Dynamic earthquake casualty number prediction method and system based on improved case reasoning-neural network
By improving the combination of case-based reasoning and neural networks, and optimizing the BP network using principal component analysis and particle swarm optimization, the problems of speed and accuracy in earthquake casualty prediction were solved, achieving efficient dynamic prediction of earthquake casualty numbers.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-14
AI Technical Summary
Existing earthquake casualty prediction methods are insufficient in terms of speed and accuracy. Traditional statistical regression models and vulnerability analysis models are difficult to capture the interactions between complex factors. Case reasoning methods lack the ability to learn and adapt to complex nonlinear relationships. Standard BP neural network prediction results are not capable of dynamic adjustment in the early stages of post-earthquake emergency response.
An improved case-based reasoning-neural network approach is adopted. Principal component analysis is used to reduce the dimensionality of earthquake data, a weighted K-NN algorithm is used to select the case set, and the initial weights and thresholds of the BP neural network are optimized by combining particle swarm optimization. Dynamic prediction is then performed by combining real-time information available after an earthquake occurs.
It improves the convergence stability and learning efficiency of earthquake casualty prediction models, enhances the real-time calibration capability of prediction results, and improves the prediction accuracy and dynamic adjustment capability in the early stages of emergency response.
Smart Images
Figure CN121860075A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of earthquake casualty prediction technology, and more specifically, to a dynamic earthquake casualty prediction method and system based on an improved case-reasoning neural network. Background Technology
[0002] Earthquakes are highly destructive natural disasters, often causing massive casualties and property damage in a short period of time. Rapidly and relatively accurately predicting the potential number of casualties from an earthquake is crucial for emergency management departments to scientifically formulate rescue plans, rationally allocate emergency resources, and minimize losses. Therefore, earthquake casualty prediction has always been a research hotspot in the field of disaster emergency management.
[0003] Traditional earthquake casualty prediction methods are mainly divided into three categories: statistical regression models based on historical experience, vulnerability analysis models based on physical mechanisms, and case-based reasoning methods. Statistical regression models predict casualties by establishing mathematical relationships between macroscopic indicators such as magnitude, intensity, and population density. However, these models often assume linear or simple nonlinear relationships, making it difficult to capture the interactions between complex factors. They also heavily rely on the distribution of historical data and have limited extrapolation capabilities. Vulnerability analysis models focus on the physical relationship between building damage and casualties. Their prediction accuracy highly depends on detailed regional building structure databases and accurate earthquake damage simulations. In actual emergency response, they are often difficult to apply quickly due to difficulties in data acquisition or computational complexity. Case-based reasoning methods predict casualties by retrieving historical cases similar to the current earthquake scenario and referencing their casualty results. While intuitive, the similarity measurement of cases often relies on expert experience or simple Euclidean distance, feature weights are subjectively set, and conclusions from historical cases are directly reused, lacking the ability to learn from and adapt to complex nonlinear relationships.
[0004] In recent years, artificial neural networks, especially backpropagation neural networks (BP networks), have been introduced into the field of earthquake casualty prediction due to their powerful nonlinear fitting capabilities. BP neural networks can build predictive models by learning complex patterns in historical data. However, standard BP neural networks also have inherent drawbacks: firstly, the initial weights and thresholds are randomly generated, which can easily lead to training getting stuck in local optima and unstable model performance; secondly, their predictions rely entirely on the historical data feature set used during training, and cannot flexibly incorporate real-time dynamic information (such as preliminary disaster reports) available shortly after an earthquake during prediction, resulting in insufficient dynamic adjustment capabilities of the prediction results in the initial post-earthquake emergency response. Summary of the Invention
[0005] The present invention provides a dynamic earthquake casualty prediction method and system based on an improved case-reasoning neural network, which can overcome some or all of the defects of the prior art.
