Urban storm flood simulation method and device

Abstract

The application discloses a kind of urban rain flood simulation method and device, method includes: obtaining the historical flood event data set of research area, historical flood event data set is divided into rating set and verification set;Flood event data set includes: flood event associated land use data, drainage pipe network topological structure data, weather data and flood data;Based on rating set and verification set, respectively with time-varying gain nonlinear city model, water and soil conservation runoff model and stormwater management model as physical constraint, with pre-constructed LSTM coupling, obtain three depth coupling models;Based on rating set and verification set, for each discharge port, the simulation result of the integrated three depth coupling models, determine integrated weight by optimization algorithm, utilize integrated three depth coupling models, obtain integrated model for executing urban rain flood simulation calculation.The application can capture the diversification characteristics of hydrological sequence, realize high-precision interpretable urban rain flood simulation.

Description

TECHNICAL FIELD

[0001] The present application relates to a city rain flood simulation method and device, belonging to the application technical field of machine learning technology in the field of water conservancy engineering. BACKGROUND

[0002] Urban flood is a recurring destructive natural disaster. With the intensification of global climate change and the rapid advancement of urbanization, the frequency and intensity of extreme events have increased significantly, causing huge population and economic losses. Urban rain flood management is facing unprecedented challenges. In this context, accurately identifying the relationship between urban rainfall and runoff, establishing an effective urban rain flood simulation model, and accurately simulating the rain flood process are crucial for urban planners and emergency management departments to predict flood conditions in advance and take measures to reduce losses.

[0003] Machine learning is a data-driven model without complex physical processes, and is widely used in nonlinear multidimensional modeling problems. So far, machine learning has been extensively studied in the field of watershed hydrology, but in the field of urban hydrology, it has mainly focused on rapid simulation of urban waterlogging or flood risk assessment.

[0004] Although machine learning models have good performance in hydrological prediction, their highly nonlinear structure makes it difficult to interpret the prediction results, which poses certain obstacles to using machine learning models for urban rain flood simulation. Standard recurrent neural networks represented by LSTM (Long Short Term Memory network) are still the core model structure of machine learning rain flood simulation at present, and the integration with physical mechanisms is more focused on loose coupling methods such as loss function modification, and most of the research is concentrated in the field of watershed hydrology. In the field of cities, how to enhance the interpretability of machine learning models in the process of rain flood simulation, while ensuring physical consistency and improving simulation accuracy, still needs to be studied.

[0005] In addition, there are various uncertainties in the process of rain flood simulation by machine learning models, including parameter uncertainty and input data uncertainty, which will directly affect the simulation accuracy of the model. SUMMARY

[0006] The present application aims to overcome the shortcomings of the prior art and provide a city rain flood simulation method and device that can capture the diversified characteristics of hydrological sequences and achieve high-precision and interpretable city rain flood simulation. To achieve the above-mentioned purpose, the present application is implemented by using the following technical solutions:

[0007] In a first aspect, the present application provides a city rain flood simulation method, comprising:

[0008] acquire a historical secondary flood event data set of a research area, divide the historical secondary flood event data set into a calibration set and a verification set; wherein the secondary flood event data set comprises: land use data associated with a secondary flood event, drainage pipe network topology data, meteorological data and secondary flood data;

[0009] Based on the calibration set and the verification set, respectively, a time-varying gain nonlinear urban model, a water and soil conservation runoff model and a stormwater management model are coupled with a pre-constructed LSTM to obtain three deep coupling models;

[0010] Based on the calibration set and the verification set, the simulation results of the three deep coupling models are integrated for each outlet, and the integrated weights are determined by an optimization algorithm, and the three deep coupling models are integrated by using the integrated weights to obtain an urban rainwater integration model; the urban rainwater integration model is used for performing urban rainwater simulation calculation.

[0011] In combination with the first aspect, optionally, based on the calibration set and the verification set, respectively, a time-varying gain nonlinear urban model, a water and soil conservation runoff model and a stormwater management model are coupled with a pre-constructed LSTM to obtain three deep coupling models, including:

[0012] Based on the calibration set and the verification set, a time-varying gain nonlinear urban model is coupled with a pre-constructed LSTM to obtain a first deep coupling model, including:

[0013] According to the physical process of the time-varying gain nonlinear urban model, the input gate, the forget gate and the output gate of the pre-constructed LSTM are reconstructed; wherein the input gate is the contribution of the current time step precipitation to the soil moisture update, the forget gate is the degree of preservation of the previous soil moisture in updating the cell state, and the output gate is the final flow output obtained by the pipe network confluence process in urban areas;

[0014] The cell state update formula in the pre-constructed LSTM is replaced by a water balance formula to update the cell state to obtain the first deep coupling model;

[0015] Based on the calibration set and the verification set, a water and soil conservation runoff model is coupled with a pre-constructed LSTM to obtain a second deep coupling model, including:

[0016] According to the physical process of the water and soil conservation runoff model, the input gate, the forget gate and the output gate of the pre-constructed LSTM are reconstructed; wherein the input gate is the contribution of the current time step precipitation and the detention volume of the catchment area to the runoff, the forget gate is the degree of preservation of the previous soil moisture in updating the cell state, and the output gate is the final flow output obtained by the pipe network confluence process in urban areas;

[0017] The cell state is updated by using a cell state update formula in the pre-constructed LSTM to obtain a second deep coupling model;

[0018] Based on the rating set and the verification set, a third deep coupling model is obtained by coupling a storm flood management model as a physical constraint with the pre-constructed LSTM, including:

[0019] A physical parameter composition vector of the storm flood management model is extracted, and the vector is embedded into the pre-constructed LSTM. The weights of the physical parameters in the input gate, the forgetting gate and the output gate of the pre-constructed LSTM are expanded.

[0020] The cell state update formula of the LSTM is adjusted in combination with the calculation constraint of the storm flood management model based on the current precipitation input and the physical parameters. The cell state is updated by using the adjusted cell state update formula.

[0021] The water balance constraint of the storm flood management model is applied at the output end of the LSTM to obtain the third deep coupling model.

