A method and system for predicting and monitoring urban flood prevention waterlogging

Abstract

This application provides a method and system for predicting and monitoring urban flooding, relating to the technical field of urban flood control. The method includes: collecting flooding-related data and physical sensing data; determining the risk classification results of each monitoring unit based on a pre-set urban risk classification knowledge base; dynamically fusing the risk classification results and physical sensing data to construct a predictive feature set; extracting high-dimensional feature vectors through a deep belief network; constructing a risk transmission network based on the drainage network topology and surface runoff paths; introducing an ensemble learning framework to optimize the high-dimensional feature vectors; and generating propagation features through risk propagation analysis under the constraints of the risk transmission network; finally, processing the propagation features using a logistic regression coupling model to obtain risk prediction results, thereby generating control parameters for pumping stations, gates, and early warning systems. This application achieves dynamic prediction and precise control of flooding risk.

Description

Technical Field

[0001] This application relates to the technical field of urban flood control, and in particular to a method and system for predicting and monitoring urban flooding. Background Technology

[0002] In the field of urban flood control and emergency management, accurate prediction and dynamic monitoring of urban flooding risks are key to enhancing urban resilience and ensuring public safety, and have significant application value.

[0003] Currently, common methods for predicting urban flooding mainly rely on hydrological and hydrodynamic numerical simulations, or on building a single early warning model based on historical statistics and real-time monitoring data. For example, sensor networks are used to collect water level information and combine it with thresholds to issue alarms.

[0004] However, existing methods often fail to fully consider the complex spatial relationships formed by drainage networks and surface runoff paths when analyzing urban flooding risks. This makes it difficult to depict the dynamic diffusion and superposition of risks along these actual paths. Consequently, risk assessment results are mostly static or limited to scattered locations and cannot effectively reflect the spatial transmission and chain effects of risks throughout the urban area. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for predicting and monitoring urban flooding, so as to solve the problem that it is difficult to achieve high-precision, dynamic, and direct-supporting urban flooding risk prediction in the existing technology.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for predicting and monitoring urban flooding and waterlogging, comprising:

[0007] Collect data related to urban flooding in the city center and physical sensing data reflecting the water accumulation situation;

[0008] Based on the aforementioned waterlogging-related data and the pre-established urban waterlogging risk classification knowledge base, the risk classification results of each monitoring unit in the central urban area are determined.

[0009] The risk classification results are dynamically correlated and fused with the physical sensing data to construct a predictive feature set for characterizing the real-time evolution of urban flooding risk. A deep belief network is used to extract high-dimensional features and mine coupling relationships in the predictive feature set to generate a high-dimensional feature vector.

[0010] Based on the drainage network topology and surface water runoff paths of the urban center, a risk transmission network is constructed with drainage network nodes and surface water runoff points as network nodes, drainage network connection lines and surface water runoff paths as edges, and flow capacity as edge weights.

[0011] An ensemble learning framework is introduced to optimize the high-dimensional feature vectors, thereby obtaining ensemble feature vectors that characterize the risk status of each network node. Under the constraints of the risk transmission network, risk propagation analysis is performed on the ensemble feature vectors to generate propagation features that characterize the direction and intensity of risk propagation. The ensemble learning framework employs a random forest model based on the Bagging ensemble method and a gradient boosting tree model based on the Boosting ensemble method.

[0012] By using a logistic regression coupling model to process the propagation characteristics, risk prediction results for each monitoring unit in the central urban area of ​​the city are obtained for a period of time in the future. Based on the risk prediction results, risk control parameters are generated to indicate the start and stop of pumping stations, gate scheduling, or the issuance of graded early warning information.

[0013] Optionally, the step of dynamically associating and fusing the risk classification results with the physical sensing data to construct a predictive feature set for characterizing the real-time evolution of urban flooding risk, and using a deep belief network to extract high-dimensional features and mine coupling relationships in the predictive feature set to generate a high-dimensional feature vector, includes:

[0014] Assign a feature vector representing the inherent vulnerability to the static risk level of each monitoring unit in the risk classification result, as the static risk feature of the corresponding monitoring unit, and organize the physical sensing data of each monitoring unit in the same time step into a dynamic situation vector representing the corresponding real-time situation.

[0015] The static risk features corresponding to each monitoring unit at each time step are concatenated with the dynamic situation vector along the feature channel dimension to obtain the fused feature vector of the corresponding monitoring unit.

[0016] The fused feature vectors generated by all monitoring units at all time steps are arranged in a structured manner according to spatial location and temporal order to form a predictive feature set;

[0017] The first feature extraction layer of the deep belief network performs sliding window aggregation of the first temporal window and spatial smoothing based on Gaussian kernel function on the predicted feature set to extract the first type of features reflecting the intensity of the heat island effect from the predicted feature set.

[0018] The second feature extraction layer of the deep belief network is used to perform sliding window difference of the second time window and spatial gradient enhancement based on the Sobel operator on the predicted feature set, so as to extract the second type of features reflecting the water accumulation and collection process from the predicted feature set.

[0019] Through the coupling interaction layer of the deep belief network, the interaction relationship between the first type of features and the second type of features is simulated and the features interact, generating coupled interaction features;

[0020] The coupling effect features are represented and learned through the feature aggregation layer of a deep belief network to generate a high-dimensional feature vector.

[0021] Optionally, the step of simulating the interaction relationship and feature interaction between the first type of features and the second type of features through the coupling interaction layer of the deep belief network to generate coupling interaction features includes:

[0022] In the coupled interaction layer of the deep belief network, the first type of features are converted into source feature maps, and the second type of features are converted into medium feature maps;

[0023] The propagation process of the driving potential represented by the source feature map is simulated by the propagation kernel function in the coupled interaction layer under the first constraint of the local water accumulation state represented by the medium feature map and the second constraint corresponding to the preset urban underlying surface attribute data, thereby generating a propagation feature map.

[0024] By coupling the gated loop unit in the interaction layer, a dynamic gate factor is generated based on the local water accumulation state corresponding to the medium feature map. The dynamic gate factor is then used to adaptively enhance the propagation feature map to generate an enhanced feature map.

[0025] By using the fusion layer in the coupling interaction layer, the enhanced feature map is superimposed with the medium feature map, and the superposition result is adjusted based on the preset attribute data of the urban underlying surface to simulate the combined effect of confluence acceleration and lag buffering, thereby generating coupling effect features.

[0026] Optionally, under the constraints of the risk transmission network, performing risk propagation analysis on the integrated feature vector to generate propagation features characterizing the direction and intensity of risk propagation includes:

[0027] The propagation process of risk state on the risk transmission network is iteratively simulated based on integrated feature vectors and graph diffusion model until the state stabilizes or reaches a preset number of iterations, and a diffusion feature vector is generated.

[0028] The diffusion feature vector is subjected to path extraction based on the maximum gradient direction and intensity quantization based on the vector norm to generate propagation features that characterize the direction and intensity of risk propagation.

[0029] Secondly, this application provides a predictive monitoring system for urban flood control and waterlogging, comprising:

[0030] The data acquisition module is used to collect data related to urban flooding in the city center and physical sensing data reflecting the water accumulation situation.

[0031] The determination module is used to determine the risk classification results of each monitoring unit in the central urban area of ​​the city based on the waterlogging-related data and a pre-set urban waterlogging risk classification knowledge base.

