Dynamic updating method for building group function loss forecast fused with social media data

Abstract

The invention discloses a building group function loss forecast dynamic updating method fusing social media data, and the method comprises the steps: constructing a disaster situation real-time monitoring module, collecting the social media data related to a disaster in real time, and extracting the local waterlogging observation information in a research region; simulating space-time evolution of building group function state probability distribution, dividing emergency response districts, and generating district function loss data sets under various rainfall scenes; constructing and training a situation awareness model, and inferring district function loss of all unobserved emergency response districts according to local waterlogging observation information; and constructing and training a space-time prediction model, replacing predicted values of the space-time prediction model at corresponding moments by using function losses of all emergency response areas obtained by observation and deduction at different moments, and outputting updated prediction results of a plurality of time steps in the future through iterative prediction. According to the invention, the problem that prediction and dynamic updating of the function loss of the building group by fusing the social media data are not realized in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of disaster management and emergency response, and in particular to a building group function loss prediction dynamic updating method fusing social media data. BACKGROUND

[0002] Flood is one of the most destructive natural disasters in the world. In recent years, the frequent occurrence of extreme flood disasters has further highlighted the urgency of flood risk management.

[0003] Although social media data has significant advantages in terms of real-time and cost, the existing technology still has obvious deficiencies in applying social media data to the dynamic updating of building group function loss prediction. Specifically: 1) Currently, the application of social media data in the flood emergency response stage is mainly focused on the disaster site; the research facing the building and social end focuses on building damage identification, rescue request extraction, economic loss estimation and sentiment analysis, but mostly stays in static evaluation, and there is no mechanism to integrate real-time observation information into future prediction update. 2) The existing few prediction update researches that fuse social media data only focus on the prediction of the flooded scene, and have not been extended to the prediction update of building group function loss, and there is a problem of insufficient calculation efficiency. 3) Due to the characteristics of social media data such as unstructured, multi-modal, spatio-temporal distribution sparse and uneven, the existing prediction dynamic updating methods based on three-dimensional variation, online learning and incremental learning are difficult to be directly applied. Therefore, at present, there is no method system that can fuse disaster social media data and support the dynamic updating of building group function loss prediction, which seriously restricts the practical value of such data in emergency response and the accuracy of the prediction results.

[0004] At present, in the aspect of fusing social media data to realize the dynamic updating of building group function loss prediction, there are still many challenges: 1) The spatio-temporal distribution of social media data is highly sparse, and how to realize global inference based on local observation is a problem to be solved. 2) The existing prediction update method fusing social media data is mainly used to update the flooded scene, relies on the probability inference of model parameter uncertainty, and needs to be coupled with a two-dimensional hydrodynamic model with high calculation cost, and there is still a lack of method for rapid prediction update of building group function loss. SUMMARY

[0005] In order to overcome the deficiencies of the above technology, the present application provides a building group function loss prediction dynamic updating method fusing social media data, which can solve the problem that the existing technology has not realized the dynamic updating of building group function loss prediction by fusing social media data, and provides an accurate, efficient and low-cost solution for the dynamic updating of building group function loss prediction in urban waterlogging disasters.

[0006] Term explanation:

[0007] 1、DEM: Digital Elevation Model, Digital Elevation Model.

[0008] 2、SWMM: Storm Water Management Model, Storm Water Management Model.

[0009] 3、WCA2D: Water Cell Automata 2D, two-dimensional water dynamic cellular automaton model.

[0010] 4、ERZs: Emergency Response Zones, Emergency Response Zones.

[0011] 5、ZFL: Zone-level functionality loss, Zone-level functionality loss.

[0012] 6、ReLU: Rectified Linear Unit, Rectified Linear Unit.

[0013] 7、LSTM: Long Short-Term Memory, Long Short-Term Memory network.

[0014] The technical solution adopted by the present application to overcome its technical problems is:

[0015] A building group function loss prediction dynamic updating method fusing social media data, comprising:

[0016] S1, constructing a disaster situation real-time monitoring module, used for collecting and processing disaster-related social media data in real time, and extracting local waterlogging observation information in the study area;

[0017] S2, preparing training data: dividing building types into four categories of high-rise, multi-story, villa and stacked villa, simulating the spatio-temporal evolution of building group function state probability distribution, dividing emergency response zones, and generating zone function loss data sets under multiple rainfall scenarios;

[0018] S3, constructing and training a situation awareness model, obtaining the observed function loss value of the observed emergency response zone at the corresponding time according to the local waterlogging observation information, and inferring the function loss of all unobserved emergency response zones;

[0019] S4, constructing and training a spatio-temporal prediction model, inputting historical observed rainfall values, future single time step predicted rainfall values and historical emergency response zone function loss values, outputting future single time step emergency response zone function loss prediction values, and obtaining prediction results for multiple time steps in the future through iterative prediction;

[0020] S5, in actual application, the local waterlogging observation information at each moment is obtained through the disaster situation real-time monitoring module, the observed emergency response area and its observation function loss value are determined, the unobserved emergency response area function loss is inferred by using the situation awareness model, and all emergency response area function losses obtained by observation and inference at different moments are used to replace the prediction values of the spatio-temporal prediction model at the corresponding moment, and the updated prediction results of multiple time steps in the future are output.

[0021] Further, in step S2, the training data includes the training data required by the situation awareness model and the spatio-temporal prediction model. In order to obtain the training data, the required input data includes meteorological variables, flood disaster environment variables and building attribute related variables. The meteorological variables include 24-hour rainfall data measured by the rainfall station, and rainfall sequences with 24-hour cumulative rainfall exceeding a preset value are selected as the analysis objects. The flood disaster environment variables include DEM data, slope direction data, curvature data, slope data, terrain humidity index data, distance to river center line and distance to pipe network outlet. The building attribute related variables include the outline, building layers, floor area and building type of each building in the building group.

[0022] Further, step S2 specifically includes:

[0023] S21, based on the meteorological variables and flood disaster environment variables of the research area, the SWMM and WCA2D models are coupled to simulate urban waterlogging inundation, and the waterlogging evolution process is obtained;

[0024] S22, a building function state probability prediction method is used to simulate the spatio-temporal evolution of the building group in different function state probability distributions under multiple rainfall events. The building function state includes I, II, III, V and VII, and the larger the value is, the more serious the building is affected.

[0025] S23, the building with the probability of being in function state I at any time under any rainfall sequence is defined as a potential affected building, and the K-means spatial clustering method is used to divide all potential affected buildings into multiple emergency response areas. The area function loss is defined as the average value of the sum of the probabilities of all potential affected buildings in the area being in function states III-VII, and thus the emergency response area function loss data set at different times under multiple rainfall scenarios is constructed.

[0026] S24, based on the rainfall sequence, the function loss time history curve of each emergency response area under each single rainfall is calculated, and the Pearson correlation coefficient matrix between areas is calculated accordingly, and the average Pearson correlation coefficient matrix under all rainfall events is used as a measure of function loss correlation.

[0027] Further, in step S3, the situational awareness model comprises two layers of multi-head graph attention layers and a ReLU activation function connected after each layer of multi-head graph attention layers, each layer of multi-head graph attention layer is configured with a plurality of attention heads and a multi-dimensional hidden layer, and an output layer of the situational awareness model is a graph attention layer with a dimension of 1.