[0022] In combination with the first aspect, optionally, the input gate, the forgetting gate and the output gate of the pre-constructed LSTM are reconfigured, including:

[0023] The input gate is the contribution of the current time step precipitation to the soil moisture update, which is represented by the following formula:

[0024] ,

[0025] Wherein, is the first deep coupling model The time period input gate vector, is the sigmod activation function of the input gate; is The precipitation in the time period; is The antecedent rainfall at the end of the time period, which is closely related to the soil moisture before the precipitation in the time period starts; is the saturated soil moisture, is the runoff reduction coefficient corresponding to the impervious pavement, is the runoff reduction coefficient corresponding to the building area; is the parameter related to the runoff calculation in the time-varying gain nonlinear urban model; is the function expression of the physical process of the water storage term increase in the time-varying gain nonlinear urban model, which is represented by the following formula:

[0026] ,

[0027] Wherein, ​is a physical formula, including the runoff generation mechanism of the time-varying gain nonlinear urban model; is a group of physical parameters in the physical formula that do not change over time, including the area proportion of impervious pavement in the entire study area , the area proportion of building area in the entire study area , the area proportion of vegetation coverage area in the entire study area , and the area proportion of bare soil area in the entire study area ; is a group of physical parameters in the physical formula that change over time, including ;

[0028] The forgetting gate is the degree of retention of the previous soil moisture content when updating the cell state, and is represented by the following formula:

[0029] ,

[0030] wherein, is the first depth-coupled model the period forgetting gate vector, is the sigmod activation function of the forgetting gate; is the dimensionless recession coefficient; is a function expression of the physical process consumed by the storage term in the time-varying gain nonlinear urban model, and is represented by the following formula:

[0031] ,

[0032] wherein, is the minimum value calculation, is the previous influence rainfall at the end of the period;

[0033] The output gate is the final flow output obtained by the pipe network confluence process of the urban area, and is represented by the following formula:

[0034] ,

[0035] wherein, is the first depth-coupled model the period output gate vector, is the sigmod activation function of the output gate; is the first depth-coupled model the period input gate vector, is the first depth-coupled model the period output gate vector, is the system conversion parameter, is the system lag time; The function expression of the physical process of outflow in the time-varying gain nonlinear urban model is represented by the following formula:

[0036] ,

[0037] is the flow of the time period, is the unit conversion coefficient, is the yield flow of the time period, is the flow of the time period. is the flow of the time period.

[0038] In combination with the first aspect, optionally, the water balance formula is used to replace the cell state update formula in the pre-constructed LSTM to update the cell state, and a first deep coupling model is obtained, represented by the following formula:

[0039] ,

[0040] is the cell state of the first deep coupling model in the time period, is the water balance formula for determining cell update, is the cell state of the first deep coupling model in the time period, is the forget gate vector of the first deep coupling model in the time period, is the cell state of the first deep coupling model in the time period, is the input gate vector of the first deep coupling model in the time period, is the input of the time period.

[0041] In combination with the first aspect, optionally, the input gate, the forget gate and the output gate of the pre-constructed LSTM are reconstructed according to the physical process of the water and soil conservation runoff model, including:

[0042] The input gate is the contribution of the precipitation and the detention volume of the catchment area to the yield flow at the current time step, represented by the following formula:

[0043] ,

[0044] is the input gate vector of the second deep coupling model in the time period, is the sigmod activation function of the input gate; is the precipitation of the time period; is the initial rainfall loss of the time period. ​​​​​​​​​​​​​​​CN is a weight coefficient represented by the runoff curve number in the second depth coupling model; is a function expression of the physical process of runoff generation in the water and soil conservation runoff model, expressed by the following formula:

[0045] ,

[0046] wherein, is the soil moisture content in the period;

[0047] The forget gate is the degree of retention of the previous soil moisture content when updating the cell state, expressed by the following formula:

[0048] ,

[0049] wherein, is the second depth coupling model the forget gate vector in the period, is the sigmod activation function of the forget gate; is the evaporation amount at the current time step; is the initial rainfall loss in the period; CN is a weight coefficient represented by the runoff curve number in the second depth coupling model; is a function expression of the physical process of determining rainfall loss in the water and soil conservation runoff model, expressed by the following formula:

[0050] ,

[0051] wherein, is the maximum value calculation;

[0052] The output gate is the final flow output converted by the pipe network confluence process in urban areas, expressed by the following formula:

[0053] ,

[0054] wherein, is the second depth coupling model the output gate vector in the period, is the sigmod activation function of the output gate; is the second depth coupling model the input gate vector in the period, is the second depth coupling model the output gate vector in the period, is a system conversion parameter, is a system lag time; is a function expression of the physical process of confluence in the time-varying gain nonlinear urban model, expressed by the following formula:

[0055] ,

[0056] wherein, is flow of the period, is a unit conversion coefficient, is yield flow of the period, is flow of the period.

[0057] In combination with the first aspect, optionally, the physical parameter vector of the storm flood management model is extracted, the vector is embedded into the pre-constructed LSTM, and the weights of the physical parameters in the input gate, the forgetting gate and the output gate of the pre-constructed LSTM are expanded, including:

[0058] The physical parameter vector of the storm flood management model is extracted , and is represented by the following formula:

[0059] ,

[0060] wherein, is the Manning coefficient of the impervious area, is the Manning coefficient of the pervious area, is the impervious coefficient, is the characteristic width coefficient, is the Manning coefficient of the pipe, is the maximum infiltration rate, is the minimum infiltration rate, is the water storage of the impervious area, is the water storage of the pervious area, is the proportion of the impervious area without water storage;

[0061] The vector is embedded into the pre-constructed LSTM, the weights of the physical parameters in the input gate, the forgetting gate and the output gate of the pre-constructed LSTM are expanded, and the expanded input gate, the forgetting gate and the output gate are represented by the following formula:

[0062] ,

[0063] wherein, are respectively the third deep coupling model the input gate vector, the forgetting gate vector and the output gate vector of the period; is a sigmod activation function; is input of the period; is hidden state of the period; are respectively weight matrix corresponding to the input gate, the forget gate, and the output gate; respectively weight matrix corresponding to the input gate, the forget gate, and the output gate; respectively weights of the physical parameters in the input gate, the forget gate, and the output gate; respectively bias vector parameters of the input gate, the forget gate, and the output gate.

[0064] With reference to the first aspect, optionally, the third deep coupling model is obtained by applying a water balance constraint of the storm flood management model at an output end of the LSTM, and the water balance constraint of the storm flood management model at the output end of the LSTM is represented by the following formula:

[0065] ,

[0066] wherein, is a third deep coupling model cell state of the time period, is a third deep coupling model cell state of the time period; is a dot product operation; respectively input gate vector and forget gate vector of the third deep coupling model of the time period, is a third deep coupling model candidate cell state of the time period, and is represented by the following formula:

[0067] ,

[0068] wherein, is a calculation constraint of the storm flood management model based on precipitation input and physical parameters of the time period.

[0069] With reference to the first aspect, optionally, the third deep coupling model is obtained by applying a water balance constraint of the storm flood management model at an output end of the LSTM, and the water balance constraint of the storm flood management model at the output end of the LSTM is represented by the following formula:

[0070] ,

[0071] wherein, is an expression of the third deep coupling model of the time period; is an output gate vector of the third deep coupling model of the time period; is a cell state of the third deep coupling model of the time period;

[0072] The third deeply coupled model outputs simulated runoff that conforms to physical laws. It can be expressed by the following formula:

[0073] ,

[0074] in, To correct the parts of the LSTM output that do not conform to the water balance constraints, a stormwater and flood management model is used.