[0032] The mining module is used to dynamically associate and fuse the risk classification results with the physical perception data, construct a predictive feature set to characterize the real-time evolution of urban flooding risk, and use a deep belief network to extract high-dimensional features and mine coupling relationships in the predictive feature set to generate a high-dimensional feature vector.

[0033] The construction module is used to construct a risk transmission network based on the drainage network topology and surface water runoff path of the urban center, with drainage network nodes and surface water runoff points as network nodes, drainage network connection lines and surface water runoff paths as edges, and flow capacity as edge weights.

[0034] The analysis module is used to introduce an ensemble learning framework to optimize the high-dimensional feature vector, obtain an ensemble feature vector representing the risk state of each network node, and perform risk propagation analysis on the ensemble feature vector under the constraints of the risk transmission network to generate propagation features representing the direction and intensity of risk propagation. The ensemble learning framework adopts a random forest model based on the Bagging ensemble method and a gradient boosting tree model based on the Boosting ensemble method.

[0035] The processing module is used to process the propagation characteristics using a logistic regression coupling model to obtain the risk prediction results of each monitoring unit in the central urban area of ​​the city for a period of time in the future. Based on the risk prediction results, risk control parameters are generated to indicate the start and stop of pumping stations, gate scheduling, or the issuance of graded early warning information.

[0036] Thirdly, this application provides an electronic device, comprising:

[0037] Memory, used to store computer programs;

[0038] A processor is configured to execute the computer program to implement the steps of a method for predicting and monitoring urban flooding as described in the first aspect above.

[0039] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the urban flood control and waterlogging prediction and monitoring method described in the first aspect above.

[0040] The urban flood control and waterlogging prediction and monitoring method provided in this application has the following beneficial effects: First, by collecting multi-source waterlogging-related data and real-time physical sensing data, this application can provide a comprehensive information foundation for risk analysis; then, based on a pre-set knowledge base, it determines the risk classification results of each monitoring unit, enabling a rapid and standardized assessment of the inherent risk level of urban areas; subsequently, it integrates static classification results with dynamic sensing data and uses a deep belief network to extract high-dimensional features, which can deeply explore the coupling relationship between static vulnerability and dynamic situation, thereby forming a comprehensive feature representation that depicts the evolution of waterlogging; next, it constructs a weighted transmission network based on drainage pipe network and surface water runoff path, which can provide an accurate physical topology for simulating the spatial diffusion of risk; then, it introduces an integrated learning framework of random forest and gradient boosting tree to optimize features and conducts propagation analysis under network constraints, which can simulate and reveal the dynamic propagation direction and intensity of risk in urban areas; finally, it processes the propagation features through a logistic regression coupling model to obtain prediction results and generate control parameters, which can directly output instructions to guide pump station and gate scheduling and graded early warning, realizing a decision-making closed loop from prediction to control.

[0041] Furthermore, this application forms a spatiotemporal feature set by concatenating static risk features with dynamic situation vectors, and uses different extraction layers of a deep belief network to analyze the characteristics of the heat island effect and water accumulation process, and then performs interactive simulation and aggregation. This can effectively separate and strengthen the core physical process features driving urban flooding from multi-source data, providing a feature foundation with high discrimination and clear physical meaning for subsequent risk propagation analysis. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating a method for predicting and monitoring urban flooding and waterlogging, provided in an embodiment of this application;

[0044] Figure 2 A schematic diagram illustrating a specific implementation of a method for predicting and monitoring urban flooding and waterlogging, as provided in this application embodiment;

[0045] Figure 3 A schematic diagram of the structure of an urban flood control and waterlogging prediction and monitoring system provided in this application embodiment;

[0046] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0047] In urban flood forecasting, existing methods often process static geographic data and dynamic monitoring data separately, making it difficult to depict the real-time evolution of risks. At the same time, the analysis process often ignores the actual transmission network formed by drainage pipe networks and surface runoff paths, which results in risk assessment being limited to a single point and failing to reflect the dynamic diffusion and superposition of risks in space. In addition, there is a lack of direct correlation between early warning results and specific control actions such as pump station and gate scheduling, making it difficult for early warnings to be quickly transformed into effective instructions.

[0048] To address this issue, this application proposes a method for predicting and monitoring urban flooding. The core of this method lies in: fusing static risk classification with real-time sensing data, and utilizing deep neural networks to uncover their inherent correlations, forming a comprehensive characteristic representing the flooding situation; then constructing a physical transmission network based on actual drainage and catchment paths, mapping the aforementioned characteristics to network nodes, and simulating the dynamic propagation process of risk in space; finally, transforming the propagation characteristics into specific risk levels through a coupled prediction model, and automatically generating control parameters that can directly guide the operation of pumping stations and gates. This method achieves a closed loop from data fusion and spatial propagation simulation to intelligent decision-making, effectively overcoming the problems of static prediction, isolated analysis, and disconnect from control in existing technologies.

[0049] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] The core of this application is to provide a method for predicting and monitoring urban flooding and waterlogging, and a flowchart of one specific implementation method is shown below. Figure 1 As shown, the method includes:

[0051] S101. Collect waterlogging-related data in the central urban area and physical sensing data reflecting the water accumulation situation.

[0052] Among them, waterlogging-related data refers to basic information used to assess the inherent waterlogging vulnerability of a region. Waterlogging-related data may include heat island effect data, surface runoff data, etc.

[0053] Physical sensing data refers to dynamic information that is captured in real time by sensing devices and directly represents the current hydrological and meteorological conditions. Physical sensing data can include precipitation, water depth, water flow velocity, etc.

[0054] In step S101, firstly, the focus is on the high-value, high-risk area of ​​the city center, and data on the heat island effect and surface runoff within this area are specifically extracted from the geographic information database and meteorological historical archives, thereby forming waterlogging-related data on the long-term causes of waterlogging.

[0055] Meanwhile, through a dense network of sensor devices deployed at key nodes of the city's drainage pipe network, flood-prone road sections, and meteorological stations, physical sensing data such as rainfall, water depth, and water flow velocity are collected in real time and automatically, thereby accurately capturing the real-time water situation dynamics within the city.

[0056] S102. Based on the aforementioned waterlogging-related data and the pre-set urban waterlogging risk classification knowledge base, determine the risk classification results of each monitoring unit in the central urban area.

[0057] The monitoring units are pre-divided based on the topology of the urban drainage network and the surface runoff area. In this embodiment, the division criteria are not limited and can be set according to the actual situation.

[0058] In one specific implementation, step S102 includes:

[0059] Step 1021: Perform coupling analysis on the heat island effect data and surface runoff data in the waterlogging-related data belonging to the same monitoring unit in the central urban area of ​​the city to generate the coupling coefficient of each monitoring unit.

[0060] In step 1021, the heat island effect data and historical surface runoff data corresponding to each monitoring unit are first extracted. Then, a mathematical model is applied to evaluate how the high-temperature underlying surface changes surface permeability, accelerates the evaporation-condensation cycle, and ultimately affects rainfall runoff efficiency. The two originally independent static data within a monitoring unit are then linked together, and a specific value characterizing their coupling strength is output as the coupling coefficient.

[0061] It should be noted that the specific implementation of this mathematical model, such as the construction method based on multiple regression or mechanistic equations, can refer to relevant technologies, and will not be elaborated on in the embodiments of this application.

