Urban rainfall flood toughness deep learning evaluation method based on machine learning

Abstract

The invention discloses an urban rainfall flood toughness deep learning evaluation method based on machine learning. The method comprises the following steps of 1, explaining time-space differentiation characteristics of urban rainfall flood; 2, analyzing and calculating urban rainfall flood toughness indexes; step 3, urban rainfall flood toughness evaluation is carried out; 4, constructing an urban rainfall flood disaster simulation model; and step 5, an evaluation prediction learning method based on machine learning. The method has the beneficial effects that the urban rainfall flood toughness can be accurately evaluated by utilizing a machine learning technology, and in-depth feasibility analysis can be carried out by virtue of a deep learning neural network model; according to the technology, multi-scale features can be learned and extracted in the space-time dimension, the nonlinear relation in urban rainfall flood toughness space-time data is effectively captured, the accuracy and stability of urban rainfall flood toughness prediction are improved, then the urban rainfall flood toughness is regulated and controlled, a corresponding urban emergency scheme is determined in time, and the overall toughness of a city is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of urban rain - flood resilience assessment, and particularly to a deep - learning assessment method for urban rain - flood resilience based on machine learning. Background Technique

[0002] Affected by the dual influence of global warming and the accelerating urbanization process, extreme rainfall events occur frequently in urban areas, resulting in a significant increase in the frequency and losses of urban flood disasters; urbanization increases the impervious area of the underlying surface, leading to an increase in surface runoff while reducing the surface runoff time; the increase in population and building density exacerbates the urban heat island effect and rain island effect, the precipitation in urban areas gradually increases, rainfall is more concentrated in the urban core area, and the frequency of extreme precipitation in the urban area shows an upward trend;

[0003] The simulation and risk assessment of urban rain - flood disasters are the basis for studying the characteristics of urban rain - floods, so as to carry out the prediction and assessment of urban rain - flood resilience and disaster warning based on machine learning, and are one of the important means for urban disaster prevention and reduction; the rain - flood process is a multi - scale fusion process including a series of links such as surface runoff generation and concentration, pipe - network and river confluence, and surface inundation;

[0004] In view of the above situation, integrating multi - source information is the research direction of future urban rain - flood resilience. The comprehensiveness and generality of rain - flood disaster simulation models in China need to be improved, and at the same time, a scientific and perfect urban rain - flood resilience management system needs to be established. Summary of the Invention

[0005] To solve the above - mentioned technical problems, the present invention provides a deep - learning assessment method for urban rain - flood resilience based on machine learning, including the following steps:

[0006] Step 1: Clarify the spatio - temporal differentiation characteristics of urban rain - floods;

[0007] Step 2: Analyze and calculate urban rain - flood resilience indicators;

[0008] Step 3: Conduct urban rain - flood resilience assessment;

[0009] Step 4: Build an urban rain - flood disaster simulation model;

[0010] Step 5: An assessment and prediction learning method based on machine learning.

[0011] As a further supplement to the technical solution, Step 1 includes the following working steps:

[0012] Step 1): Investigate the current situation of urban rain - flood resilience sources;

[0013] Step 2): Determine the main rain - flood periods.

[0014] For further supplement to this technical solution, it is determined that the SCS runoff curve number method (Soil Conservation Service curve number) is used to estimate the relationship between rainfall and runoff during the main rainstorm period, and the SWMM model is used to simulate the urban rainfall runoff process.

[0015] For further supplement to this technical solution, step 2 includes the following working steps:

[0016] Step 1): Calculate the maximum potential soil water holding capacity and infiltration rate. Soil water storage and drainage capacity are the main factors determining whether the land can absorb precipitation and maintain its original state.

[0017] Step 2): Calculate precipitation runoff, total runoff volume and outflow rate. If the rainfall exceeds the infiltration rate of the soil, surface runoff will occur.

[0018] Step 3): Calculate urban assets and execution speed.

[0019] For further supplement to this technical solution, step 3 includes the following working steps:

[0020] Step 1): Use the ArcGIS spatial analysis tool to couple flood information with social engineering information through the ArcGIS platform to analyze urban rainstorm resilience.