[0075] In conjunction with the first aspect, optionally, based on the calibration set and validation set, for each outfall, the simulation results of integrating three deeply coupled models are used to determine the integration weights through an optimization algorithm. The three deeply coupled models are then integrated using these integration weights to obtain an integrated urban stormwater model, including:

[0076] The simulation results for each outlet, integrating three deeply coupled models, are expressed by the following formula:

[0077] ,

[0078] in, The integrated sub-catchment flow rate, These are the weights of the three deeply coupled models. These are the initial outflow rates for the three deeply coupled models;

[0079] Using genetic algorithms to adjust weights The calibration is performed, and the objective function of the calibration is to verify the actual flow rate of the sub-catchment area and the flow rate of the integrated sub-catchment area. The ensemble weights are obtained by minimizing the Nash efficiency coefficients among the groups.

[0080] By integrating three deeply coupled models using integrated weights, an integrated urban stormwater model is obtained.

[0081] Secondly, the present invention provides an urban stormwater simulation device, comprising:

[0082] Acquisition Module: Used to acquire historical secondary flood event datasets for the study area, dividing the historical secondary flood event datasets into calibration and validation sets; the secondary flood event datasets include: land use data associated with secondary flood events, drainage network topology data, meteorological data, and secondary flood data;

[0083] Deeply Coupled Module: Based on the calibration set and validation set, the module couples the time-varying gain nonlinear urban model, the soil and water conservation runoff generation model, and the rainstorm flood management model with a pre-built LSTM to obtain three deeply coupled models.

[0084] The integrated module is used for integrating simulation results of the three deep coupling models for each outlet based on the rating set and the verification set, determining integrated weights by an optimization algorithm, integrating the three deep coupling models by using the integrated weights, and obtaining an urban rain flood integrated model; and the urban rain flood integrated model is used for performing urban rain flood simulation calculation.

[0085] Compared with the prior art, the urban rain flood simulation method and device provided by the embodiment of the present application has the following beneficial effects:

[0086] The historical secondary flood event data set of the research area is obtained, and the historical secondary flood event data set is divided into a rating set and a verification set; the secondary flood event data set includes land use data, drainage pipe network topological structure data, meteorological data and secondary flood data associated with the secondary flood event; three deep coupling models are obtained by coupling the time-varying gain nonlinear urban model, the water and soil conservation runoff model and the rainstorm flood management model based on the rating set and the verification set as physical constraints and the LSTM constructed in advance;

[0087] The integrated module is used for integrating simulation results of the three deep coupling models for each outlet based on the rating set and the verification set, determining integrated weights by an optimization algorithm, integrating the three deep coupling models by using the integrated weights, and obtaining an urban rain flood integrated model; and the urban rain flood integrated model is used for performing urban rain flood simulation calculation; the result output by the urban rain flood integrated model obtained by the embodiment of the present application is determined by the results of the three models, can capture the diversified characteristics of the hydrological sequence, obtain the optimal outflow result, and finally realize effective reduction of model uncertainty;

[0088] The embodiment of the present application first realizes deep fusion of two integrated urban rain flood models, one distributed urban rain flood model and LSTM, solves the problem that a single model cannot balance the calculation efficiency and spatial heterogeneity by optimizing the integrated weights, verifies the feasibility of the machine learning urban rain flood model based on the physical mechanism remodeling neurons, and lays a foundation for subsequent expansion to other models. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is a process schematic diagram of obtaining an urban rain flood integrated model in the urban rain flood simulation method in the embodiment 1 of the present application;

[0090] Figure 2is a schematic diagram of reconstructing the input gate, the forget gate and the output gate of the LSTM in the urban rain flood simulation method in Embodiment 1 of the present application;

[0091] Figure 3 is a schematic diagram of embedding the vector of physical parameters into the LSTM in the urban rain flood simulation method in Embodiment 1 of the present application;

[0092] Figure 4 is a schematic diagram of a study area in Embodiment 2 of the present application. DETAILED DESCRIPTION

[0093] The present application will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0094] Embodiment 1

[0095] The present embodiment provides an urban rain flood simulation method, comprising: performing urban rain flood simulation calculation by using an urban rain flood integrated model.

[0096] As shown in Figure 1 , the specific steps of obtaining the urban rain flood integrated model are as follows.

[0097] Step 1: Obtain a historical secondary flood event data set of a study area, and divide the historical secondary flood event data set into a calibration set and a validation set.

[0098] The secondary flood event data set includes land use data associated with secondary flood events, drainage pipe network topology data, meteorological data and secondary flood data.

[0099] Step 2: Based on the calibration set and the validation set, respectively, a time-varying gain nonlinear urban model, a water and soil conservation runoff model and a stormwater management model are coupled with a pre-constructed LSTM as physical constraints to obtain three deep coupling models.

[0100] Step 2.1: Based on the calibration set and the validation set, a time-varying gain nonlinear urban model is coupled with a pre-constructed LSTM as a physical constraint to obtain a first deep coupling model.

[0101] In some embodiments, the time-varying gain nonlinear urban model is represented as Time Variant Gain Model_Urban or TVGM_Urban, and the first deep coupling model is represented as P-LSTM-TVGM_Urban.

[0102] Step 2.1.1: As shown in Figure 2 , according to the physical process of the time-varying gain nonlinear urban model, the input gate, the forget gate and the output gate of the pre-constructed LSTM are reconstructed.

[0103] Based on the physical processes of the time-varying gain nonlinear urban model, the input gate, forget gate, and output gate of LSTM are reconstructed. The hydrological equation of the time-varying gain nonlinear urban model is used to replace the activation function in the LSTM neuron. Weights and bias parameters with practical physical meaning are defined. The cell state is analogous to the water storage term in the hydrological model. The cell state is updated through the consumption, increase, and outflow of the water storage term.

[0104] Specifically, the input gate is the contribution of precipitation at the current time step to soil moisture renewal, the forget gate is the degree to which previous soil moisture content is retained during cell renewal, and the output gate is the final flow output obtained from the conversion of the urban area's pipe network confluence process.

[0105] The input gate represents the contribution of precipitation at the current time step to soil moisture renewal, expressed by the following formula:

[0106] Equation (1)

[0107] In equation (1), For the first deep coupling model Time-phase input gate vector, The sigmoid activation function for the input gate; for Rainfall during the period, in mm; for The early impact of rainfall at the end of the period, and The soil moisture content before the onset of precipitation is closely related to the time period, in mm; The water content of saturated soil is expressed in mm. This is the runoff reduction factor corresponding to impermeable pavement. This is the runoff reduction factor corresponding to the building area; These are parameters related to runoff calculation in the time-varying gain nonlinear urban model.