[0062] Step 1022: Using the grading rules in the pre-set urban flooding risk grading knowledge base, the coupling coefficient is matched to obtain the preliminary risk level of each monitoring unit. The grading rules define the correspondence between the coupling coefficient range and the risk level.

[0063] In step 1022, the coupling coefficient can be a numerical value. Each calculated coupling coefficient is automatically compared with the classification rules stored in the urban waterlogging risk classification knowledge base. This process assigns a preliminary risk level label, such as "low", "medium" or "high", to each monitoring unit based on which predefined threshold range the value of the coupling coefficient falls into, thereby completing the first conversion from continuous numerical values ​​to discrete risk levels.

[0064] It should be noted that the threshold matching logic involved in this automatic comparison process can be referred to in relevant technologies, and will not be elaborated on in the embodiments of this application.

[0065] Step 1023: Based on the topographic elevation difference between adjacent monitoring units and the connectivity of the drainage network, the preliminary risk level is spatially verified and adjusted to generate the risk classification results for each monitoring unit in the central urban area.

[0066] In step 1023, topographic and pipeline topology data reflecting spatial correlation are introduced, and the above preliminary levels are verified and corrected for rationality. For example, if a monitoring unit with a preliminary level of "medium risk" is significantly lower than a surrounding "high risk" unit, and the drainage pipes between the two are directly connected, the actual risk of the monitoring unit may be increased because the monitoring unit is very likely to receive water runoff from higher places.

[0067] Conversely, if a "high-risk" unit has an independent and unobstructed drainage outlet, its risk level may be appropriately lowered. Then, by traversing all adjacent monitoring units and making adjustments, a risk classification result that is spatially consistent, uniform, and more in line with the actual physical process is finally output.

[0068] It should be noted that the specific implementation algorithm used for spatial verification and adjustment here can refer to relevant spatial analysis techniques, and will not be described in detail in the embodiments of this application.

[0069] This application quantifies the inherent coupling relationship between heat island and runoff data and maps it to a preliminary risk level. Then, it combines topography and pipeline spatial structure for verification and adjustment. This generates a risk base map that accurately reflects the static vulnerability of different urban areas and conforms to hydrological logic, thus laying a reliable foundation for subsequent real-time prediction by integrating dynamic data.

[0070] S103. Dynamically associate and fuse the risk classification results with the physical sensing data to construct a predictive feature set for characterizing the real-time evolution of urban flooding risk. Use a deep belief network to extract high-dimensional features and mine coupling relationships in the predictive feature set to generate a high-dimensional feature vector.

[0071] The deep belief network is configured to extract a first type of feature related to the spatiotemporal distribution of urban heat island intensity and a second type of feature related to the spatiotemporal evolution of water accumulation and collection processes, and to perform interactive analysis on the first and second types of features. Furthermore, this embodiment does not limit the structural design, parameter design, or training process of the deep belief network; it can be configured according to actual conditions.

[0072] In one specific implementation, such as Figure 2 As shown, step S103 includes:

[0073] Step 1031: Assign a feature vector representing the inherent vulnerability of each monitoring unit in the risk classification result as the static risk feature of the corresponding monitoring unit, and organize the physical sensing data of each monitoring unit in the same time step into a dynamic situation vector representing the corresponding real-time situation.

[0074] Among them, static risk characteristics refer to a vector composed of numerical values, which uses embedding coding technology to map discrete static risk levels into a continuous vector representation that can be processed by a computer; dynamic situation vector refers to a vector composed of multiple physical sensing data collected within the same time step arranged in a predetermined order.

[0075] In step 1031, firstly, for each monitoring unit corresponding to the risk classification result, a pre-trained embedding layer is used to convert the static risk level of each monitoring unit into a dense vector of fixed dimensions, which serves as the static risk feature of the unit. At the same time, the precipitation, water depth, and water flow velocity, and other physical sensing data collected by each monitoring unit within the same time step are spliced ​​and standardized according to a unified feature order to form the dynamic situation vector of the unit.

[0076] For example, the static risk level of monitoring unit A is "high risk". Then, through the embedding layer mapping, "high risk" is encoded into a three-dimensional vector, and the static risk feature vector VsA is obtained. Then, within the same time step t1, the precipitation collected by the monitoring unit is 10 mm, the water depth is 5 cm, and the water flow velocity is 0.5 m / s. These three values ​​are then arranged in order to obtain the dynamic situation vector VdA(t1).

[0077] Step 1032: Concatenate the static risk features corresponding to each monitoring unit at each time step with the dynamic situation vector along the feature channel dimension to obtain the fused feature vector of the corresponding monitoring unit.

[0078] In step 1032, for each monitoring unit at each time step, its corresponding static risk feature vector and corresponding dynamic situation vector are concatenated in the feature dimension. Specifically, the element sequences of the two vectors are concatenated end to end and combined into a higher-dimensional fusion feature vector.

[0079] For example, for monitoring unit A at time step t1, its static risk feature vector VsA and dynamic situation vector VdA are concatenated to obtain the fused feature vector FA(t1), and the length of this vector is 6, with the first three elements coming from static features and the last three elements coming from dynamic data.

[0080] Step 1033: Arrange the fused feature vectors generated by all monitoring units at all time steps in a structured manner according to spatial location and temporal order to form a predicted feature set.

[0081] In step 1033, the fused feature vectors generated by all monitoring units at all time steps are systematically organized into a three-dimensional tensor data structure according to the spatial numbering order of the monitoring units and the order of the time steps. The first dimension of this tensor corresponds to the spatial index of the monitoring unit, the second dimension corresponds to the time step sequence, and the third dimension corresponds to the length of the fused feature vector. In this way, a predictive feature set containing spatial distribution, temporal evolution and multi-dimensional feature information is constructed.

[0082] For example, suppose there are three monitoring units in the system, numbered A, B, and C, with two time steps, t1 and t2, and each fused feature vector has a length of 6. Then the predicted feature set X is a tensor with a shape of 3×2×6. In this way, all the data is integrated into a structured data block.

[0083] Step 1034: Through the first feature extraction layer of the deep belief network, the predicted feature set is subjected to sliding window aggregation of the first temporal window and spatial smoothing processing based on Gaussian kernel function, so as to extract the first type of features reflecting the intensity of the heat island effect from the predicted feature set.

[0084] The first time-series window refers to a fixed-length time interval, which is used to slide across the time series to extract aggregated information within the time period. In this embodiment, the size of the fixed length is not limited and can be set according to the actual situation. Sliding window aggregation refers to performing operations such as averaging on the features of all time steps covered by the window within each sliding window.

[0085] Spatial smoothing based on Gaussian kernel function refers to using a weight matrix generated by a two-dimensional Gaussian function to perform a weighted average of the features of spatially adjacent monitoring units in order to filter out noise and enhance spatial continuity.

[0086] In step 1034, the first feature extraction layer of the deep belief network takes the predicted feature set tensor as input. This layer first slides a first time-series window of a preset size in the time dimension and calculates the average value of the feature values ​​within the window along the time axis, thereby compressing the time-series information into the aggregated features of each window. Subsequently, in the spatial dimension, this layer applies a two-dimensional Gaussian kernel with a preset standard deviation to perform a convolution operation on the aggregated features of all monitoring units at the same time, so that the feature value of each unit is affected by the features of its neighboring units, and finally outputs a smoothed feature map, which is the first type of feature reflecting the cumulative effect of the large-scale thermal environment.

