Post-disaster population flow prediction method and device

Abstract

The application provides a post-disaster population flow prediction method and device, comprising the following steps: obtaining population flow data of a target area according to a predetermined space-time fine granularity; obtaining an input population flow dynamic graph according to the population flow data of the target area; wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of a disaster peak period of the target area and an average normal population flow graph of a non-disaster period of the target area; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training a sample area population flow data based on a coupled dynamic graph neural ordinary differential equation. The application adopts the coupled dynamic graph neural ordinary differential equation to construct the population flow prediction model, and realizes the post-disaster population flow prediction with high accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data processing, and particularly relates to a post-disaster population flow prediction method and device. BACKGROUND

[0002] With the acceleration of urbanization, the frequent extreme events (disasters) occurring all over the world have caused an undeniable impact on the life safety of urban residents. Therefore, there is an urgent need to understand the urban resilience, that is, the ability of the city to quickly recover its functions after being impacted by extreme events, among which the population flow within the city plays a crucial role, mainly in two aspects: on the one hand, the urban population flow reflects the behavior of urban residents to obtain essential resources (such as food and money) from residential areas to other functional areas, and is a key indicator to measure the satisfaction of residents in obtaining necessities of life; on the other hand, the urban population flow has a complex relationship with the normal operation of urban infrastructure such as transportation facilities and office buildings, thereby reflecting the recovery status of urban infrastructure. Generally speaking, cities with poor resilience usually need more time to recover to normal urban population flow, thereby causing greater impact on the life of urban residents and bringing greater economic losses. Therefore, accurately predicting the recovery process of post-disaster urban population flow helps to find high-risk urban communities, design better emergency response strategies, and ultimately build a more intelligent and resilient city.

[0003] However, predicting the post-disaster recovery process of urban population flow is also a difficult task, facing the following challenges: first, the post-disaster recovery of urban population flow has a complex pattern, which is intricately intertwined with different disaster situations and its pre-disaster population flow characteristics, however, the underlying mechanism is still unclear, and it is difficult to accurately model using only observable data, and the change pattern of population flow under normal circumstances is not helpful for the change pattern of population flow under abnormal circumstances. Second, the recovery process of population flow in different urban areas is not independent of each other, and the interregional population flow also has a significant impact on its recovery process. For urban areas with a large net population inflow, the recovery speed will be faster due to the larger number of active population, thereby being conducive to the reconstruction and recovery process, and the highly dynamic interregional population interaction process further increases the difficulty of effectively modeling and predicting the population flow recovery dynamics.

[0004] The existing modeling methods for the post-disaster recovery process of urban population flow are mainly based on mathematical modeling methods, that is, making an oversimplified assumption about the urban flow, and constructing a mathematical model with a very limited number of parameters. Although these models provide valuable knowledge about the post-disaster recovery of urban population flow and have good interpretability, the strong simplification assumption and limited model parameters limit their ability to fully capture the diversification and complex recovery pattern of post-disaster urban traffic. In addition, all these models fail to effectively simulate the impact of interregional population flow on its recovery.

[0005] Meanwhile, the rapid development of data-driven deep learning techniques represented by graph neural networks (GNN) and neural ordinary differential equations (Neural ODE) further enhances our ability to model the dynamics of co-evolving nodes and edges in graph structures, and a number of algorithms for learning graph data with temporal characteristics have emerged, that is, to model multivariate time series and their interactions in the form of dynamic graphs. It is feasible to use mainstream dynamic graph prediction methods to predict the post-disaster urban population flow recovery process, but the prediction results have large errors, because the mainstream dynamic graph prediction method uses the graph of a period of time in the past to predict the graph of a period of time in the future, which relies on the graph signal evolution pattern learned from the historical graph sequence. However, this pattern only reflects the normal data fluctuations and has little help for the abnormal population flow recovery process after a disaster.

[0006] In summary, the prior art has the problem of low accuracy. SUMMARY

[0007] The present application provides a post-disaster population flow prediction method and device to solve the problem of low accuracy in the prior art and achieve high-accuracy post-disaster population flow prediction.

[0008] The present application provides a post-disaster population flow prediction method, comprising the following steps:

[0009] Obtaining population flow data of a target area according to a predetermined spatiotemporal granularity;

[0010] Obtaining an input population flow dynamic graph according to the population flow data of the target area; wherein the input population flow dynamic graph at least includes an initial abnormal population flow graph of the target area at a disaster peak period and an average normal population flow graph of the target area during a non-disaster period;

[0011] Inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result;

[0012] The population flow prediction model is obtained by training based on a coupled dynamic graph neural ordinary differential equation using population flow data of a sample area.

[0013] The present application provides a post-disaster population flow prediction method based on a coupled dynamic graph neural ordinary differential equation, which uses population flow data of a sample area to train the population flow prediction model, comprising:

[0014] Obtaining population flow data of a sample area;

[0015] obtain an initial abnormal population flow graph of a sample area disaster peak period, an average normal population flow graph of the sample area during a non-disaster period, and a real abnormal population flow graph of a sample area recovery stage according to population flow data of the sample area;

[0016] construct a basic prediction model based on a coupled dynamic graph neural ordinary differential equation;

[0017] train the basic prediction model by taking the initial abnormal population flow graph of the sample area disaster peak period and the average normal population flow graph of the sample area during the non-disaster period as inputs and taking the real abnormal population flow graph of the sample area recovery stage as a label until the model converges, to obtain a population flow prediction model.