[0028] Further, in step S3, training the situational awareness model specifically comprises:

[0029] (1) The input variable of the situational awareness model is , wherein represents the total number of emergency response zones, 1 represents the dimension of the function loss value, represents the dimension of the node-related attribute, the node-related attribute includes the proportion of four building types, represents the dimension of the environmental variable, the environmental variable includes DEM, curvature, distance to the pipe network outlet, and distance to the river center line; the observed emergency response zone is defined as the zone in which more than a set proportion of potential affected buildings are located within the observed waterlogging coverage, and the observed function loss value is calculated according to the observed water depth; for each unobserved emergency response zone, the prior function loss value is calculated based on all observed function loss values by inverse distance weighting method; the output of the situational awareness model is , which represents the predicted function loss value of all emergency response zones at time t, and only the prediction result for the unobserved emergency response zone is finally retained;

[0030] (2) All rainfall sequences contained in the meteorological variable are divided into training set, validation set and test set according to a preset proportion; from the function loss value spatial distribution of each time step generated by the training set rainfall sequence in step S2, data is extracted in turn, and the corresponding node-related attribute variable and environmental variable are spliced to construct a complete node feature vector; the situational awareness model is independently trained under a plurality of preset observation proportions, and the training loss is calculated.

[0031] (3) The situational awareness model trained under a plurality of preset observation proportions is verified and tested, and the validation loss and test loss are calculated respectively.

[0032] Further, in step (2), when training the situational awareness model, a random mask strategy is introduced to simulate the actual situation that only part of the emergency response zones can be observed, specifically including:

[0033] (1) For the spatial distribution results of the function loss values of all emergency response subarea nodes at each time step generated by the rainfall sequence contained in the training set through step S2, a preset observation proportion of emergency response subarea nodes is randomly selected, the function loss values obtained by simulation are retained, and the remaining nodes are subjected to mask processing, that is, the function loss values of the remaining nodes are shielded;

[0034] (2) The prior function loss values of all masked nodes are calculated according to step (1) to fill the feature vectors of the corresponding masked subarea nodes, and the function loss values of all emergency response subarea nodes are spliced with the node-related attribute variables and environmental variables to form the input of the situation awareness model;

[0035] (3) The masked nodes are predicted through the situation awareness model, and after the prediction is completed, only the prediction results of the masked subarea nodes and the simulation values are calculated by the mean square error loss function.

[0036] Further, in step S4, the spatio-temporal prediction model includes a rainfall feature extraction module, a spatio-temporal feature capturing module, and an emergency response subarea function loss prediction module; the rainfall feature extraction module is an input module of the spatio-temporal prediction model, used for processing rainfall data containing multiple historical time steps and 1 future time step, and splicing the processed rainfall data with the emergency response subarea function loss tensor corresponding to the historical time step along the time dimension; the spatio-temporal feature capturing module is used for processing time series data with a graph structure so that each graph node can fuse features from the corresponding spatio-temporal domain; the emergency response subarea function loss prediction module is used for outputting emergency response subarea function loss prediction results with a time step number consistent with the input future rainfall time step number.

[0037] Further, the rainfall feature extraction module in the spatio-temporal prediction model includes two types of input data, respectively containing a rainfall data tensor containing historical time steps and future F =1 time steps , and an emergency response subarea function loss tensor corresponding to the historical hours , wherein is an integer and , represents the batch size used in the training of the spatio-temporal prediction model, represents that the input channel size is 1, corresponding to the feature dimension of the emergency response subarea function loss, and each time step contains function loss value information of emergency response subarea nodes; the rainfall feature extraction module includes an LSTM module and a fully connected layer connected after the LSTM module, and the rainfall data tensor is input into the LSTM module and processed by the fully connected layer to obtain a tensor , the , and along the time dimension, forming the final input tensor .

[0038] Further, the spatio-temporal feature capturing module includes two-level serial spatio-temporal convolution modules and an output layer, the two-level serial spatio-temporal convolution modules are used to extract deep spatio-temporal features in the input sequence step by step, and the output layer is used for feature compression and output mapping, and outputs the functional loss prediction result of the future single time step; each spatio-temporal convolution module includes two gated temporal convolution layers and a spatial graph convolution layer located between the two gated temporal convolution layers, wherein the gated temporal convolution layer is used to perform convolution operation on the time series graph signal in the time dimension, and combines different activation functions to realize dynamic feature modeling, which adopts one-dimensional causal convolution and introduces nonlinearity through a gated linear unit, and the spatial graph convolution adopts Chebyshev polynomial approximation strategy for approximate calculation; the output layer includes two gated convolution layers and a two-dimensional convolution layer.

[0039] Further, the emergency response area functional loss prediction module includes a fully connected layer with input dimension and output dimension of 1, and the tensor output by the spatio-temporal feature capturing module is compressed in the second dimension after passing through the fully connected layer with input dimension and output dimension of 1, to obtain the functional loss prediction result of the future 1 time step .

[0040] Further, in step S4, the training of the spatio-temporal prediction model specifically includes:

[0041] (1) all rainfall sequences contained in the meteorological variables are divided into training set, validation set and test set according to a preset proportion;

[0042] (2) for the rainfall sequence contained in the training set and the functional loss value space distribution of all emergency response area nodes at each time step generated by the rainfall sequence through step S2, a sliding time window mechanism is used to construct a training sample pair, the input of each training sample pair includes the rainfall time series data of the historical time steps and the future 1 time step, and the functional loss value of each emergency response area in the historical time steps, and the output is the functional loss prediction value of each emergency response area in the future 1 time step;

[0043] (3) before the rainfall value corresponding to the time step, fill zero values, and supplement the functional loss value of the emergency response area corresponding to the time step as zero value, and the spatio-temporal prediction model predicts the functional loss value of the emergency response area corresponding to the future 1 time step The functional loss of each emergency response area at each time step is calculated using the mean square error loss function. Then, the sliding time window is advanced by one time step, and the above process is repeated. For each 24-hour rainfall sequence, 24 training sample pairs are generated. All rainfall sequences in the training set are traversed, and the average mean square error function of all generated sample pairs is used as the training loss of the spatiotemporal prediction model.

[0044] (4) Validate and test the trained spatiotemporal prediction model, and calculate the validation loss and test loss respectively.

[0045] Furthermore, in step S4, the prediction results for multiple future time steps are obtained through iterative forecasting, specifically including:

[0046] Let the rainfall sequence be denoted as When predicting the functional loss value at the first time step, fill in the gaps before the first rainfall value. There are zero values, denoted as And supplement this The functional loss value of the emergency response area at each time point. Both values ​​are zero; this is achieved by inputting the rainfall sequence. and the corresponding historical time of emergency response area functional loss value get Prediction results corresponding to time steps ;

[0047] When predicting the functional loss value at the second time step, the rainfall sequence used becomes , Will Prediction results corresponding to time steps As part of the input, namely, the historical moment-in-time emergency response area functional loss value used. ,get Prediction results corresponding to time steps ;

[0048] This process is repeated 24 times until the time-series prediction results of functional loss for all emergency response areas under the entire 24-hour rainfall sequence are generated.

[0049] The beneficial effects of the present application are: 1. The intelligent update framework of multi-module cooperation is constructed, and the technical system of real-time perception and function prediction of urban waterlogging disaster under the driving of social media data is perfected. 2. In order to solve the limitation of spatio-temporal distribution of observation information, the graph attention network and the random mask training strategy are innovatively introduced, the situation awareness model adapting to different observation ratios is constructed, and the global inference of regional function loss based on sparse local observation information is realized. 3. Facing the timeliness demand of emergency response, the spatio-temporal prediction model fusing the spatio-temporal graph convolution structure and the long short-term memory network is established, the input mechanism based on the prior knowledge of physical simulation is constructed, and the efficient dynamic update of building group function loss and evacuation demand forecast is realized. 4. The present application dynamically couples real-time observation and physical simulation prior knowledge, and provides an auxiliary tool with accuracy and real-time performance for urban waterlogging emergency response. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The flowchart of the building group function loss prediction dynamic update method fusing social media data according to the present application.

[0051] Figure 2 The architecture diagram of the situation awareness model according to the embodiment of the present application.

[0052] Figure 3 The training strategy diagram of the situation awareness model according to the embodiment of the present application.

[0053] Figure 4 The architecture diagram of the spatio-temporal prediction model according to the embodiment of the present application.