[0087] For example, assuming the first time window size is 3, that is, averaging the features over three consecutive time steps, and assuming that a certain feature value for monitoring unit A in the three consecutive time steps is 2.0, 2.2, and 2.1 respectively, then the value after sliding window aggregation is (2.0+2.2+2.1) / 3=2.1; then in the spatial smoothing process, the standard deviation of the Gaussian kernel is set to 1.0, and a 3×1 weight matrix [0.25, 0.5, 0.25] is generated; then for monitoring unit A, its aggregated feature values ​​with neighboring units B and C are 2.1, 1.9, and 2.0 respectively, then the smoothed feature value is 2.1×0.25+1.9×0.5+2.0×0.25=1.975, thus obtaining an element in the first type of feature corresponding to monitoring unit A as 1.975.

[0088] Step 1035: Through the second feature extraction layer of the deep belief network, the predicted feature set is subjected to sliding window difference of the second time window and spatial gradient enhancement based on the Sobel operator, so as to extract the second type of features reflecting the water accumulation and collection process from the predicted feature set.

[0089] Among them, spatial gradient enhancement based on the Sobel operator refers to using the Sobel operator to calculate the spatial gradient of the feature map in the horizontal and vertical directions, so as to highlight the abrupt change regions in space by the gradient magnitude.

[0090] In step 1035, the second feature extraction layer of the deep belief network also receives the predicted feature set. In the time dimension, this layer slides a second time window of a preset size and calculates the difference between the feature values ​​of adjacent time steps within the window to capture the instantaneous change trend of key indicators such as water depth and flow velocity. Then, in the spatial dimension, this layer applies the Sobel operator to the feature map at each time point to calculate its spatial gradient and adds the calculated gradient magnitude to the original feature map to enhance the signal strength of areas with significant spatial changes such as confluence paths and water boundaries. Finally, it outputs a second type of feature that can reflect the dynamic generation and confluence process of local water accumulation.

[0091] For example, assuming the second time window size is 2, which calculates the difference between two adjacent time steps, and assuming that for the water depth characteristic of monitoring unit A, the values ​​at times t1 and t2 are 5.0 and 5.8 respectively, then the value obtained by sliding window differencing is... Subsequently, in spatial gradient enhancement, the Sobel horizontal direction operator is used. Calculations are performed, assuming the difference eigenvalues ​​of monitoring unit A and its adjacent monitoring units B and C are 0.8, 0.5, and 0.7 respectively, then the approximate value of the horizontal gradient is... The absolute value of 0.1 is taken as the gradient magnitude contribution. Finally, this gradient magnitude is added to the original feature value to obtain the enhanced feature value of 0.8 + 0.1 = 0.9, that is, one element of the second type of feature corresponding to monitoring unit A is 0.9.

[0092] Step 1036: Through the coupling interaction layer of the deep belief network, the interaction relationship between the first type of features and the second type of features is simulated and the features interact to generate coupling interaction features.

[0093] Step 1036 may specifically include the following steps:

[0094] Step a1: In the coupled interaction layer of the deep belief network, the first type of features are converted into source feature maps, and the second type of features are converted into medium feature maps.

[0095] The source feature map refers to the first type of feature representation obtained after feature transformation, which is used to characterize the large-scale external driving potential energy that causes urban flooding; the medium feature map refers to the second type of feature representation obtained after feature transformation, which abstractly represents the local environmental state within the city that receives and responds to the aforementioned driving potential energy. Furthermore, this embodiment does not limit the representation of the two types of maps.

[0096] In step a1, the coupling interaction layer first uses two independent linear transformation layers to process the first type of features and the second type of features respectively. These linear transformations are implemented through learnable weight matrices, with the aim of adjusting the channel dimension of the features and extracting a more abstract representation. After the transformation, the first type of features are converted into source feature maps, and the second type of features are converted into medium feature maps. Both are consistent in spatial size and represent the driving factor and the carrying medium, respectively.

[0097] For example, in the first type of features, the feature value of monitoring unit A is f1 = 1.975, and in the second type of features, the corresponding feature value of monitoring unit A is f2 = 0.9. Then, a linear transformation layer multiplies each feature by a learnable weight and adds a bias. For example, for the source feature map, the weight ws = 0.8 and the bias bs = 0.1, then the transformed value of the source feature map in monitoring unit A is SA = ws × f1 + bs = 0.8 × 1.975 + 0.1 = 1.68. Similarly, for the medium feature map, the weight wm = 1.2 and the bias... Then the value of the transformed medium characteristic map in cell A is .

[0098] Step a2: Simulate the propagation process of the driving potential represented by the source feature map under the first constraint of the local water accumulation state represented by the medium feature map and the second constraint corresponding to the preset urban underlying surface attribute data by simulating the propagation kernel function in the coupled interaction layer, and generate a propagation feature map.

[0099] The propagation kernel function is a learnable parameterized function used to simulate diffusion or propagation phenomena in a physical field. This embodiment does not limit the expression of this function.

[0100] The first constraint refers to the impact of the local water accumulation state, expressed by the medium characteristic map, on the potential energy propagation path and efficiency. The second constraint refers to the regulating effect of the surface characteristics, described by the preset urban underlying surface attribute data, on potential energy propagation. The attribute data includes underlying surface roughness, permeability data, etc.

[0101] In step a2, the propagation kernel function in the coupled interaction layer takes the source feature map, medium feature map, and underlying surface attribute data as input. The core of this function is a conditional convolution operation, and its convolution kernel weights are not fixed, but dynamically generated based on the current medium feature map and underlying surface data. Specifically, a weight generation network takes the medium feature map and underlying surface data as input and outputs the parameters of the convolution kernel. Then, this dynamically generated convolution kernel is used to perform a convolution operation on the source feature map to obtain the propagation feature map. This allows the propagation mode of potential energy to be adaptively adjusted according to the real-time water accumulation and static surface attributes.

[0102] For example, taking monitoring unit A and its adjacent units B and C as an example, assuming that the source feature values ​​of units A, B, and C are SA=1.68, SB=1.50, and SC=1.60 respectively, the weight generation network generates a 3×1 convolution kernel weight θ=[0.2, 0.6, 0.2] based on the medium feature value MA=0.88 and the underlying surface attribute value UA=0.6 of monitoring unit A. Then, the propagation feature value PA of the propagation feature map in monitoring unit A is obtained by convolution calculation: PA=0.2×SB+0.6×SA+0.2×SC=1.628.

[0103] Step a3: By using the gated loop unit in the coupling interaction layer, a dynamic gating factor is generated based on the local water accumulation state corresponding to the medium feature map. The dynamic gating factor is then used to adaptively enhance the propagation feature map to generate an enhanced feature map.

[0104] In step a3, to further refine the propagation results, the medium feature map is input into a gated loop unit module. The gated loop unit module then processes the feature sequence of each spatial location sequentially. Internally, it uses update gate and reset gate mechanisms to calculate the hidden state at the current time step based on the current input and the hidden state at the previous time step. This hidden state is used as a dynamic gating factor. Then, the dynamic gating factor is multiplied element-wise with the propagation feature map to achieve adaptive enhancement or suppression of the propagation features, thereby obtaining an enhanced feature map.

