Frontiers in Business, Economics and Management 
ISSN: 2766-824X | Vol. 15, No. 3, 2024 

 

186 

Post-earthquake Emergency Supplies Prediction Based 
on BP Neural Network 

Zhihang Liu, Bo Dong 

School of Emergency Management and Safety Engineering, North China University of Science and Technology, Tangshan 063200, China 

 

Abstract: The prediction of earthquake casualty population is a typical complex prediction system, which needs to 
comprehensively consider a variety of factors such as the earthquake damage itself, the population distribution in the affected 

area and its environment. Aiming at the prediction of emergency supplies in earthquake disasters, this paper collects historical 

earthquake data, including key factors such as magnitude, depth of epicenter, population density of the affected area, constr ucts 

the input layer of BP neural network, and utilizes the self-learning ability of the network for training to realize the prediction of 

the number of people injured in an earthquake, which in turn indirectly predicts the demand for emergency supplies. The 

experimental results show that the prediction model of the number of earthquake injuries based on BP neural network has 

excellent performance in prediction accuracy, and the operational coefficient of determination of the model R^2 reaches 0.93337, 

which indicates that the prediction results of the model have ahigh degree of correlation and accuracy with the actual values, and 

it is able to provide a more accurate and reliable reference for the rescue work of the emergency supplies after the earthquake. 

Keywords: Bp neural network, Casualty prediction, Emergency supplies, Earthquake disaster. 

 

1. Introduction 

China, as a country located at the intersection of the Pacific 

Rim seismic belt and the Eurasian seismic belt, is prone to 

frequent earthquakes as a result of the frequent extrusion of 

the crustal plates. Earthquakes are one of the most serious 

natural disasters that China faces, and given the country's vast 

size, large population and dense urban distribution, the 

destructive power of a large-scale earthquake would be 

incalculable, with the potential to cause a large number of 

deaths and injuries as well as irreversible damage to property. 

Since the beginning of 2000 to 2023, just more than twenty 

years, China and its neighboring regions have experienced 

more than a hundred earthquakes of magnitude 6.0 or above, 

and each disaster is accompanied by heavy losses [1]. In 

particular, the occurrence of several major earthquakes in 

recent years has not only brought a deep shock to the people 

of the country, but also exposed many challenges in the 

process of disaster relief. After an earthquake, cities are 

instantly turned into ruins, roads are damaged and 

communication is interrupted, which brings great obstacles to 

rescue teams. In such emergencies, there is a need to quickly 

find trapped people and at the same time ensure that the basic 

needs of the victims are met. However, due to various 

unfavorable factors, such as the difficulty of obtaining 

information and the poor deployment of supplies, it is difficult 

for the relevant departments to accurately grasp the situation 

of the population casualties and the needs of the victims, 

which makes the supply of supplies a shortcoming in the 

rescue work, exacerbates the difficulty of the rescue work, 

and seriously threatens the survival and safety of the victims. 

After the earthquake, the affected areas faced severe 

survival challenges, and were in urgent need of all kinds of 

emergency supplies to guarantee the basic living needs of the 

affected people. However, due to the suddenness and 

complexity of earthquake disasters, the mobilization and 

distribution of materials in the early stages often face many 

difficulties. On the one hand, the lack of resources such as 

manpower, financial resources and means of transportation 

makes the transportation and distribution of materials 

inefficient; on the other hand, the differences in the degree of 

damage in different disaster areas also lead to the diversity 

and variability of the demand for materials. If the materials 

are not properly distributed, it may not only lead to the 

backlog and waste of materials in some areas, but also cause 

serious shortage of materials in other areas, which may 

threaten people's lives and safety. Therefore, it is particularly 

important to make scientific and reasonable prediction and 

distribution of post-earthquake emergency supplies. This will 

not only help to improve the efficiency and fairness of the use 

of materials, but also provide strong support and guarantee for 

the subsequent emergency relief work. 

2. Post-earthquake Emergency Supplies 
Forecasting Study 

2.1. Identification of Data Indicators 

The casualties caused by earthquakes are affected by a 

variety of parameters including, but not limited to, the 

magnitude of the earthquake, the exact time of the earthquake, 

the intensity of the fortification, the number of people affected 

by the earthquake, the severity of the destruction of the 

housing stock, and the level of economy in the area. 