[0006] According to an improved case-based reasoning neural network-based dynamic earthquake casualty prediction method of the present invention, the method includes the following steps: Step 1: Collect and build a dataset of earthquake casualties; Step 2: Use principal component analysis to simplify the earthquake casualty data set and establish a case set; Step 3: Use the weighted K-NN algorithm to filter the case set and collect historical cases with high similarity to the prediction target to build training samples; Step 4: Obtain real-time indicators of the predicted target as correction parameters. Real-time indicators include the number of affected people, the rate of house damage, the number of deaths on the day of the earthquake, and the number of injuries on the day of the earthquake. Step 5: Input the training samples and correction parameters into the PSO-BP neural network prediction model to predict the number of casualties.
[0007] 2. The dynamic earthquake casualty prediction method based on improved case reasoning-neural network according to claim 1, characterized in that: in step 1, the dataset includes seven indicators and the number of deaths and injuries corresponding to each case; the seven indicators are earthquake occurrence time period, magnitude, epicenter intensity, population density, seismic fortification intensity, forecast level and season.
[0008] 3. The dynamic earthquake casualty prediction method based on improved case reasoning-neural network according to claim 2 is characterized in that: in step 2, principal component analysis is performed on the 7 indicators of the earthquake casualty dataset to calculate the contribution rate, eigenvalue and cumulative contribution rate of each principal component, and the top 6 principal components whose cumulative contribution rate reaches the set threshold are extracted to form a case set.
[0009] 4. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 3, characterized in that: step 3 specifically includes: (1) Feature normalization; Mini-maximum normalization is performed on the first 6 principal components:
[0010] (2) Calculation of weighted Euclidean distance; Let the target case be x*=( x* 1, x* 2,..., x* 6), The historical cases in the case set are x=( x 1, x 2,..., x 6); Case similarity is calculated based on case distance. Weighted Euclidean distance is used to calculate the distance between the target case and all cases in the case set, with weighting coefficients... Given the contribution rate of each principal component, the weighted Euclidean distance is:
[0011] (3) Sort in ascending order of distance; Calculate the distance between the target case and all N historical cases in the case set, sort them by D in ascending order, and select the top k nearest neighbor cases as training samples.
[0012] 5. The dynamic earthquake casualty prediction method based on the improved case reasoning-neural network according to claim 4, characterized in that: in step 5, the network structure and network parameters of the BP neural network are designed first; then the initial weights and thresholds of the BP neural network are optimized using the particle swarm optimization algorithm (PSO) to obtain the PSO-BP neural network prediction model.
[0013] 6. The dynamic earthquake casualty prediction method based on the improved case-reasoning neural network according to claim 5, characterized in that: in the network structure design of the BP neural network, the number of nodes in the input layer is 6 or 10; the number of nodes in the output layer of the neural network is 2; the number of nodes L in the hidden layer is determined using the following empirical formula (1): (1) In the formula, p and q are the number of input nodes and output nodes, respectively. p is 6 or 10, q is 2, a is a constant between [1, 10], and L takes values in the range of [4, 15].
[0014] 7. The dynamic earthquake casualty prediction method based on the improved case-reasoning neural network according to claim 6, characterized in that: in the network parameter design of the BP neural network, the transfer function between the input layer and the hidden layer adopts the Tansig function in the sigmoid function, and the transfer function between the hidden layer and the output layer adopts the Purelin function in the linear function.
[0015] 8. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 7, characterized in that: when optimizing the initial weights and thresholds of the BP neural network using the particle swarm optimization (PSO) algorithm, the output of the PSO algorithm is used as the initial weights and thresholds of the BP neural network, and then the network is trained. That is, based on the determined neural network structure, a mapping is established between the two. The fitness function of the PSO algorithm is the output error of the BP neural network, and the initial weights and thresholds of the BP neural network after iteration are the optimal solution of the PSO algorithm; the specific process of the PSO-BP neural network prediction model is as follows: Step 1: Initialize the population, including: the initial position of each particle in the domain [V min V max Randomly selected from [X], with the initial position in the domain [X]. min X max Randomly select from [the data], the optimal position p j (t) represents the initial position; Step 2: Calculate the fitness of the particle at this time. The fitness function uses the mean squared error (MSE). Record the current position and find the individual optimal and population optimal values of the current particle. Step 3: Compare the particle's fitness with the individual optimal and the global optimal. If the particle's fitness is less than the individual optimal, then the current value is taken as the individual optimal value. Otherwise, the global optimal value of the particle swarm is the individual optimal value at this time. Step 4: Use the optimal individual obtained by the PSO algorithm through updating position and velocity as the initial weights and threshold of the BP neural network to train the BP neural network. End the training when the network reaches the maximum number of iterations, and then output the result. Otherwise, go to Step 2 to continue iterating until convergence. Step 5: Output the optimized weights and thresholds.