[0105] For example, for monitoring unit A, its medium characteristic value MA=0.88 is used as the input of the gated loop unit module. Assuming the previous hidden state hprev=0.5, update gate weight wz=0.7, reset gate weight wr=0.5, candidate hidden state weight wh=0.6, and biases are ignored, then the update gate output is z=σ(wz×MA+hprev), where σ is the Sigmoid function, and the reset gate output is r=σ(wr×MA+hprev); the candidate hidden state is... =tanh(wh×MA+r×hprev), where tanh is a commonly used non-linear activation function in neural networks that takes a real input and compresses and transforms the real number to a value between tanh and hprev. A value between 1 and 0; the current hidden state is 1. This hidden state is the dynamic gating factor GA. Since the propagation characteristic value PA of monitoring unit A is 1.628, the enhanced characteristic value EA = PA × GA = 1.073 when GA is 0.659.

[0106] Step a4: The enhanced feature map and the medium feature map are superimposed through the fusion layer in the coupling interaction layer, and the superposition result is adjusted based on the preset attribute data of the urban underlying surface to simulate the combined effect of confluence acceleration and lag buffering, and generate coupling effect features.

[0107] The attribute data also includes the flow coefficient of the underlying surface.

[0108] In step a4, the fusion layer adds the enhanced feature map to the original medium feature map element by element to obtain a superimposed feature map. In order to simulate the acceleration or buffering effect of different underlying surfaces on water flow, the preset underlying surface attribute data generates a spatial adjustment coefficient matrix through a fully connected layer, and multiplies the coefficient matrix element by element with the superimposed feature map to make the final scene adaptation adjustment of the feature values, thereby outputting the coupling effect feature.

[0109] For example, based on the underlying surface attribute value UA=0.6 of monitoring unit A, and through a linear transformation, an adjustment coefficient αA=wα×UA+bα is generated. Assuming coefficient wα=0.8 and coefficient bα=0.5, then αA=0.8×0.6+0.5=0.98. Finally, the coupling effect characteristic value of monitoring unit A is CA=(EA+MA)×αA≈1.914.

[0110] Step 1037: Through the feature aggregation layer of the deep belief network, the coupling effect features are represented and learned to generate a high-dimensional feature vector.

[0111] In step 1037, the feature aggregation layer takes the coupling effect features as input. This layer first converts the multidimensional feature map into a one-dimensional feature vector through a flattening operation. Then, it performs nonlinear transformation and dimensionality reduction through two fully connected layers. Specifically, the first fully connected layer uses the ReLU activation function to map the coupling effect features to a higher dimension. The second fully connected layer further refines and compresses it into a fixed-length vector, which is the final high-dimensional feature vector, which condenses all kinds of risk information extracted and coupled in the previous steps.

[0112] This application not only extracts the core features of the heat island effect and the water accumulation process, but also simulates the nonlinear interaction between the two, ultimately generating a high-dimensional feature representation containing rich spatiotemporal correlations and causal semantics, thus laying a solid data foundation for subsequent accurate risk propagation simulation.

[0113] S104. Based on the drainage network topology and surface water runoff path of the urban center, construct a risk transmission network with drainage network nodes and surface water runoff points as network nodes, drainage network connection lines and surface water runoff paths as edges, and flow capacity as edge weights.

[0114] Among them, the drainage network topology refers to the spatial data structure that describes the interconnection between urban underground drainage pipes. It clarifies how the pipes are connected to each other through nodes such as inspection wells and pumping stations to form a network. The surface runoff path refers to the natural runoff channels formed by surface water flow under the drive of terrain slope during rainfall. These paths determine the spatial trajectory of surface runoff from the point of generation to the drainage outlet or receiving water body.

[0115] In step S104, drainage network data of the target urban area is first extracted from the urban geographic information system and drainage facility database to obtain the spatial location and topological relationship of all pipelines and their connecting nodes. At the same time, based on the high-precision digital elevation model, the water runoff path and key water collection points in the urban area are calculated by hydrological analysis algorithm. Then, these two types of physical nodes are merged to form a set of network nodes for the risk transmission network.

[0116] Then, based on the physical connection relationship of the drainage pipe network, directed edges are established between all directly connected pipe network nodes, and the direction of the edges is consistent with the actual designed flow direction of water in the pipes; at the same time, based on the flow direction of the surface water runoff path, directed edges are also established between adjacent water runoff points, and the direction of the edges is consistent with the surface water flow direction. Thus, an initial network diagram containing both the city's "underground pipe network" and "surface runoff" water flow paths is completed.

[0117] Finally, each edge in the initial network graph is assigned a weight. For pipe edges, the weight is determined based on the pipe's design flow capacity, such as by calculating it using the Manning formula. For surface water catchment path edges, the weight is determined by estimating the maximum allowable flow rate based on parameters such as surface permeability, slope, and roughness. The weights of all edges together constitute the weight matrix of the network, ultimately generating a weighted directed graph, which serves as a risk transmission network that acts as a carrier for the dynamic propagation of risk.

[0118] This application constructs a unified physical topology network representing urban water flow paths and quantifies the conduction intensity of each path by flow capacity, thereby abstracting complex hydrogeographic entities into a computable graph structure, which lays a solid structural foundation for subsequent accurate simulation of the dynamic spatial diffusion process of urban flooding risk.

[0119] S105. An ensemble learning framework is introduced to optimize the high-dimensional feature vector to obtain an ensemble feature vector representing the risk state of each network node. Under the constraints of the risk transmission network, risk propagation analysis is performed on the ensemble feature vector to generate propagation features representing the direction and intensity of risk propagation. The ensemble learning framework adopts a random forest model based on the Bagging ensemble method and a gradient boosting tree model based on the Boosting ensemble method.

[0120] In one specific implementation, step S105 includes:

[0121] Step 1051: Iteratively simulate the propagation process of the risk state on the risk transmission network based on the integrated feature vector and graph diffusion model until the state stabilizes or the preset number of iterations is reached, and generate the diffusion feature vector.

[0122] Step 1051 may specifically include the following steps:

[0123] Step b1: Map the integrated feature vector to the initial state vector of the network nodes in the risk transmission network.

[0124] Among them, the ensemble feature vector refers to the feature vector obtained after optimization by the random forest and gradient boosting tree models. It integrates the prediction results of multiple models and can more robustly represent the initial risk state of each network node.

[0125] In step b1, the ensemble feature vector of each network node is first obtained from the output of the ensemble learning framework. These vectors contain multi-dimensional features. In order to simplify the propagation simulation, the key dimensions are selected or compressed into scalars through linear transformation as the initial state values ​​of the nodes. Then, the initial state values ​​of each network node are organized into a vector according to the numbering order of the network nodes, i.e., the initial state vector.

[0126] For example, suppose there are three network nodes in the risk transmission network, numbered N1, N2, and N3, and the ensemble feature vectors of the three network nodes are EN1=[0.85, 0.12], EN2=[0.70, 0.25], and EN3=[0.90, 0.08], respectively. If the first element of each vector is selected as the initial state value, then the initial state vector Vinit=[0.85, 0.70, 0.90], and corresponds to nodes N1, N2, and N3 respectively.