[0018] According to the post-disaster population flow prediction method provided by the application, the basic prediction model comprises a decoding module and an encoding module;

[0019] Correspondingly, the initial abnormal population flow graph of the sample area disaster peak period and the average normal population flow graph of the sample area during the non-disaster period are taken as inputs, and the real abnormal population flow graph of the sample area recovery stage is taken as a label to train the basic prediction model until the model converges, to obtain a population flow prediction model, specifically comprising:

[0020] According to the initial abnormal population flow graph of the sample area disaster peak period and the average normal population flow graph of the sample area during the non-disaster period, the encoding module is used to encode the interval population flow and the intra-district population flow to obtain an encoding result;

[0021] The decoding module is used to decode the encoding result to obtain a population flow prediction value at multiple time points in the post-disaster recovery process;

[0022] According to the population flow prediction value and the real abnormal population flow graph of the sample area recovery stage, a pre-set loss function is used for back propagation to update model parameters until the model converges, to obtain a population flow prediction model.

[0023] According to the post-disaster population flow prediction method provided by the application, the encoding module comprises:

[0024] ,

[0025] ,

[0026] ,

[0027] wherein, is an attenuation coefficient vector with a dimension of is an attenuation coefficient vector with a dimension of is a function based on , represents encoded abnormal intra-regional population flow encoding with dimension , represents the population flow in the region on the th day, is the population flow in the normal region after the same encoding, represents the attribute value of each point of the normal population flow graph, represents the transformation matrix, represents the activation function, is used to represent whether an edge exists, represents the vector coefficient with dimension , represents the region with inter-regional population flow between regions , and Softmax represents the Softmax activation function, represents an MLP, represents encoded abnormal inter-regional population flow encoding, represents the population flow from region to region on the th day, represents a multi-layer perception, represents the convolution operator of the GCN, represents the degree matrix of . According to the post-disaster population flow prediction method provided by the application, the decoding module comprises:

[0028] ,

[0029] ,

[0030] ,

[0031] ,

[0032] wherein, and respectively represent the population flow in the th region and the predicted value of the population flow from region to region on the th day in the recovery period, represents encoded abnormal intra-regional population flow encoding, ​The decoding function represented by the MLP stacked by two linear layers and two RELU activation functions.

[0033] In the case that the target area and the sample area are the same, the real abnormal population flow graph of the sample area in the recovery stage is the real abnormal population flow graph of the sample area in the initial recovery stage.

[0034] In the case that the target area and the sample area are different, the real abnormal population flow graph of the sample area in the recovery stage is the real abnormal population flow graph of the sample area in the full recovery stage.

[0035] The application further provides a post-disaster population flow prediction device, comprising the following modules.

[0036] A data unit is configured to acquire population flow data of the target area according to a predetermined spatio-temporal granularity.

[0037] A conversion unit is configured to obtain an input population flow dynamic graph according to the population flow data of the target area, wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of the target area in a disaster peak period and an average normal population flow graph of the target area in a non-disaster period.

[0038] A prediction unit is configured to input the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result.

[0039] The population flow prediction model is obtained by training the population flow data of the sample area based on a coupled dynamic graph neural differential equation.

[0040] The application further provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the post-disaster population flow prediction method according to any one of the above embodiments when executing the program.

[0041] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the post-disaster population flow prediction method according to any one of the above embodiments.

[0042] The application further provides a computer program product comprising a computer program, wherein the computer program is executable on a processor to implement the post-disaster population flow prediction method according to any one of the above embodiments.

[0043] The application provides a post-disaster population flow prediction method and device, which comprises the following steps: obtaining population flow data of a target area according to a predetermined space-time fine granularity; obtaining an input population flow dynamic graph according to the population flow data of the target area; wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of a disaster peak period of the target area and an average normal population flow graph of a non-disaster period of the target area; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training a sample area population flow data based on a coupled dynamic graph neural ordinary differential equation. The application adopts the coupled dynamic graph neural ordinary differential equation to construct the population flow prediction model, can effectively model the diversified and highly volatile urban mobility in different regions, and jointly considers the urban mobility before and after the disaster, solves the problem that the prior art relies on an oversimplified assumption, so that the influence of the population flow between regions cannot be effectively captured, the normal flow change mode cannot be used to predict the abnormal flow recovery process after the disaster, and the accuracy is low, and realizes the post-disaster population flow prediction with high accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0045] Figure 1 is one of the flow schematic diagrams of the post-disaster population flow prediction method provided by the application.

[0046] Figure 2 is the second flow schematic diagram of the post-disaster population flow prediction method provided by the application.

[0047] Figure 3 is the third flow schematic diagram of the post-disaster population flow prediction method provided by the application.