[0021] Step 2): By evaluating land use changes and their impacts on parameters such as the maximum potential water holding capacity and infiltration rate, use the scientific computing foundation of python (numpy library) combined with the advanced scientific computing library (scipy library) to calculate the Pearson correlation coefficient and p-value, and calculate the urban rainstorm resilience using the Pearson correlation coefficient, so as to reflect the system's performance, recovery time and recovery efficiency, and evaluate the spatio-temporal evolution of urban rainstorm resilience.

[0022] For further supplement to this technical solution, step 4 includes the following working steps:

[0023] Step 1): Collect urban rainfall data, surface water accumulation data and underground pipe network monitoring data.

[0024] Step 2): Establish an urban rainstorm disaster model by coupling the SWMM and Infoworks ICM-2D models.

[0025] For further supplement to this technical solution, step 1) includes rainfall period, surface water accumulation monitoring equipment and underground pipe network detection equipment.

[0026] For further supplement to this technical solution, step 5 includes the following working steps:

[0027] Step 1): Use the NLP text processing method in machine learning algorithms to extract data and establish a database.

[0028] Step 2): Input the extracted data into the improved deep learning-based neural network model;

[0029] Step 3): Introduce the MIV algorithm to filter and purify the extracted text data;

[0030] Step 4): Use the deep learning-based neural network model to predict and evaluate the urban stormwater resilience.

[0031] Among them, in step 5, the mean square error is selected as the index to evaluate the prediction effect of the model, and the resilience change trend is analyzed in combination with simulation scenarios (such as changes in rainstorm intensity and infrastructure damage).

[0032] As a further supplement to this technical solution, in step three, the prediction and evaluation of urban stormwater resilience uses the random forest-entropy weight method to determine the weights, and then a feasible construction strategy is proposed based on the prediction and evaluation of the deep learning-based neural network model under machine learning.

[0033] As a further supplement to this technical solution, set urban stormwater objectives and spatio-temporal differentiation characteristics, verify and estimate the urban stormwater runoff process, the recovery ability of the underlying surface, and rainstorm disaster simulation, and finally use the learning method of machine learning to predict and evaluate the urban stormwater resilience.

[0034] Its beneficial effects are as follows: Using machine learning technology can accurately evaluate urban stormwater resilience and can conduct in-depth feasibility analysis with the help of the deep learning-based neural network model; this technology can learn and extract multi-scale features in the spatio-temporal dimension, effectively capture the non-linear relationships in urban stormwater resilience spatio-temporal data, and improve the accuracy and stability of urban stormwater resilience prediction, thereby regulating urban stormwater resilience, timely determining corresponding urban emergency plans, and enhancing the overall resilience of the city. Brief Description of the Drawings

[0035] Figure 1 is the urban stormwater disaster risk map of the present invention;

[0036] Figure 2 is the coupling urban stormwater resilience evaluation and prediction framework model diagram of the present invention based on deep learning and neural network;

[0037] Figure 3 is the LSTM structure diagram of the present invention;

[0038] Figure 4 is the CNN-LSTM model prediction flow chart of the present invention. Detailed Embodiment

[0039] For the convenience of those skilled in the art to understand this technical solution more clearly, the following will be combined with the attached Figures 1-4Elaborate on the technical solution of the present invention:

[0040] A deep learning evaluation method for urban rain - flood resilience based on machine learning, comprising the following steps:

[0041] Step 1, clarify the spatio - temporal differentiation characteristics of urban rain - flood;

[0042] Step 2, analyze and calculate urban rain - flood resilience indicators;

[0043] Step 3, conduct urban rain - flood resilience assessment;

[0044] Step 4, construct an urban rain - flood disaster simulation model;

[0045] Step 5, an assessment and prediction learning method based on machine learning.