[0054] Figure 5 The training strategy diagram of the spatio-temporal prediction model according to the embodiment of the present application.

[0055] Figure 6 The spatial distribution diagram of the emergency response area according to the embodiment of the present application.

[0056] Figure 7 The number distribution statistical diagram of the four building types in each emergency response area according to the embodiment of the present application.

[0057] Figure 8 The average Pearson correlation coefficient matrix between the emergency response areas according to the embodiment of the present application.

[0058] Figure 9 The diagram of the hourly prediction and measured rainfall and the number of available social media data from 16:00 on September 5 to 16:00 on September 6 according to the embodiment of the present application.

[0059] Figure 10The application scenario schematic diagram of the building group function loss prediction dynamic updating method fusing social media data is shown in the embodiment of the application.

[0060] Figure 11 The water depth distribution result schematic diagram of the observed waterlogging accumulation point after local expansion at different moments is shown in the embodiment of the application.

[0061] Figure 12 The function loss value space-time distribution diagram of the emergency response area at 23:00 on September 5 and the corresponding statistical histogram are shown in the embodiment of the application.

[0062] Figure 13 The function loss value space-time distribution diagram of the emergency response area at 00:00 on September 6 and the corresponding statistical histogram are shown in the embodiment of the application.

[0063] Figure 14 The function loss value space-time distribution diagram of the emergency response area at 01:00 on September 6 and the corresponding statistical histogram are shown in the embodiment of the application.

[0064] Figure 15 The function loss value space-time distribution diagram of the emergency response area at 02:00 on September 6 and the corresponding statistical histogram are shown in the embodiment of the application.

[0065] Figure 16 The prediction updating result schematic diagram after fusing social media data at different report starting moments is shown in the embodiment of the application. DETAILED DESCRIPTION

[0066] In order to better understand the above technical solutions, the exemplary embodiments of the application will be described in detail below with reference to the accompanying drawings, which are only exemplary embodiments of the application, and it should be understood that the application can also be realized in various forms and is not limited to the embodiments described herein. These embodiments are intended to enable those skilled in the art to better understand and more thoroughly understand the application.

[0067] The application discloses a building group function loss prediction dynamic updating method fusing social media data, as shown in the figure, comprising the following steps: Figure 1

[0068] S1, a disaster real-time monitoring module is constructed, which is used for collecting and processing disaster-related social media data in real time, and extracting local waterlogging observation information in a research area;

[0069] S2, training data is prepared: building types are divided into four categories of high-rise, multi-storey, villa and stacked villa, the space-time evolution of building group function state probability distribution is simulated, emergency response areas are divided, and function loss data sets of the areas under multiple rainfall scenarios are generated;

[0070] ​S3, build and train the situational awareness model, obtain the observed function loss value of the observed emergency response area at the corresponding time according to the local waterlogging observation information, and infer the function loss of all unobserved emergency response areas;

[0071] S4, build and train the spatiotemporal prediction model, input including historical observation rainfall value, future single time step forecast rainfall value and historical emergency response area function loss value, output future single time step emergency response area function loss prediction value, and obtain prediction result of future multiple time steps through iteration prediction;

[0072] S5, in actual application, obtain local waterlogging observation information at each time through disaster real-time monitoring module, determine the observed emergency response area and its observation function loss value, infer the function loss of unobserved emergency response area by using situational awareness model, and replace the prediction value of spatiotemporal prediction model at corresponding time with all emergency response area function loss obtained by observation and inference at different time, output updated prediction result of future multiple time steps.

[0073] The steps S1 to S5 will be described in detail below.

[0074] S1, build a disaster real-time monitoring module, which is used to collect and process disaster related social media data in real time, and extract local waterlogging observation information in the study area. Specifically, it includes the following:

[0075] (1) Collect multi-source social media data related to disasters in real time through the disaster real-time monitoring module, and the data acquisition channels include at least various social platforms, information platforms and news platforms. The retrieval process is based on the set of keywords and time range. Due to the differences in the web structures of different platforms, the data acquisition varies. Some platform data is collected automatically by using Scrapy framework based on Python language to develop a crawler program; while some other platform data is manually selected and downloaded by browsing the posts related to disaster.

[0076] (2) Process the collected original multi-source social media data, including at least cleaning, deduplication and denoising operations, then classify the disaster relatedness, form structured social media data and store, and the structured social media data at least includes three types of multi-modal data such as text, picture and video.

[0077] (3) Fine-grained water depth information mining is performed on the obtained multi-modal data, including geographic location information extraction, water depth information extraction and local area water depth expansion, so as to extract the local waterlogging observation information in the study area.

[0078] Specifically, in the process of extracting geographic location information, two methods are used according to different data sources:

[0079] The first, structured location information direct extraction: for the text data with latitude and longitude coordinates, directly extract its structured geographic field.

[0080] The second, based on the picture content of the geographical sign back method: for the text description is not complete but contains visual clues of pictures or video data, extract the landmark geographical elements in the picture or video, such as landmark buildings, street signs or attached text address information, and match the extracted landmark geographical elements with the street view picture data provided by the virtual earth software, so as to reverse the specific geographical position. And all the extracted geographical position is converted into specific latitude and longitude coordinates.

[0081] Specifically, in the extraction process of water depth information, for the pictures or videos showing the flood scene, according to the grade conversion rule between different reference objects and water depth, the sub grade mapping rule of submerged water depth suitable for the embodiment is formulated; based on the sub grade mapping rule of submerged water depth, at least through the following way to identify water depth: based on the grade identification method of picture or video: for the submerged scene identified in the picture or video, it is divided into different severity grades first, and then converted into corresponding quantitative water depth value, this embodiment selects three typical reference objects of human body, shared bicycle and car, and defines the water depth value corresponding to different submerged severity grades for each type of reference object, as shown in table 1.

[0082] Table 1 sub grade mapping rule of submerged water depth

[0083]

[0084] Specifically, in the local area water depth expansion process, after identifying the waterlogging point data from the social media data, it is assumed that the water surface elevation of the adjacent area within the preset range around the waterlogging point is consistent with that of the grid where the waterlogging point is located, the observed water depth information of the waterlogging point is expanded to the surrounding area, so as to obtain the flood water depth distribution of the local area. In this embodiment, at each observation time, a buffer zone with a radius of 300m is constructed for all waterlogging points, and combined with DEM data, the pixel points with a height difference of not more than 3m in the buffer zone are identified as adjacent areas with the same waterlogging water surface elevation; then, according to the water depth and elevation information of the waterlogging point and the elevation information of each adjacent pixel meeting the condition, the waterlogging water depth of all adjacent pixels meeting the condition is calculated, so as to obtain the flood water depth distribution in the local range. For the pixel conflict problem caused by the overlapping of adjacent waterlogging point buffer zones in the regional expansion process, this embodiment adopts the maximum water depth value as the conservative estimation strategy to ensure the risk controllability of the result in emergency response.

[0085] S2, preparing training data: dividing the building types into four categories of high-rise, multi-story, villa and stacked villa, simulating the spatio-temporal evolution of the probability distribution of building group function state, dividing the emergency response area, and generating the area function loss data set under various rainfall scenarios.

[0086] In this embodiment, the training data includes the training data required by the situation awareness model and the spatio-temporal prediction model. In order to obtain the training data, the required input data includes meteorological variables, flood disaster environment variables and building attribute related variables. Among them, the meteorological variables include 24-hour rainfall data measured by the rainfall station. In this embodiment, the 24-hour cumulative rainfall exceeding the preset value of 350 mm is selected as the analysis object, and each rainfall sequence is regarded as a uniform rainfall event. The uniform rainfall event refers to a rainfall event in which the rainfall in each hour is uniformly distributed in space within the study area. The flood disaster environment variables include DEM data, slope direction data, curvature data, slope data, terrain humidity index data, distance to river center line and distance to pipe network outlet; the building attribute related variables include the contour, building layers, floor area and building type of each building in the building group.

