Probabilistic prediction method of building function under flood disasters based on deep learning algorithm
Through deep learning algorithms and FFC models, combining hydrotopographic features and rainfall sequences, the functional probability distribution of buildings under flood disasters is predicted, which solves the roughness and inaccuracy of evacuation demand prediction in the existing technology, and provides a more refined Emergency guidance before disasters.
Patent Information
- Application Number
- CN202411118929.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-08-15
AI Technical Summary
The existing flood warning system is based on rough rainfall information, resulting in unnecessary evacuation and waste of resources. It lacks technology that can quickly and accurately predict the probability distribution of building functions, affecting pre-disaster emergency decisions.
The probability prediction method of building functions under flood disasters based on deep learning algorithm is adopted. The FFC model is used to integrate static hydrotopographic features and dynamic rainfall time series to predict the probability distribution of buildings in different functional states, and provide pre-disaster emergency guidance.
It achieves rapid and accurate acquisition of building function probability distribution in the pre-disaster emergency stage, improves the accuracy and refinement of evacuation demand forecasts, helps decision makers quantify uncertainty, and is suitable for flood vulnerability assessments of different building types.
Smart Images

Figure CN118643579B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of disaster prevention and mitigation, and in particular to a method for predicting the probability of building functions under flood disasters based on a deep learning algorithm. Background Art
[0002] Coastal areas are one of the areas in the world that are very vulnerable to strong tropical cyclones. The heavy rainfall brought by typhoons usually causes serious waterlogging disasters, which can cause serious casualties and economic losses in urban areas. So far, few people have paid attention to the emergency management issues in the pre-disaster emergency stage for specific events, such as pre-disaster emergency evacuation planning. The important prerequisite for emergency evacuation planning is to accurately and quickly forecast the evacuation needs. However, there is still a lack of methods to quickly and scientifically determine the emergency evacuation needs.
[0003] At present, the determination of pre-disaster evacuation needs still mainly relies on the early warning system. The existing flood warning system usually determines whether to issue corresponding warning information based on meteorological and hydrological variables such as rainfall, water level or water depth inundation map. However, the spatial resolution of the warning based on rainfall information is too coarse, which may cause a large number of unnecessary personnel to evacuate and waste precious emergency resources. At the same time, in order to obtain fine-scale water depth inundation map information, it is usually necessary to rely on very time-consuming and costly hydrological and hydrodynamic models. This time cost often cannot meet the actual needs in the emergency stage. In addition, it is inaccurate to determine the evacuation needs based solely on water depth inundation maps, because this will ignore the characteristics of the building itself. Even under the same rainfall or flood depth, different building types have different flood vulnerabilities due to their own structures, which leads to different evacuation needs for different building types. "Building function" is a concept proposed by Almufti and Willford et al. in 2013. The concept of "building function" goes beyond traditional structural safety and focuses more on the ability of the structure to achieve its intended purpose, including the safety of the building structure and the availability of infrastructure (water supply and power supply). Since "building function" can be a more comprehensive indicator of building applicability, the evacuation demand (mainly the number of people to be evacuated) generated by each single building under the action of a disaster can be more appropriately obtained based on this indicator. However, there is currently no technology that can quickly and accurately obtain the spatiotemporal prediction results of the probability distribution of building functions in the pre-disaster emergency stage of urban waterlogging disasters. The lack of this technology will make it impossible to provide decision makers with accurate spatiotemporal distribution results of evacuation demand in the pre-disaster emergency stage, which will seriously affect the formulation of subsequent evacuation plans. Summary of the invention
[0004] In order to overcome the deficiencies of the above technologies, the present invention provides a method for predicting the probability of building functions under flood disasters based on a deep learning algorithm. The present invention aims to predict the probability of all residential buildings in the study area being in different building functions based on the predicted rainfall time series when facing typhoon-induced rainstorm disasters; the FFC model proposed in the present invention has input features including the rainfall time series of the current and past several hours, hydrological environmental parameters, and the spatial distribution of the probability values of all buildings being in different functions at the previous moment, and the target variable is the spatial distribution of the probability values of all buildings being in different functions at the current moment. The present invention integrates the static hydrological and topographic spatial characteristics and the dynamic rainfall time series, and can achieve rolling continuous predictions for multiple moments in the future, thereby providing guidance for emergency actions before disasters.
[0005] Terminology explanation:
[0006] DEM: Digital Elevation Model.
[0007] FFC model: Flood Functionality Cast model, a building function prediction model under flood disasters.
[0008] ArcGIS: A geographic information system software used to manage, analyze, and present geospatial data.
[0009] TWI: Topographic Wetness Index, topographic wetness index.
[0010] CNN: Convolutional Neural Networks, convolutional neural network.
[0011] LSTM: Long Short Term Memory, long short-term memory unit.
[0012] Conv: Convolution, convolution.
[0013] BN1d: Batch Normalization 2d, one-dimensional batch normalization.
[0014] BN2d: Batch Normalization 2d, two-dimensional batch normalization.
[0015] LeakyReLU: Leaky Rectified Linear Unit, leaky corrected linear unit.
[0016] ConvTranspose2d: Convolutional Transpose 2D, two-dimensional convolutional transpose layer.
[0017] AvgPooling: Average Pooling, average pooling.
[0018] Dropout: Dropout module is a regularization method commonly used in deep learning. When training a neural network, the Dropout module randomly sets some elements of the input data to 0 according to the specified probability, that is, it randomly "discards" the activation of some neurons and their connections in each iteration.
[0019] Dropout rate: In deep learning, Dropout is a technique used to prevent neural networks from overfitting. It reduces the model's dependence on specific training samples by randomly "discarding" certain neurons in the neural network with a certain probability (i.e., the Dropout rate) during the training process, thereby improving the model's generalization ability; the Dropout rate refers to the probability that each neuron is temporarily removed from the network in each training batch.
[0020] Upsampling: A technique for increasing the resolution of an image or signal.
[0021] Early Stopping Mechanism: Early stopping mechanism is a strategy commonly used in training deep learning models to prevent model overfitting. The basic idea is to stop the training process when the performance on the validation set no longer improves, even if the performance of the model on the training set may continue to improve.
[0022] KL divergence: Kullback-Leibler, also known as KL divergence, is a method for measuring the difference between two probability distributions.
[0023] Confusion Matrix: Confusion Matrix is a tool for evaluating the performance of classification models. It helps analyze the accuracy and other performance indicators of the classifier by showing the distribution of model predictions and true labels. Each row of the confusion matrix represents an actual category, and each column represents a predicted category.
[0024] The technical solution adopted by the present invention to overcome the technical problems is:
[0025] A method for predicting building function probability under flood disasters based on a deep learning algorithm comprises the following steps:
[0026] S1. Collect basic data and historical rainstorm time series data of the study area;
[0027] S2. Process the data collected in step S1: randomly divide the time series data of several historical rainstorms into three rainfall event sets according to a preset ratio, namely, a rainfall event set for training and verification, a rainfall event set for optimizing the attention coefficients of affected buildings, and a rainfall event set for testing, wherein the number of rainfall event sets for optimizing the attention coefficients of affected buildings and the number of rainfall event sets for testing are both smaller than the number of rainfall event sets for training and verification; then use the urban waterlogging physical simulation model to obtain the spatiotemporal distribution of the flooded water depth, obtain the water depth experienced by each building at different times based on the spatiotemporal distribution of the flooded water depth, and calculate the water depth of each building at different times. The water depths experienced at different times are input into the building function probability prediction model based on physical conditions to obtain the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall; then the building surface vector with the information of the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall is first converted into a building point vector by finding the centroid, and then converted into a raster layer with the same dimension as the DEM based on the spatial position relationship. Each raster layer corresponds to a possible function. In each raster layer, the value of the grid where the building point vector is located is the probability value of the corresponding function, and the value of the other grids where there is no building point vector is 0;
[0028] S3. Using the data collected in step S1 and processed in step S2, construct an FFC model based on a deep learning algorithm, and train and verify the FFC model; optimize the affected building attention coefficient based on the performance on the rainfall event set used to optimize the affected building attention coefficient, and obtain the optimal affected building attention coefficient for each building type; and test the optimized FFC model on the rainfall event set used for testing;
[0029] S4. Input the predicted rainfall time series into the FFC model tested for each building type, perform multiple rolling iterative forecasts, and obtain the probability distribution results of each building in the study area being in different functions within a certain time range in the future.