The aim of this paper is to construct a prediction method 

that can quickly provide a basis for decision-making in the 

mobilization and dispatch of emergency supplies in the early 

stages of an earthquake. Therefore, the selected data needs to 

meet the requirements of rapid access and easy analysis. After 

comprehensive consideration of various aspects, this paper 

selects the following seven parameters [2] as the key 

indicators for assessing the number of earthquake casualties: 

the magnitude, the time of the earthquake, the intensity of 

defense, the intensity of damage, the number of people 

affected, the destruction of housing, and the local economic 

development level, and the specific content of each predictive 

indicator is described below. 

 



 

187 

(1) Magnitude 

The magnitude of an earthquake is a classification of the 

magnitude of the energy released by the earthquake source. 

The greater the magnitude, the greater the energy released by 

the source, and the greater the impact on the surface structure. 

The greater the magnitude, the greater the destructive power 

to the surface structure and buildings on the ground, and 

therefore, the greater the possibility of casualties. Therefore, 

the possibility of casualties is also greater. 

(2) Time of occurrence 

According to the analysis results of relevant literature, the 

number of casualties in earthquakes occurring between 0:00 

a.m. and 6:00 a.m. is relatively high, so this paper marks this 

time period as 2, and other times as 1. 

(3) Intensity of defense 

Intensity of defense reflects the ability of buildings to resist 

the damage caused by seismic waves, which directly 

determines the stability and safety of buildings in earthquakes. 

In an area with a high intensity of defense, buildings may be 

less damaged in an earthquake, thus reducing the risk of 

casualties. 

(4) Damage Intensity 

Damage intensity is the actual observed degree of damage 

caused by the earthquake to the ground and buildings, and 

also directly reflects the actual impact of the earthquake, 

which is closely related to the casualties. Generally speaking, 

the higher the damage intensity, the more serious the 

casualties. 

(5) Affected Population 

The affected population refers to the group of people who 

are affected and may suffer casualties in an earthquake, and is 

one of the important indicators for assessing earthquake 

casualties. Densely populated areas tend to cause greater 

casualties in earthquakes. In addition, data on the affected 

population can be collected and organized by statistical 

departments or related organizations, and the data is relatively 

accurate and reliable. 

(6) Destruction of housing 

Housing is a basic place for people to live, and the degree 

of damage to housing directly affects people's life safety and 

quality of life, and heavy housing destruction often leads to a 

large number of casualties. Housing destruction refers to the 

extent of housing damage in an earthquake, including 

collapse, cracks, etc., which can be obtained through on-site 

investigation, remote sensing technology and feedback from 

residents. 

(7) Level of economic development 

The level of local economic development has a certain 

correlation with earthquake casualties. Regions with higher 

levels of economic development usually have better 

infrastructures and emergency response capabilities, and are 

able to respond more effectively to earthquake disasters and 

reduce casualties. 

2.2. Data Processing 

2.2.1. Data Standardization 
Data standardization can scale the data to a uniform range 

of scales, reducing the differences between variables and 

facilitating the convergence of machine learning algorithms. 

Especially in BP neural network, its training process based on 

gradient descent method is very sensitive to the scale of input 

data. The standardization process can accelerate the 

adjustment process of weights and improve the training 

efficiency of the model, and the following are the specific 

steps of data standardization. 

(1) Calculate the mean value 

For all the collected indicator data, first calculate the mean 

value of its data values, and the calculation formula is: 

𝑀𝑒𝑎𝑛 =
1

𝑛
∑ 𝑥𝑖

𝑛
𝑖=1                (1) 

(2) Calculating the standard deviation 

Average the square of the difference from the mean for each 

data point and then take the square root of that average, and 

the calculation formula is: 

𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝐷𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛 = √
1

𝑛
∑ (𝜆𝑖 − 𝑀𝑒𝑎𝑛)

2𝑛
𝑖=1     (2) 

(3) Standardizing data 

For each case data point collected, subtract the mean and 

divide by the standard deviation, and the calculation formula 

is: 

𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝑖𝑧𝑒𝑑𝐷𝑎𝑡𝑎 =
𝑥𝑖−𝑀𝑒𝑎𝑛

𝑆𝑡𝑎𝑛𝑑𝑎𝑟𝑑𝐷𝑒𝑣𝑖𝑎𝑡𝑖𝑜𝑛
       (3) 

2.2.2. Calculate the Pearson correlation Coefficient 
Matrix 

By calculating the Pearson correlation coefficients between 

each feature and the number of injuries, this subsection can 

clearly demonstrate the linear relationship between features 

such as magnitude, time of occurrence, intensity of defense, 

and population affected by the earthquake and the number of 

injuries, in order to better understand the impact of 

earthquakes on human casualties. The Pearson correlation 

coefficient helps to select the features that are most helpful in 

predicting the number of casualties and improves the 

accuracy of the model. The calculation formula is shown 

below: 

𝑟𝑥𝑦 =
∑ (𝑥𝑖−�̅�)(𝑦𝑖−�̅�)

𝑛
𝑖=1

√∑ (𝑥𝑖−�̅�)
2𝑛

𝑗=1
√∑ (𝑦𝑖−�̅�)

2𝑛
𝑗=1

           (4) 

Where 𝑟𝑥𝑦   denotes the Pearson correlation coefficient 

between feature 𝑥  and feature 𝑦 , 𝑛  denotes the number of 
observations, 𝑥𝑖   and 𝑦𝑖   denote the 𝑖 th observation of 
feature 𝑥  and feature 𝑦 , respectively, and �̅�  and �̅�  denote 
the mean values of feature 𝑥 and feature 𝑦, respectively. 

2.3. Constructing a BP Neural Network 
Prediction Model 

2.3.1. Model Structure Identification 
From the theoretical point of view, when the transfer 

function of the hidden layer of the three-layer BP neural 

network utilizes the S-type function, and the transfer function 

of the output layer adopts the linear function, it is able to 

approximate the arbitrary continuous function with arbitrary 

accuracy. Based on this, the BP neural network used in this 

paper contains three layers, specifically the input layer, a 

hidden layer and the output layer. Moreover, the transfer 

function between the input layer and the hidden layer is a 

linear activation function, while the transfer function between 

the hidden layer and the output layer is a sigmoid activation 

function. 

 



 

188 

(1) Determine the number of neurons in the input layer 

After various studies, the seven indicators of earthquake 

magnitude, occurrence time, defense intensity, damage 

intensity, affected population, housing destruction and local 

economic development level are selected as the input 

parameters of the model, so the input layer neurons are 

determined to be seven. 

(2) Determine the output layer neurons 

In this paper, the number of people injured in the 

earthquake is chosen as the output of the network, so the 

output layer neuron is 1. 

(3) Determine the hidden layer neurons 

The determination of the number of hidden layer neurons 

is not standardized and is often based on the volume of 

samples collected and human experience. Setting the number 

of nodes too little may result in the model prediction accuracy 

not being able to meet the expected goal, while setting too 

much may improve the accuracy to a certain extent, but the 

overly complex model is prone to fall into the dilemma of 

overfitting. Therefore, in this paper, the implicit layer node 

settings are determined by the principle of trial-and-error 

algorithm, and the number of nodes ranges from 8 to 20, and 

the number of neurons in the implicit layer is determined to 

be 17 through the experiments of multiple model training and 

observation of the error. 

(4) Other parameter settings of the model 

Smaller batch sizes usually speed up training and help the 

model generalize better, but may also lead to unstable training, 

therefore, the number of data samples used for training in 

each batch was set to be 3. In addition, a higher number of 

loops may help the model learn more information, but may 

also lead to overfitting. Therefore, it was specified that the 

model was changed to run 500 cycles. 

2.3.2. Steps to Build a BP Neural Model 
BP (Back Propagation) neural network, proposed by 

scientists led by Rumelhart and McClelland in 1986, aims to 

solve the problem of multilayer neural network implicit layer 

connection weights learning, is a multilayer feed-forward 

neural network, whose core feature lies in the training through 

the back propagation of error algorithm. The sample data is 

input from the input layer, and is continuously calculated 

forward according to the parameters set in each layer until it 

reaches the output layer, and the back propagation is carried 

out according to the error between the output value and the 

expected value, and the weights and thresholds of the 

connections between the layers are adjusted and updated until 

the final output result meets the expected value. The structure 

of the BP neural network with three layers studied in this 

paper is shown in figure 1. 