[0108] The input gate can be analogized to the runoff calculation process in the physical model, where precipitation is converted into runoff after infiltration through the soil layer. The time-varying gain nonlinear urban model is meticulously classified into bare soil areas, vegetation-covered areas, built-up areas, and impermeable pavement areas to simulate real runoff conditions. Therefore, in equation (1), The functional expression of the physical process added to the water storage term in the time-varying gain nonlinear urban model. It can be expressed by the following formula:

[0109] Equation (2)

[0110] In equation (2), The physical formula includes the runoff generation mechanism of the time-varying gain nonlinear urban model;

[0111] in formula (2), is a group of physical parameters in the physical formula that do not change with time, including the area proportion of impervious pavement in the entire study area (unit: %), the area proportion of building area in the entire study area (unit: %), the area proportion of vegetation coverage area in the entire study area (unit: %), and the area proportion of bare soil area in the entire study area (unit: %);

[0112] in formula (2), is a group of physical parameters in the physical formula that change with time, including the saturated soil moisture , the runoff reduction coefficient corresponding to the impervious pavement , the runoff reduction coefficient corresponding to the building area , the time-varying gain nonlinear urban model parameter related to runoff calculation .

[0113] The soil moisture in the time-varying gain nonlinear urban model will be gradually consumed over time due to factors such as infiltration and evaporation, which can be analogized as the forgetting gate in LSTM, so the forgetting gate of the first deep coupling model is the degree of preservation of the previous soil moisture when updating the cell state, which is expressed by the following formula:

[0114] , formula (3),

[0115] in formula (3), is the forgetting gate vector of the first deep coupling model in the time period, is the sigmod activation function of the forgetting gate; is the dimensionless runoff coefficient; is the function expression of the physical process of consumption of the water storage term in the time-varying gain nonlinear urban model , which is expressed by the following formula:

[0116] , formula (4),

[0117] in formula (4), is the minimum value calculation; is the previous influencing rainfall at the end of the time period, which is expressed by the following formula:

[0118] , formula (5).

[0119] In the time-varying gain nonlinear urban model, the pipe network confluence process of the urban area refers to the flow and confluence process of the yield in the pipe system, which determines the final flow. Similar to the role of the output gate in LSTM, therefore, the output gate is the final flow output of the pipe network confluence process of the urban area, which is expressed by the following formula:

[0120] , formula (6),

[0121] In formula (6), is the first deep coupling model the time period output gate vector, is the sigmod activation function of the output gate; is the first deep coupling model the time period input gate vector, is the first deep coupling model the time period output gate vector, is the system conversion parameter, is the system lag time; is the function expression of the physical process of outflow in the time-varying gain nonlinear urban model , which is expressed by the following formula:

[0122] , formula (7),

[0123] In formula (7), is the flow of the time period, is the unit conversion coefficient; is the yield of the time period, mm; is the flow of the time period.

[0124] In this way, the first deep coupling model can output the flow process of the research area outlet by equivalent simulation of the pipe network confluence.

[0125] Step 2.1.2: Replace the cell state update formula in the pre-constructed LSTM with the water balance formula to update the cell state, and obtain the first deep coupling model.

[0126] The cell state of the first deep coupling model is analogous to the soil moisture update process in the time-varying gain nonlinear urban model, so the cell state is updated by replacing the cell state update formula in the pre-constructed LSTM with the water balance formula, and the first deep coupling model is obtained, which is expressed by the following formula:

[0127] , formula (8),

[0128] In formula (8), For the first deep coupling model Cell state during a given time period To determine the water balance formula for cell renewal, For the first deep coupling model Forgetting gate vector for time period, For the first deep coupling model Cell state during a given time period For the first deep coupling model Time-phase input gate vector, for Input of time period.

[0129] Step 2.2: Based on the calibration set and validation set, the water and soil conservation runoff generation model is coupled with the pre-constructed LSTM to obtain the second deep coupling model, using the water and soil conservation runoff generation model as the physical constraint.

[0130] In some embodiments, the soil and water conservation runoff model is represented as the Soil Conservation Service (SCS), specifically the USDA Soil and Water Conservation Service's soil and water conservation runoff model. In some embodiments, the second deep coupling model is represented as P-LSTM-SCS.

[0131] Step 2.2.1: As Figure 2 As shown, based on the physical process of the soil and water conservation runoff generation model, the input gate, forget gate, and output gate of the pre-constructed LSTM are reconstructed.

[0132] Specifically, the input gate is the contribution of precipitation and water retention in the catchment area at the current time step to the flow output, the forget gate is the degree to which the previous soil moisture content is retained when the cell state is renewed, and the output gate is the final flow output obtained from the transformation of the urban area's pipe network confluence process.

[0133] The input gate represents the contribution of precipitation and catchment retention at the current time step to the runoff, expressed by the following formula:

[0134] Equation (9)

[0135] In equation (9), For the second deep coupling model Time-phase input gate vector, The sigmoid activation function for the input gate; for Rainfall during the period, in mm; for Initial rainfall loss for the period, mm; CN represents the weighting coefficient in the second deep coupling model for the number of runoff curves.

[0136] The input gate can be compared to the runoff calculation process of the water and soil conservation runoff model, that is, after deducting the initial rainfall loss, the runoff is converted into runoff through soil layer penetration. The water and soil conservation runoff model uses the runoff curve number CN to represent the underlying surface conditions and the quantitative relationship between the variables, which can reflect the dynamic response of rainfall runoff. Therefore, in formula (9), f is the function expression of the physical process of runoff calculation in the water and soil conservation runoff model , which is represented by the following formula:

[0137] , formula (10),

[0138] In formula (10), the function expression is the runoff generated in the period, mm; is the soil moisture in the period, which is represented by the following formula:

[0139] , formula (11).

[0140] In the water and soil conservation runoff model, the initial loss of rainfall usually determines the actual rainfall that affects the current soil moisture change. This process is analogous to the forget gate in LSTM, so the forget gate is the degree of preservation of the previous soil moisture in updating the cell state, which is represented by the following formula:

[0141] , formula (12),

[0142] In formula (12), is the second deep coupling model the forget gate vector in the period, is the sigmod activation function of the forget gate; is the precipitation in the period, mm; is the evaporation in the current time step, mm; is the initial rainfall loss in the period, mm; is the weight coefficient represented by the dimensionless reduction coefficient dc in the coupling model; is the function expression of the physical process that determines rainfall loss in the water and soil conservation runoff model , which is represented by the following formula: , formula (13),

[0143] In formula (13), is the maximum value calculation.