[0030] Further, in step S1, the basic data includes DEM, land use data, river network vector data, drainage network vector data, residential building vector data and raster feature layer data derived from DEM of the study area; wherein the drainage network vector data at least includes inspection wells, outlets, pipe sections, pipe diameter, pipe roughness, and inspection well bottom elevation information; the raster feature layer data derived from DEM at least includes slope, aspect, curvature, and terrain humidity index;
[0031] Based on the river network vector data, drainage network vector data and the DEM range of the study area, the raster feature layer of the distance to the river network and the raster feature layer of the distance to the outlet are obtained respectively. The DEM, the raster feature layer data derived from the DEM, the raster feature layer of the distance to the river network and the raster feature layer of the distance to the outlet are combined together and called the flood condition factor layer; and all the raster feature layers contained in the flood condition factor layers are standardized with a mean of 0 and a standard deviation of 1.
[0032] Furthermore, in step S1, the acquisition of historical rainstorm time series data is specifically as follows:
[0033] Get The rainfall data during a typhoon, assuming that the duration of the original rainfall data is hours, of which , ;
[0034] Using the sliding window method, based on The time interval is hours, and the duration of each game is The hourly rainfall data is divided into durations of The hourly rainfall time series will be The rainfall time series with hourly total rainfall greater than the preset threshold is used as the uniform rainfall event for the input of the urban waterlogging physical simulation model, where , ,Uniform rainfall events refer to rainfall events in which the hourly rainfall in the study area is spatially uniformly distributed.
[0035] Further, in step S2, the urban waterlogging physical simulation model is coupled based on the drainage network model and the two-dimensional surface runoff model. The building function probability prediction model based on physical conditions is used to calculate the probability of buildings of different building types being in different functions at different water depth levels, and the building function categories are divided into five categories: I, II, III, V, and VII; specifically, function I means that all kinds of needs of all residential units in the current building can be met; function II means that the advanced needs of the lowest residential unit in the current building cannot be met, while all kinds of needs of other residential units can be met, wherein the advanced needs at least include needs related to office and cleaning; function III means that the basic needs of the lowest residential unit in the current building cannot be met, while all kinds of needs of other residential units can be met, wherein the basic needs at least include basic physiological needs for diet; function V means that the safety needs of the lowest residential unit in the current building cannot be met, while all kinds of needs of other residential units can be met, wherein the safety needs at least include life safety needs; function VII means that the safety needs of all residential units in the current building cannot be met. The probability of buildings of different building types being in different functions at different water depth levels is obtained according to the vulnerability curve corresponding to each function.
[0036] Furthermore, in step S3, before training the FFC model, the input features and output variables required by the FFC model are determined, as follows:
[0037] The input features required to generate the FFC model are: For each rainfall event set used for training and validation, the duration length is The hourly rainfall time series is obtained by using a sliding time window method with an interval of 1 hour and a duration of The hourly rainfall time series is used as the time feature, where For odd numbers Round up; train each building type separately. First, create a corresponding building mask layer for each building type. The dimension of the building mask layer is the same as the DEM feature layer. The value of the grid containing the building centroid is 1, and the value of the grid not containing the building centroid is 0. Then, all flood condition factor layers, the building mask layer of the corresponding building type, and the intercepted duration length are The probability spatial distribution feature layers of each building in different functions corresponding to the second to last time step in the hourly rainfall time series are combined to form the input features;
[0038] Generate the target variable required for the FFC model: The intercept duration length is The probability spatial distribution feature layer of each building in different functions corresponding to the penultimate time step in the hourly rainfall time series is used as the target variable;
[0039] The duration of each interception will be The corresponding input features obtained from the hourly rainfall time series are combined with the corresponding target variables to form a sample pair, and all sample pairs are divided into sample pairs for training and sample pairs for verification according to a preset ratio, wherein the number of sample pairs for training is greater than the number of sample pairs for verification.
[0040] Further, in step S3, the FFC model includes at least an encoder module, a temporal feature extraction module and a decoder module;
[0041] The encoder module includes four convolution units and an average pooling layer set after each convolution unit. Each convolution unit includes a convolution layer, a two-dimensional batch normalization, and a leakage corrected linear unit in turn. The size of the input feature map is converted from height × width × number of channels to H × W × (8 + number of possible feature classes) to height × width × number of channels ,when and When it is not divisible and Round down;
[0042] In the time feature extraction module, for each duration of The rainfall time series of each hour is first extracted through two LSTM models. Then, the hidden features output by the LSTM model are normalized through one-dimensional batch normalization. Then, a fully connected layer is used to convert the data dimensions (i.e., rainfall time steps × number of channels × height × width) into , which is then processed by a leaky rectified linear unit and a dropout module and reshaped into Tensor modules, the dimensions of each tensor are height × width × number of channels. , and then Each tensor in the tensor module is individually connected to the output of the encoder module along the channel dimension, and each combined tensor is used as the input of the decoder module;
[0043] The decoder module includes four upsampling layers and two-dimensional convolutional transposition units respectively arranged after each upsampling layer. Each two-dimensional convolutional transposition unit includes a two-dimensional convolutional transposition layer, a two-dimensional batch normalization and a leakage corrected linear unit. Each output result of the decoder module is processed by a convolutional layer and a leakage corrected linear unit to convert the number of channels into the dimension matching the output result of the FFC model. The time step corresponds to The output of the decoder module, traversing After the tensor module corresponding to each time step is obtained at each time step and the result of the encoder module is combined and input into the decoder module, the results of all time steps are stacked together in the channel dimension, and then passed through the convolution layer, the leaky rectified linear unit and the softmax function, so that the output duration length used is The predicted results of the probability spatial distribution of each building in different functions corresponding to the penultimate time step in the hourly rainfall time series.
[0044] Furthermore, when the FFC model is trained in step S3, in order to solve the problem of extremely unbalanced distribution of disaster-affected building samples and non-disaster-affected building samples under the action of disasters, the custom loss function is as follows:
[0045] In the above formula, Represents the number of all grids in a single channel, represents the number of channels of samples in the target variable, Indicates the number of samples contained in each batch, and They represent the building function probability prediction model based on physical conditions and the probability prediction model based on FFC model in a batch. The sample The first The probability value at each grid, is a natural constant, It is the affected building attention coefficient, which is used to strengthen the FFC model's attention to buildings that produce functional losses;
[0046] Then, the Adam optimizer is used to optimize the various parameters in the encoder module, decoder module and temporal feature extraction module in the FFC model by minimizing the loss.
[0047] Further, in step S3, the FFC model is trained and verified; the affected building attention coefficient is optimized based on the performance on the rainfall event set for optimizing the affected building attention coefficient, and the optimal affected building attention coefficient for each building type is obtained; and the optimized FFC model is tested on the rainfall event set for testing, specifically including the following:
[0048] Before training each building type, preset the affected building attention coefficient The loss function is clear at this time, and the FFC model training and verification can be carried out. After each training, the parameter combination in the encoder module, decoder module and time feature extraction module with the lowest average loss value on the sample pairs used for verification is selected as the final choice of this training, thereby completing the training of the FFC model;
[0049] The concentrated duration of rainfall events based on several fields for optimizing the affected building attention factor is The trained FFC model is evaluated on the rainfall event set used to optimize the affected building attention coefficients: for each rainfall event, the previous The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment in the hourly rainfall time series; statistics on the use of different affected building attention coefficients in the custom loss function The trained FFC model is used to select the affected building attention coefficients in several rainfall events. The average value of the kl divergence of all buildings at all times is taken, and the corresponding value is the smallest when the average value of the kl divergence is taken. As the optimal affected building attention coefficient; for each building type, the model training result obtained by applying the loss function corresponding to the optimal affected building attention coefficient is used as the final model parameter;
[0050] The duration of the concentrated rainfall events based on several fields used for testing is The optimal FFC model is tested on the rainfall event set for testing: for each rainfall event, the previous The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment in the hourly rainfall time series; then based on the simulation results and prediction results of the probability distribution of each building in different functions at each moment in all rainfall events, a scatter plot is drawn and fitted to obtain an indicator to measure the goodness of fit. , if The performance of the FFC model meets the preset requirements.
