LR 38.49883 0.843325 1.143505 0.5908 SVR 32.53639 0.646973 0.897589 0.747876 RF 29.11144 0.655437 0.888789 0.752795 MLP 29.79366 0.686948 0.879671 0.757841 CNN 28.3722 0.648002 0.829664 0.784591 LSTM 29.91617 0.665828 0.848097 0.774913 CNN-LSTM 26.00633 0.505325 0.696929 0.848002 CNN-MAFP 24.4063 0.47212 0.64357 0.87039

[0173] Through comparison, it is found that the hybrid prediction model CNN-MAFP proposed in the present invention shows better prediction performance among the four indicators proposed compared with the regression model and the deep learning model, especially the R2 can reach above 0.87. Therefore, it is believed that the model is effective in predicting earthquake emergency needs and road section integrity.

[0174] The above-mentioned embodiments only express several implementation methods of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for designing an emergency material transportation network, characterized in that: include: Obtain historical datasets of sudden disasters; A feature extraction module is constructed by two stacked one-dimensional convolutional layers 1DCNN. The end of the feature extraction module is connected to a multi-head attention mechanism consisting of multiple independent attention heads, and a feature projection module is added to the end of the multi-head attention mechanism to construct a hybrid prediction model CNN-MAFP. The hybrid prediction model CNN-MAFP is trained using the historical dataset of sudden disasters; The real-time disaster data of the area to be tested is input into the trained hybrid prediction model CNN-MAFP. The feature extraction module is used to extract features from the real-time disaster data to obtain local features that characterize the temporal pattern and spatial distribution of the disaster. The attention weights of the local features are obtained through the multi-head attention mechanism, and the attention weights are concatenated and weighted summed to model the global dependency between local features, thereby obtaining high-dimensional features that reflect the development of the disaster and the scope of impact. The feature projection module is used to reduce the high-dimensional features to the target output space, and the post-disaster emergency demand and road section integrity measurement function after the disaster in the area to be tested are obtained. According to the post-disaster emergency demand and road section integrity measurement function, a cross-regional network for transporting emergency supplies is determined; wherein the cross-regional network includes disaster points, alternative facility points, emergency warehouse points and connecting sections between nodes in different regions.

2. The method for designing an emergency material transportation network according to claim 1, characterized in that: The historical disaster data set includes disaster characteristic data, socio-economic characteristic data and disaster loss characteristic data, and the matrix is ​​expressed as M=(Dis i,m ,Soc i,n ,Los i,k ),in: Among them, disaster data represents the mth disaster characteristic value of disaster site i, social and economic data Represents the nth socioeconomic characteristic value of disaster site i, disaster loss data Represents the kth disaster loss characteristic value of disaster-affected point i.

3. The method for designing an emergency material transportation network according to claim 1, characterized in that: The regional correction factor K is introduced into the post-disaster emergency demand p and seasonal coefficient Q p , which specifically include: Regional correction factor K p The calculation formula is expressed as: Among them, x pi represents the weight of the i-th reference indicator, and n represents the total number of indicators; By regional correction factor K p and seasonal coefficient Q p , the demand for emergency supplies is ranked according to the influence of region and season to obtain the post-disaster emergency demand.

4. The method for designing an emergency material transportation network according to claim 1, characterized in that: The road section integrity calculation function includes the degree of self-damage of the road body after the disaster and the quantified value of the earthquake damage to the roadbed and pavement of the earthquake intensity. The specific calculation formula is: Among them, P rij Indicates the degree of road self-damage after the disaster, k i represents the average earthquake damage index, X i represents the quantified value of earthquake damage to roadbed and pavement caused by earthquake intensity, σ represents the dispersion coefficient of earthquake damage, and P bij It represents the impact coefficient of the collapse of buildings along the street on the traffic capacity of the road section.

5. The method for designing an emergency material transportation network according to claim 1, characterized in that: The feature extraction module is used to extract features from real-time disaster data, specifically including: X′=ReLU(convld(X)); X″=ReLU(convld(X′)); in, is the input of historical disaster data, and It is the output value after the channel dimension is increased for the feature. represents a real matrix.

6. The method for designing an emergency material transportation network according to claim 1, characterized in that: The attention weights of local features are obtained through the multi-head attention mechanism, specifically including: The weight vector X″=[x1″,x2″,...,x n ″], the query matrix Q, key matrix K and value matrix V are obtained through three linear transformations, and the calculation formula includes: in, and Represents the learned weight matrix; The query, key, and value matrices are divided into multiple headers and the computation process is expressed as: split(Q,K,V)=[head1,…,head n ]; Among them, head i represents the i-th head of the query matrix Q, key matrix K and value matrix V; Determine the attention of each head, for the i-th head: in, represents the stable gradient, d k Represents the dimensionality of the query and key; The concatenated polyhedron output is mapped to the desired output dimension by a linear transformation, which is given by: MultiHead(Q,K,V)=Concat(head1,…,head n )W 0 ; (X″′,W score )=MultiHead(Q,K,V); in, represents the characteristic output value, W score represents the attention score matrix, W 0 Represents the output projection matrix, Concat(head1,…,head n ) indicates the splicing of output signals from multiple heads. represents a real matrix.

7. The method for designing an emergency material transportation network according to claim 1, characterized in that: The step of reducing the high-dimensional features to the target output space through the feature projection module specifically includes: The input features are projected into a higher-dimensional space through the feature projection layer ProjectionLayer, and then the odd and even features are separated and added after dimensionality reduction operations. Finally, a single-dimensional output is obtained through linear transformation. The calculation formula is: WITH odd =ReLU(W1 x odd +b1); <h2 style=";text-align:left;direction:ltr">Z<h2 style=";text-align:left;direction:ltr"> even <h2 style=";text-align:left;direction:ltr"> =ReLU(W2·x<h2 style=";text-align:left;direction:ltr"> even <h2 style=";text-align:left;direction:ltr"> +b2); Z=Z odd +Z even ; Among them, x even Indicates odd features, x odd Represents even characteristics, Z even represents the odd features after dimensionality reduction, Z odd represents the even feature after dimensionality reduction operation, W represents the linear transformation coefficient, b represents the linear transformation constant, and ReLU() represents the linear transformation; The single-dimensional output of the feature projection layer ProjectionLayer is mapped to the target output dimension through linear transformation for the final loss prediction. The calculation formula is: The predicted value of the model output Compared with the actual disaster loss y, the model parameters are optimized by minimizing the mean square error MSE loss function, and the calculation formula is: in, represents the predicted value of the model output, y represents the actual disaster loss, and N is the total number.

8. The method for designing an emergency material transportation network according to claim 1, characterized in that: The cross-regional network determined for transporting emergency supplies specifically includes: Minimize the cost of site selection and resource pre-positioning before disaster. The formula is: Minimize the transportation cost of emergency supplies after a disaster. The formula is: By combining the minimization of pre-disaster site selection and resource pre-positioning costs, the minimization of post-disaster emergency material transportation costs and the reference distribution, a stochastic optimization model for emergency network design is obtained, which is expressed as: in, represents the fixed cost of construction of emergency facility i, represents the unit cost of materials pre-placed at emergency facility i, y i represents a 0-1 decision variable, h i represents a continuous decision variable, y represents the actual disaster loss, and N represents the size of the data-driven data set.

Images