[0016] 9. A dynamic earthquake casualty prediction system based on an improved case-reasoning neural network, characterized in that: it adopts the dynamic earthquake casualty prediction method based on an improved case-reasoning neural network as described in any one of claims 1-8.
[0017] The beneficial effects of this invention are as follows: This invention combines an improved case-based reasoning (CBR) approach with an optimized neural network (PSO-BP). First, Principal Component Analysis (PCA) is used to reduce the dimensionality of the original high-dimensional features, extracting the comprehensive principal components that best explain the changes in casualties. This effectively eliminates redundant information and noise between indicators, providing a cleaner and more fundamental input for subsequent models. Second, a weighted K-NN algorithm is used for case selection. Euclidean distance, weighted by the contribution rate of the principal components, accurately measures case similarity, ensuring that the selected training samples are highly relevant to the current prediction target, providing high-quality, targeted learning samples for the neural network. Finally, Particle Swarm Optimization (PSO) is used to optimize the initial weights and thresholds of the BP neural network. This effectively overcomes the shortcomings of traditional BP neural networks, such as being prone to getting trapped in local minima, slow convergence speed, and high randomness in results. This allows the network to start training from a better starting point, significantly improving the model's convergence stability, learning efficiency, and final prediction accuracy.
[0018] Based on preliminary modeling using historical cases, key real-time information that can be quickly obtained shortly after an earthquake (such as the number of affected people, building damage rate, and the number of casualties on the day of the earthquake) is used as correction parameters and directly input into the pre-trained PSO-BP neural network. This allows the model to dynamically integrate the latest disaster information and calibrate and correct the preliminary predictions based on historical patterns in real time. This feature greatly enhances the practical value of the prediction system in the early stages of emergency response, enabling the prediction results to be continuously optimized as disaster information is updated, providing increasingly accurate dynamic references for rescue command and decision-making over time. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating a dynamic earthquake casualty prediction method based on an improved case-reasoning neural network, as shown in the embodiment.
[0020] Figure 2 This is a schematic diagram comparing the simulation prediction results in the example. Detailed Implementation
[0021] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.
[0022] Example like Figure 1 As shown, this embodiment provides a dynamic earthquake casualty prediction method based on an improved case-based reasoning neural network, which includes the following steps: Step 1: Collect and establish an earthquake casualty dataset; the dataset includes seven indicators (X1~X7) and the number of deaths and injuries corresponding to each case; the seven indicators are: earthquake occurrence time (X1), magnitude (X2), epicentral intensity (X3), population density (X4), seismic fortification intensity (X5), forecast level (X6), and season (X7); the reasons for the selection of each indicator are as follows: 1) Time period of earthquake occurrence (X1) The time of an earthquake is an important factor affecting the number of casualties: earthquakes at night cause more casualties than earthquakes during the day. Earthquakes occurring between 0:00 and 6:00 AM have a higher number of casualties. Therefore, this period with the highest number of casualties is marked as 2, and the other times are marked as 1.
[0023] 2) Magnitude (X2) Magnitude indicates the amount of energy released at the earthquake's source.
[0024] 3) Epicenter Intensity (X3) Epicentral intensity is often used to measure the impact and damage of an earthquake on buildings in the affected area, and it is positively correlated with the casualty rate in the disaster area.
[0025] 4) Population density (X4) (people / km) 2 ) There is a positive correlation between population density and casualties in disaster areas.
[0026] 5) Seismic fortification intensity (X5) There is a negative correlation between seismic fortification intensity and earthquake casualty rate; reasonable seismic design can effectively reduce the damage caused by earthquakes.