[0144]

[0145] ​​​​The output gate embodies the confluence process in the urban stormwater model, and determines how the yield flow is transformed into the final flow output. Similarly, a linear system response equation similar to the lag algorithm is used to simulate the overall confluence process. Therefore, the output gate is the final flow output transformed by the pipe network confluence process in urban areas, which is represented by the following formula:

[0146] , formula (14),

[0147] In formula (14), is the second deep coupling model the period output gate vector, is the sigmod activation function of the output gate; is the second deep coupling model the period input gate vector, is the second deep coupling model the period output gate vector, is the system transformation parameter, is the system lag time; is the function expression of the physical process of confluence in the time-varying gain nonlinear urban model , which is represented by the following formula:

[0148] , formula (15),

[0149] In formula (15), is the flow in the period, is the unit conversion coefficient; is the yield flow in the period, mm; is the flow in the period.

[0150] Step 2.2.2: Update the cell state using the pre-constructed cell state update formula in the LSTM to obtain the second deep coupling model.

[0151] Step 2.3: Based on the rating set and the verification set, the stormwater management model is coupled with the pre-constructed LSTM based on the physical constraint of the stormwater management model to obtain the third deep coupling model.

[0152] In some embodiments, the stormwater management model is represented as Storm Water Management Model or SWMM, and the third deep coupling model is represented as P-LSTM-SWMM.

[0153] Step 2.3.1: Extract the physical parameter composition vector of the stormwater management model, such as Figure 3The vector is embedded into the pre-constructed LSTM, and the weights of the physical parameters in the input gate, the forgetting gate and the output gate of the pre-constructed LSTM are expanded.

[0154] The physical parameters of the storm flood management model are embedded into the input weight of the LSTM, and are transmitted to the cell state update formula through the weighting mechanism of the input gate, so that the model directly receives information related to the physical process, and the physical constraints provided by the storm flood management model are combined in each time step, thereby enhancing the perception ability of the LSTM to the physical law.

[0155] The physical parameters of the storm flood management model are extracted to form a vector , which is expressed by the following formula:

[0156] , formula (16),

[0157] In formula (16), is the Manning coefficient of the impervious area, is the Manning coefficient of the pervious area, is the impervious coefficient, is the characteristic width coefficient, is the Manning coefficient of the pipe, is the maximum infiltration rate, is the minimum infiltration rate, is the water storage of the impervious area, is the water storage of the pervious area, is the proportion of the impervious area without water storage.

[0158] In this embodiment, the physical parameters of the storm flood management model are the key parameters extracted through the feature layer.

[0159] The vector is embedded into the pre-constructed LSTM, and the weights of the physical parameters in the input gate, the forgetting gate and the output gate of the pre-constructed LSTM are expanded, and the expanded input gate, the forgetting gate and the output gate are expressed by the following formula:

[0160] , formula (17),

[0161] In formula (17), are respectively the third deep coupling model the input gate vector, the forgetting gate vector and the output gate vector of the time period; is a sigmod activation function; is the input of the time period; is the hidden state of the time period; are respectively weight matrix corresponding to the input gate, the forget gate, and the output gate; respectively weight matrix corresponding to the input gate, the forget gate, and the output gate; respectively weights of the physical parameters in the input gate, the forget gate, and the output gate; respectively offset vector parameters of the input gate, the forget gate, and the output gate.

[0162] Step 2.3.2: Adjust the cell state update formula of the LSTM based on the current precipitation input and the calculation constraints of the physical parameters of the storm flood management model, and update the cell state using the adjusted cell state update formula.

[0163] By defining the input weight, the recursion weight, and the bias through the self-defined gating mechanism, the initialization of the weight, the gradient calculation, and the injection of the physical information are realized, and the cell state is updated combined with the physical constraints. By combining the physical information with each gating unit, the update accuracy of the cell state and the hidden state is improved through dynamic adjustment, so as to optimize the weight adjustment in the training process and further improve the learning ability of the model.

[0164] The adjusted cell state update formula is represented by the following formula:

[0165] , formula (18),

[0166] In formula (18), is the third depth-coupled model cell state of the time period, is the third depth-coupled model cell state of the time period; is the dot product operation; respectively input gate vector and forget gate vector of the third depth-coupled model time period; is the third depth-coupled model candidate cell state of the time period, which is represented by the following formula:

[0167] , formula (19),

[0168] In formula (19), is the calculation constraint of the storm flood management model based on precipitation input and physical parameters of the time period.

[0169] Step 2.3.3: Apply the water balance constraint of the storm flood management model at the output end of the LSTM to obtain the third depth-coupled model.

[0170] The water balance constraint of the storm flood management model is applied at the output end of the LSTM, which is represented by the following formula:

[0171] , formula (20),

[0172] In formula (20), is a third deep coupling model Expression of the time period; is a third deep coupling model Time period output gate vector; is a third deep coupling model Cell state of the time period.

[0173] The result of the output of the third deep coupling model is a simulated runoff conforming to the physical law , which is expressed by the following formula:

[0174] , formula (21),

[0175] In formula (21), is to correct the part of the result of the output of the LSTM that does not conform to the water balance constraint by the storm flood management model.

[0176] In step 2 of the embodiment, the runoff generation mechanism of the time-varying gain nonlinear urban model, the runoff curve number of the soil and water conservation runoff model, and the physical parameters of the storm flood management model are remodeled to the gating mechanism of the LSTM, realizing the remodeling of the physical mechanism at the neuron level, so that the hydrological process has physical meaning; It not only overcomes the low calculation efficiency of traditional urban rainwater model, but also can improve the interpretability of the intermediate process of the machine learning model.

[0177] Step 3: Based on the calibration set and the verification set, the simulation results of the three deep coupling models for each outlet are integrated, the integration weight is determined by the optimization algorithm, the three deep coupling models are integrated by using the integration weight, and the urban rainwater integrated model is obtained.

[0178] Step 3.1: The simulation results of the three deep coupling models for each outlet are integrated by the following formula:

[0179] ,

[0180] Wherein, is the integrated sub-catchment flow, are the weights of the three deep coupling models, respectively, are the initial outlet flows of the three deep coupling models, respectively.

[0181] Step 3.2: The weight is calibrated by using the genetic algorithm, and the objective function of the calibration is the actual flow of the sub-catchment in the verification set and the integrated flow of the sub-catchment The Nash efficiency coefficient between them is minimized to obtain the integrated weight.

[0182] Step 3.3: Integrate the three deep-coupled models to obtain the urban stormwater integrated model using the integrated weight.

[0183] The output of the urban stormwater integrated model obtained in step 3 of the embodiment is jointly determined by the results of the three models, can capture the diversified characteristics of hydrological sequences, obtain optimal outflow results, and finally realize effective reduction of model uncertainty.

[0184] As a further improvement, the embodiment also provides a framework of an urban stormwater integrated model.