[0048] Figure 4 is the fourth flow schematic diagram of the post-disaster population flow prediction method provided by the application.

[0049] Figure 5 is the structural schematic diagram of the post-disaster population flow prediction device provided by the application.

[0050] Figure 6 is the structural schematic diagram of the electronic device provided by the application. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] The following is combined Figures 1-4 This invention describes a method for predicting post-disaster population movement. Figure 1 This is one of the flowcharts illustrating the post-disaster population movement prediction method provided by the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0053] Step 110: Obtain population flow data for the target area based on a predetermined spatiotemporal fine granularity.

[0054] It should be noted that the objective of this invention is to predict multiple population flow maps during the post-disaster recovery process, given at least one abnormal population flow map on the day of the disaster and at least one real normal population flow map. In this context, the target area refers to the city or region to be predicted, and the population flow data for the target area includes both the abnormal and real normal population flow data for that city or region. This population flow data does not require any normalization and is presented solely on an individual basis.

[0055] Spatiotemporal fine-grainedness includes temporal fine-grainedness and spatial fine-grainedness. This invention does not impose limitations on temporal or spatial fine-grainedness. Temporal granularity can be a day or other time intervals; spatial granularity can be a street, a county, or even a province. After determining the spatial and temporal granularity, population flow data should also be aggregated to the corresponding spatiotemporal granularity.

[0056] In some embodiments, population movement data for the target area includes initial abnormal population movement data during the peak of a disaster and average normal population movement data during non-disaster periods.

[0057] It should be noted that this invention focuses on the daily changes in population flow; therefore, the population flow map presented here... Representing the The population flow map for each day is shown, where each point represents a region, the features on the point represent population flow within that region, and the lines connecting the points, i.e., the edges, represent population flow between regions. The relevant concepts of population flow attribute values ​​are defined as follows:

[0058] Definition 1 (Population movement within an area): The first The area of ​​the sky The intra-population flow available represents the total number of people who only moved within the region on this day.

[0059] Definition 2 (Inter-population flow): The number of people who moved from region to region on day . The inter-population flow available represents the total number of people who came from region to region on this day.

[0060] It is observed that the urban population flow will decrease to the minimum value on the day of disaster and slowly recover to normal in the following days, so, unlike the traditional dynamic graph prediction method based on a sequence of past graphs to predict a sequence of future graphs, the present application focuses on starting from the abnormal population flow graph when the disaster comes, and according to the normal population flow graph, to predict the population flow graph during the recovery period. The normal population flow graph and the prediction problem of the post-disaster population flow recovery dynamics are defined as follows:

[0061] Definition 3 (Normal population flow graph): Given a sequence of normal population flow graphs during a period of non-disaster , the normal population flow graph is defined as , specifically, the attribute value of each point, i.e. the normal intra-population flow in the region, is defined as , and the attribute value of each edge, i.e. the normal inter-population flow between regions, is defined as .

[0062] Definition 4 (Prediction problem of post-disaster population flow recovery dynamics): Given the abnormal population flow graph on the first day of disaster and the normal population flow graph , predict the population flow graph in the following days of post-disaster recovery .

[0063] Step 120: Obtain an input population flow dynamic graph according to the population flow data of the target region; wherein the input population flow dynamic graph at least includes an initial abnormal population flow graph of the target region during the peak period of disaster and an average normal population flow graph of the target region during a period of non-disaster.

[0064] In step 120, the population flow data of the target region is converted into an input population flow dynamic graph. Based on the description of step 110, the population flow data of the target region includes initial abnormal population flow data in the disaster peak period and average normal population flow data in the non-disaster period, and accordingly, the input population flow dynamic graph includes an initial abnormal population flow graph of the target region in the disaster peak period and an average normal population flow graph of the target region in the non-disaster period.

[0065] Step 130: inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training based on a coupled dynamic graph neural ordinary differential equation using population flow data of a sample region.

[0066] In step 130, the obtained input population flow dynamic graph is inputted into the population flow prediction model to output the population flow prediction result of the target region. The population flow prediction model mentioned in the present application is a coupled dynamic graph neural ordinary differential equation network (CDGON), as shown in Figure 2 As shown in the figure, the present application uses a physically inspired machine learning paradigm to construct a neural ODE function guided by a spatiotemporal decay model (ST Decay Model), which can effectively model the diversified and highly volatile urban mobility in different regions and jointly consider the urban mobility before and after the disaster, and at the same time, introduce another set of coupled neural ordinary differential equation functions to describe the dynamics of inter-regional population flow as edges in the formed dynamic graph, to represent the complex post-disaster urban mobility. The two sets of neural ODE functions jointly model the joint evolution process of nodes and edges in the dynamic graph of intra-regional and inter-regional population flow. In terms of specific calculation process, the network model first encodes the normal and abnormal intra-regional population flow and the abnormal inter-regional population flow, then considers the joint evolution of points and edges guided by the normal intra-regional population flow, solves the neural ordinary differential equation inspired by the ST Decay Model to obtain the population flow encoding sequence at multiple time points in the post-disaster recovery process, and finally decodes to realize the prediction of the post-disaster population flow recovery process.