[0087] Further, step S2 specifically includes:

[0088] S21, based on the meteorological variables and flood disaster environment variables of the study area, coupling SWMM and WCA2D model to simulate urban waterlogging inundation, and obtaining the waterlogging evolution process.

[0089] S22, using the building functionality state forecasting method (Building Functionality State Forecasting, referred to as "BFS-Forecasting method") to forecast the probability of building function state, combining the meteorological variables, flood disaster environment variables and building attribute related variables of the study area, simulating the spatio-temporal evolution scenario of the entire building group in the study area under several rainfall events in different function state probability distribution, wherein the BFS-Forecasting method is the existing method.

[0090] In this embodiment, the building function state includes five types of I, II, III, V and VII, and the larger the value is, the more serious the building is affected. Specifically, the function state I means that all types of demands of all residential units in the current building can be met; the function state II means that the advanced demands of the bottom residential units in the current building cannot be met, while all types of demands of other residential units can be met, wherein the advanced demands at least include demands related to office and cleaning; the function state III means that the basic demands of the bottom residential units in the current building cannot be met, while all types of demands of other residential units can be met, wherein the basic demands at least include basic physiological demands for food; the function state V means that the safety demands of the bottom residential units in the current building cannot be met, while all types of demands of other residential units can be met, wherein the safety demands at least include life safety demands; and the function state VII means that the safety demands of all residential units in the current building cannot be met. The high-rise building refers to a building with a number of floors greater than or equal to 7, and the multi-storey building refers to a building with a number of floors between 4 and 6. Generally, for high-rise buildings and multi-storey buildings, the bottom residential unit refers to the first floor resident; for a villa, the bottom residential unit refers to the entire residential unit, which may include several floors; for a stacked villa, the bottom residential unit refers to the lower stack, and if a stack includes two floors, it refers to the first floor and the second floor.

[0091] Specifically, in the BFS-Forecasting method, by considering the water depth resistance distribution of components such as electronic socket, bottom switch, middle switch, refrigerator, kitchen countertop, water heater, air conditioner, computer, etc., the vulnerability curves of different building types corresponding to different function states are deduced by using Monte Carlo simulation, so as to obtain the probability distribution of the entire building group in the study area in different function states at different times under several rainfall events.

[0092] S23, the probability of any building in any function state I under any rainfall sequence at any time is not the highest value corresponding to the probability in all possible states, then the building is defined as a potential affected building, and then a spatial clustering method of K-means is used to divide all potential affected buildings into multiple emergency response zones, which are set as ERZs. In the spatial clustering process based on K-means, the standard for determining the number of clusters is that the average number of buildings contained in each emergency response zone is between 10 and 15.

[0093] When the building is in the function state III to VII, evacuation demand is generated, the zone function loss is defined as the average of the sum of the probabilities of all potential affected buildings in the zone in the function state III to VII, denoted as ZFL, and the expression is as shown in formula (1), and thus the emergency response zone function loss data set at different times under multiple rainfall scenarios is constructed:

[0094] (1)

[0095] In formula (1), represents the functional loss value of the i-th emergency response sub-area at the time t, represents the number of potentially affected buildings in the i-th emergency response sub-area, represents the corresponding building index, represents the building function state category that may occur, represents the probability value of the i-th building having the function state category at the time t.

[0096] S24, based on the rainfall sequence, calculate the functional loss time curve of each emergency response sub-area under the action of each single field rainfall, and calculate the Pearson correlation coefficient matrix between each emergency response sub-area under the action of the field rainfall, then take the average of the Pearson correlation coefficient matrix obtained under all rainfall field, and take the average value as the measure of the functional loss correlation between the emergency response sub-areas.

[0097] S3, construct and train a situational awareness model (SA), obtain the observed functional loss value of the observed emergency response sub-area at the corresponding time according to the local waterlogging observation information, and infer the functional loss of all unobserved emergency response sub-areas.

[0098] In this embodiment, as shown in Figure 2 , the situational awareness model includes two layers of multi-head graph attention layers and ReLU activation functions connected after each layer of multi-head graph attention layers for enhancing the non-linear expression ability, each layer of multi-head graph attention layer is configured with 16 attention heads and a hidden layer with a dimension of 128, and the output layer of the situational awareness model is a graph attention layer with a dimension of 1. The adjacency matrix used in each multi-head graph attention layer and the graph attention layer in the situational awareness model is determined according to the following method: based on the average Pearson correlation coefficient matrix obtained in step S24, first set a correlation coefficient threshold , , the value range of which is 0.7-1, when the functional loss correlation coefficient between any two sub-areas is greater than or equal to the correlation coefficient threshold, the corresponding element in the adjacency matrix used in the situational awareness model is 1, otherwise, it is set to 0.

[0099] In step S3 of this embodiment, as shown in Figure 3 , training the situational awareness model specifically includes: ​​​​​​

[0100] (1) Each emergency response subarea node contains three types of feature vectors, which are the functional loss value of the emergency response subarea, the node-related attribute variable and the environmental variable, wherein the node-related attribute variable includes the proportion of the four building types in each emergency response subarea, and the environmental variable includes DEM, curvature, distance to the pipe network outlet and distance to the river center line. Each index in the environmental variable is represented by the average value of the corresponding index at the centroid of all buildings in each emergency response subarea.

[0101] The input variable of the situational awareness model is , which is composed of three types of features, i.e., the functional loss value of the emergency response subarea, the node-related attribute variable and the environmental variable, in the feature dimension, wherein represents the total number of emergency response subareas, 1 represents the dimension of the functional loss value, represents the dimension of the node-related attribute, represents the dimension of the environmental variable, as shown in Figure 2 ; The emergency response subarea is divided into observed emergency response subarea and unobserved emergency response subarea, and the observed emergency response subarea is defined as the subarea in which more than a set proportion of potential affected buildings are located within the observed waterlogging coverage range. The set proportion is preferably 70%. The observed functional loss value is calculated using the BFS-Forecasting method and formula (1) according to the observed water depth, and only the buildings located within the observed waterlogging coverage range are considered in formula (1); for each unobserved emergency response subarea, the prior functional loss value is calculated based on all observed functional loss values by the inverse distance weighting method, as follows:

[0102] (2)

[0103] In formula (2), represents the prior functional loss value of the th emergency response subarea, represents the set of observed emergency response subareas, represents the Euclidean distance between the th emergency response subarea and the th emergency response subarea centroid, represents the observed functional loss value of the th emergency response subarea, represents the power weight parameter of the distance, which is 1. The output of the situational awareness model is , which represents the predicted functional loss value of all emergency response subareas at time . Finally, only the prediction results of the unobserved emergency response subareas are retained; for the observed emergency response subarea, the observed functional loss value is directly retained.

[0104] (2) For all 24-hour rainfall series contained in the meteorological variable, the training set, the validation set and the test set are divided according to the preset proportion. In the training process, the data is extracted from the spatial distribution of the function loss value of each time step generated by the rainfall series in the training set through step S2 in turn, and is spliced with the corresponding node-related attribute variables and environmental variables to construct a complete node feature vector. Then, the situational awareness model is independently trained under a plurality of preset observation ratios, and the training loss is calculated.

[0105] (3) The situational awareness models trained under several preset observation ratios are verified and tested, and the validation loss and the test loss are calculated respectively.

[0106] In this embodiment, 90% of the rainfall series is used for training and verification of the situational awareness model (and further divided into a training set and a validation set in a ratio of 7:3), and the remaining 10% of the rainfall series is used for model testing and performance evaluation. For each rainfall series used for model training, verification and testing, the hourly emergency response functional loss distribution scenarios obtained based on the BFS-Forecasting method are summarized.