[0051] Furthermore, for each rainfall, The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment under the hourly rainfall time series include:
[0052] For a duration of When predicting the probability distribution of all buildings in different functions in the first hour, the previous The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is combined with the forecast rainfall at the current moment as the rainfall time series input. At the same time, it is considered that the probability that all buildings are in function I at the moment before the current moment is 100%, and the probability of being in other functions is 0%;
[0053] When predicting the probability distribution of all buildings in different functions in the second hour, The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is spliced with the forecast rainfall at the current moment and the previous moment as the rainfall time series input. At the same time, the prediction result of the probability spatial distribution feature layer of each building with different functions obtained by the FFC model in the previous round is used as the probability spatial distribution feature layer of each building with different functions at the previous moment.
[0054] By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0055] When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0056] By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0057] When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted results of the probability spatial distribution feature layers of each building with different functions are obtained in hours as the probability spatial distribution feature layers of each building with different functions at the moment before the current moment.
[0058] Furthermore, in step S4, the predicted rainfall time series is input into the FFC model tested for each building type, and multiple rolling iterative predictions are performed to obtain the probability distribution results of each building in the study area in different functions within a certain time range in the future, including:
[0059] Set before The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment under the hourly rainfall time series include:
[0060] For a duration of When predicting the probability distribution of all buildings in different functions in the first hour, the previous The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is combined with the forecast rainfall at the current moment as the rainfall time series input. At the same time, it is considered that the probability that all buildings are in function I at the moment before the current moment is 100%, and the probability of being in other functions is 0%;
[0061] When predicting the probability distribution of all buildings in different functions in the second hour, The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is spliced with the forecast rainfall at the current moment and the previous moment as the rainfall time series input. At the same time, the prediction result of the probability spatial distribution feature layer of each building with different functions obtained by the FFC model in the previous round is used as the probability spatial distribution feature layer of each building with different functions at the previous moment.
[0062] By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0063] When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0064] By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0065] When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted results of the probability spatial distribution feature layers of each building with different functions are obtained in hours as the probability spatial distribution feature layers of each building with different functions at the moment before the current moment.
[0066] The beneficial effects of the present invention are:
[0067] The present invention can quickly obtain the spatiotemporal evolution process of the probability distribution of building functions at the building level in the pre-disaster emergency stage of urban waterlogging disasters. The prediction results are very refined, and the uncertainty of the building's own resistance under the action of flood disasters is taken into account. The given building function probability distribution results are conducive to providing more information and uncertainty quantification for decision makers; in addition, the FFC model proposed by the present invention can predict the results of the affected buildings with a very small proportion, solving the problem of extremely unbalanced distribution of samples of affected buildings and non-affected buildings under the action of disasters. The present invention is conducive to improving the prediction accuracy of the probability distribution of building functions, thereby providing a good basis for determining the subsequent evacuation needs. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] Figure 1 This is a principle flow chart of the method for predicting the probability of building function under flood disasters based on a deep learning algorithm according to an embodiment of the present invention.
[0069] Figure 2 This is a diagram showing the spatial distribution of residential buildings of different building types and the study area described in an embodiment of the present invention.
[0070] Figure 3 These are the characteristic graphs contained in the flood condition factor layer described in the embodiment of the present invention.
[0071] Figure 4 Schematic diagram of vulnerability curves and exceedance probabilities of various functions of a high-rise building according to an embodiment of the present invention.
[0072] Figure 5 FIG. 1 is a schematic diagram of vulnerability curves and exceedance probabilities of various functions of a multi-story building according to an embodiment of the present invention.
[0073] Figure 6 It is a schematic diagram of vulnerability curves and exceedance probabilities of various functions of the villa according to the embodiment of the present invention.
[0074] Figure 7 FIG. 1 is a schematic diagram of vulnerability curves and exceedance probabilities of various functions of the duplex villa according to an embodiment of the present invention.
[0075] Figure 8 A schematic diagram of generating input features and target variables required for the FFC model described in an embodiment of the present invention.
[0076] Fig. 9 It is a schematic diagram of the architecture of the FFC model described in an embodiment of the present invention.
[0077] Fig.10 This is a histogram of the hourly average rainfall actually measured at sites within the study area described in the embodiment of the present invention.
[0078] Fig.11 Schematic diagram of the spatial distribution of the 24h KL divergence average value described in an embodiment of the present invention.
[0079] Fig.12 This is a bar chart comparing label values and predicted values of different functional probability distributions at different times on a single high-rise building when the 24-hour k1 divergence average value is 0.021 according to an embodiment of the present invention.
[0080] Fig.13 This is a bar chart comparing label values and predicted values at different times in different functional probability distributions on a single multi-story building when the 24-hour k1 divergence average value is 0.037 as described in an embodiment of the present invention.
[0081] Fig.14 This is a bar chart comparing label values and predicted values of probability distributions of different functional states at different times on a single villa building when the 24h k1 divergence average value is 0.014 as described in an embodiment of the present invention.
[0082] Fig.15 This is a bar chart comparing label values and predicted values of different functional probability distributions at different times on a single duplex building when the 24h k1 divergence average value is 0.017 as described in an embodiment of the present invention.
[0083] Fig.16 This is a schematic diagram of a confusion matrix consisting of label values and predicted values of representative functions at different times taking a high-rise building as an example according to an embodiment of the present invention.
[0084] Fig.17 This is a time series diagram of the number of buildings with different functions of all building types at different times predicted by the FFC model described in an embodiment of the present invention.
[0085] Fig.18 This is a spatial distribution diagram of all building types with different functions under the most severe disaster scenario predicted by the FFC model described in an embodiment of the present invention.
[0086] Fig.19 This is a time series variation diagram of the label values of the number of buildings with different functions of all building types at different times according to an embodiment of the present invention.
[0087] Fig. 20 This is a spatial distribution diagram of label values of all building types in different functions under the most severe disaster scenario described in an embodiment of the present invention. DETAILED DESCRIPTION
[0088] In order to facilitate those skilled in the art to better understand the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The following is only exemplary and does not limit the protection scope of the present invention.
[0089] The present invention discloses a method for predicting the probability of building function under flood disasters based on a deep learning algorithm, comprising the following steps:
[0090] S1. Collect basic data and historical rainstorm time series data of the study area;
[0091] S2. Process the data collected in step S1: randomly divide the time series data of several historical rainstorms into three rainfall event sets according to a preset ratio, namely, a rainfall event set for training and verification, a rainfall event set for optimizing the affected building attention coefficient, and a rainfall event set for testing, wherein the number of rainfall event sets for optimizing the affected building attention coefficient and the number of rainfall event sets for testing are both less than the number of rainfall event sets for training and verification; then use the calibrated urban waterlogging physical simulation model and the building function probability prediction model based on physical conditions to jointly obtain the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall, and then convert the building surface vector with the information of the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall into a building point vector by finding the centroid, and then convert it into a raster layer with the same dimension as the DEM based on the spatial position relationship, each raster layer corresponds to a possible function, and the value of the grid where the building point vector is located in each raster layer is the probability value of the corresponding function, and the value of the other grids where no building point vector is located is 0;
[0092] S3. Using the data collected in step S1 and processed in step S2, construct an FFC model based on a deep learning algorithm, and train and verify the FFC model; optimize the affected building attention coefficient based on the performance on the rainfall event set used to optimize the affected building attention coefficient, and obtain the optimal affected building attention coefficient for each building type; and test the optimized FFC model on the rainfall event set used for testing;
[0093] S4. Input the predicted rainfall time series into the FFC model tested for each building type, perform multiple rolling iterative forecasts, and obtain the probability distribution results of each building in the study area being in different functions within a certain time range in the future.
[0094] In order to better understand the above technical solution, the exemplary embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. These are only exemplary embodiments of the present invention. However, it should be understood that the present invention can also be implemented in various forms and is not limited to the embodiments described herein. These embodiments are intended to enable those skilled in the art to understand the present invention more clearly and thoroughly.
[0095] The present embodiment describes a method for predicting building function probability under flood disasters based on a deep learning algorithm, such as Figure 1 As shown, the following steps are included:
[0096] S1. Collect basic data and historical rainstorm time series data of the study area.