 
Figure 1. BP neural network structure diagram 

The construction steps of the BP neural network model are 

as follows: 

(1) Forward propagation 

The output corresponding to the 𝑖th input layer node is: 

𝑎𝑖
0 = 𝑥𝑖                     (5) 

The inputs and outputs of the 𝑗th node of the hidden layer 
are: 

𝐼𝑗 = ∑ 𝑤𝑖𝑗 𝑎𝑖
0𝑛

𝑖=1 + 𝜃𝑗              (6) 

𝑎𝑗
1 = 𝑓(𝐼𝑗 )                 (7) 

(2) Objective function of BP network 

Assuming that there are 𝑁 samples, the error of the input 
and output corresponding to the 𝑖th sample is: 

𝐸𝑖 =
1

2
∑ (𝑑𝑖𝑘 − 𝑦𝑖𝑘 )

2𝑚
𝑘=1             (8) 

𝐸 = ∑ 𝐸𝑖
𝑁
𝑖=1 =

1

2
∑ ∑ (𝑑𝑖𝑘 − 𝑦𝑖𝑘 )

2𝑚
𝑘=1

𝑁
𝑖=1       (9) 

In the above two formulas, m is the number of output nodes, 

𝑑𝑖𝑘  is the 𝑘th expected value of the 𝑖th sample, and 𝑦𝑖𝑘  is 
the 𝑘 th neural network predicted output value of the ith 
sample. 

(3) Gradient descent method is used to train the network, 

and the weight coefficient update formula is: 

𝛥𝑊 = −
𝜇

𝑁

𝜕𝐸

𝜕𝑤
                 (10) 

(4) Adjusting the weights and thresholds for each layer of 

the network: 

𝑊 (𝑙) = 𝑊 (𝑙) −
𝜇

𝑁
∑

𝜕𝐸(𝑖)

𝜕𝑊(𝑙)
𝑁
𝑖=1            (11) 

𝑏(𝑙) = 𝑏(𝑙) −
𝜇

𝑁
∑

𝜕𝐸(𝑖)

𝑏(𝑙)
𝑁
𝑖=1             (12) 

Where is the rate of change of 
𝜇

𝑁
, i.e., the step size; 

𝜕𝐸(𝑖)

𝜕𝑊(𝑙)
 is 

the derivative of the error equation. 

2.3.3. Performance Evaluation Indicators 
Performance evaluation metrics play a vital role in 

evaluating BP neural networks, which can help researchers 

understand the performance of the model for model 

optimization, comparing different models, and decision 

making, as well as reflecting potential problems in the model, 

such as overfitting or underfitting. Evaluation metrics usually 

use mean square error, mean absolute error, and coefficient of 

determination, as shown in the following equations: 



 

189 

(1) Mean Squared Error (MSE) 

𝑀𝑆𝐸 =
1

𝑛
∑ (𝑦𝑖 − �̂�𝑖 )

2𝑛
𝑖=1              (13) 

(2) Mean Absolute Error (MAE) 

𝑀𝐴𝐸 =
1

𝑛
∑ |𝑦𝑖 − �̂�𝑖 |

𝑛
𝑖=1               (14) 

(3) Coefficient of Determination (R^2) 

𝑅2 = 1 −
∑ (𝑦𝑖−�̂�𝑖)

2𝑛
𝑖=1

∑ (𝑦𝑖−�̅�)
2𝑛

𝑖=1

               (15) 

In the three evaluation indicator formulas above, 𝑛 is the 
sample size, 𝑦𝑖  is the true value, �̂�𝑖  is the predicted value, 
and �̅� is the average of the true values.  

2.4. Estimated Material Requirements 

(1) Daily consumables 

According to CDC data, the minimum daily requirement 

per person in an emergency is 4183kJ (1000kcal) of energy 

and 1000ml (1kg) of drinking water. In this paper, bread, 

which is a representative of daily consumable relief foods, is 

stipulated to be about 1000 kcal for 500g of bread among the 

common bread types. Therefore, the total weight of food 

consumed by the general population in an emergency is 

1.5kg/d. 