[0107] Preferably, when training the situational awareness model, a random mask strategy is introduced to simulate the actual situation that only part of the emergency response area can be observed, which specifically includes:

[0108] (1) For the function loss value spatial distribution results of all emergency response area nodes at each time step generated by the rainfall series contained in the training set through step S2, a preset observation ratio number of emergency response area nodes is randomly selected, and the simulated function loss value is retained. The remaining nodes are masked, that is, the function loss values of the remaining nodes are shielded.

[0109] (2) According to step (1), the prior function loss value of all masked nodes is calculated and used to fill the feature vector of the corresponding masked area node. The function loss value of all emergency response area nodes and their corresponding node-related attribute variables and environmental variables are spliced to form the input of the situational awareness model.

[0110] (3) The masked nodes are predicted by the situational awareness model. After the prediction is completed, only the difference between the prediction result of the masked area node and the simulation value is calculated by the mean square error loss function.

[0111] To enhance the applicability of the situational awareness model under different observation ratio scenarios, the situational awareness model is independently trained under five pre-defined observation ratio settings (10%, 20%, 30%, 40%, and 50%) in this embodiment, so as to improve the robustness and prediction accuracy of the model under different information availability levels. Based on the situational awareness models trained under the above five pre-defined observation ratios, the model test results are used to draw scatter plots and perform fitting to obtain the index measuring the goodness of fit . If the values of the scatter plots obtained under each pre-defined observation ratio are all greater than 0.9, the selected value meets the requirements, otherwise the value is increased by 0.05, and the above steps are repeated until the condition that the values are all greater than 0.9 is met. In this embodiment, the value is preferably set to 0.95.

[0112] The training process of the above situational awareness model is implemented based on the PyTorch 2.0 framework in a Python environment, and the mean squared error (MSE) is selected as the loss function. The batch size during training is set to 1. The Adam (Adaptive Moment Estimation) optimizer is used for optimization of the parameters of the situational awareness model. This algorithm estimates the first moment (mean) and second moment (uncentered variance) of the gradient to dynamically adjust the learning rate of each parameter, thereby improving the convergence speed and enhancing the stability of the training process. The initial learning rate is set to 0.001. To further improve the training effect of the situational awareness model, an adaptive learning rate adjustment strategy is adopted: after completing 5 training rounds, the learning rate is reduced by 70%. In addition, an early stopping mechanism is introduced to save training time and prevent overfitting. The specific settings are as follows: if the average loss of the situational awareness model on the validation set does not further decrease for 20 consecutive rounds, the training process is terminated in advance.

[0113] S4, a spatiotemporal forecasting (STF) model is constructed and trained, with the input including historical observed rainfall values, future single-time-step forecasted rainfall values, and historical emergency response zone functional loss values, and the output being the future single-time-step emergency response zone functional loss prediction value, and the prediction result for multiple time steps in the future being obtained through iterative forecasting.

[0114] In this embodiment, as Figure 4As shown, the spatiotemporal prediction model includes a rainfall feature extraction module, a spatiotemporal feature capture module, and an emergency response area functional loss prediction module. The rainfall feature extraction module, serving as the input module to the spatiotemporal prediction module, processes rainfall data containing multiple historical time steps and one future time step, and concatenates the processed rainfall data with the corresponding historical time step's emergency response area functional loss tensor along the time dimension. The spatiotemporal feature capture module processes time-series data with a graph structure, enabling each graph node to incorporate features from its corresponding spatiotemporal domain. The emergency response area functional loss prediction module outputs the emergency response area functional loss prediction result with the same number of time steps as the input future rainfall time steps.

[0115] Specifically, the rainfall feature extraction module in the spatiotemporal prediction model includes two types of input data: one containing historical data and the other containing other data. A time step and the future F =Rainfall data tensor at 1 time step and corresponding history Functional loss tensor of emergency response area for 24 hours ,in, Integer and , This indicates the batch size used in training the spatiotemporal prediction model. This indicates that the input channel size is 1, corresponding to the feature dimension of functional loss in the emergency response area, and each time step contains... Functional loss information for each emergency response zone node. The rainfall feature extraction module includes an LSTM module and a fully connected layer following the LSTM module, and the rainfall tensor... First, after dimensional rearrangement, it becomes... The input is then fed into an LSTM module, where the input feature dimension is 1, the hidden state dimension is set to 64, and the number of stacked layers is 1. Then, the hidden state of the last time step is extracted from the LSTM output, and after adding a dimension, it becomes... After processing through a fully connected layer, a tensor is obtained. ;Will and By concatenating the data along the time dimension, the final input tensor is formed. .

[0116] Specifically, the spatio-temporal feature capturing module includes two-level serial spatio-temporal convolution modules and an output layer, the two-level serial spatio-temporal convolution modules are used to extract deep spatio-temporal features in the input sequence step by step, the output layer is used for feature compression and output mapping, and a functional loss prediction result for a future single time step is output. Each spatio-temporal convolution module includes two gated temporal convolution layers and a spatial graph convolution layer between the two gated temporal convolution layers.

[0117] The gated temporal convolution layer is used to perform convolution operation on the time series signal in the time dimension, and dynamic feature modeling is realized by combining different activation functions, which adopts one-dimensional causal convolution, the kernel width is , and nonlinearity is introduced through a gated linear unit. Specifically, for each node in the graph, the gated temporal convolution operation checks adjacent elements in the input time series in each sliding, and no additional padding is performed, which will cause the length of the time series to be reduced after each convolution operation.

[0118] In the spatial graph convolution, a Chebyshev polynomial approximation strategy is adopted for approximate calculation:

[0119] (3)

[0120] In formula (3), denotes a convolution kernel, denotes a graph convolution operator, denotes a graph signal, is a normalized graph Laplacian matrix, where is an identity matrix, is a degree matrix, is an adjacency matrix, which is the average Pearson correlation coefficient matrix obtained by using step S24; is a scaled Laplacian matrix, is the largest eigenvalue of , and denotes the th Chebyshev polynomial of order . denotes a convolution kernel parameter, is an index of the Chebyshev polynomial approximation order, is the Chebyshev polynomial approximation order, which also corresponds to the receptive field size of the graph convolution kernel.

[0121] In the first-level spatio-temporal convolution module, the convolution kernel size of the first layer gated temporal convolution layer is 3, the input channel ( ) is 1, and the output channel ( The input and output channels of the spatial graph convolution layer are 32, and the Chebyshev polynomial approximation order is 3; the convolution kernel size of the first layer of the gated time convolution layer in the second level spatio-temporal convolution module is 3, the input channel is 32, the output channel is 64, and the activation function is a Relu function.

[0122] The convolution kernel size of the first layer of the gated time convolution layer in the second level spatio-temporal convolution module is 3, the input channel is 64, the output channel is 128, and the activation function is a GLU function; the input and output channels of the spatial graph convolution layer are 128, and the Chebyshev polynomial approximation order is 3. The convolution kernel size of the first layer of the gated time convolution layer in the second level spatio-temporal convolution module is 3, the input channel is 64, the output channel is 128, and the activation function is a GLU function; the input and output channels of the spatial graph convolution layer are 128, and the Chebyshev polynomial approximation order is 3.

[0123] The output layer includes two gated time convolution layers and a two-dimensional convolution layer. The convolution kernel size of the first layer of the gated time convolution layer is 3, the input channel is 256, the output channel is 256, and the activation function is a GLU function; the convolution kernel size of the second layer of the gated time convolution layer is 1, the input channel is 256, the output channel is 256, and the activation function is a Sigmoid function. The input channel number of the two-dimensional convolution layer is 256, the output channel number is 1, and the convolution kernel size is 1. The size of the tensor finally output by the output layer is 1.

[0124] Specifically, in the emergency response functional loss prediction module, first, the is input into a full connection layer with the input dimension and the output dimension being 1, and the second dimension is compressed to obtain the functional loss prediction result of the future 1 time step.