[0097] Specifically, the basic data include DEM of the study area, land use data, river network vector data, drainage network vector data, residential building vector data and raster feature layer data derived from DEM.
[0098] Among them, in this embodiment, the land use data includes land use type information presented in the form of a grid. The river network vector data includes vector data representing the centerline of each river. The drainage network vector data at least includes inspection wells, outlets, pipe sections, pipe diameters, pipe roughness, and inspection well bottom elevation information. The drainage network vector data is used to model the physical simulation model of urban waterlogging. The residential building vector data at least includes information such as building outlines, number of building floors, and building area. The study area of this embodiment and the spatial distribution of residential buildings of different building types are shown in the figure. Figure 2 The grid feature layer data derived from DEM at least includes slope, aspect, curvature, and terrain wetness index (TWI). Part of the grid feature layer data derived from DEM is obtained using the toolbox built into ArcGIS software, and the other part is obtained through SAGA GIS software.
[0099] Based on the river network vector data, drainage network vector data and the range of the DEM of the study area, the raster feature layer of the distance to the river network and the raster feature layer of the distance to the outlet are obtained respectively. The DEM, the raster feature layer data derived from the DEM, the raster feature layer of the distance to the river network and the raster feature layer of the distance to the outlet are combined together and called the flood condition factor layer, such as Figure 3 As shown. All grid feature layers contained in all flood condition factor layers are standardized with a mean of 0 and a standard deviation of 1.
[0100] Specifically, the acquisition of historical rainstorm time series data is as follows:
[0101] Get The rainfall data during a typhoon, assuming that the duration of the original rainfall data is hours, of which , ,here is an integer. For the convenience of calculation, It is usually an integer.
[0102] Using the sliding window method, based on The time interval is hours, and the duration of each game is The hourly rainfall data is divided into durations of The hourly rainfall time series will be The total rainfall in the hour is greater than the preset threshold The rainfall time series of is used as the uniform rainfall event input for the urban waterlogging physical simulation model, where , , and Generally, integers are used for ease of calculation. A uniform rainfall event refers to a rainfall event in which the hourly rainfall in the study area is uniformly distributed in space.
[0103] Preferably, in this embodiment, , , , , Specifically, rainfall data during three typhoons are obtained. The rainfall data can be station data or forecast grid data. Assume that the duration of the original rainfall data is 72 hours. This embodiment takes three typhoons that have occurred in a certain area in history. Assume that the three typhoons are Typhoon 1, Typhoon 2, and Typhoon 3.
[0104] Then, the sliding window method is used to divide each rainfall data with a duration of 72 hours into a rainfall time series with a duration of 24 hours based on a 12-hour time interval. Since most residential buildings are built on high terrain and the buildings themselves have a certain elevation, only when the total rainfall in 24 hours reaches a certain level can the scope and depth of water accumulation caused by the overflow of the sewer system affect the applicability of residential buildings. Therefore, this embodiment selects rainfall time series with a total rainfall of more than 350 mm in 24 hours, a total of 608 records, and each record is used as a uniform rainfall event as the input of the urban waterlogging physical simulation model.
[0105] S2. Process the data collected in step S1.
[0106] Specifically, several historical rainstorm time series data are randomly divided into three rainfall event sets according to a preset ratio, namely, a rainfall event set for training and verification, a rainfall event set for optimizing the affected building attention coefficient, and a rainfall event set for testing. Among them, the number of rainfall event sets for optimizing the affected building attention coefficient and the number of rainfall event sets for testing are both smaller than the number of rainfall event sets for training and verification. In this embodiment, 488 rainfall event sets for training and verification, 60 rainfall event sets for optimizing the affected building attention coefficient, and 60 rainfall event sets for testing are preferably randomly selected from 608 historical rainstorm time series data in a ratio of 8:1:1.
[0107] Then, the urban waterlogging physical simulation model is used to obtain the spatiotemporal distribution of the flooded water depth. Based on the spatiotemporal distribution of the flooded water depth, the water depth experienced by each building at different times is obtained. The water depth experienced by each building at different times is input into the building function probability prediction model based on physical conditions to obtain the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall. Then, the building surface vector with the information of the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall is first converted into a building point vector by finding the centroid, and then converted into a raster layer with the same dimension as the DEM based on the spatial position relationship. Each raster layer corresponds to a possible function. In each raster layer, the value of the grid where the building point vector is located is the probability value of the corresponding function, and the value of the other grids where there is no building point vector is 0.
[0108] Among them, the physical simulation model of urban waterlogging is based on the coupling of the drainage network model and the two-dimensional surface runoff model. Specifically, the drainage network model adopts Pyswmm, and the two-dimensional surface runoff model adopts WCA2D, that is, by coupling Pyswmm and WCA2D together, a dynamic simulation of flooding in the study area is performed. Among them, SWMM is a free hydrological and hydraulic model, which has been widely used in the simulation of urban rainwater runoff. PySwmm is a secondary development module built based on SWMM software using Python language. Based on Pyswmm, batch simulation of different rainfall data can be achieved. In this embodiment, a certain flood-prone area is selected as the study area, such as Figure 2 As shown in the figure, Pyswmm 0.6.2 is used to simulate the rainfall runoff process and the water flow process in the drainage network. According to the distribution of the drainage network system, there are 700 inspection wells, 74 outlets and 700 pipe sections; the entire catchment area is divided into 700 sub-catchments. Combined with the land use data in the study area and the "Urban Waterlogging Prevention and Control Planning Standards (Draft for Comments)", the surface slope and impermeability of each sub-catchment area were determined based on ArcGIS10.8 software. WCA2D (Weighted Cellular Automata 2D Inundation Model), the full name is weighted cellular automaton two-dimensional inundation simulation model, is a cellular automaton model based on weight conversion rules, which is very efficient in simulating two-dimensional hydrodynamic processes.
[0109] Specifically, the building function probability prediction model based on physical conditions is used to calculate the probability of buildings of different building types being in different functions at different water depth levels, and the building function categories are divided into 5 categories: I, II, III, V, and VII. The larger the number corresponding to the function category, the more serious the disaster situation; specifically, function I means that all kinds of needs of all residential units in the current building can be met; function II means that the advanced needs of the residential unit at the bottom of the current building cannot be met, while all kinds of needs of other residential units can be met, among which the advanced needs at least include needs related to office and cleaning; function III means that the basic needs of the residential unit at the bottom of the current building cannot be met, while all kinds of needs of other residential units can be met, among which the basic needs at least include basic physiological needs for food; function V means that the safety needs of the residential unit at the bottom of the current building cannot be met, while all kinds of needs of other residential units can be met, among which the safety needs at least include life safety needs; function VII means that the safety needs of all residential units in the current building cannot be met.
[0110] A high-rise building refers to a building with 7 or more floors, and a multi-story building refers to a building with 4 to 6 floors. Generally speaking, for high-rise buildings and multi-story buildings, the bottom residential unit refers to the first-floor residents; for villas, the bottom residential unit is the entire residential unit, which may include several floors; for duplex villas, the bottom residential unit refers to the lower duplex, and if a duplex contains 2 floors, it is the first and second floors. In this embodiment, the vulnerability curves of different building types corresponding to different functions and the corresponding exceedance probabilities are as follows: Figures 4 to 7 As shown, Figure 4 , Figure 5 , Figure 6 , Figure 7 The vulnerability curves and corresponding exceedance probabilities of high-rise buildings, multi-story buildings, villas, and duplex villas in different functions are respectively represented. The vulnerability curve of a certain function represents the probability of the building being in the function and the function corresponding to the more serious disaster situation under a certain water depth level. For example, the vulnerability curve corresponding to function I is the probability of the building being in the state of I, II, III, V, VII, that is, 100%, the vulnerability curve corresponding to function II is the probability of the building being in the state of II, III, V, VII, the vulnerability curve corresponding to function III is the probability of the building being in the state of III, V, VII, and so on; the exceedance probability of a certain function refers to the corresponding value on the vulnerability curve corresponding to the function under a certain water depth level. Among them, the possible functional categories of high-rise buildings are I, II, III, V, VII, the possible functional categories of multi-story buildings are I and VII, the possible functional categories of villas are I, III, V, and the possible functional categories of duplex villas are I, III, V, and VII.