[0127] Step b2: Based on the initial state vector and the edge weights in the risk propagation network, the risk state is iteratively simulated using a graph diffusion model to simulate the propagation process of the risk state on the risk propagation network until the state stabilizes or the preset number of iterations is reached, thereby obtaining the risk state value of each network node.

[0128] Among them, the graph diffusion model refers to a mathematical model based on graph theory, which is used to simulate the propagation process of information or risk in a graph structure. It updates the state value of a node iteratively, so that it is affected by the state of its neighboring nodes and the weight of the connecting edges, and eventually reaches a stable state.

[0129] In step b2, the graph diffusion model uses an iterative update algorithm. In each iteration, the risk state value of each node is updated according to the state of its neighboring nodes and the weight of the connecting edges. The update formula can be referred to relevant technologies. The iteration process continues until the state changes of all nodes are less than the set threshold or the preset number of iterations is reached, and finally the stable state values ​​of all nodes are obtained.

[0130] For example, suppose the connection relationship of the risk transmission network is N1 connected to N2, N2 connected to N3, and N3 connected to N1, with corresponding edge weights of 0.6, 0.7, and 0.5, respectively. Assume the diffusion coefficient α is 0.5, the preset number of iterations is 3, and then the risk state value of each node is obtained after each iteration according to the update formula, until the risk state value of each node after the last iteration is obtained. For example, the risk state values ​​of nodes N1, N2, and N3 are 0.79, 0.84, and 0.83, respectively.

[0131] Step b3: Spatially aggregate the risk state values ​​of all network nodes to generate a diffusion feature vector.

[0132] For example, the risk status values ​​of the three nodes, 0.79, 0.84, and 0.83, are concatenated according to the node numbering order N1, N2, and N3 to generate the diffusion feature vector D = [0.79, 0.84, 0.83].

[0133] Step 1052: Extract the path based on the maximum gradient direction and quantize the intensity based on the vector norm of the diffusion feature vector to generate a propagation feature that characterizes the direction and intensity of risk propagation.

[0134] Among them, path extraction based on the maximum gradient direction refers to analyzing the spatial variation trend of the diffusion feature vector and finding the direction with the maximum gradient by calculating the gradient of the state values ​​of adjacent nodes, thereby determining the main path of risk propagation; intensity quantification based on vector norm refers to calculating the norm of the diffusion feature vector and using it as a quantitative indicator of the overall risk intensity.

[0135] In step 1052, the gradient of the state values ​​between each pair of adjacent network nodes is first calculated based on the spatial adjacency relationship of the network nodes. Then, starting from the initial high-risk node, the path is traced along the direction of the maximum gradient until the gradient no longer increases or reaches the boundary, thereby extracting the main path of risk propagation. At the same time, the L2 norm of the entire diffusion feature vector is calculated, which is the square root of the sum of the squares of the state values ​​of all nodes, and is used as the overall measure of risk intensity. Finally, the extracted path information and intensity values ​​are combined to generate the propagation feature.

[0136] For example, if the adjacency relationship of network nodes is N1 adjacent to N2, and N2 adjacent to N3, then the gradient from N1 to N2 is calculated as follows: The gradient from N2 to N3 is Since 0.05 > The direction of the maximum gradient is from N1 to N2, therefore the main path of risk propagation is from N1 to N2; then intensity quantization is performed, i.e., the L2 norm is calculated. ≈1.42, therefore the propagation characteristics include the propagation path and the propagation intensity, where the propagation path can be from N1 to N2 and the propagation intensity can be 1.42.

[0137] This step enables quantitative analysis and visual representation of the spatial transmission process of urban flooding risk, thus providing a scientific basis for accurately locating the source of risk and predicting the scope of impact.

[0138] S106. Using a logistic regression coupling model, the propagation characteristics are processed to obtain the risk prediction results of each monitoring unit in the central urban area of ​​the city for a period of time in the future. Based on the risk prediction results, risk control parameters are generated to indicate the start and stop of pumping stations, gate scheduling, or the release of graded early warning information.

[0139] In one specific implementation, step S106 includes:

[0140] Step 1061: Using the logistic regression unit in the logistic regression coupling model, perform linear weighted summation and sigmoid function activation processing on the propagation features to obtain the probability of waterlogging occurrence for each monitoring unit in the future.

[0141] The logistic regression unit refers to a standard logistic regression model, which performs a linear weighted summation of the input features and then maps the result to a probability value between 0 and 1 using the Sigmoid function. The probability of waterlogging is a numerical value output by the model, representing the likelihood of a waterlogging event occurring in a certain monitoring unit in the future. The closer the value is to 1, the greater the likelihood of occurrence.

[0142] It should be noted that this embodiment does not impose specific limitations on the structural design, parameter design, and training process of the logistic regression coupling model, and can be set accordingly based on the actual situation.

[0143] In step 1061, the logistic regression unit first performs a linear weighted summation on the propagation features; then, the result of this weighted summation is input into the Sigmoid function for activation, and the Sigmoid function maps any real number to the range of 0 to 1, thereby outputting a probability value, which is the probability of waterlogging occurring in the monitoring unit.

[0144] For example, suppose the propagation characteristics of monitoring unit A consist of two parts: one part is the propagation intensity of 1.42 from step 1052, and the second part is the risk state value of node N1 of this unit of 0.79 from step 1051, and set the weight coefficient vector of the logistic regression unit as follows. The bias term is The result of linear weighted summation is: Then input 0.452 into the Sigmoid function, and the calculation formula is as follows: The probability of flooding in monitoring unit A is approximately 0.611.

[0145] Step 1062: Using the hierarchical classification unit in the logistic regression coupling model, and based on the pre-set urban flooding risk classification knowledge base, the probability of flooding occurrence is mapped to a dynamic risk level to generate a risk prediction result.

[0146] In step 1062, the classification unit stores or accesses a pre-set urban flood risk classification knowledge base, which defines the risk level corresponding to different probability intervals; then the classification unit compares the probability of flooding occurrence of each monitoring unit with these intervals to determine its risk level, thereby generating a risk prediction result for each monitoring unit.

[0147] For example, the probability of flooding in monitoring unit A is 0.611. Assuming the pre-set knowledge base stipulates that a probability less than 0.3 is low risk, 0.3 to 0.7 is medium risk, and greater than 0.7 is high risk, since 0.611 falls within the range of 0.3 to 0.7, monitoring unit A is classified as medium risk.

[0148] Step 1063: Input the risk prediction results, the drainage network topology, and the surface water runoff path into the collaborative decision-making network. Through the strategy generation module in the collaborative decision-making network, candidate control actions are generated based on a preset rule base or reinforcement learning strategy.

[0149] Among them, the collaborative decision-making network refers to a comprehensive decision-making system that takes into account risk prediction, pipeline topology and water catchment path to formulate control strategies; the strategy generation module is a component in the collaborative decision-making network that is responsible for generating possible control actions based on input information.

[0150] In step 1063, the risk prediction results, drainage network topology, and surface water runoff paths are used as inputs to the strategy generation module. The strategy generation module contains a preset rule base or a trained reinforcement learning strategy model. If based on the rule base, the strategy generation module will match the current risk status and network status according to the rules and generate a corresponding list of control actions. If based on the reinforcement learning strategy, the strategy generation module will use the strategy network to directly output a series of action suggestions based on the input information. These generated control actions are the candidate control actions.