[0067] It should be noted that the purpose of the spatiotemporal decay model (ST Decay Model) is to fit the recovery curve of population flow after the peak of large-scale natural disasters such as typhoons and cold waves, and its formula can be expressed as follows:

[0068] ,

[0069] wherein, denotes the normal population flow indicator, is a function of time, referred to as the temporal decay term, and denotes the severity of the disaster at the th day in the th region, denotes the weighted sum of the severity of the disaster in the nearby regions of the th region. Since the severity of the disaster tends to decrease during the recovery period, the spatio-temporal decay model indicates that the abnormal population flow during the disaster will eventually tend to be normal as time elapses. How to capture the change process of the decay factor is crucial.

[0070] The population flow prediction model is further described below. In some embodiments, the population flow prediction model is obtained by training based on a coupled dynamic graph neural ordinary differential equation using population flow data of a sample region, specifically including:

[0071] Step 210: obtaining population flow data of a sample region;

[0072] Step 220: obtaining an initial abnormal population flow graph of a sample region at a disaster peak period, an average normal population flow graph of the sample region during a non-disaster period, and a real abnormal population flow graph of the sample region at a recovery stage according to the population flow data of the sample region;

[0073] Step 230: constructing a basic prediction model based on a coupled dynamic graph neural ordinary differential equation;

[0074] Step 240: training the basic prediction model by taking the initial abnormal population flow graph of the sample region at the disaster peak period and the average normal population flow graph of the sample region during the non-disaster period as inputs and taking the real abnormal population flow graph of the sample region at the recovery stage as a label until the model converges, to obtain a population flow prediction model.

[0075] Specifically, in step 210, population flow data of a sample region is obtained. It needs to be noted that the population flow data of the sample region obtained is for reference. The population flow data of the sample region includes initial abnormal population flow data of the sample region at a disaster peak period, average normal population flow data of the sample region during a non-disaster period, and real abnormal population flow data of the sample region at a recovery stage.

[0076] It can be understood that the method provided by the present application can use the population flow data of a plurality of cities in a certain region in the early recovery stage to train the model, so as to extrapolate the population flow in the later recovery stage of the same batch of cities (that is, the target region and the sample region are the same), as shown in FIG. 2a; or use the population flow data of a plurality of cities in a certain region in the whole recovery stage to train the model, and use the trained model to predict the post-disaster population flow recovery in another batch of cities in another region (that is, the target region and the sample region are different), which is equivalent to using the experience of other cities in dealing with disasters to predict the recovery process of the target city, as shown in FIG. 2b. Figure 3 Figure 4

[0077] On this basis, in some embodiments, when the target region and the sample region are the same, the real abnormal population flow graph of the sample region in the recovery stage is the real abnormal population flow graph in the early recovery stage of the sample region.

[0078] When the target region and the sample region are different, the real abnormal population flow graph of the sample region in the recovery stage is the real abnormal population flow graph in the whole recovery stage of the sample region.

[0079] Specifically, the population flow data of a plurality of cities in a certain region in the early recovery stage is used to extrapolate the population flow in the later recovery stage of the same batch of cities, that is, the target region and the sample region are the same, and the real abnormal population flow graph of the sample region in the recovery stage can be only the population flow graph obtained from the real abnormal population flow data in the early recovery stage.

[0080] With the experience of population flow recovery mode of other cities or regions after encountering the same type of disaster to help predict the whole process recovery cycle population flow of a part of cities or regions after encountering a certain type of disaster, that is, the target region and the sample region are different, the 3 parts of data to be collected are similar to but not completely consistent with the case where the target region and the sample region are the same, the difference lies in that not only the real abnormal population flow data in the early recovery stage is needed, but also the population flow graph obtained from the real abnormal population flow data in the whole recovery stage is needed.

[0081] In step 220, the collected data is divided into a third part: the initial abnormal population flow data of the sample region in the disaster peak period, the average normal population flow data in the non-disaster period, and the real abnormal population flow data in the recovery stage, which is converted into a population flow graph to obtain the initial abnormal population flow graph of the sample region in the disaster peak period, the average normal population flow graph of the sample region in the non-disaster period, and the real abnormal population flow graph of the sample region in the recovery stage.

[0082] ​​In step 230, the construction of the base prediction model based on the coupled dynamic graph neural ordinary differential equation is performed. In this step, based on the ST Decay Model, the present application proposes a coupled dynamic graph neural ordinary differential equation network CDGON, which can start from the initial abnormal population flow after the disaster, combine the normal population flow, and realize the prediction of the recovery dynamics of the post-disaster population flow. The present application will illustrate the base prediction model, i.e., the coupled dynamic graph neural ordinary differential equation network CDGON, in the subsequent embodiments.

[0083] In step 240, the initial abnormal population flow graph of the sample area disaster peak period and the average normal population flow graph of the sample area non-disaster period constructed in step 220 are input into the model, and the real abnormal population flow graph of the sample area recovery stage is used as a label to supervise the update of the model parameters. After multiple iterations of updating the parameters to make the model converge, the population flow prediction model can be obtained.