[0111] Figures 4 to 7 In the figure, it is assumed that the height of the first floor is 0.35 meters. Except that the vulnerability curve corresponding to function I is a straight line parallel to the x-axis at y=1, and the vulnerability curve corresponding to function VII is a straight line perpendicular to the x-axis at the outdoor water depth of 0.85 meters, the basic expressions of the vulnerability curves corresponding to the other functions are as follows:
[0112]
[0113] In the above formula, represents the standard normal cumulative distribution function, Indicates the outdoor water depth experienced by a building at a certain moment. represents the logarithmic median of water depth resistance, Represents the logarithmic standard deviation of water depth resistance. The values of the parameters in the vulnerability curve expression corresponding to each possible function are shown in Table 1.
[0114] Table 1 Parameters in the vulnerability function curve expression corresponding to each possible function
[0115]
[0116] The specific parameter values can be modified according to the actual situation. Table 1 only gives the parameter values of an example vulnerability curve.
[0117] The probability of being in different functions at different water depth levels can be obtained based on the vulnerability curve corresponding to each function. That is, when the vulnerability curve corresponding to each function is clear, the probability of being in different functions at a certain water depth level can be calculated. The outdoor water depth experienced at a certain moment is The probability of being in a certain function FS when The vulnerability curve corresponding to a certain function FS is The corresponding value is Figure 4 For example, then:
[0118]
[0119]
[0120]
[0121]
[0122]
[0123] S3. Using the data collected in step S1 and processed in step S2, construct an FFC model based on a deep learning algorithm, and train the FFC model. Based on the performance of the FFC model on the rainfall event set used for testing, optimize the affected building attention coefficient to obtain the optimal affected building attention coefficient for each building type.
[0124] In this embodiment, before training the FFC model, the input features and output variables required by the FFC model are first determined, such as Figure 8 As shown, the details are as follows:
[0125] Generate the input features required for the FFC model: For each rainfall time series with a duration of 24 hours in the rainfall event set used for training and verification, the sliding time window method is used to intercept the rainfall time series with a duration of 12 hours at an interval of 1 hour as the time feature; each building type is trained separately. First, a corresponding building mask layer is created for each building type. The dimension of the building mask layer is the same as the DEM feature layer. The value of the raster containing the building centroid is 1, and the value of the raster not containing the building centroid is 0. Then, all the flood condition factor layers, the building mask layers of the corresponding building types, and the probability spatial distribution feature layers of each building in different functions corresponding to the second to last time step in the intercepted rainfall time series with a duration of 12 hours are combined to form the input features.
[0126] Generate the target variable required for the FFC model: take the probability spatial distribution feature layer of each building in different functions corresponding to the penultimate time step in the intercepted rainfall time series with a duration of 12 hours as the target variable.
[0127] A total of 6344 training sample pairs were generated from the 24-hour rainfall sequences of 488 rainfall event sets for training and validation, and were divided into sample pairs for training and sample pairs for validation in a ratio of 7:3.
[0128] In this embodiment, the FFC model includes at least an encoder module, a temporal feature extraction module and a decoder module. Fig. 9 shown.
[0129] The encoder module includes four convolution units and an average pooling layer respectively arranged after each convolution unit. Each convolution unit includes a convolution layer (Conv), a two-dimensional batch normalization (BN2d) and a leaky rectified linear unit (LeakyReLU) in sequence. In this embodiment, the size of the convolution kernel in each convolution layer is 3×3, and the padding is padding = 1. The output dimensions of the four convolution layers are 64, 128, 256, and 512 respectively. Each convolution unit is followed by an average pooling layer (AvgPooling) of size 2×2. The size of the input feature map is reduced from height×width×number of channels (104×222×(8+number of possible functional categories)) to height×width×number of channels (6×13×512) through the encoder module.
[0130] In the time feature extraction module, for each rainfall time series with a duration of 12 hours, the time features are first extracted through two LSTM models. The number of layers (num_layers) of each LSTM model is 2, and the dropout rate (Dropout rate) is set to 0.4. After being processed by the first LSTM model, the number of hidden features output is 12, and after being processed by the second LSTM model, the number of hidden features output is 6. Then, the hidden features output by the LSTM model are normalized by one-dimensional batch normalization (BN1d), and then the data dimension is changed to 12×3×6×13 (number of rainfall time steps×number of channels×height×width) through a fully connected layer. Then, after being processed by the leaky rectified linear unit (LeakyReLU) and the dropout module (Dropout), it is reshaped into 12 tensor modules, each with a dimension of 6×13×3 (height×width×number of channels). Then, each tensor in the 12 tensor modules is individually connected to the output of the encoder module along the channel dimension, and each combined tensor is used as the input of the decoder module.
[0131] The decoder module includes four upsampling layers (Upsampling) and two-dimensional convolutional transpose units respectively arranged after each upsampling layer. Each two-dimensional convolutional transpose unit includes a two-dimensional convolutional transpose layer (ConvTranspose2d), a two-dimensional batch normalization (BN2d) and a leaky rectified linear unit (LeakyReLU). The output dimensions of the four two-dimensional convolutional transpose units are 512, 256, 128, and 64 respectively. The height and width of the data after the first upsampling in the decoder module are the same as the height and width of the data before the fourth average pooling in the encoder module, and so on. Each output result of the decoder module is processed by a convolution layer with a convolution kernel size of 3×3 and padding = 1 and a leaky rectified linear unit (LeakyReLU) to change the number of channels (64 dimensions) to the dimension matching the output result of the FFC model (i.e., the number of possible functional categories). 12 time steps correspond to the output results of 12 decoder modules. After traversing 12 time steps to obtain the tensor module corresponding to each time step and the result of the encoder module and inputting the result obtained by the decoder module (at this time, the output result dimension corresponding to each time step is (104×222×the number of possible functional categories), the results obtained by all time steps are stacked together in the channel dimension, and then passed through a convolution layer with a convolution kernel size of 3×3 and padding = 1, a leaky rectified linear unit (LeakyReLU) and a softmax function, thereby outputting the probability spatial distribution prediction results of each building in different functions corresponding to the penultimate time step in the rainfall time series with a duration of 12 hours.
[0132] Furthermore, when training the FFC model, considering the reality, the damage data set of buildings is highly unbalanced, and the number of buildings not affected by disasters is often much higher than the number of buildings affected by disasters. For pre-disaster forecasts, accurately predicting the spatiotemporal distribution of affected buildings is the key to carrying out emergency actions. In order to solve the problem of extremely unbalanced distribution of samples of damaged buildings and samples of unaffected buildings under the action of disasters, the custom loss function of this embodiment is as follows:
[0133]
[0134] In the above formula, Represents the number of all grids in a single channel, represents the number of channels of samples in the target variable, Indicates the number of samples contained in each batch, and They represent the building function probability prediction model based on physical conditions and the probability prediction model based on FFC model in a batch. The sample The first The probability value at each grid, is a natural constant, It is the affected building attention coefficient, which is used to strengthen the FFC model's attention to buildings that have functional losses.
[0135] Then, the Adam optimizer is used to optimize the parameters of the encoder module, decoder module and temporal feature extraction module in the FFC model by minimizing the loss. Specifically, the learning rate is set to 0.0006, the batch size is 16, and an early stopping mechanism with a tolerance step of 20 for validation loss not decreasing is set during the training of the FFC model. That is, if the validation loss does not decrease for 20 consecutive training rounds, the FFC model stops training.