[0151] For example, the risk prediction results show that monitoring unit A is of medium risk, and a pumping station PS1 is connected to the drainage network node N1 corresponding to the network topology. Then, according to a rule in the preset rule base: "If monitoring unit A is of medium risk and its associated pumping station PS1 is currently closed, then generate an action to turn on pumping station PS1", that is, generate a candidate control action: turn on pumping station PS1.

[0152] Step 1064: Through the effect inference module in the collaborative decision-making network, simulate the change in the risk of urban flooding in the central urban area after executing each of the candidate control actions, and obtain the risk control effect of each action.

[0153] In step 1064, the effect simulation module receives candidate control actions and current data including risk prediction results, pipeline topology, and water catchment paths. Then, for each candidate control action, the effect simulation module simulates the execution of the action in a built-in hydrological and hydraulic model or a simplified risk propagation model. For example, after the effect simulation starts a pumping station, it recalculates the risk value of the affected area based on the changes in water flow in the pipeline network, and calculates the risk control effect of the action by comparing the risk values ​​before and after the action is executed.

[0154] For example, for the candidate control action "starting pump station PS1", the effect inference module simulates the start of the pump station in the model. Assuming that after the simulation, the probability of waterlogging in monitoring unit A decreases from 0.611 to 0.401 and the risk level changes from medium risk to low risk, the risk control effect of this action can be quantified as: the probability decreases by 0.210 and the risk level decreases by one level.

[0155] Step 1065: Based on the risk control effect, the strategy optimization module in the collaborative decision-making network, combined with the drainage network topology and the surface water runoff path, optimizes each of the candidate control actions to generate risk control parameters for pumping stations, gates and early warning points.

[0156] Among them, the strategy optimization module refers to the component in the collaborative decision-making network that is responsible for the final decision. It selects the optimal action or combination of actions based on the risk control effect of each action and system constraints.

[0157] In step 1065, the strategy optimization module receives all candidate control actions and their corresponding risk control effects. The strategy optimization module comprehensively considers the magnitude of the effect, the cost of action execution, the physical constraints of pipeline operation, and the impact of surface water runoff paths. It selects one or more optimal actions through multi-objective optimization algorithms or heuristic rules. Finally, the selected actions are converted into specific, deployable control parameters to generate the final risk control parameters.

[0158] For example, suppose that in addition to "starting pump station PS1", there is another candidate action "issue a level 3 warning in monitoring unit A", which has the risk control effect of prompting the public to avoid risks, but cannot quantify the reduction of probability; then the strategy optimization module compares the effects and costs of the two actions, and considering that starting pump station PS1 can directly reduce physical risks, it selects this action; the final risk control parameter is: the start / stop status of pump station PS1 is set to "start".

[0159] This application achieves a closed loop from risk prediction to proactive control through the above steps, thereby improving the accuracy and timeliness of urban flood control emergency response.

[0160] Figure 3 This is a schematic diagram illustrating a specific implementation of an urban flood control and waterlogging prediction and monitoring system provided in this application. (Refer to...) Figure 3 The system may include:

[0161] The data acquisition module 31 is used to collect data related to urban flooding in the central urban area and physical sensing data reflecting the water accumulation situation.

[0162] The determination module 32 is used to determine the risk classification results of each monitoring unit in the central urban area based on the waterlogging-related data and the pre-set urban waterlogging risk classification knowledge base.

[0163] The mining module 33 is used to dynamically associate and fuse the risk classification results with the physical perception data, construct a predictive feature set to characterize the real-time evolution of urban flooding risk, and use a deep belief network to extract high-dimensional features and mine coupling relationships in the predictive feature set to generate a high-dimensional feature vector.

[0164] The construction module 34 is used to construct a risk transmission network based on the drainage network topology and surface water runoff path of the urban center, with drainage network nodes and surface water runoff points as network nodes, drainage network connection lines and surface water runoff paths as edges, and flow capacity as edge weights.

[0165] Analysis module 35 is used to introduce an ensemble learning framework to optimize the high-dimensional feature vector, obtain an ensemble feature vector representing the risk state of each network node, and perform risk propagation analysis on the ensemble feature vector under the constraints of the risk transmission network to generate propagation features representing the direction and intensity of risk propagation. The ensemble learning framework adopts a random forest model based on the Bagging ensemble method and a gradient boosting tree model based on the Boosting ensemble method.

[0166] The processing module 36 is used to process the propagation features using a logistic regression coupling model to obtain the risk prediction results of each monitoring unit in the central urban area of ​​the city for a period of time in the future. Based on the risk prediction results, risk control parameters are generated to indicate the start and stop of pumping stations, gate scheduling, or the release of graded early warning information.

[0167] The urban flood control and waterlogging prediction and monitoring system of this application embodiment is used to implement the aforementioned urban flood control and waterlogging prediction and monitoring method. Therefore, the specific implementation of the urban flood control and waterlogging prediction and monitoring system can be found in the embodiment section of the urban flood control and waterlogging prediction and monitoring method above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0168] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described methods for predicting and monitoring urban flooding.

[0169] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for predicting and monitoring urban flooding.

[0170] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0171] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the urban flood control and waterlogging prediction and monitoring method.

[0172] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0173] The above provides a detailed description of the urban flood control and waterlogging prediction and monitoring method and system provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for predicting and monitoring urban flooding and waterlogging, characterized in that, include: Collect data related to urban flooding in the city center and physical sensing data reflecting the water accumulation situation; Based on the aforementioned waterlogging-related data and the pre-established urban waterlogging risk classification knowledge base, the risk classification results of each monitoring unit in the central urban area are determined. The risk classification results are dynamically correlated and fused with the physical sensing data to construct a predictive feature set for characterizing the real-time evolution of urban flooding risk. A deep belief network is used to extract high-dimensional features and mine coupling relationships in the predictive feature set to generate a high-dimensional feature vector. Based on the drainage network topology and surface water runoff paths of the urban center, a risk transmission network is constructed with drainage network nodes and surface water runoff points as network nodes, drainage network connection lines and surface water runoff paths as edges, and flow capacity as edge weights. An ensemble learning framework is introduced to optimize the high-dimensional feature vectors, thereby obtaining ensemble feature vectors that characterize the risk status of each network node. Under the constraints of the risk transmission network, risk propagation analysis is performed on the ensemble feature vectors to generate propagation features that characterize the direction and intensity of risk propagation. The ensemble learning framework employs a random forest model based on the Bagging ensemble method and a gradient boosting tree model based on the Boosting ensemble method. By using a logistic regression coupling model to process the propagation characteristics, the risk prediction results of each monitoring unit in the central urban area of ​​the city are obtained in the future. Based on the risk prediction results, risk control parameters are generated to indicate the start and stop of pumping stations, gate scheduling, or to issue graded early warning information. The risk classification results for each monitoring unit within the central urban area are determined based on the aforementioned waterlogging-related data and a pre-established urban waterlogging risk classification knowledge base, including: The heat island effect data and surface runoff data in the waterlogging-related data belonging to the same monitoring unit in the central urban area are coupled and analyzed to generate the coupling coefficient of each monitoring unit. Using the grading rules in the pre-set urban waterlogging risk grading knowledge base, the coupling coefficients are matched to obtain the preliminary risk level of each monitoring unit. The grading rules define the correspondence between the coupling coefficient range and the risk level. Based on the topographic elevation difference between adjacent monitoring units and the connectivity of the drainage network, the preliminary risk level is spatially verified and adjusted to generate risk classification results for each monitoring unit in the central urban area. The step of dynamically associating and fusing the risk classification results with the physical sensing data to construct a predictive feature set for characterizing the real-time evolution of urban flooding risk includes: Assign a feature vector representing the inherent vulnerability to the static risk level of each monitoring unit in the risk classification result, as the static risk feature of the corresponding monitoring unit, and organize the physical sensing data of each monitoring unit in the same time step into a dynamic situation vector representing the corresponding real-time situation. The static risk features corresponding to each monitoring unit at each time step are concatenated with the dynamic situation vector along the feature channel dimension to obtain the fused feature vector of the corresponding monitoring unit. The fused feature vectors generated by all monitoring units at all time steps are arranged in a structured manner according to spatial location and temporal order to form a predicted feature set.