[0027] 6) Forecast Level (X6) There is a negative correlation between forecast accuracy and earthquake casualty rate; a higher forecast accuracy can effectively reduce the damage caused by earthquakes. This embodiment divides the forecast accuracy into three levels: Level 1 indicates no forecast, Level 2 indicates a forecast with some deviation, and Level 3 indicates an accurate forecast.
[0028] 7) Seasons (X7) Season is an important factor affecting the number of casualties. The number of casualties caused by earthquakes varies in different seasons. In this embodiment, the seasons are divided into four levels: spring (March-May) is level 1, summer (June-August) is level 2, autumn (September-November) is level 3, and winter (December-February of the following year) is level 4. Based on the above indicators, the original dataset of 103 earthquake groups of magnitude 6 and above in China is selected as shown in Table 1 below.
[0029] Table 1 Earthquake-related data ; Step 2: Simplify the earthquake casualty data set using Principal Component Analysis (PCA) to establish a case set. Since the indicators selected in the original dataset may exhibit correlation, PCA is used to reduce the attributes of the seven indicators in the original dataset to avoid this phenomenon. In practice, due to the different physical meanings and dimensions of the indicators, the original dataset is standardized as shown in Table 2. PCA is then performed on the standardized dataset to calculate the contribution rate, eigenvalues, and cumulative contribution rate of each principal component, as shown in Table 3. Since the cumulative contribution rate of the first six principal components has reached 98.2%, meeting the set threshold and accuracy requirements, the first six principal components can be selected to form the case set.
[0030] Table 2 Principal Component Coefficients ; Table 3. Principal Component Eigenvalues and Contribution Rates ; Step 3: Use the weighted K-NN algorithm to filter the case set and collect historical cases with high similarity to the prediction target to build training samples. The core idea of this hybrid modeling method, which combines Case-Based Reasoning (CBR), K-Nearest Neighbors (K-NN), and a backpropagation neural network, is to retrieve the k most similar cases to the target earthquake from the historical earthquake case database using K-NN, and then train the backpropagation neural network using only these k highly correlated samples. This significantly reduces the amount of training data and improves training efficiency while maintaining prediction accuracy.
[0031] The distance metric used is weighted Euclidean distance, specifically: (1) Feature normalization; Mini-maximum normalization is performed on the first 6 principal components (input variables):
[0032] (2) Calculation of weighted Euclidean distance; Let the target case be x*=( x* 1, x* 2,..., x* 6), The historical cases in the case set are x=( x 1, x 2,..., x 6); Case similarity is calculated based on case distance. Weighted Euclidean distance is used to calculate the distance between the target case and all cases in the case set, with weighting coefficients... The contribution rates of each principal component are shown in Table 4 below; the weights of each dimension are... =[0.304,0.233,0.148,0.134,0.114,0.068]; then the weighted Euclidean distance is:
[0033] Table 4. Contribution Ratio of Principal Components ; (3) Sort in ascending order of distance; Calculate the distance between the target case and all N historical cases in the case set, sort them by D in ascending order, and select the top k nearest neighbor cases (k is a hyperparameter, such as k=10, 20, 40, 40 is selected in this embodiment) as training samples (approximate case set), as shown in Table 5. The training samples are used to train the subsequent PSO-BP neural network prediction model. The input layer has 6 features (X1~X6), and the output layer has 2 targets (number of deaths Z1, number of injuries Z2). After selection, the training samples are small, convergence is fast, and the risk of overfitting is low.
[0034] Table 5 Training Samples (Approximate Case Set) ; Step 4: Obtain real-time indicators of the predicted targets as correction parameters. Real-time indicators include the number of affected people (Y1), the rate of house destruction (Y2), the number of deaths on the day of the earthquake (Y3), and the number of injuries on the day of the earthquake (Y4). By sorting out statistical data and disaster assessment reports, the above historical statistical data are obtained, as shown in Table 6. Among them, there are few relevant reports before 2003, and there are no reports on the number of casualties on the day and the next day. Missing data is indicated by -.