[0136] Further, in step S3, the FFC model is trained and verified; the affected building attention coefficient is optimized based on the performance on the rainfall event set for optimizing the affected building attention coefficient, and the optimal affected building attention coefficient for each building type is obtained; and the optimized FFC model is tested on the rainfall event set for testing, specifically including the following:
[0137] Before training each building type, the attention coefficient of the affected buildings is preset through artificial experience. The loss function is clear at this time, and the FFC model training and verification can be carried out. After each training, the parameter combination in the encoder module, decoder module and time feature extraction module with the lowest loss value of the sample pair used for verification is selected as the final choice of this training, thereby completing the training of the FFC model;
[0138] Based on the rainfall time series with a duration of 24 hours in the 60 rainfall events used to optimize the affected building attention coefficient, the trained FFC model is evaluated on the rainfall event set used for testing: Since the FFC model needs the rainfall in the previous 12 hours (including the current moment) and the probability distribution of all buildings in different functions at the moment before the current moment as input when predicting the probability distribution of all buildings in different functions at the current moment, it is necessary to perform 24 rolling iterations for each rainfall. Through 24 rolling iterations, the prediction results of the probability spatial distribution feature layer of each building in different functions at the current moment obtained by the FFC model are used as the model input for the prediction at the next moment, thereby obtaining the probability distribution results of all buildings in different functions at each moment under the complete 24-hour rainfall time series. Specifically:
[0139] For a 24-hour forecast rainfall time series, when predicting the probability distribution of all buildings in different functions in the first hour, the forecast rainfall series of the first 11 hours is padded with 0, that is, the rainfall in each hour of the first 11 hours is considered to be 0, and it is spliced with the forecast rainfall at the current moment as the rainfall time series input. At the same time, it is considered that the probability of all buildings in function I at the moment before the current moment is 100%, and the probability of being in other functions is 0%;
[0140] When predicting the probability distribution of all buildings with different functions in the second hour, the forecast rainfall sequence of the first 10 hours is padded with 0, that is, the rainfall per hour in the first 10 hours is considered to be 0, and it is spliced with the forecast rainfall at the current moment and the previous moment as the rainfall time series input. At the same time, the prediction result of the probability spatial distribution feature layer of each building with different functions obtained by the FFC model in the previous round is used as the probability spatial distribution feature layer of each building with different functions at the previous moment;
[0141] Similarly, when predicting the probability distribution of all buildings in different functions at the 12th hour, the forecast rainfall sequence of the previous 12 hours is input as the rainfall time series, where the previous 12 hours include the hour corresponding to the current moment, and the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model obtained in the previous round at the 11th hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0142] When predicting the probability distribution of all buildings in different functions at the 13th hour, the forecast rainfall sequence of the previous 12 hours is input as the rainfall time series. The previous 12 hours here include the hour corresponding to the current moment. At the same time, the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model in the previous round at the 12th hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment.
[0143] Similarly, when predicting the probability distribution of all buildings in different functions at the 23rd hour, the forecast rainfall sequence of the previous 12 hours is used as the rainfall time series input, where the previous 12 hours include the hour corresponding to the current moment, and the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model obtained in the previous round at the 22nd hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0144] When predicting the probability distribution of all buildings in different functions at the 24th hour, the forecast rainfall sequence for the previous 12 hours is input as the rainfall time series, where the previous 12 hours include the hour corresponding to the current moment. At the same time, the FFC model obtained in the previous round is used to predict the probability spatial distribution feature layer of each building in different functions at the 23rd hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment.
[0145] Finally, statistics are used in the custom loss function to determine the attention coefficients of different affected buildings. The trained FFC model obtained is the average value of the kl divergence of all buildings at all times in the 60 rainfall events used to optimize the affected building attention coefficients, and the corresponding value when the kl divergence average value is the smallest As the optimal affected building attention coefficient; for each building type, the model training results obtained by applying the loss function corresponding to the optimal affected building attention coefficient are used as the final model parameters. After experimental calculation, the optimal affected building attention coefficients of the four building types are: high-rise building 3.5, multi-story building 2.5, villa 2.5, duplex villa 3. For each building type, the model training results obtained by applying the loss function corresponding to the optimal affected building attention coefficient are used as the final model parameters for evaluation and subsequent application on the rainfall event set used for testing.
[0146] Based on the rainfall time series with a duration of 24 hours in the 60 rainfall events used for testing, the selected FFC model is tested on the rainfall event set used for testing: Since the FFC model needs the rainfall in the previous 12 hours (including the current moment) and the probability distribution of all buildings with different functions at the moment before the current moment as input when predicting the probability distribution of all buildings with different functions at the current moment, it is necessary to perform 24 rolling iterations for each rainfall. Through 24 rolling iterations, the prediction results of the probability spatial distribution feature layer of each building with different functions at the current moment obtained by the FFC model each time are used as the model input for the prediction at the next moment, thereby obtaining the probability distribution results of all buildings with different functions at each moment in the complete 24-hour rainfall time series. Specifically:
[0147] For a 24-hour forecast rainfall time series, when predicting the probability distribution of all buildings in different functions in the first hour, the forecast rainfall series of the first 11 hours is padded with 0, that is, the rainfall in each hour of the first 11 hours is considered to be 0, and it is spliced with the forecast rainfall at the current moment as the rainfall time series input. At the same time, it is considered that the probability of all buildings in function I at the moment before the current moment is 100%, and the probability of being in other functions is 0%;
[0148] When predicting the probability distribution of all buildings with different functions in the second hour, the forecast rainfall sequence of the first 10 hours is padded with 0, that is, the rainfall per hour in the first 10 hours is considered to be 0, and it is spliced with the forecast rainfall at the current moment and the previous moment as the rainfall time series input. At the same time, the prediction result of the probability spatial distribution feature layer of each building with different functions obtained by the FFC model in the previous round is used as the probability spatial distribution feature layer of each building with different functions at the previous moment;
[0149] Similarly, when predicting the probability distribution of all buildings in different functions at the 12th hour, the forecast rainfall sequence of the previous 12 hours is input as the rainfall time series, where the previous 12 hours include the hour corresponding to the current moment, and the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model obtained in the previous round at the 11th hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0150] When predicting the probability distribution of all buildings in different functions at the 13th hour, the forecast rainfall sequence of the previous 12 hours is input as the rainfall time series. The previous 12 hours here include the hour corresponding to the current moment. At the same time, the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model in the previous round at the 12th hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment.
[0151] Similarly, when predicting the probability distribution of all buildings in different functions at the 23rd hour, the forecast rainfall sequence of the previous 12 hours is used as the rainfall time series input, where the previous 12 hours include the hour corresponding to the current moment, and the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model obtained in the previous round at the 22nd hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0152] When predicting the probability distribution of all buildings in different functions at the 24th hour, the forecast rainfall sequence for the previous 12 hours is input as the rainfall time series, where the previous 12 hours include the hour corresponding to the current moment. At the same time, the FFC model obtained in the previous round is used to predict the probability spatial distribution feature layer of each building in different functions at the 23rd hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment.
[0153] Then, based on the simulation results and prediction results of the probability distribution of each building in different functions at each moment in all rainfall events, a scatter plot was drawn and fitted to obtain an index to measure the goodness of fit. , we found that the optimal FFC models for the four building types are The results were all greater than 0.7, meeting the preset requirements.
[0154] Step S4: Input the predicted rainfall time series into the FFC model tested for each building type, perform multiple rolling iterative predictions, and obtain the probability distribution results of each building in the study area being in different functions within a certain time range in the future.
[0155] In this embodiment, the rainfall in the previous 12 hours and the probability distribution of all buildings in different functions at the moment before the current moment are assumed as input, wherein the rainfall in the previous 12 hours includes the rainfall at the current moment, and 24 rolling iterations are performed on each rainfall. Through 24 rolling iterations, the prediction result of the probability spatial distribution feature layer of each building in different functions at the current moment obtained by the FFC model each time is used as the model input for the prediction at the next moment, thereby obtaining the probability distribution results of all buildings in different functions at each moment under the complete 24-hour rainfall time series, specifically including:
[0156] For a 24-hour forecast rainfall time series, when predicting the probability distribution of all buildings in different functions in the first hour, the forecast rainfall series of the first 11 hours is padded with 0, that is, the rainfall in each hour of the first 11 hours is considered to be 0, and it is spliced with the forecast rainfall at the current moment as the rainfall time series input. At the same time, it is considered that the probability of all buildings in function I at the moment before the current moment is 100%, and the probability of being in other functions is 0%;
[0157] When predicting the probability distribution of all buildings with different functions in the second hour, the forecast rainfall sequence of the first 10 hours is padded with 0, that is, the rainfall per hour in the first 10 hours is considered to be 0, and it is spliced with the forecast rainfall at the current moment and the previous moment as the rainfall time series input. At the same time, the prediction result of the probability spatial distribution feature layer of each building with different functions obtained by the FFC model in the previous round is used as the probability spatial distribution feature layer of each building with different functions at the previous moment;
[0158] Similarly, when predicting the probability distribution of all buildings in different functions at the 12th hour, the forecast rainfall sequence of the previous 12 hours is input as the rainfall time series, where the previous 12 hours include the hour corresponding to the current moment, and the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model obtained in the previous round at the 11th hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0159] When predicting the probability distribution of all buildings in different functions at the 13th hour, the forecast rainfall sequence of the previous 12 hours is input as the rainfall time series. The previous 12 hours here include the hour corresponding to the current moment. At the same time, the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model in the previous round at the 12th hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment.