(2) Common medicines 

This paper focuses on the impact of the number of injured 

people on the demand for rescue supplies in the drug category, 

abstracting it into routine medical supplies and first aid 

supplies, such as bandages, anti-inflammatory drugs, iodine 

vapors, tetanus, and so on. They are packaged into one-person 

portions and distributed to the injured in the disaster area, 

stipulating that the demand of each injured person is fixed at 

120g/d. 

3. Case Studies 

The purpose of this section is to demonstrate, through an 

example analysis, how the constructed prediction method can 

quickly predict the emergency supplies required by the 

victims at each disaster site when responding to a large-scale 

earthquake, so as to provide a reference for the rapid and 

effective mobilization and supply of emergency supplies in 

the first time after the earthquake. To this end, the 7.0 

magnitude earthquake that occurred in Ya'an, Sichuan 

Province on April 20, 2013 was chosen as a case study in this 

chapter [3]. This earthquake is a major natural disaster that 

occurred in China in recent years, and we will use the 

constructed principal component BP neural network 

earthquake casualty prediction model and the method of 

estimating the demand for earthquake emergency supplies to 

analyze the demand for emergency supplies in the severely 

affected areas in the earthquake. 

3.1. Data Collection 

3.1.1. Yaan Earthquake Data 
According to the China Earthquake Network, at 8:02:46 

Beijing time on April 20, 2013, an earthquake with a 

magnitude of 7.0 occurred in Lushan County, Ya'an City, 

Sichuan Province, at a depth of 13 kilometers, with a 

maximum intensity of 9 degrees, and affected an area of about 

18,682 square kilometers. The latest situation of the 

earthquake damage coming from the Ya'an City Earthquake 

Relief Command shows that as of 12:00 on the 23rd, the 

number of deaths in the city has reached 173, 22 are missing, 

and the number of injuries has reached 10,974, of which 968 

are serious injuries [4]. Specific data on the damage in Ya'an 

is shown in Table 1.

Table 1. Statistics on the damage in Ya'an 

Region 
Magnit

ude 

Occurren

ce Time 

Fortification 

Intensity 

Damage 

Intensity 

Affected 

Population 

Destroyed 

Housing 

GDP/ 

billion 

Number 

of Injured 

Number 

of Deaths 

Yucheng 7 1 7 7 356800 125579 12.51 1109 15 

Lushan 7 1 6 9 112000 127710 2.65 5537 120 

Baoxing 7 1 6 6 58954 57731 1.91 2503 26 

Tianquan 7 1 7 7 155200 18705 3.78 811 5 

Mingshan 7 1 6 7 279900 14000 5.74 607 2 

Yingjing 7 1 7 7 149000 7865 4.36 341 2 

Shimian 7 1 6 6 124400 899 5.88 39 2 

Hanyuan 7 1 6 6 325000 807 6.07 35 1 

3.1.2. BP Neural Network Training Set Data Collection 
and Processing 

In this paper, 39 earthquakes of magnitude 6 or above that 

occurred in China in the past 30 years are selected as the 

training sample set based on news reports, thesis data, and the 

"National Center for Sharing Earthquake Science Data", etc. 

Table 2 shows some of the statistical information of the 

earthquakes.

 

 

 

 

 

 

 

 

 

 



 