[0185] The physical mechanisms of different urban stormwater models (including but not limited to the time-varying gain nonlinear urban model, the soil and water conservation runoff model, and the stormwater management model mentioned in step 2) are embedded into the input gate, the forget gate, and the output gate of the LSTM model, all the gating mechanisms are remodeled based on the physical process, the connection between neurons is defined by the numerical solution of the physical process equation, and each gate has a set of adjustable parameters. Compared with the traditional LSTM model, the weight and bias variables defined under this framework are parameters with actual physical meaning, which are adjusted during training combined with physical information to update the gating mechanism. The specific formula is as follows:

[0186] , formula (22),

[0187] In formula (22), is the input gate vector of the time period; is the input of the time period; is the hidden state of the time period; is the forget gate vector of the time period; is the output gate vector of the time period; are sigmod activation functions of the input gate, the forget gate, and the output gate, respectively; are function expressions representing physical processes in different urban stormwater models, respectively; correspond to the weight matrix of the input gate, the forget gate, and the output gate, respectively; correspond to the weight matrix of the input gate, the forget gate, and the output gate, respectively; are weights of physical parameters in the input gate, the forget gate, and the output gate, respectively; are offset vector parameters of the input gate, the forget gate, and the output gate, respectively.

[0188] ​​​​​​​The cell state in the neural network is updated with different urban rain flood models as physical constraints, which is expressed by the following formula:

[0189] , formula (23),

[0190] In formula (23), is the cell state at the t time, is a function that determines cell update, and the expression form is different when different urban rain flood models are used as constraints.

[0191] In the traditional LSTM, the input gate determines the "importance" of the current input, the forget gate controls the "retention degree" of the previous memory, and the output gate determines the amount of information passed down from the cell state. After the coupling process provided in this embodiment, the obtained model can not only learn the statistical characteristics of the data, but also consider the mechanism of the physical process, and each time step can reflect the relationship between the physical process and the data in a more accurate way, embodying the coupling of data driving and physical process.

[0192] This embodiment first realizes the deep integration of two integrated urban rain flood models, one distributed urban rain flood model and LSTM, solves the problem that a single model cannot balance the calculation efficiency and spatial heterogeneity by optimizing the integrated weight, and verifies the feasibility of the machine learning urban rain flood model based on the remodeling of neurons with physical mechanisms, which lays a foundation for subsequent expansion to other models.

[0193] Embodiment 2:

[0194] Based on the urban rain flood simulation method provided in embodiment 1, this embodiment selects a certain administrative district of a certain city as a research area, as shown in Figure 4 , to simulate urban rain flood.

[0195] The meteorological data and secondary flood data of the research area include 6 rain gauges with 1-minute interval rainfall data and 102 flow gauges with 5-minute interval flow data, and the time range is from January 1, 2019 to December 31, 2022. After data cleaning, there are a total of 12614 400 rainfall data and 42 888 960 flow data. The land use data and drainage pipe network topology data of the research area are obtained from the city planning and design research institute of the research area.

[0196] The historical secondary flood event data of the research area is shown in Table 1.

[0197] Table 1 Historical secondary flood event data of the research area

[0198]

[0199] The traditional LSTM model, the first deep coupling model in Embodiment 1, the second deep coupling model and the third deep coupling model are used to simulate the historical secondary flood event data of the research area according to the drainage partition, and the simulation results shown in Table 2 are obtained.

[0200] Table 2 Simulation results of all drainage partitions on different models

[0201]

[0202] Table 2 intuitively shows the average NSE performance of different drainage partitions on four models, which not only reflects the adaptability and advantages and disadvantages of the four models in different drainage partitions, but also further verifies the potential of the coupling model. The average NSE of the LSTM model, the first deep coupling model P-LSTM-TVGM_Urban, the second deep coupling model P-LSTM-SCS and the third deep coupling model P-LSTM-SWMM is 0.583, 0.670, 0.632 and 0.608 respectively. The average NSE of the first deep coupling model P-LSTM-TVGM_Urban is improved by 14.92% compared with the pure data-driven LSTM model, the second deep coupling model P-LSTM-SCS is improved by 8.40%, and the third deep coupling model P-LSTM-SWMM is improved by 4.28%.

[0203] The coupling model is obviously better than the physical model and the pure data-driven model in terms of overall runoff fitting effect and peak fitting effect, which shows that the combination of physical mechanism and data-driven can significantly improve the model performance. From the perspective of model stability, the smaller the coefficient of variation (CV) is, the more stable the model is. Table 2 shows the CV values of the four models, and the CV values from high to low are: LSTM model, P-LSTM-TVGM_Urban, P-LSTM-SCS, P-LSTM-SWMM. The CV values of the coupling model show a clear downward trend compared with the pure data-driven machine learning model, which shows that the coupling model with physical mechanism can effectively improve the reliability of the simulation results.

[0204] The weights of the genetic algorithm in this embodiment are calibrated , and the objective function of the calibration is to minimize the Nash efficiency coefficient (NSE) between the measured outflow of the subcatchment and .

[0205] In summary, after the coupling process provided in Embodiment 1, the obtained model can not only learn the statistical characteristics of the data, but also consider the mechanism of the physical process, and at each time step, it can reflect the relationship between the physical process and the data in a more accurate way, embodying the coupling of data-driven and physical process.

[0206] Embodiment 3:

[0207] The embodiment provides a city rain flood simulation device, comprising:

[0208] The acquisition module is configured to acquire a historical secondary flood event data set of a research area, and divide the historical secondary flood event data set into a calibration set and a verification set.

[0209] The deep coupling module is configured to, based on the calibration set and the verification set, respectively take a time-varying gain nonlinear city model, a water and soil conservation runoff model and a storm flood management model as physical constraints, and couple the LSTM to obtain three deep coupling models.

[0210] The integration module is configured to, based on the calibration set and the verification set, integrate simulation results of the three deep coupling models for each outlet, determine an integrated weight by using an optimization algorithm, and integrate the three deep coupling models by using the integrated weight to obtain a city rain flood integrated model.

[0211] Those skilled in the art will understand that embodiments of the application can be provided as methods, systems or computer program products. Therefore, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0212] The application is described with reference to flowcharts and / or block diagrams of the method, device (system) and computer program product according to embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing apparatus to produce a machine, so that the instructions executed by the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 An apparatus for performing each flow or multiple flows and / or blocks Figure 1 An apparatus for performing each flow or multiple flows and / or blocks

[0213] These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0214] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, so that the instructions executed on the computer or other programmable data processing devices provide processes for implementing the flow Figure 1 one or more processes and / or blocks Figure 1 the steps of the function specified in the one or more blocks.

[0215] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative, but not restrictive, and those of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, which are all within the protection of the present application.