[0035] Table 6 Real-time Indicator Parameters for Approximate Case Sets ; Step 5: Input the training samples and correction parameters into the PSO-BP neural network prediction model to predict the number of casualties.
[0036] This embodiment first completes the design of the BP neural network structure and network parameters; then, it uses the particle swarm optimization algorithm (PSO) to optimize the initial weights and thresholds of the BP neural network, thus obtaining the PSO-BP neural network prediction model.
[0037] Network architecture design Three-layer networks, consisting of only an input layer, hidden layers, and an output layer, are currently one of the preferred network structures for solving various practical problems using multi-layer neural networks. This is because they are the simplest, offer good parallelism and low computational cost, and can arbitrarily approximate a nonlinear continuous function. Therefore, this model also uses a three-layer network for modeling.
[0038] The dimensionality of the actual problem determines the number of neurons in the input and output layers. In the static prediction model, there are 6 prediction indicators (X1~X6), while in the dynamic prediction model, there are 12 prediction indicators (X1~X6, Y1~Y4). That is, the input variables of the BP neural network are 6 or 10, so the number of nodes in the input layer is 6 or 10. In this embodiment, the prediction focus when predicting the demand for emergency supplies after an earthquake is the number of deaths and injuries caused by the earthquake, so the number of nodes in the output layer of the neural network is 2. The number of hidden layer nodes L can directly affect the network training effect, and is generally determined by the following empirical formula (1): (1) In the formula, p and q are the number of input nodes and output nodes, respectively. p is 6 or 10, q is 2, and a is a constant between [1, 10]. Therefore, the value range of L is [4, 15]. By gradually increasing the number of neurons using the trial-and-error method, after multiple calculations, it was determined that the training effect is optimal when the number of hidden nodes is 10.
[0039] Network parameter design The transfer functions for each layer were determined. The transfer function from the input layer to the hidden layer was the Tansig function from the sigmoid function group, and the transfer function from the hidden layer to the output layer was the Purelin function from the linear function group. The other network parameters, after debugging, are shown in Table 7.
[0040] Table 7 Network Parameter Design ; PSO optimizes the BP neural network. Most scholars who combine the PSO algorithm and the BP neural network algorithm mainly use the former to optimize the topology of the latter, including optimizing the network weights, number of nodes, and number of hidden layers. However, this method is difficult to implement and has a relatively slow convergence speed, making it unsuitable for predicting emergency resources in sudden situations. Therefore, the optimization approach adopted in this embodiment mainly utilizes the PSO algorithm to optimize the weights and thresholds of the BP neural network during error backpropagation. Specifically, the output of the PSO algorithm is used as the initial weights and thresholds of the BP neural network, and the network is then trained. This establishes a mapping between the two based on the determined neural network structure. The fitness function of the PSO algorithm is the output error of the BP neural network, and therefore, the initial weights and thresholds of the BP neural network after iteration are the optimal solution of the PSO algorithm. The specific process of the PSO-BP neural network prediction model is as follows: Step 1: Initialize the population, including: the initial position of each particle in the domain [V min V max Randomly selected from [X], with the initial position in the domain [X]. min X max Randomly select from [the data], the optimal position p j (t) represents the initial position; the initialization parameters are shown in Table 8 below, including the number of particles, and the domain of the inertial weight w is [W]. min W max Maximum number of iterations T max Social coefficient C2, cognitive coefficient C l ; Table 8 PSO Algorithm Parameter Settings ; Step 2: Calculate the fitness of the particle at this time. The fitness function uses the mean squared error (MSE). Record the current position and find the individual optimal and population optimal values of the current particle. Step 3: Compare the particle's fitness with the individual optimal and the global optimal. If the particle's fitness is less than the individual optimal, then the current value is taken as the individual optimal value. Otherwise, the global optimal value of the particle swarm is the individual optimal value at this time. Step 4: Use the optimal individual obtained by the PSO algorithm through updating position and velocity as the initial weights and threshold of the BP neural network to train the BP neural network. End the training when the network reaches the maximum number of iterations, and then output the result. Otherwise, go to Step 2 to continue iterating until convergence. Step 5: Output the optimized weights and thresholds.