190 

Table 2. Training set data 

Region Magnitude 
Occurrence 

Time 

Fortification 

Intensity 

Damage 

Intensity 

Affected 

Population 

Destroyed 

Housing 

GDP/ 

billion 

Number 

of Injured 

Number 

of Deaths 

Lixian 6.1 1 7 8 103054 36500 0.63 359 30 

Panzhihua 6.1 2 7 8 130311 7455 5.16 145 5 

Ninglang 6.2 1 7 7 127870 17500 10.74 208 5 

Dangxiong 6.6 1 8 8 899 171 3.84 19 9 

Jiangyou 8.0 1 7 7 870000 200000 2.87 19278 437 

Zitong 8.0 1 6 7 312807 16517 1.33 2489 25 

Youxian 8.0 1 6 7 450000 112987 14.25 482 85 

Xiaojin 8.0 1 6 8 79982 12888 12.73 452 27 

Fucheng 8.0 1 7 8 545465 38298 1.48 2038 86 

Luojiang 8.0 1 6 8 203000 24596 5.07 307 19 

… … … … … … … … … … 

Lasa 6.1 2 7 9 130300 7466 12.18 132 5 

Lijiang 7.0 1 7 7 1075000 959000 5.18 16912 309 

Ludian 6.1 1 7 8 130311 7455 0.34 145 5 

After the collected data were processed according to the 

standardization method in section 2.2.1, the data in the data 

table were converted to standard normal distribution as a way 

to eliminate the difference in magnitude between different 

features [5], which is beneficial for the training of the network. 

Some of the data after the standardization process is shown in 

Table 3 below.

 

Table 3. Standardized processing of training set data 

Region Magnitude 
Occurrence 

Time 

Fortification 

Intensity 

Damage 

Intensity 

Affected 

Population 

Destroyed 

Housing 

GDP/ 

billion 

Number 

of Injured 

Number 

of Deaths 

1 -2.286 -0.549 0.483 -0.194 -0.546 -0.427 -0.024 -0.560 -0.425 

2 -2.13 -0.549 0.483 -0.194 -0.553 -0.372 0.989 -0.557 -0.425 

3 -1.506 -0.549 2.526 -0.194 -0.880 -0.467 -0.263 -0.567 -0.424 

4 0.677 -0.549 0.483 -0.194 1.361 0.632 -0.439 0.442 -0.338 

5 0.677 -0.549 -1.558 -0.194 -0.075 -0.377 -0.721 -0.438 -0.421 

6 0.677 -0.549 -1.558 -0.194 0.278 0.153 1.628 -0.543 -0.409 

7 0.677 -0.549 0.483 0.483 -0.676 -0.397 1.351 -0.544 -0.420 

8 0.677 -0.549 -1.558 -1.558 0.524 -0.257 -0.693 -0.461 -0.409 

9 0.677 -0.549 -1.558 -0.194 -0.359 -0.333 -0.041 -0.552 -0.422 

10 0.677 -0.549 0.483 -0.194 -0.723 -0.421 1.149 -0.567 -0.407 

… … … … … … … … … … 

37 0.677 1.77 0.483 1.447 0.442 0.268 -0.253 1.382 1.851 

38 0.677 -0.549 0.483 1.447 0.180 0.094 0.962 1.166 0.799 

39 0.677 -0.549 0.483 1.447 -0.237 -0.112 0.216 0.259 0.539 

3.2. Analysis of experimental results 

3.2.1. Forecast of earthquake injuries 
BP simulation experiments are carried out using the above 

standardized processed data, the dataset contains 39 seismic 

case samples, and all the later simulation experiments use the 

data after this processing. Firstly, the Pearson correlation 

coefficient between the data is calculated and the correlation 

heat map is plotted as shown in figure 2 below. 

 
Figure 2. Heat map of Pearson's correlation coefficient 

For the BP neural network model parameters are set as 

shown in Table 4 below. 

 

 



 

191 

Table 4. Model parameters 

Trainin

g 

Epochs 

Batc

h 

Size 

Lear

ning 

Rate 

Number 

of 

Neurons 

Hidden 

Layer 

Activation 

Function 

Output 

Layer 

Activation 

Function 

500 3 0.01 17 Linear Sigmoid 

 

The loss function curve for the output of the algorithm is 

shown in figure 3. 

 
Figure 3. Graph of loss function 

The loss function pictures visualize the change in loss 

during training and testing of the two different models. It can 

be seen from the figure that with the increase of training 

rounds, the loss of both the training and testing sets is 

gradually decreasing, which indicates that the model is 

gradually learning and optimizing its performance. The 

metrics of this prediction model are shown in the table below. 

Table 5. Performance evaluation indicators 

Evaluation Metrics Experimental Values 

MSE 1.56954 

MAE 0.00404 

R2 0.93337 

 

The output values predicted by the model and the real test 

set target values are reduced to the original scale by a specific 

scaling transformation for more intuitive comparison and 

analysis. Figure 4 shows the predicted and actual graphs after 

reduction to the original scale. 