Claims

1. A method of urban storm flood simulation, characterized in that, The method comprises the following steps: obtain a historical secondary flood event data set of a research area, and divide the historical secondary flood event data set into a calibration set and a verification set; wherein the secondary flood event data set comprises land use data, drainage pipe network topology data, meteorological data and secondary flood data associated with the secondary flood event; based on the calibration set and the verification set, a time-varying gain nonlinear urban model, a water and soil conservation runoff model and a stormwater management model are coupled with a pre-constructed LSTM to obtain three deep coupling models, including: based on the calibration set and the verification set, a time-varying gain nonlinear urban model is coupled with a pre-constructed LSTM to obtain a first deep coupling model, including: According to the physical process of the time-varying gain nonlinear urban model, the input gate, the forget gate and the output gate of the pre-constructed LSTM are reconstructed; wherein the input gate is the contribution of the current time step precipitation to the soil moisture update, the forget gate is the degree of preservation of the previous soil moisture in updating the cell state, and the output gate is the final flow output obtained by the pipe network confluence process in urban areas; The cell state update formula in the pre-constructed LSTM is replaced by the water balance formula to update the cell state, and a first deep coupling model is obtained; based on the calibration set and the verification set, a water and soil conservation runoff model is coupled with a pre-constructed LSTM to obtain a second deep coupling model, including: According to the physical process of the water and soil conservation runoff model, the input gate, the forget gate and the output gate of the pre-constructed LSTM are reconstructed; wherein the input gate is the contribution of the current time step precipitation and the detention volume of the catchment area to the runoff, the forget gate is the degree of preservation of the previous soil moisture in updating the cell state, and the output gate is the final flow output obtained by the pipe network confluence process in urban areas; The cell state update formula in the pre-constructed LSTM is replaced by the water balance formula to update the cell state, and a first deep coupling model is obtained; based on the calibration set and the verification set, a time-varying gain nonlinear urban model is coupled with a pre-constructed LSTM to obtain a first deep coupling model, including: The physical parameter vector of the stormwater management model is extracted and embedded into the pre-constructed LSTM, and the weights of the physical parameters in the input gate, the forget gate and the output gate of the pre-constructed LSTM are expanded; The cell state update formula of the LSTM is adjusted based on the calculation constraint of the stormwater management model based on the current precipitation input and the physical parameters, and the cell state is updated by using the adjusted cell state update formula; The water balance constraint of the stormwater management model is applied to the output end of the LSTM to obtain a third deep coupling model; based on the calibration set and the verification set, the simulation results of the three deep coupling models are integrated for each outlet, the integration weight is determined by an optimization algorithm, and the three deep coupling models are integrated by using the integration weight to obtain a city rainwater integration model; the city rainwater integration model is used for performing city rainwater simulation calculation.

2. The urban storm flood simulation method of claim 1, wherein, The input gate, the forget gate and the output gate of the pre-constructed LSTM are reconstructed, including: The input gate is the contribution of the current time step precipitation to the soil moisture update, and is represented by the following formula: ， in, For the first deep coupling model Time-phase input gate vector, The sigmoid activation function for the input gate; for Rainfall during a given period; for The early impact of rainfall at the end of the period, and It is closely related to the soil moisture content before the onset of precipitation in a given period; This represents the saturated soil moisture content. This is the runoff reduction factor corresponding to impermeable pavement. This is the runoff reduction factor corresponding to the building area; These are parameters related to runoff calculation in the time-varying gain nonlinear urban model. The functional expression for the physical process added to the water storage term in the time-varying gain nonlinear urban model is represented by the following equation: ， wherein, is a physical equation that includes the runoff generation mechanism of the time-varying gain nonlinear urban model; is a group of physical parameters in the physical equation that do not vary with time, including the area proportion of impervious surface in the entire study area , the area proportion of building area in the entire study area , the area proportion of vegetation coverage area in the entire study area , and the area proportion of bare soil area in the entire study area ; is a group of physical parameters in the physical equation that vary with time, including ; The forget gate is the degree of retention of the previous soil moisture when updating the cell state, and is represented by the following formula: ， wherein, is a first depth coupling model is a period forget gate vector, is a sigmod activation function for the forget gate; is a dimensionless recession coefficient; is a function expression of the physical process of the storage term consumption in the time-varying gain nonlinear urban model, expressed by: ， wherein, is calculated for the minimum value, is the antecedent rainfall at the end of the period; The output gate is the final flow output obtained by the pipe network confluence process in urban areas, and is represented by the following formula: ， wherein, is a first depth coupled model a period output gate vector, is a sigmod activation function for the output gate; is a first depth coupled model a period input gate vector, is a first depth coupled model a period output gate vector, is a system conversion parameter, is a system lag time; is a function expression of the physical process of outflow in the time-varying gain nonlinear urban model, expressed by: ， wherein is the flow rate of the period, is the unit conversion factor, is the production flow rate of the period, is the flow rate of the period.

3. The urban storm flood simulation method of claim 2, wherein, The water balance equation is used to replace the pre-constructed LSTM cell state update formula to update the cell state, and a first deep coupling model is obtained, represented by the following formula: ， wherein, is a first deep coupling model cell states for the period, is a water balance equation that determines cell renewal, is a first deep coupling model forget gate vector for the period, is a first deep coupling model cell states for the period, is a first deep coupling model input gate vector for the period, is input for the period.

4. The urban storm flood simulation method of claim 1, wherein, The input gate is the contribution of the current time step precipitation to the soil moisture update, and is represented by the following formula: The forget gate is the degree of retention of the previous soil moisture when updating the cell state, and is represented by the following formula: ， wherein, is a second depth coupling model is a period input gate vector, is a sigmod activation function for the input gate; is a is a period precipitation; is a is a period initial rainfall loss; is a weight coefficient represented by a runoff curve number CN in the second depth coupling model; is a function expression of a physical process for runoff calculation in the water and soil conservation runoff model, represented by the following formula: ， wherein, is period soil moisture content; The output gate is the final flow output obtained by the pipe network confluence process in urban areas, and is represented by the following formula: ， wherein, is the second depth coupling model is the time period forgetting gate vector, is the sigmod activation function for the forgetting gate; is the current time step evaporation; is the is the time period initial rainfall loss; is the weight coefficient represented by the damping coefficient dc in the coupling model; is the function expression of the physical process determining the rainfall loss in the water and soil conservation runoff model, which is represented by the following formula: ， wherein is the maximum value calculation; The physical parameter vector of the storm flood management model is extracted, and the vector is embedded into the pre-constructed LSTM to expand the weights of the physical parameters in the input gate, the forget gate and the output gate of the pre-constructed LSTM, including: ， wherein, is a second deep coupling model a period output gate vector, is a sigmod activation function for the output gate; is a second deep coupling model a period input gate vector, is a second deep coupling model a period output gate vector, is a system conversion parameter, is a system lag time; is a function expression of the physical process of confluence in the time-varying gain nonlinear urban model, expressed by: ， wherein, is flow rate of the period, is a unit conversion factor, is flow rate of the period, is flow rate of the period.