[0084] Further, in some embodiments, during the training process, since the prediction target of the present application is to realize the accurate prediction of the post-disaster population flow recovery process, in this process, considering that the numerical value of the inter-regional population flow is smaller than that of the intra-regional population flow, in order to take into account the intra-regional and inter-regional population flow and realize accurate prediction of both, the loss function of the present application uses a hyperparameter to amplify the weight of the edge prediction, so as to ensure that the model does not discard the prediction of the inter-regional population flow, which is as follows

[0085]

[0086] wherein, represents the number of days required for post-disaster recovery. In each training, by inputting the initial abnormal population flow within and between regions and the normal population flow within the region, the population flow prediction value in the future few days (i.e., the recovery period) can be obtained, the loss is calculated between the real value, and finally the gradient is calculated, the back propagation is carried out, and the model parameters are updated.

[0087] In use, for the same city extension prediction, i.e., when the target region and the sample region are the same, the initial abnormal population flow graph of the sample area disaster peak period and the average normal population flow graph of the sample area non-disaster period are continued to be used to input the model, the integration step number of the differential equation is extended, and the extrapolated population flow prediction value in the later recovery stage is generated. For cross-city migration prediction, the initial abnormal population flow graph of the disaster peak period of the target region and the average normal population flow graph of the target region during the non-disaster period are required to input the model, and the model will generate the population flow data of the corresponding length of the recovery stage according to the set parameters.

[0088] ​The following describes a basic prediction model. In some embodiments, the basic prediction model includes a decoding module and an encoding module;

[0089] Correspondingly, the initial abnormal population flow graph of the sample area in the disaster peak period and the average normal population flow graph of the sample area in the non-disaster period are taken as inputs, and the real abnormal population flow graph of the sample area in the recovery stage is taken as a label to train the basic prediction model until the model converges, to obtain a population flow prediction model, specifically including:

[0090] According to the initial abnormal population flow graph of the sample area in the disaster peak period and the average normal population flow graph of the sample area in the non-disaster period, the encoding module is used to encode the inter-regional population flow and the intra-regional population flow, to obtain an encoding result;

[0091] The decoding module is used to decode the encoding result, to obtain a population flow prediction value at multiple time points in the post-disaster recovery process;

[0092] According to the population flow prediction value and the real abnormal population flow graph of the sample area in the recovery stage, a pre-set loss function is used for back propagation, to update the model parameters until the model converges, to obtain a population flow prediction model.

[0093] Specifically, in the specific construction process, the same encoder needs to be used for the normal and abnormal population flow graphs, so that they can be aligned in a high-dimensional space. After obtaining the high-dimensional embedding of the initial abnormal population flow and the normal population flow, a neural ordinary differential equation inspired by the ST Decay Model is solved to obtain the population flow at multiple future time points, and finally a same decoder is used to obtain the final prediction result.

[0094] The encoding and decoding process of the relevant values in the population flow graph is as shown in Figure 2 The neural ordinary differential equation inspired by the ST Decay Model, i.e., the encoding module, is described in detail, which mainly includes two parts: a physics-informed dynamic module (PIDM) and a node-edge interaction module (NEIM).

[0095] In the population flow graph defined before, there are both edge features and node features, which represent the inter-regional and intra-regional population flows, respectively. However, specifically, due to the large volatility and sparsity of the inter-regional population flow, for example, before the disaster, some people went from to to The land, and after the disaster, will appear Heading to The population of the area has increased dramatically, whereas normally there would be almost no one there. Heading to However, the population flow data within a region is more stable and denser than the population flow data between regions, and has a more obvious recovery process. Therefore, this invention only applies PIDM to the population flow within a region.

[0096] Some existing research has used the negative logarithmic exponential function. To represent the time decay term in the ST Decay Model However, this invention argues that the spatial decay term also decays over time; therefore, it is assumed here that the entire spatiotemporal decay factor... Substituting the values ​​and taking the derivative, we get the following formula:

[0097] ,

[0098] in, The attenuation coefficient is represented by this symbol. Based on this, the PIDM of this invention is designed as follows:

[0099] ,

[0100] in, express The encoded population movement code within the non-normal area is denoted by the following dimension: Similarly Then it is After and Population flow coding within the normal area following the same encoder.

[0101] Since the patterns and rates of change for each component of the population mobility code may differ, the dimension is also [missing information]. attenuation coefficient vector It was used to replace the original At the same time, in order to avoid the occurrence of The square term in The resulting numerical explosion, here we use a pre-defined basis function The original population flow coding vector item Convert to a scalar. Finally. This represents a transformation matrix, and an additional one is added. Activation function.

[0102] On this basis, it is obviously insufficient to only model the population flow within the region, although the population flow between regions has greater volatility and more complex characteristics, and it is obviously difficult to use traditional methods to develop modeling, while data-driven methods represented by deep learning naturally have an advantage in solving such modeling. Therefore, in addition to designing ordinary differential equations for population flow within the region, ordinary differential equations for population flow between regions also need to be designed. Before this, considering the topological structure of the population flow graph in actual situations is dynamic, therefore, the possible edge values at each time can be calculated from its encoding, if the edge value is less than , it means that this edge does not exist, as follows:

[0103] ,

[0104] ,

[0105]

[0106] wherein, is used to indicate whether the edge exists, note that at this time does not represent the predicted population flow value from to , the ordinary differential equation of the edge value, i.e. the design of the neural ordinary differential equation of the edge in NEIM, is as follows:

[0107] ,

[0108] wherein, represents the vector splicing operation, represents the abnormal inter-regional population flow encoding after encoding.