 
Figure 4. Comparison of predicted and actual values 

In order to visualize the comparison between the model 

data and the real values, the specific data are organized into 

the following table. 

Table 6. Statistics of predicted and real values 

Region Actual Value Predicted Value 

Yucheng 1109 1098 

Lushan 5537 5780 

Baoxing 2503 2110 

Tianquan 811 699 

Mingshan 607 560 

Yingjing 341 495 

Shimian 39 52 

Hanyuan 35 45 

ToTal 10982 10839 

3.2.2. Forecasting of Requirements for Emergency 
Supplies 

Based on the data projected in section 3.2.1, the demand 

for emergency supplies is calculated, and the demand for 

daily consumables as well as general pharmaceuticals in each 

region is shown in the following statistical table: 

Table 7. Forecast of emergency supplies 

Region 
Daily 

Consumables (t) 

General 

Pharmaceuticals (kg) 

Yucheng 535.2 131.7 

Lushan 168 693.6 

Baoxing 88.4 253.2 

Tianquan 232.8 83.8 

Mingshan 419.8 67.2 

Yingjing 223.5 59.4 

Shimian 186.6 6.2 

Hanyuan 186.6 5.4 

Total 2040.9 1300.7 

4. Summary and outlook 

China is in an earthquake-prone area, and earthquake 

disaster, as a serious natural disaster, has caused serious 

economic losses and also seriously jeopardized people's lives. 

It is of great significance to predict the number of injuries for 

the first time after the earthquake and reasonably to provide 

data reference to the related organizations of emergency 

rescue. 

In this paper, after analyzing many factors affecting 

earthquake casualties, according to the principle of being able 

to obtain direct or using rapid assessment methods for the first 

time after the earthquake, seven parameters such as the 

magnitude, the time of the earthquake, the seismic intensity, 

the intensity of earthquake damage, the affected population, 

the collapse of buildings, and the level of economic 

development are selected as the indicators for predicting the 

number of people injured in each affected point of the 

earthquake. In addition, the article constructs a three-layer BP 

neural network model and divides the collected samples into 

training sample sets and test sample sets. By continuously 

training and testing the samples, it proves that the network 

meets the error requirements and verifies the reasonableness 

of the constructed model. The experimental results of the 

model show that the coefficient of determination 𝑅^2 
reaches 0.93337, proving that the neural network model has 

high accuracy in predicting the target variable. 

Although the BP neural network model constructed in this 

paper achieved high accuracy in predicting the number of 

injuries of people at earthquake-affected sites, the data used 

in this paper may be limited by factors such as sample size, 

data quality and coverage. The complexity and diversity of 



 

192 

earthquake hazards may result in some important influencing 

factors not being included in the model, thus affecting the 

predictive ability of the model. In addition, although the 

model achieved high accuracy on the training set, its 

generalization ability on unknown data or newly occurring 

seismic events has not been fully validated. Therefore, in 

future research, it is necessary to improve the predictive 

ability and generalization performance of the model by 

collecting more and more comprehensive seismic hazard data, 

including real-time and historical data, as well as integrating 

data from different sources. 

References 

[1] Guo Jinfen. Research on Emergency Material Demand 
Forecasting Methods for Large-scale Earthquakes [D]. Tianjin 
University, 2012. 

[2] Fan Rui. Research on Large-scale Earthquake Emergency 
Material Demand Forecasting Based on PSO-BP Neural 
Network [D]. Beijing Jiaotong University, 2020. DOI: 
10.26944/d.cnki.gbfju.2020.001371. 

[3] Lan Hengxing, Zhou Chenghu, Gao Xing, et al. Evaluation and 
Countermeasure Suggestions on Secondary Geological 
Disasters in the Lushan Earthquake-stricken Area of Ya'an, 
Sichuan [J]. Progress in Geographical Sciences, 2013, 32(04): 
499-504. 

[4] Communist Party Members' Network. The death toll in Ya'an 
City due to the earthquake as of 12:00 on April 23 has reached 
173 [EB/OL]. 2013. 

[5] He Xinyun, Wang Chao. Research on Flood Control 
Emergency Material Demand Forecasting Based on BP Neural 
Network [J]. Information and Computer (Theoretical Edition), 
2022, 34(24): 84-86 + 91.