[0125] In this embodiment, as shown in Figure 5 , the training spatio-temporal prediction model specifically includes:

[0126] (1) All rainfall sequences contained in the meteorological variables are divided into a training set, a validation set and a test set according to a preset ratio. Specifically, first, the training set and the validation set are divided according to a ratio of 9:1, and then the training set and the validation set are further divided according to a ratio of 7:3.

[0127] (2) For the rainfall sequence contained in the training set and the functional loss value space distribution of all emergency response area nodes at each time step generated by the step S2, a sliding time window mechanism is used to construct a training sample pair. The input of each training sample pair includes the rainfall time series data of the historical time steps and the future 1 time step, and the rainfall time series data of the historical ​​​the function loss value of each emergency response subarea at the time step, and the output is the function loss prediction value of each emergency response subarea at the future 1 time step. Preferably, here .

[0128] (3) In this embodiment, in order to construct standardized training sample pairs, for each 24-hour rainfall sequence and its corresponding subarea function loss value sequence , first, the rainfall sequence is forward filled: 12 zero value time steps are introduced before the start, denoted as , and all emergency response subarea function loss values at the corresponding time steps are set to 0 to construct the complete initial state. When generating training sample pairs, time steps to are used as time input features of the spatiotemporal prediction model, to are used as spatial feature inputs. The spatiotemporal prediction model predicts the function loss distribution of each emergency response subarea at time step accordingly, denoted as , and compares it with the simulation value to calculate the loss. Then, the time window is moved forward by 1 hour, and the above process is repeated until all sample pairs are generated. For each 24-hour rainfall sequence, this strategy will generate 24 sample pairs. By traversing all rainfall sequences in the training set, the average value of the mean square error function of all generated sample pairs is taken as the training loss of the spatiotemporal prediction model.

[0129] (4) The trained spatiotemporal prediction model is verified and tested, and the verification loss and test loss are calculated respectively. For the rainfall sequences contained in the verification set and the function loss value spatial distribution of all emergency response subarea nodes at each time step generated by step S2, the verification loss of the spatiotemporal prediction model is calculated in the same way as above; for the rainfall sequences contained in the test set and the function loss value spatial distribution of all emergency response subarea nodes at each time step generated by step S2, the test loss of the spatiotemporal prediction model is calculated in the same way as above.

[0130] In this embodiment, the spatio-temporal prediction model is also implemented based on the PyTorch 2.0 framework in a Python environment. During the training process, the mean square error is used as the loss function to measure the deviation between the predicted results and the actual values. The training batch size is set to 64. The model parameter optimization uses the Adam optimizer, which estimates the first moment (mean) and second moment (uncentered variance) of the gradient simultaneously to achieve adaptive adjustment of the learning rate of each parameter, thereby improving the convergence speed and enhancing the stability of the training process. The initial learning rate is set to 0.001 as the starting parameter configuration for model optimization. An adaptive learning rate adjustment strategy is also used in the training process of the spatio-temporal prediction model, which adjusts the learning rate to 70% of the current value after every 5 training rounds. In addition, an early stopping mechanism is introduced to save model training time, and the tolerance value is set to 20, i.e., if the average loss of the model on the validation set does not continue to decrease for 20 consecutive training rounds, the training is terminated.

[0131] In step S4, the prediction results for multiple future time steps are obtained through iterative prediction, specifically including:

[0132] When real-time prediction of the functional loss of the emergency response area at each time corresponding to a certain 24-hour rainfall sequence is performed, the selected generation prediction strategy is used, and the rainfall sequence is denoted as .

[0133] When predicting the functional loss value at the first time step, 12 zero values are filled before the first rainfall value, preferably, the 12 zero values are , denoted as , and the functional loss values of the emergency response area corresponding to the 12 time points are , which are also zero values. By inputting the rainfall sequence and the historical time point functional loss values of the emergency response area , the prediction result corresponding to the time step is obtained . ;

[0134] When predicting the functional loss value at the second time step, the rainfall sequence used becomes , and the prediction result corresponding to the time step is used as part of the input, i.e., the historical time point functional loss value of the emergency response area is , and the prediction result corresponding to the time step is obtained . ;

[0135] Similarly, this process is repeated 24 times until the functional loss time series prediction results of all emergency response areas under the entire 24-hour rainfall sequence are generated.

[0136] S5、Actual application, through the disaster real-time monitoring module to obtain the local waterlogging observation information at each moment, determine the observed emergency response area and its observation function loss value, use the situation awareness model to infer the unobserved emergency response area function loss, and replace the prediction value of the corresponding time and space prediction model with all the emergency response area function loss inferred at different time, output the updated future multi-time prediction result. Specifically, the following steps are included:

[0137] S51、Through the disaster real-time monitoring module to obtain the local waterlogging observation information at different time, and collect the local waterlogging observation information within 30 minutes before and after each whole point time into the local waterlogging observation information corresponding to the whole point time.

[0138] S52、Based on the local waterlogging observation information at different whole point time extracted by the disaster real-time monitoring module, when more than 70% of the potential affected buildings in a certain emergency response area are within the coverage range of the local waterlogging observation information at a certain time, it is considered as an observed emergency response area, and the proportion of the observed emergency response area at this time is obtained, that is, the observation proportion. Then, calculate the observation function loss value corresponding to all observed emergency response areas at the corresponding time, then calculate the prior function loss value of all unobserved emergency response areas through formula (2), and splice the function loss value corresponding to all emergency response area nodes with its corresponding node related attribute variable and environmental variable to build complete node feature information. Select the situation awareness model trained at the observation proportion closest to the observation proportion at this time, input the complete node feature information built above into the situation awareness model to infer the global emergency response area function loss at this time, wherein for the unobserved emergency response area, the function loss value prediction result output by the situation awareness model is used, and for the observed emergency response area, the observation function loss value is directly retained to ensure that the output result is consistent with the result based on the observation data.

[0139] S53、Suppose the measured rainfall process starts at , if after , at a certain whole point time , the proportion of the corresponding observed emergency response area is > 5%, the time point for prediction update is set to half an hour later, because at this time, the collection of local waterlogging observation information representing is completed, and the rainfall input sequence used by the time and space prediction model at this time is "measured hourly rainfall during to ". The 24-hour rainfall sequence is spliced by the forecasted rainfall at each time point after the time point. First, the iterative forecast strategy described in step S4 is used to obtain the forecasted value of the functional loss of the emergency response patch at each time point after the time point , and then the global functional loss of the emergency response patch at each time point is obtained based on the situational awareness model in step S52 , and is replaced by , and the forecasted update result of the functional loss of the emergency response patch at and other time points is obtained by using the iterative forecast strategy.

[0140] The following takes a rainstorm event occurring in a coastal city (referred to as XX City) from September 5 to September 6 of a certain year as an example to demonstrate the actual application effect of the building group functional loss forecast dynamic updating method fusing social media data described in this embodiment.

[0141] The measured rainfall data of 13 typhoon events in the province where XX City belongs is used in the meteorological variable, and the rainfall sequences with 24-hour cumulative rainfall exceeding 350 mm are selected as the analysis objects, and a total of 752 rainfall sequences are obtained for subsequent research. The 24-hour cumulative rainfall of these rainfall sequences is between 350.9 mm and 890.6 mm. Based on the spatio-temporal evolution results of the aggregated building functional state, a total of 553 potential affected buildings are identified in the research area, and these potential affected building groups constitute the main research objects for subsequent analysis. Then, the spatial clustering method based on K-means is used to divide all the potential affected buildings into 50 emergency response patches, and the spatial distribution of each emergency response patch is shown in Figure 6 . Figure 7 The number distribution statistics of the four building types in each emergency response patch are shown, and the average number of buildings in each emergency response patch is about 11 at this time.