2. The method according to claim 1, characterized in that, A deep belief network is used to perform high-dimensional feature extraction and coupling relationship mining on the predicted feature set to generate a high-dimensional feature vector, including: The first feature extraction layer of the deep belief network performs sliding window aggregation of the first temporal window and spatial smoothing based on Gaussian kernel function on the predicted feature set to extract the first type of features reflecting the intensity of the heat island effect from the predicted feature set. The second feature extraction layer of the deep belief network is used to perform sliding window difference of the second time window and spatial gradient enhancement based on the Sobel operator on the predicted feature set, so as to extract the second type of features reflecting the water accumulation and collection process from the predicted feature set. Through the coupling interaction layer of the deep belief network, the interaction relationship between the first type of features and the second type of features is simulated and the features interact, generating coupled interaction features; The coupling effect features are represented and learned through the feature aggregation layer of a deep belief network to generate a high-dimensional feature vector.

3. The method according to claim 2, characterized in that, The coupling interaction layer of the deep belief network simulates the interaction between the first type of features and the second type of features, generating coupled interaction features, including: In the coupled interaction layer of the deep belief network, the first type of features are converted into source feature maps, and the second type of features are converted into medium feature maps; The propagation process of the driving potential represented by the source feature map is simulated by the propagation kernel function in the coupled interaction layer under the first constraint of the local water accumulation state represented by the medium feature map and the second constraint corresponding to the preset urban underlying surface attribute data, thereby generating a propagation feature map. By coupling the gated loop unit in the interaction layer, a dynamic gate factor is generated based on the local water accumulation state corresponding to the medium feature map. The dynamic gate factor is then used to adaptively enhance the propagation feature map to generate an enhanced feature map. By using the fusion layer in the coupling interaction layer, the enhanced feature map is superimposed with the medium feature map, and the superposition result is adjusted based on the preset attribute data of the urban underlying surface to simulate the combined effect of confluence acceleration and lag buffering, thereby generating coupling effect features.

4. The method according to claim 1, characterized in that, Under the constraints of the risk transmission network, the integrated feature vector is subjected to risk propagation analysis to generate propagation features characterizing the direction and intensity of risk propagation, including: The propagation process of risk state on the risk transmission network is iteratively simulated based on integrated feature vectors and graph diffusion model until the state stabilizes or reaches a preset number of iterations, and a diffusion feature vector is generated. The diffusion feature vector is subjected to path extraction based on the maximum gradient direction and intensity quantization based on the vector norm to generate propagation features that characterize the direction and intensity of risk propagation.

5. The method according to claim 4, characterized in that, The iterative simulation of the propagation process of risk state on the risk transmission network based on integrated feature vectors and graph diffusion models continues until the state stabilizes or a preset number of iterations is reached, generating diffusion feature vectors, including: The integrated feature vector is mapped to the initial state vector of the network nodes in the risk transmission network; Based on the initial state vector and the edge weights in the risk propagation network, a graph diffusion model is used to iteratively simulate the propagation process of the risk state on the risk propagation network until the state stabilizes or the preset number of iterations is reached, thereby obtaining the risk state value of each network node. Spatial aggregation of the risk status values ​​of all network nodes is performed to generate a diffusion feature vector.

6. The method according to claim 1, characterized in that, The propagation characteristics are processed using a logistic regression coupling model to obtain risk prediction results for each monitoring unit in the urban central area over a future period. Based on these risk prediction results, risk control parameters are generated to indicate pump station start / stop, gate scheduling, or the issuance of tiered early warning information, including: By using the logistic regression unit in the logistic regression coupling model, the propagation features are subjected to linear weighted summation and sigmoid function activation to obtain the probability of waterlogging occurrence for each monitoring unit in the future period. By using the hierarchical classification unit in the logistic regression coupling model and based on the pre-set urban waterlogging risk classification knowledge base, the probability of waterlogging occurrence is mapped to a dynamic risk level to generate risk prediction results. The risk prediction results, the drainage network topology, and the surface water runoff path are input into the collaborative decision-making network. The strategy generation module in the collaborative decision-making network generates candidate control actions based on a preset rule base or reinforcement learning strategy. By using the effect inference module in the collaborative decision-making network, the changes in the risk of urban flooding in the central urban area after executing each of the candidate control actions are simulated, and the risk control effect of each action is obtained. Based on the risk control effect, the strategy optimization module in the collaborative decision-making network, combined with the drainage network topology and the surface water runoff path, optimizes each of the candidate control actions to generate risk control parameters for pumping stations, gates and early warning points.

7. A predictive monitoring system for urban flood control and waterlogging, characterized in that, A method for predicting and monitoring urban flooding as described in claim 1 includes: The data acquisition module is used to collect data related to urban flooding in the city center and physical sensing data reflecting the water accumulation situation. The determination module is used to determine the risk classification results of each monitoring unit in the central urban area of ​​the city based on the waterlogging-related data and a pre-set urban waterlogging risk classification knowledge base. The mining module is used to dynamically associate and fuse the risk classification results with the physical perception data, construct a predictive feature set to characterize the real-time evolution of urban flooding risk, and use a deep belief network to extract high-dimensional features and mine coupling relationships in the predictive feature set to generate a high-dimensional feature vector. The construction module is used to construct a risk transmission network based on the drainage network topology and surface water runoff path of the urban center, with drainage network nodes and surface water runoff points as network nodes, drainage network connection lines and surface water runoff paths as edges, and flow capacity as edge weights. The analysis module is used to introduce an ensemble learning framework to optimize the high-dimensional feature vector, obtain an ensemble feature vector representing the risk state of each network node, and perform risk propagation analysis on the ensemble feature vector under the constraints of the risk transmission network to generate propagation features representing the direction and intensity of risk propagation. The ensemble learning framework adopts a random forest model based on the Bagging ensemble method and a gradient boosting tree model based on the Boosting ensemble method. The processing module is used to process the propagation characteristics using a logistic regression coupling model to obtain the risk prediction results of each monitoring unit in the central urban area of ​​the city for a period of time in the future. Based on the risk prediction results, risk control parameters are generated to indicate the start and stop of pumping stations, gate scheduling, or the issuance of graded early warning information.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of a method for predicting and monitoring urban flooding as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables a method for predicting and monitoring urban flooding as described in any one of claims 1 to 6.

Images