[0160] Similarly, when predicting the probability distribution of all buildings in different functions at the 23rd hour, the forecast rainfall sequence of the previous 12 hours is used as the rainfall time series input, where the previous 12 hours include the hour corresponding to the current moment, and the prediction result of the probability spatial distribution feature layer of each building in different functions obtained by the FFC model obtained in the previous round at the 22nd hour is used as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment;
[0161] When predicting the probability distribution of all buildings in different functions at the 24th hour, the forecast rainfall sequence for the previous 12 hours is input as the rainfall time series, where the previous 12 hours include the hour corresponding to the current moment. At the same time, the FFC model obtained in the previous round is used to predict the probability spatial distribution feature layer of each building in different functions at the 23rd hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment.
[0162] Finally, the probability distribution results of all buildings in different functions at each moment in the next 24-hour rainfall time series in the study area are obtained.
[0163] Taking the rainstorm event that occurred in a certain country F city on September 5 of a certain year as an example to verify the performance of the building function probability prediction method under flood disasters based on the deep learning algorithm described in this embodiment, the hourly average rainfall measured at the stations in the study area is as follows: Fig.10 As shown, the total is 24 hours. For the four building types, the trained FFC model is used to predict the probability distribution of each building in different functions at different times for the measured 24-hour rainfall. In this embodiment, the probability spatial distribution result of all buildings in different functional states at different times of each rainfall is obtained by using the urban waterlogging physical simulation model and the building function probability prediction model based on physical conditions, which is called the label value, and the result obtained by using the trained FFC model is called the predicted value.
[0164] After a 24-hour rolling forecast for the four building types, the following is obtained: Fig.11 The spatial distribution diagram of the average value of the KL divergence for 24 hours is shown. Figures 12 to 15 The following are the bar charts comparing the label values and the predicted values at different times for the four building types, namely, high-rise buildings, multi-story buildings, villas, and duplex villas, which are randomly selected.
[0165] Assuming that the function with the highest probability value of a building at a certain moment is defined as the representative function of the building at that moment, the confusion matrix consisting of the label value and the predicted value of the representative function state at different moments of a high-rise building is as follows: Fig.16 As shown. Fig.16It can be found that at different times, most of the label values and predicted values are around the diagonal of the confusion matrix and its two sides, which indicates that the FFC model described in this embodiment has good prediction ability.
[0166] from Figures 17 to 20 It can be seen that the FFC model described in this embodiment can achieve good prediction for the rolling iterative prediction of the next 24 hours, and can comprehensively obtain the number of buildings in different representative functional states in the entire study area and the spatial distribution characteristics of buildings in different representative functional states.
[0167] The above only describes the basic principles and preferred embodiments of the present invention. Those skilled in the art may make many changes and improvements based on the above description, and these changes and improvements should fall within the protection scope of the present invention.
Claims
1. A method for predicting building function probability under flood disasters based on deep learning algorithm, characterized in that: The steps include: S1. Collect basic data and historical rainstorm time series data of the study area; The basic data include DEM, land use data, river network vector data, drainage network vector data, residential building vector data and raster feature layer data derived from DEM in the study area; the drainage network vector data at least includes inspection wells, outlets, pipe sections, pipe diameter, pipe roughness, and inspection well bottom elevation information; the raster feature layer data derived from DEM at least includes slope, aspect, curvature, and terrain humidity index; Based on the river network vector data, drainage network vector data and the DEM range of the study area, the raster feature layer of the distance to the river network and the raster feature layer of the distance to the outlet are obtained respectively. The DEM, the raster feature layer data derived from the DEM, the raster feature layer of the distance to the river network and the raster feature layer of the distance to the outlet are combined together and called the flood condition factor layer; and each raster feature layer contained in all flood condition factor layers is standardized with a mean of 0 and a standard deviation of 1; S2. Process the data collected in step S1: randomly divide the time series data of several historical rainstorms into three rainfall event sets according to a preset ratio, namely, a rainfall event set for training and verification, a rainfall event set for optimizing the attention coefficients of affected buildings, and a rainfall event set for testing, wherein the number of rainfall event sets for optimizing the attention coefficients of affected buildings and the number of rainfall event sets for testing are both smaller than the number of rainfall event sets for training and verification; then use the urban waterlogging physical simulation model to obtain the spatiotemporal distribution of the flooded water depth, obtain the water depth experienced by each building at different times based on the spatiotemporal distribution of the flooded water depth, and calculate the water depth of each building at different times. The water depths experienced at different times are input into the building function probability prediction model based on physical conditions to obtain the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall; then the building surface vector with the information of the spatial distribution results of the probability of all buildings being in different functions at different times of each rainfall is first converted into a building point vector by finding the centroid, and then converted into a raster layer with the same dimension as the DEM based on the spatial position relationship. Each raster layer corresponds to a possible function. In each raster layer, the value of the grid where the building point vector is located is the probability value of the corresponding function, and the value of the other grids where there is no building point vector is 0; Among them, the physical simulation model of urban waterlogging is based on the coupling of the drainage network model and the two-dimensional surface runoff model; the building function probability prediction model based on physical conditions is used to calculate the probability of buildings of different building types being in different functions at different water depth levels, and the building function categories are divided into 5 types: I, II, III, V, and VII; specifically, function I means that all kinds of needs of all residential units in the current building can be met; function II means that the advanced needs of the residential units at the bottom of the current building cannot be met, while all kinds of needs of other residential units can be met, where the advanced needs at least include needs related to office and cleaning; function III means that the basic needs of the residential units at the bottom of the current building cannot be met, while all kinds of needs of other residential units can be met, where the basic needs at least include basic physiological needs for food; function V means that the safety needs of the residential units at the bottom of the current building cannot be met, while all kinds of needs of other residential units can be met, where the safety needs at least include life safety needs; function VII means that the safety needs of all residential units in the current building cannot be met; the probability of buildings of different building types being in different functions at different water depth levels is obtained according to the vulnerability curve corresponding to each function; S3. Using the data collected in step S1 and processed in step S2, construct an FFC model based on a deep learning algorithm, and train and verify the FFC model; optimize the affected building attention coefficient based on the performance on the rainfall event set used to optimize the affected building attention coefficient, and obtain the optimal affected building attention coefficient for each building type; and test the optimized FFC model on the rainfall event set used for testing; Before training the FFC model, the input features and output variables required by the FFC model are determined as follows: The input features required to generate the FFC model are: For each rainfall event set used for training and validation, the duration length is The hourly rainfall time series is obtained by using a sliding time window method with an interval of 1 hour and a duration of The hourly rainfall time series is used as the time feature, where For odd numbers Round up; train each building type separately. First, create a corresponding building mask layer for each building type. The dimension of the building mask layer is the same as the DEM feature layer. The value of the grid containing the building centroid is 1, and the value of the grid not containing the building centroid is 0. Then, all flood condition factor layers, the building mask layer of the corresponding building type, and the intercepted duration length are The probability spatial distribution feature layers of each building in different functions corresponding to the second to last time step in the hourly rainfall time series are combined to form the input features; Generate the target variable required for the FFC model: The intercept duration length is The probability spatial distribution feature layer of each building in different functions corresponding to the penultimate time step in the hourly rainfall time series is used as the target variable; The duration of each interception will be The corresponding input features obtained from the hourly rainfall time series are combined with the corresponding target variables to form a sample pair, and all sample pairs are divided into sample pairs for training and sample pairs for verification according to a preset ratio, wherein the number of sample pairs for training is greater than the number of sample pairs for verification; The FFC model includes at least an encoder module, a temporal feature extraction module, and a decoder module; The encoder module includes four convolution units and an average pooling layer set after each convolution unit. Each convolution unit includes a convolution layer, a two-dimensional batch normalization, and a leakage corrected linear unit in turn. The size of the input feature map is converted from height × width × number of channels to H × W × (8 + number of possible feature classes) to height × width × number of channels ,when and When it is not divisible and Round down; In the time feature extraction module, for each duration of The rainfall time series of each hour is first extracted through two LSTM models. Then, the hidden features output by the LSTM model are normalized through one-dimensional batch normalization. Then, a fully connected layer is used to convert the data dimensions (i.e., rainfall time steps × number of channels × height × width) into , which is then processed by a leaky rectified linear unit and a dropout module and reshaped into Tensor modules, the dimensions of each tensor are height × width × number of channels. , and then Each tensor in the tensor module is individually connected to the output of the encoder module along the channel dimension, and each combined tensor is used as the input of the decoder module; The decoder module includes four upsampling layers and two-dimensional convolutional transposition units respectively arranged after each upsampling layer. Each two-dimensional convolutional transposition unit includes a two-dimensional convolutional transposition layer, a two-dimensional batch normalization and a leakage corrected linear unit. Each output result of the decoder module is processed by a convolutional layer and a leakage corrected linear unit to convert the number of channels into the dimension matching the output result of the FFC model. The time step corresponds to The output of the decoder module, traversing After the tensor module corresponding to each time step is obtained at each time step and the result of the encoder module is combined and input into the decoder module, the results of all time steps are stacked together in the channel dimension, and then passed through the convolution layer, the leaky rectified linear unit and the softmax function, so that the output duration length used is The predicted results of the probability spatial distribution of each building in different functions corresponding to the penultimate time step in the hourly rainfall time series; S4. Input the predicted rainfall time series into the FFC model tested for each building type, perform multiple rolling iterative forecasts, and obtain the probability distribution results of each building in the study area being in different functions within a certain time range in the future.