[0109] It can be found that through the edge self-evolution function and the link point information aggregation function , this equation can learn the change pattern from as many factors as possible that may affect the changes of inter-regional population flow.

[0110] Since the population flow between regions will have different degrees of impact on the population flow within the region, for example, a large number of people leaving a city to escape will return to the city after the disaster, and the number of people returning determines the speed of recovery of the population flow within the city.

[0111] Therefore, the ordinary differential equation for the population flow within the region adds an interactive function with the population flow between regions on the basis of the original PIDM, that is, the encoding module can be expressed as follows:

[0112] ,

[0113] in, and Similarly, it also has dimensions of vector coefficients, Indicates with region Regions where inter-regional population movement exists. Similar to ordinary differential equations, This also consists of two parts: self-evolution and edge information aggregation. Specifically, here is given... and The calculation method is as follows:

[0114] ,

[0115] ,

[0116] in, By A single-layer GCN with convolution operators, and a single-layer multilayer perceptron. The method is obtained by using the maximum absolute value normalization method, each The values ​​of the components are restricted to This allows the present invention to capture unexpected attenuations during the recovery of population movement. This represents an MLP (Multilayer Perceptron). express Activation function. express The degree matrix.

[0117] Furthermore, by solving the ordinary differential equations for population movement codes within and between regions, the codes for all population movements within and between regions during the post-disaster recovery process are obtained, yielding the coding results. The integration process is as follows:

[0118]

[0119]

[0120] By performing integration at equal intervals, the population movement code for the recovery process after a disaster is obtained, yielding the following coding results: as well as The coding results include codes for population movement within non-normal areas and codes for population movement between non-normal areas.

[0121] Finally, through the decoding process of the following decoding module, the final population flow prediction value can be obtained. It is worth noting that... are also used in the decoding process of edges to guarantee consistency of the final topology and integration process as follows, and , respectively represent the predicted values of the population flow within the th region and the population flow between the th region and the th region on the th day of the recovery period. The decoding module comprises:

[0122] ,

[0123] ,

[0124]

[0125] wherein, and respectively represent the predicted values of the population flow within the th region and the population flow between the th region and the th region on the th day of the recovery period, represents the encoded population flow within the abnormal region, represents the decoding function represented by the MLP obtained by stacking the two linear layers and the two RELU activation functions. The post-disaster population flow prediction method provided by the present application comprises the following steps: obtaining population flow data of a target region according to a predetermined spatio-temporal granularity; obtaining an input population flow dynamic graph according to the population flow data of the target region; wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of the target region at a disaster peak period and an average normal population flow graph of the target region during a non-disaster period; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training a sample region population flow data based on a coupled dynamic graph neural ordinary differential equation. The present application adopts a coupled dynamic graph neural ordinary differential equation to construct a population flow prediction model, which can effectively model the diversified and highly volatile urban mobility in different regions, and jointly consider the urban mobility before and after the disaster, solve the problem that the prior art relies on an oversimplified assumption, so as to fail to effectively capture the influence of the population flow between regions, and it is difficult to use the flow change pattern under normal circumstances to predict the recovery process of the abnormal flow after the disaster, resulting in low accuracy, and realize high-accuracy post-disaster population flow prediction.

[0126]

[0127] ​The post-disaster population flow prediction device provided by the present application is described below, and the post-disaster population flow prediction device described below can be correspondingly referred to the post-disaster population flow prediction method described above. As shown in the following description, the device comprises: Figure 5

[0128] A data unit 510 is configured to obtain population flow data of a target region according to a predetermined spatio-temporal granularity;

[0129] A conversion unit 520 is configured to obtain an input population flow dynamic graph according to the population flow data of the target region; wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of a disaster peak period of the target region and an average normal population flow graph of a non-disaster period of the target region;

[0130] A prediction unit 530 is configured to input the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result;

[0131] The population flow prediction model is obtained by training based on a coupled dynamic graph neural ordinary differential equation and using population flow data of a sample region.

[0132] According to the post-disaster population flow prediction device provided by the present application, the population flow prediction model is obtained by training based on a coupled dynamic graph neural ordinary differential equation and using population flow data of a sample region, and specifically comprises:

[0133] Obtaining population flow data of a sample region;

[0134] Obtaining an initial abnormal population flow graph of a disaster peak period of the sample region, an average normal population flow graph of a non-disaster period of the sample region, and a real abnormal population flow graph of a recovery stage of the sample region according to the population flow data of the sample region;

[0135] Constructing a basic prediction model based on a coupled dynamic graph neural ordinary differential equation;

[0136] Training the basic prediction model by taking the initial abnormal population flow graph of the disaster peak period of the sample region and the average normal population flow graph of the non-disaster period of the sample region as inputs and taking the real abnormal population flow graph of the recovery stage of the sample region as a label until the model converges, to obtain a population flow prediction model.