[0142] Based on the obtained 752 24-hour rainfall sequences, first, under the action of a single rainfall, the Pearson correlation coefficient matrix between the emergency response patches is calculated according to the functional loss time curve of each emergency response patch. Subsequently, the average value of the correlation coefficient matrix obtained under all rainfall events is taken as the measurement of the functional loss correlation between the emergency response patches, and the average Pearson correlation coefficient matrix between the emergency response patches is shown in Figure 8 .

[0143] ​A web crawler program was developed using the Scrapy framework based on the Python language to automate data collection through keywords. For example, each piece of data from a social media platform includes metadata such as user ID, text content, timestamp, and image attachments. In addition to the data obtained through web crawling, to further expand the amount of information that can reflect real-time disaster conditions, this embodiment also obtained relevant data from citizen assistance forms on certain information platforms and various news platforms. During the data collection process, "heavy rain," "torrential rain," and "waterlogging" were consistently used as search keywords.

[0144] Figure 9 The data displays hourly forecasts and measured rainfall, as well as the activity of available social media data, from 16:00 on September 5th to 16:00 on September 6th. Considering that the time resolution of both measured rainfall data and simulated flood scenarios is 1 hour, while the time resolution of social media data is typically 1 minute or 1 second, this embodiment assumes that the social media data collected within 30 minutes before and after each hour corresponds to the flooding scenario at that hour.

[0145] Figure 10 This example demonstrates a typical application scenario of the dynamic update method for predicting functional loss in building clusters that integrates social media data. Prior to the extreme rainstorm event triggered by Typhoon TF1 on September 5-6, the study area (XX City) experienced another heavy rainfall event triggered by a typhoon (denoted as TF2) from July 29-30 of that year. Figure 10 As shown on the left, this event provided the foundation for the construction and calibration of the urban flooding physical model in this embodiment. In the event of July 29th, the urban flooding model constructed within the study area was initially calibrated and validated based on the flooding location information released by the XX City Traffic Police Department. Subsequently, based on the calibrated urban flooding model, the spatiotemporal evolution of functional loss values ​​in various emergency response areas under different rainfall scenarios was simulated. These simulation results are considered to contain the spatiotemporal correlation between functional loss values ​​in different emergency response areas, and this embodiment uses them as the dataset for training the situational awareness model and the spatiotemporal forecasting model.

[0146] like Figure 10 As shown on the right, when the research perspective focuses on the rainstorm event in XX City on September 5th, the dynamic update method for predicting the functional loss of building clusters proposed in this embodiment, which integrates social media data, aims to integrate real-time observation data to minimize the systematic bias introduced by insufficient calibration data in terms of quantity or quality, thereby improving the accuracy of predicting future functional losses in each emergency response area.

[0147] The trigger time point of the forecast update is set to every whole hour and a half hour in this embodiment. Taking the forecast update result published at 23:30 on September 5 as an example, the collection of social media data related to waterlogging scenarios around 23:00 has been completed at this time. Based on these observation data, and using the situational awareness model to infer the global emergency response functional loss value at 23:00, the functional loss value at 23:00 in the original spatio-temporal prediction model is replaced. Subsequently, combined with the rainfall input sequence constructed by "measured rainfall before 23:00" and "forecasted rainfall after 23:00", the spatio-temporal prediction model is used to generate updated forecast results of emergency response functional loss at each future time.

[0148] This study combines Figure 9 As shown in the change of hourly rainfall and the number of available social media data, the period from 22:00 on September 5 to 02:00 on September 6 is selected as a typical time window to demonstrate the dynamic updating effect of the functional loss prediction results of the fusion of social media data. Figure 11 The observed water depth spatial distribution scenario generated by combining the waterlogging accumulation points extracted from social media data and the local area water depth expansion algorithm is shown.

[0149] Figures 12-15 The functional loss value spatio-temporal distribution map of the emergency response area and its corresponding statistical histogram at four time points from 23:00 on September 5 to 02:00 on September 6 are shown. At each time point, the functional loss value of the emergency response area is composed of two sources: the observed emergency response area marked with a blue thick border, whose functional loss value is the average of the probabilities of all buildings in the observed water depth coverage range being in functional state III to functional state VII; the unobserved emergency response area is inferred by the situational awareness model trained under the observation ratio closest to the current observation ratio. From the statistical histogram results, it can be seen that the distribution of the functional loss value of each area gradually shifts to the high value interval, reflecting the trend of the overall functional loss level of each area rising over time.

[0150] Figure 16 The forecast update results after fusing social media data at different reporting times are shown, and the related forecast results are marked with a blue outer frame. To evaluate the prediction accuracy at different reporting times, Figure 16The function loss results of the local observed area at 00:00 on September 6, 02:00 on September 6 and 04:00 on September 6 are also shown in the figure, all identified by red frames, and the mean and standard deviation of the function loss calculated for the local observed area at different times are listed in the table at the lower left. The results show that the closer the reporting time is to the target time point, the closer the prediction results of the local observed area at different times are to the observation results, thereby verifying the positive role of the fusion of social media data in improving the prediction accuracy.

[0151] The above only describes the basic principles and preferred embodiments of the present application, and those skilled in the art can make many changes and improvements based on the above description, which should belong to the protection scope of the present application.

Claims

1. A method for dynamically updating the prediction of functional loss of building complexes by integrating social media data, characterized in that, include: S1. Construct a real-time disaster monitoring module to collect and process disaster-related social media data in real time and extract local waterlogging observation information within the study area; S2. Prepare training data: Divide building types into four categories: high-rise, multi-story, villa and stacked villa. Simulate the spatiotemporal evolution of the probability distribution of functional status of building groups, divide emergency response areas, and generate functional loss datasets of areas under various rainfall scenarios. S3. Construct and train a situational awareness model. Based on local flooding observation information, obtain the observation function loss value of the locally observed emergency response area at the corresponding time, and infer the function loss of all unobserved emergency response areas. S4. Construct and train a spatiotemporal prediction model. The input includes historical observed rainfall values, forecasted rainfall values ​​for a single future time step, and historical functional loss values ​​for emergency response areas. The output is the predicted value of functional loss for emergency response areas for a single future time step. The prediction results for multiple future time steps are obtained through iterative forecasting. S5. In practical applications, the local waterlogging observation information at each moment is obtained through the real-time disaster monitoring module. The observed emergency response areas and their observation function loss values ​​are determined. The function loss of the unobserved emergency response areas is inferred using the situational awareness model. The predicted values ​​of the corresponding spatiotemporal prediction model are replaced with the function losses of all emergency response areas observed and inferred at different times. The updated forecast results for multiple future time steps are then output.

2. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 1, characterized in that, In step S2, the training data includes the training data required for both the situational awareness model and the spatiotemporal prediction model. To obtain the training data, the required input data includes meteorological variables, flood-prone environmental variables, and building attribute-related variables. Among them, meteorological variables include 24-hour rainfall data measured by rain gauges, and rainfall sequences with 24-hour cumulative rainfall exceeding preset values ​​are selected as the analysis objects. Flood-prone environmental variables include DEM data, slope aspect data, curvature data, slope data, topographic humidity index data, distance to the river centerline, and distance to the pipeline outlet. Building attribute-related variables include the outline, number of floors, floor area, and building type of each building in the building complex.

3. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 2, characterized in that, Step S2 specifically includes: S21. Based on meteorological variables and flood-prone environmental variables of the study area, urban waterlogging inundation simulation is carried out by coupling SWMM and WCA2D models to obtain the waterlogging evolution process. S22. Using a probabilistic prediction method for building functional states, simulate the spatiotemporal evolution of the probability distribution of different functional states of a building complex during multiple rainfall events; the building functional states include five types: I, II, III, V, and VII, with larger values ​​indicating a more severe impact on the buildings; S23. Define the building whose probability of being in functional state I at any time under any rainfall sequence is not the highest as a potentially affected building. Use the K-means spatial clustering method to divide all potentially affected buildings into multiple emergency response zones. Define the functional loss of a zone as the average of the sum of the probabilities of all potentially affected buildings in functional states III to VII within that zone. In this way, construct a dataset of functional loss of emergency response zones at different times under various rainfall scenarios. S24. Based on the rainfall sequence, calculate the time history curve of functional loss for each emergency response area under each single rainfall event, and calculate the Pearson correlation coefficient matrix between the areas accordingly. Use the average Pearson correlation coefficient matrix under all rainfall events as a measure of functional loss correlation.

4. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 1, characterized in that, In step S3, the situational awareness model includes two multi-head graph attention layers and ReLU activation functions connected to each multi-head graph attention layer. Each multi-head graph attention layer is configured with several attention heads and multi-dimensional hidden layers. The output layer of the situational awareness model is a graph attention layer with a dimension of 1.

5. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 1, characterized in that, Step S3, training the situational awareness model specifically includes: (1) The input variables of the situational awareness model are ,in, This represents the total number of emergency response zones, and 1 represents the dimension of the functional loss value. This represents the dimensions of node-related attributes, which include the proportion of the four building types. The dimensions of environmental variables are represented, including DEM, curvature, distance to the pipeline outlet, and distance to the river centerline. The observed emergency response area is defined as a region where potentially affected buildings exceeding a set proportion are located within the observed flood coverage area; the observed function loss value is calculated based on the observed water depth. For each unobserved emergency response area, the prior function loss value is calculated using an inverse distance weighting method based on all observed function loss values. The output of the situational awareness model is... ,express The prediction function loss value of all emergency response areas at any given time is used, and only the prediction results for the unobserved emergency response areas are retained. (2) For all rainfall sequences included in the meteorological variables, divide them into training set, validation set and test set according to a preset ratio; extract data from the spatial distribution of functional loss values ​​of each time step generated by step S2 in the rainfall sequence of the training set, and concatenate them with the corresponding node-related attribute variables and environmental variables to construct a complete node feature vector; train the situational awareness model independently under a preset number of observation ratios, and calculate the training loss. (3) Verify and test the situational awareness models trained under several preset observation ratios, and calculate the verification loss and test loss respectively.

6. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 5, characterized in that, In step (2), when training the situational awareness model, a random masking strategy is introduced to simulate the situation where only a portion of the emergency response area is observable in reality. Specifically, this includes: (1) For the spatial distribution results of the functional loss values ​​of all emergency response area nodes at each time step generated by step S2 for the rainfall sequence contained in the training set, randomly select a preset observation ratio of emergency response area nodes, retain the simulated functional loss values, and perform masking on the remaining nodes, that is, mask the functional loss values ​​of the remaining nodes. (2) Calculate the prior functional loss value for all masked nodes according to step (1), and use it to fill the feature vector of the corresponding masked area node. Then, concatenate the functional loss value of all emergency response area nodes with their corresponding node-related attribute variables and environmental variables to form the input of the situational awareness model. (3) The mask nodes are predicted by the situational awareness model. After the prediction is completed, the loss is calculated only between the predicted results and the simulated values ​​of the masked area nodes by the mean square error loss function.

7. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 1, characterized in that, In step S4, the spatiotemporal prediction model includes a rainfall feature extraction module, a spatiotemporal feature capture module, and an emergency response area functional loss prediction module. The rainfall feature extraction module serves as the input module for the spatiotemporal prediction module. It processes rainfall data containing multiple historical time steps and one future time step and concatenates the processed rainfall data with the emergency response area functional loss tensor of the corresponding historical time step along the time dimension. The spatiotemporal feature capture module processes time series data with a graph structure, enabling each graph node to fuse features from the corresponding spatiotemporal domain. The emergency response area functional loss prediction module outputs the emergency response area functional loss prediction result with the same number of time steps as the input future rainfall time steps.

8. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 7, characterized in that, The rainfall feature extraction module in the spatiotemporal prediction model includes two types of input data: one containing historical data and the other containing other data. A time step and the future F =Rainfall data tensor at 1 time step and corresponding history Functional loss tensor of emergency response area for 24 hours ,in, Integer and , This indicates the batch size used in training the spatiotemporal prediction model. This indicates that the input channel size is 1, corresponding to the feature dimension of functional loss in the emergency response area, and each time step contains... Functional loss value information for each emergency response zone node; the rainfall feature extraction module includes an LSTM module and a fully connected layer connected after the LSTM module, and the rainfall data tensor. The input is fed into an LSTM module and processed by a fully connected layer to obtain a tensor. ,Will and By concatenating the data along the time dimension, the final input tensor is formed. .

9. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 7, characterized in that, The spatiotemporal feature capture module comprises a two-stage cascaded spatiotemporal convolutional module and an output layer. The two-stage cascaded spatiotemporal convolutional module is used to progressively extract deep spatiotemporal features from the input sequence. The output layer is used for feature compression and output mapping, and outputs the functional loss prediction result for a single future time step. Each spatiotemporal convolutional module includes two gated temporal convolutional layers and a spatial graph convolutional layer located between the two gated temporal convolutional layers. The gated temporal convolutional layer is used to perform convolution operations on the time series signal in the time dimension and combines different activation functions to achieve dynamic feature modeling. It adopts one-dimensional causal convolution and introduces nonlinearity through gated linear units. The spatial graph convolution uses a Chebyshev multinomial approximation strategy for approximate calculation. The output layer includes two gated convolutional layers and one two-dimensional convolutional layer.

10. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 8, characterized in that, The emergency response area function loss prediction module includes a fully connected layer with both input and output dimensions of 1. The tensor output by the spatiotemporal feature capture module is passed through the fully connected layer with both input and output dimensions of 1, and the second dimension is compressed to obtain the function loss prediction result for the next time step. .

11. The method for dynamically updating the prediction of functional loss of building complexes by integrating social media data according to claim 8, characterized in that, Step S4, training the spatiotemporal prediction model specifically includes: (1) Divide all rainfall sequences included in the meteorological variables into training set, validation set and test set according to a preset ratio; (2) For the rainfall sequence included in the training set and the spatial distribution of the functional loss values ​​of all emergency response area nodes at each time step generated by step S2, a sliding time window mechanism is used to construct training sample pairs. The input of each training sample pair includes historical data. Rainfall time-series data at one time step and one future time step, as well as historical data. The functional loss values ​​of each emergency response area at each time step are output as the predicted functional loss values ​​of each emergency response area at the next time step. (3) In Fill in the time step before the rainfall value. A zero value, and supplement The functional loss value of the emergency response area corresponding to each time step is zero, and the spatiotemporal prediction model predicts... The functional loss of each emergency response area at each time step is calculated using the mean square error loss function. Then, the sliding time window is advanced by one time step, and the above process is repeated. For each 24-hour rainfall sequence, 24 training sample pairs are generated. All rainfall sequences in the training set are traversed, and the average mean square error function of all generated sample pairs is used as the training loss of the spatiotemporal prediction model. (4) Validate and test the trained spatiotemporal prediction model, and calculate the validation loss and test loss respectively.

12. The method for dynamic updating of building complex function loss prediction by integrating social media data according to claim 8 or 11, characterized in that, In step S4, prediction results for multiple future time steps are obtained through iterative forecasting, specifically including: Let the rainfall sequence be denoted as When predicting the functional loss value at the first time step, fill in the gaps before the first rainfall value. There are zero values, denoted as And supplement this The functional loss value of the emergency response area at each time point. Both values ​​are zero; this is achieved by inputting the rainfall sequence. and the corresponding historical time of emergency response area functional loss value get Prediction results corresponding to time steps ; When predicting the functional loss value at the second time step, the rainfall sequence used becomes , Will Prediction results corresponding to time steps As part of the input, namely, the historical moment-in-time emergency response area functional loss value used. ,get Prediction results corresponding to time steps ; This process is repeated 24 times until the time-series prediction results of functional loss for all emergency response areas under the entire 24-hour rainfall sequence are generated.

Images