[0041] This embodiment provides a dynamic earthquake casualty prediction system based on an improved case-reasoning neural network, which adopts the aforementioned dynamic earthquake casualty prediction method based on an improved case-reasoning neural network.
[0042] Simulation test To fully verify the performance improvement effect of K-NN case selection, dynamic index introduction, and PSO-optimized BP neural network (PSO-BP) on earthquake casualty prediction model, a multi-factor ablation experiment was designed, and the experimental scheme is shown in Table 9.
[0043] Table 9. Experimental Design Scheme for Multifactor Ablation ; According to the above experimental plan, the seismic data in Tables 1, 5, and 6 were imported into the M1~M4 neural network for training. Four sets of data were randomly selected as the validation set, and the remaining data were used as the training set. The simulation test results are as follows. Figure 2 As shown.
[0044] The complement of the mean absolute percentage error (MAPE) is used as the "prediction accuracy" index: prediction accuracy = 1 - MAPE × 100%. The prediction accuracy of each model is shown in Table 10.
[0045] Table 10 Model Prediction Accuracy ; In the ablation experiment of this embodiment, M1, as the baseline model (using full data, static indices X1~X6, and a standard BP network), achieved prediction accuracies of 57.90% for deaths and 64.60% for injuries, reflecting the limitations of traditional static modeling methods. After introducing the K-NN sample selection strategy (M2), the model was trained using only k=40 historical samples most similar to the current earthquake. With other settings unchanged, the prediction accuracy significantly improved: the accuracy for deaths increased by 23.88 percentage points to 81.78%, and the accuracy for injuries increased by 10.69 percentage points to 75.29%. This indicates that learning based on similar cases can effectively filter out noisy samples and significantly improve the model's adaptability to specific seismic events.
[0046] Further replacing the standard BP network with PSO-BP (M3), which utilizes the particle swarm optimization algorithm to optimize the initial network weights to avoid local optima, resulted in a continued steady improvement in model performance. The prediction accuracy for deaths and injuries increased to 88.02% and 79.47%, respectively, representing improvements of 6.24% and 4.18% compared to M2. These results validate the effectiveness of optimizing the network initialization strategy in improving model convergence quality, especially demonstrating stable gains when handling high-dimensional nonlinear relationships.
[0047] Building upon this, the M4 model introduces dynamic indicators (Y1~Y4, such as early disaster feedback information) to construct a "static + dynamic" fused feature system, while keeping other configurations unchanged. The results show that the accuracy of death prediction improved to 91.74%, and the accuracy of injury prediction improved even more significantly to 85.80%, representing increases of 3.72% and 6.33% respectively compared to M3. This difference indicates that dynamic indicators contribute more to injury prediction because they rely more heavily on real-time disaster response information (building damage rate, number of affected people, etc.), while death prediction is more strongly influenced by static factors such as initial magnitude and site conditions.
[0048] In summary, the progressive improvements from M1 to M4 achieved synergistic effects: K-NN screening brought the greatest performance leap, PSO-BP provided robust optimization, and dynamic indicators further mined information on disaster evolution. These three strategies jointly constructed a highly efficient prediction framework "based on similar scenarios, incorporating real-time feedback, and optimized training," ultimately improving the prediction accuracy of deaths and injuries by 33.84% and 21.20%, respectively, fully validating the effectiveness and engineering value of the proposed method.
[0049] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.
Claims
1. A dynamic earthquake casualty prediction method based on an improved case-based reasoning neural network, characterized by: Includes the following steps: Step 1: Collect and build a dataset of earthquake casualties; Step 2: Use principal component analysis to simplify the earthquake casualty data set and establish a case set; Step 3: Use the weighted K-NN algorithm to filter the case set and collect historical cases with high similarity to the prediction target to build training samples; Step 4: Obtain real-time indicators of the predicted target as correction parameters. Real-time indicators include the number of affected people, the rate of house damage, the number of deaths on the day of the earthquake, and the number of injuries on the day of the earthquake. Step 5: Input the training samples and correction parameters into the PSO-BP neural network prediction model to predict the number of casualties.
2. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 1, characterized in that: In step 1, the dataset includes seven indicators and the number of deaths and injuries corresponding to each case; the seven indicators are the time period of the earthquake, magnitude, epicenter intensity, population density, seismic fortification intensity, forecast level, and season.
3. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 2, characterized in that: In step 2, principal component analysis is performed on the seven indicators of the earthquake casualty data set to calculate the contribution rate, eigenvalue, and cumulative contribution rate of each principal component. The top six principal components whose cumulative contribution rate reaches the set threshold are then used to form a case set.
4. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 3, characterized in that: Step 3 specifically involves: (1) Feature normalization; Mini-maximum normalization is performed on the first 6 principal components: ； (2) Calculation of weighted Euclidean distance; Let the target case be x*=( x* 1, x* 2,..., x* 6), The historical cases in the case set are x=( x 1, x 2,..., x 6); Case similarity is calculated based on case distance. Weighted Euclidean distance is used to calculate the distance between the target case and all cases in the case set, with weighting coefficients... Given the contribution rate of each principal component, the weighted Euclidean distance is: ； (3) Sort in ascending order of distance; Calculate the distance between the target case and all N historical cases in the case set, sort them by D in ascending order, and select the top k nearest neighbor cases as training samples.
5. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 4, characterized in that: In step 5, the network structure and parameters of the BP neural network are designed first; then the initial weights and thresholds of the BP neural network are optimized using the particle swarm optimization (PSO) algorithm to obtain the PSO-BP neural network prediction model.
6. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 5, characterized in that: In the design of the BP neural network structure, the number of nodes in the input layer is 6 or 10; the number of nodes in the output layer is 2; and the number of nodes L in the hidden layer is determined using the following empirical formula (1): (1)； In the formula, p and q are the number of input nodes and output nodes, respectively. p is 6 or 10, q is 2, a is a constant between [1, 10], and L ranges from [4, 15].
7. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 6, characterized in that: In the design of BP neural network parameters, the transfer function from the input layer to the hidden layer adopts the Tansig function among sigmoid functions, and the transfer function from the hidden layer to the output layer adopts the Purelin function among linear functions.
8. The dynamic earthquake casualty prediction method based on improved case-reasoning neural network according to claim 7, characterized in that: When optimizing the initial weights and thresholds of a backpropagation (BP) neural network using the Particle Swarm Optimization (PSO) algorithm, the output of the PSO algorithm is used as the initial weights and thresholds of the BP neural network for training. This involves establishing a mapping between the PSO and BP neural network based on the determined neural network structure. The fitness function of the PSO algorithm represents the output error of the BP neural network, and the initial weights and thresholds of the BP neural network after iteration represent the optimal solution for the PSO algorithm. The specific process of the PSO-BP neural network prediction model is as follows: Step 1: Initialize the population, including: the initial position of each particle in the domain [V min V max Randomly selected from [X], with the initial position in the domain [X]. min X max Randomly select from [the data], the optimal position p j (t) represents the initial position; Step 2: Calculate the fitness of the particle at this time. The fitness function uses the mean squared error (MSE). Record the current position and find the individual optimal and population optimal values of the current particle. Step 3: Compare the particle's fitness with the individual optimal value and the global optimal value; if the particle's fitness is less than the individual optimal value, then the current value is taken as the individual optimal value; otherwise, the global optimal value of the particle swarm is the individual optimal value at this time. Step 4: Use the optimal individual obtained by the PSO algorithm through updating position and velocity as the initial weights and threshold of the BP neural network to train the BP neural network. End the training when the network reaches the maximum number of iterations, and then output the result. Otherwise, go back to Step 2 to continue iterating until convergence. Step 5: Output the optimized weights and thresholds.
9. A dynamic earthquake casualty prediction system based on an improved case-based reasoning neural network, characterized in that: It employs the dynamic earthquake casualty prediction method based on an improved case-reasoning neural network as described in any one of claims 1-8.