[0137] According to the post-disaster population flow prediction device provided by the present application, the basic prediction model comprises a decoding module and an encoding module;

[0138] ​Correspondingly, the initial abnormal population flow graph of the sample area in the disaster peak period and the average normal population flow graph of the sample area in the non-disaster period are taken as inputs, and the real abnormal population flow graph of the sample area in the recovery stage is taken as a label to train the basic prediction model until the model converges, to obtain a population flow prediction model, specifically including:

[0139] According to the initial abnormal population flow graph of the sample area in the disaster peak period and the average normal population flow graph of the sample area in the non-disaster period, interval population flow and intra-district population flow are encoded based on the encoding module to obtain an encoding result;

[0140] Based on the decoding module, the encoding result is decoded to obtain a population flow prediction value at multiple time points in the post-disaster recovery process;

[0141] According to the population flow prediction value and the real abnormal population flow graph of the sample area in the recovery stage, a pre-set loss function is used for back propagation to update model parameters until the model converges, to obtain a population flow prediction model.

[0142] According to the population flow prediction model provided by the application, the encoding module comprises:

[0143] ,

[0144] ,

[0145] ,

[0146] Among them, is a decay coefficient vector with a dimension of , is a function based on , represents the encoded abnormal intra-district population flow, and the dimension is , represents the population flow in the district on the th day, is the normal intra-district population flow after the same encoding mode as and , represents the attribute value of each point of the normal population flow graph, represents a transformation matrix, represents an activation function, is used to represent whether an edge exists, represents a vector coefficient with a dimension of , represents a region regions with inter-regional population flow, Softmax represents a Softmax activation function, represents an MLP, represents encoded abnormal inter-regional population flow encoding, represents the first day of the region to the region population flow, represents a multi-layer perception, represents the convolution operator of the GCN, represents the degree matrix of

[0147] According to the post-disaster population flow prediction device provided by the application, the decoding module comprises:

[0148] ,

[0149] ,

[0150] ,

[0151] wherein, and respectively represent the population flow in the first region and the population flow from the region to the region in the recovery period on the first day, represents encoded abnormal intra-regional population flow encoding, represents the decoding function represented by the MLP obtained by stacking two linear layers and two RELU activation functions.

[0152] According to the post-disaster population flow prediction device provided by the application, in the case that the target region and the sample region are the same, the real abnormal population flow graph of the sample region in the recovery stage is the real abnormal population flow graph of the sample region in the early recovery stage.

[0153] In the case that the target region and the sample region are different, the real abnormal population flow graph of the sample region in the recovery stage is the real abnormal population flow graph of the sample region in the full recovery stage.

[0154] The application provides a post-disaster population flow prediction device, which comprises the following steps: obtaining population flow data of a target area according to a predetermined spatiotemporal granularity; obtaining an input population flow dynamic graph according to the population flow data of the target area; wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of a disaster peak period of the target area and an average normal population flow graph of a non-disaster period of the target area; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training a sample area population flow data based on a coupled dynamic graph neural ordinary differential equation.

[0155] Figure 6 An example of an entity structure diagram of an electronic device is shown in Figure 6 As shown, the electronic device can include a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other through the communications bus 640. The processor 610 can invoke the logical instructions in the memory 630 to execute a post-disaster population flow prediction method, which comprises the following steps: obtaining population flow data of a target area according to a predetermined spatiotemporal granularity; obtaining an input population flow dynamic graph according to the population flow data of the target area; wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of a disaster peak period of the target area and an average normal population flow graph of a non-disaster period of the target area; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training a sample area population flow data based on a coupled dynamic graph neural ordinary differential equation.

[0156] Further, the logic instructions in the memory 630 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0157] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the post-disaster population flow prediction method provided by the above-mentioned methods. The method comprises: obtaining population flow data of a target region according to a predetermined spatio-temporal granularity; obtaining an input population flow dynamic graph according to the population flow data of the target region; wherein the input population flow dynamic graph at least includes an initial abnormal population flow graph of a disaster peak period of the target region and an average normal population flow graph of a non-disaster period of the target region; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training based on a coupled dynamic graph neural ordinary differential equation using population flow data of a sample region.

[0158] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the post-disaster population flow prediction method provided by the above-mentioned methods. The method comprises: obtaining population flow data of a target region according to a predetermined spatio-temporal granularity; obtaining an input population flow dynamic graph according to the population flow data of the target region; wherein the input population flow dynamic graph at least includes an initial abnormal population flow graph of a disaster peak period of the target region and an average normal population flow graph of a non-disaster period of the target region; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training based on a coupled dynamic graph neural ordinary differential equation using population flow data of a sample region.