2. The method for predicting building function probability under flood disasters based on deep learning algorithm according to claim 1 is characterized in that: In step S1, the acquisition of historical rainstorm time series data is as follows: Get The rainfall data during a typhoon, assuming that the duration of the original rainfall data is hours, of which , ; Using the sliding window method, based on The time interval is hours, and the duration of each game is The hourly rainfall data is divided into durations of The hourly rainfall time series will be The rainfall time series with hourly total rainfall greater than the preset threshold is used as the uniform rainfall event for the input of the urban waterlogging physical simulation model, where , ,Uniform rainfall events refer to rainfall events in which the hourly rainfall in the study area is spatially uniformly distributed.
3. The method for predicting building function probability under flood disasters based on deep learning algorithm according to claim 2 is characterized in that: When training the FFC model in step S3, in order to solve the problem of extremely unbalanced distribution of disaster-affected building samples and non-disaster-affected building samples, the custom loss function is as follows: In the above formula, Represents the number of all grids in a single channel, represents the number of channels of samples in the target variable, Indicates the number of samples contained in each batch, and They represent the building function probability prediction model based on physical conditions and the probability prediction model based on FFC model in a batch. The sample The first The probability value at each grid, is a natural constant, It is the affected building attention coefficient, which is used to strengthen the FFC model's attention to buildings that produce functional losses; Then, the Adam optimizer is used to optimize the various parameters in the encoder module, decoder module, and temporal feature extraction module in the FFC model by minimizing the loss.
4. The method for predicting building function probability under flood disasters based on deep learning algorithm according to claim 3 is characterized in that: In step S3, the FFC model is trained and verified; the affected building attention coefficient is optimized based on the performance on the rainfall event set for optimizing the affected building attention coefficient, and the optimal affected building attention coefficient for each building type is obtained; and the optimized FFC model is tested on the rainfall event set for testing, specifically including the following: Before training each building type, preset the affected building attention coefficient The loss function is clear at this time, and the FFC model training and verification can be carried out. After each training, the parameter combination in the encoder module, decoder module and time feature extraction module with the lowest average loss value on the sample pairs used for verification is selected as the final choice of this training, thereby completing the training of the FFC model; The concentrated duration of rainfall events based on several fields for optimizing the affected building attention factor is The trained FFC model is evaluated on the rainfall event set used to optimize the affected building attention coefficients: for each rainfall event, the previous The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment under the hourly rainfall time series; Statistics using different affected building attention coefficients in custom loss functions The trained FFC model is used to select the affected building attention coefficients in several rainfall events. The average value of the kl divergence of all buildings at all times is taken, and the corresponding value is the smallest when the average value of the kl divergence is taken. As the optimal affected building attention coefficient; for each building type, the model training result obtained by applying the loss function corresponding to the optimal affected building attention coefficient is used as the final model parameter; The duration of the concentrated rainfall events based on several fields used for testing is The optimal FFC model is tested on the rainfall event set for testing: for each rainfall event, the previous The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment in the hourly rainfall time series; then based on the simulation results and prediction results of the probability distribution of each building in different functions at each moment in all rainfall events, a scatter plot is drawn and fitted to obtain an indicator to measure the goodness of fit. , if The performance of the FFC model meets the preset requirements.
5. The method for predicting building function probability under flood disasters based on deep learning algorithm according to claim 4 is characterized in that: For each rainfall, The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment under the hourly rainfall time series include: For a duration of When predicting the probability distribution of all buildings in different functions in the first hour, the previous The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is combined with the forecast rainfall at the current moment as the rainfall time series input. At the same time, it is considered that the probability that all buildings are in function I at the moment before the current moment is 100%, and the probability of being in other functions is 0%; When predicting the probability distribution of all buildings in different functions in the second hour, The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is spliced with the forecast rainfall at the current moment and the previous moment as the rainfall time series input. At the same time, the prediction result of the probability spatial distribution feature layer of each building with different functions obtained by the FFC model in the previous round is used as the probability spatial distribution feature layer of each building with different functions at the previous moment. By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment; When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment; By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment; When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted results of the probability spatial distribution feature layers of each building with different functions are obtained in hours as the probability spatial distribution feature layers of each building with different functions at the moment before the current moment.
6. The method for predicting building function probability under flood disasters based on a deep learning algorithm according to any one of claims 2 to 5, characterized in that: In step S4, the predicted rainfall time series is input into the FFC model tested for each building type, and multiple rolling iterative predictions are performed to obtain the probability distribution results of each building in different functions within a certain time range in the future in the study area, including: Set before The hourly rainfall and the probability distribution of all buildings in different functions at the previous moment are used as input, where the previous The hourly rainfall includes the rainfall at the current moment, and each rainfall is Rolling iteration forecast, through rolling iteration Each time, the FFC model predicts the probability spatial distribution feature layer of each building in different functions at the current moment as the model input for prediction at the next moment, thereby obtaining a complete The probability distribution results of all buildings in different functions at each moment under the hourly rainfall time series include: For a duration of When predicting the probability distribution of all buildings in different functions in the first hour, the previous The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is combined with the forecast rainfall at the current moment as the rainfall time series input. At the same time, it is considered that the probability that all buildings are in function I at the moment before the current moment is 100%, and the probability of being in other functions is 0%; When predicting the probability distribution of all buildings in different functions in the second hour, The forecast rainfall sequence of the previous hour is padded with 0, that is, The rainfall in each hour is 0, which is spliced with the forecast rainfall at the current moment and the previous moment as the rainfall time series input. At the same time, the prediction result of the probability spatial distribution feature layer of each building with different functions obtained by the FFC model in the previous round is used as the probability spatial distribution feature layer of each building with different functions at the previous moment. By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment; When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment; By analogy, when predicting When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted result of the probability spatial distribution feature layer of each building in different functions is obtained in one hour as the probability spatial distribution feature layer of each building in different functions at the moment before the current moment; When the prediction When all buildings are in different functional probability distributions for one hour, The hourly forecast rainfall series is used as the rainfall time series input. The hour includes the hour corresponding to the current time, and the FFC model obtained in the previous round is applied to the The predicted results of the probability spatial distribution feature layers of each building with different functions are obtained in hours as the probability spatial distribution feature layers of each building with different functions at the moment before the current moment.