5. The urban storm flood simulation method of claim 1, wherein, The vector is embedded into the pre-constructed LSTM to expand the weights of the physical parameters in the input gate, the forget gate and the output gate of the pre-constructed LSTM, and the expanded input gate, the forget gate and the output gate are represented by the following formula: extracting a physical parameter composition vector of a storm flood management model is expressed by the following equation: ， wherein, is the Manning coefficient for impervious areas, is the Manning coefficient for pervious areas, is the impervious coefficient, is the characteristic width coefficient, is the Manning coefficient for pipes, is the maximum infiltration rate, is the minimum infiltration rate, is the water storage for impervious areas, is the water storage for pervious areas, is the proportion of impervious areas without water storage; The cell state update formula of the LSTM is adjusted based on the calculation constraints of the storm flood management model based on the current precipitation input and physical parameters, and the cell state is updated using the adjusted cell state update formula, wherein the adjusted cell state update formula is represented by the following formula: ， wherein, are third deep coupling models are period input gate vectors, forget gate vectors, output gate vectors; is a sigmod activation function; is is an input of the period; is is a hidden state of the period; are are weight matrices corresponding to the input gate, the forget gate, the output gate; are are weight matrices corresponding to the input gate, the forget gate, the output gate; are weights of physical parameters in the input gate, the forget gate, the output gate; are offset vector parameters of the input gate, the forget gate, the output gate.

6. The urban storm flood simulation method of claim 1, wherein, The water balance constraint of the storm flood management model is applied to the output end of the LSTM to obtain a third deep coupling model, wherein the water balance constraint of the storm flood management model is applied to the output end of the LSTM by the following formula: ， wherein, is a third deep coupling model cell state of the time period, is a third deep coupling model cell state of the time period; is a dot product operation; is a third deep coupling model time period input gate vector, forget gate vector; is a third deep coupling model time period candidate cell state, represented by: ， wherein, is a storm flood management model based on the calculation of the period precipitation input and physical parameters.

7. The urban storm flood simulation method of claim 1, wherein, Based on the calibration set and the verification set, the simulation results of the three deep coupling models for each outlet are integrated by an optimization algorithm to determine the integrated weights, and the three deep coupling models are integrated using the integrated weights to obtain a city rain flood integrated model, including: ， wherein, is a third depth coupling model expression of the period; is a third depth coupling model period output gate vector; is a third depth coupling model cell state of the period; The result output by the third deep coupling model is a simulated runoff conforming to physical laws represented by the following formula: ， wherein, To correct the part of the results of the output of the LSTM that does not meet the water balance constraint by the storm flood management model.

8. The urban storm flood simulation method of claim 1, wherein, The simulation results of the three deep coupling models for each outlet are integrated by the following formula: The three deep coupling models are integrated using the integrated weights to obtain a city rain flood integrated model. ， wherein, is the integrated subcatchment flow, are the weights of the three depth-coupled models, respectively, are the initial outlet flows of the three depth-coupled models, respectively. The weights are calibrated using a genetic algorithm, with a target function of minimizing the Nash efficiency coefficient between the actual flow of the sub-catchment in the validation set and the integrated flow of the sub-catchment The integrated weights are obtained. ​ Including:

9. A city rain flood simulation device based on the city rain flood simulation method according to claim 1, characterized by, An acquisition module is configured to acquire a historical secondary flood event data set of a research area, and divide the historical secondary flood event data set into a calibration set and a verification set; wherein the secondary flood event data set includes land use data, drainage pipe network topology data, meteorological data and secondary flood data associated with a secondary flood event; A deep coupling module is configured to couple a pre-constructed LSTM based on the calibration set and the verification set, respectively, with a time-varying gain nonlinear urban model, a water and soil conservation runoff model and a storm flood management model as physical constraints, to obtain three deep coupling models; The input gate is the contribution of the current time step precipitation to the soil moisture update, and is represented by the following formula: The forget gate is the degree of retention of the previous soil moisture when updating the cell state, and is represented by the following formula: The output gate is the final flow output obtained by the pipe network confluence process in urban areas, and is represented by the following formula: The water balance equation is used to replace the pre-constructed LSTM cell state update formula to update the cell state, and a first deep coupling model is obtained, represented by the following formula: The input gate is the contribution of the current time step precipitation to the soil moisture update, and is represented by the following formula: The forget gate is the degree of retention of the previous soil moisture when updating the cell state, and is represented by the following formula: The output gate is the final flow output obtained by the pipe network confluence process in urban areas, and is represented by the following formula: The physical parameter vector of the storm flood management model is extracted, and the vector is embedded into the pre-constructed LSTM to expand the weights of the physical parameters in the input gate, the forget gate and the output gate of the pre-constructed LSTM, including: The vector is embedded into the pre-constructed LSTM to expand the weights of the physical parameters in the input gate, the forget gate and the output gate of the pre-constructed LSTM, and the expanded input gate, the forget gate and the output gate are represented by the following formula: The cell state update formula of the LSTM is adjusted based on the calculation constraints of the storm flood management model based on the current precipitation input and physical parameters, and the cell state is updated using the adjusted cell state update formula, wherein the adjusted cell state update formula is represented by the following formula: The water balance constraint of the storm flood management model is applied to the output end of the LSTM to obtain a third deep coupling model, wherein the water balance constraint of the storm flood management model is applied to the output end of the LSTM by the following formula: Based on the calibration set and the verification set, the simulation results of the three deep coupling models for each outlet are integrated by an optimization algorithm to determine the integrated weights, and the three deep coupling models are integrated using the integrated weights to obtain a city rain flood integrated model, including: The simulation results of the three deep coupling models for each outlet are integrated by the following formula: The three deep coupling models are integrated using the integrated weights to obtain a city rain flood integrated model. Including: An acquisition module is configured to acquire a historical secondary flood event data set of a research area, and divide the historical secondary flood event data set into a calibration set and a verification set; wherein the secondary flood event data set includes land use data, drainage pipe network topology data, meteorological data and secondary flood data associated with a secondary flood event; A deep coupling module is configured to couple a pre-constructed LSTM based on the calibration set and the verification set, respectively, with a time-varying gain nonlinear urban model, a water and soil conservation runoff model and a storm flood management model as physical constraints, to obtain three deep coupling models; The integrated module is configured to integrate simulation results of the three deep coupling models for each outlet based on the rating set and the verification set, determine integrated weights by using an optimization algorithm, integrate the three deep coupling models by using the integrated weights, and obtain an urban rain flood integrated model; the urban rain flood integrated model is configured to perform urban rain flood simulation calculation.

Images