[0159] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0160] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0161] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A post-disaster population flow prediction method, characterized by, The method comprises the following steps: acquiring population flow data of a target region according to a predetermined spatiotemporal granularity; obtaining an input population flow dynamic graph according to the population flow data of the target region; wherein the input population flow dynamic graph at least comprises an initial abnormal population flow graph of a disaster peak period of the target region and an average normal population flow graph of a non-disaster period of the target region; inputting the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training a sample region population flow data based on a coupled dynamic graph neural ordinary differential equation; training the population flow prediction model based on a coupled dynamic graph neural ordinary differential equation using sample region population flow data, specifically comprising: acquiring sample region population flow data; obtaining an initial abnormal population flow graph of a disaster peak period of the sample region, an average normal population flow graph of a non-disaster period of the sample region, and a real abnormal population flow graph of a recovery stage of the sample region according to the sample region population flow data; constructing a basic prediction model based on a coupled dynamic graph neural ordinary differential equation; training the basic prediction model by taking the initial abnormal population flow graph of the disaster peak period of the sample region and the average normal population flow graph of the non-disaster period of the sample region as inputs and taking the real abnormal population flow graph of the recovery stage of the sample region as a label until the model converges, to obtain a population flow prediction model.

2. The post-disaster population flow prediction method according to claim 1, wherein The basic prediction model comprises a decoding module and an encoding module; correspondingly, training the basic prediction model by taking the initial abnormal population flow graph of the disaster peak period of the sample region and the average normal population flow graph of the non-disaster period of the sample region as inputs and taking the real abnormal population flow graph of the recovery stage of the sample region as a label until the model converges, to obtain a population flow prediction model, specifically comprising: encoding the interval population flow and the intra-district population flow based on the encoding module according to the initial abnormal population flow graph of the disaster peak period of the sample region and the average normal population flow graph of the non-disaster period of the sample region to obtain an encoding result; decoding the encoding result based on the decoding module to obtain population flow prediction values at multiple time points in the post-disaster recovery process; updating the model parameters by using a pre-set loss function to perform back propagation according to the population flow prediction values and the real abnormal population flow graph of the recovery stage of the sample region until the model converges, to obtain a population flow prediction model.

3. The post-disaster population flow prediction method according to claim 2, characterized by, The encoding module comprises: ， ， ， in, The dimension is The attenuation coefficient vector, Based on The function, express The coded population movement code for non-normal areas has the following dimensions: , Indicates the first The area of ​​the sky Population movement within the country, Then it is After and Population mobility coding within the normal area after using the same coding method This represents the attribute value of each point in the normal population flow map. Represents the transformation matrix. This represents the activation function. Used to indicate whether an edge exists. The dimension is vector coefficients, Indicates with region For regions where there is inter-regional population movement, Softmax represents the Softmax activation function. Represents an MLP, express Encoded non-normal inter-regional population movement. Indicates the first The area of ​​the sky To the area Population movement, This represents a multilayer perceptron. This represents the convolution operator of GCN. express The degree matrix.

4. The post-disaster population flow prediction method according to claim 2, wherein The decoding module comprises: ， ， ， where, and denote the population flow within the th region and the population flow between the th region and the th region at the th day of the recovery period, respectively, denote the encoded abnormal region population flow encoding, denote the decoding function represented by the MLP composed of two linear layers and two RELU activation functions. 5.The post-disaster population flow prediction method of claim 1, wherein, in the case that the target region and the sample region are the same, the real abnormal population flow graph of the recovery stage of the sample region is a real abnormal population flow graph of an early recovery stage of the sample region; in the case that the target region and the sample region are different, the real abnormal population flow graph of the recovery stage of the sample region is a real abnormal population flow graph of a full recovery stage of the sample region.

6. A post-disaster population flow prediction device characterized by comprising: The method comprises the following steps: a data unit is configured to acquire population flow data of a target region according to a predetermined spatiotemporal granularity; a conversion unit configured to obtain an input population flow dynamic graph according to population flow data of the target region; wherein the input population flow dynamic graph at least includes an initial abnormal population flow graph of the target region in a disaster peak period and an average normal population flow graph of the target region in a non-disaster period; a prediction unit configured to input the input population flow dynamic graph into a pre-trained population flow prediction model to obtain a population flow prediction result; wherein the population flow prediction model is obtained by training a sample region population flow data based on a coupled dynamic graph neural ordinary differential equation; the population flow prediction model is obtained by training a sample region population flow data based on a coupled dynamic graph neural ordinary differential equation, and specifically includes: obtaining population flow data of a sample region; obtaining an initial abnormal population flow graph of the sample region in a disaster peak period, an average normal population flow graph of the sample region in a non-disaster period, and a real abnormal population flow graph of the sample region in a recovery stage according to the population flow data of the sample region; constructing a basic prediction model based on a coupled dynamic graph neural ordinary differential equation; training the basic prediction model by taking the initial abnormal population flow graph of the sample region in the disaster peak period and the average normal population flow graph of the sample region in the non-disaster period as inputs and taking the real abnormal population flow graph of the sample region in the recovery stage as a label until the model converges to obtain a population flow prediction model.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the post-disaster population flow prediction method of any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the post-disaster population flow prediction method of any one of claims 1 to 5.

9. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to realize the post-disaster population flow prediction method of any one of claims 1 to 5.

Images