[0046] 1. Clarify the spatio - temporal differentiation characteristics of urban rain - flood

[0047] 1.1 Investigate the current situation of urban rain - flood resilience sources

[0048] 1.1.1 Temporal variation of rain - flood

[0049] The characteristics of urban rain - flood include not only temporal changes but also topographic differences and land - use distribution characteristics in space. The temporal characteristics of urban rain - flood are closely related to the duration and intensity of rainfall. The longer the rainfall duration and the greater the intensity, the more rain - flood volume is generated. By analyzing rainfall data in different time periods, such as hourly rainfall, daily rainfall, etc., the variation law of rainfall intensity over time can be understood, and then the generation and development trend of rain - flood can be inferred.

[0050] There are differences in rainfall conditions in different seasons and years. In some areas, rainfall is concentrated and intense in summer, which is a high - incidence period of urban rain - flood; while rainfall is scarce in winter, and the rain - flood risk is relatively low. Inter - annually, extreme rainfall events may occur in some years due to factors such as climate anomalies, leading to severe urban rain - flood disasters. For example, in the southeastern coastal areas of China affected by typhoons, rainfall is frequent and intense in summer, and the urban rain - flood problem is relatively prominent. The rainfall recurrence period refers to how many years on average a certain rainfall intensity appears in a long period, which reflects the rarity of rainfall events. Rainfall with different recurrence periods has different impacts on urban rain - flood. The larger the recurrence period, the higher the rainfall intensity, the more rain - flood volume is generated, and the greater the pressure on the urban drainage system. For example, when the rainfall recurrence period increases from 5 years to 50 years, the peak reduction rate of waterlogging depth in a certain area of Ningbo drops from 0.49% to 0.10%, indicating that the effect of reducing waterlogging depth caused by high - recurrence - period rainfall weakens.

[0051] 1.1.2 Spatial variation of rain - flood

[0052] The topography and geomorphology of a city have a significant impact on the spatial distribution of rainstorms and floods. Low-lying areas are prone to waterlogging, while highlands are relatively less affected by rainstorms and floods. In addition, the slope of the city also affects the flow velocity and direction of rainstorms and floods; in areas with a larger slope, the rainwater flow velocity increases, and the confluence time shortens, which may lead to concentrated waterlogging in the downstream low-lying areas. Different land use types have different abilities to infiltrate, store, and discharge rainwater, thus affecting the spatial distribution of rainstorms and floods. For example, impervious surfaces in the city, such as roads, squares, and building roofs, have weak rainwater infiltration ability, and most of the rainwater is converted into surface runoff, increasing the amount of rainstorms and floods and the peak flow rate, which is prone to cause waterlogging; while areas with good water permeability, such as green spaces, wetlands, and parks, can effectively absorb and store rainwater, reducing the pressure of rainstorms and floods. The layout and drainage capacity of the urban drainage system are the key factors affecting the spatial distribution of rainstorms and floods. Areas with dense drainage pipe networks, larger pipe diameters, and reasonable drainage pump station configurations can effectively transport and discharge rainwater, reducing the occurrence of waterlogging; on the contrary, areas with insufficient drainage facilities or aging and disrepair are prone to poor rainwater drainage, resulting in waterlogging. The hydrogeological conditions of the city, such as the groundwater level and soil permeability, also affect the spatial distribution of rainstorms and floods. In areas with a higher groundwater level, the soil water content is saturated, rainwater infiltration is difficult, the surface runoff increases, and waterlogging is easily formed; while in areas with good soil permeability, rainwater can quickly infiltrate into the ground, reducing surface runoff and lowering the risk of rainstorms and floods.

[0053] 1.2 Determine the main rainstorm and flood periods

[0054] Based on the analysis of the temporal and spatial changes of urban rainstorms and floods, determine the main rainstorm and flood periods, collect information such as historical rainfall, rainfall intensity, and frequency, and statistically analyze their seasonal distribution, interannual changes, and the frequency and intensity of extreme rainfall events, so as to identify the periods with the most concentrated rainfall and the greatest intensity. Secondly, the topography and geomorphology of the city and the drainage system capacity are also key factors. It is necessary to understand the elevation, slope changes, water body distribution, and the improvement degree of the drainage system to judge whether it can effectively cope with different intensities of rainfall. At the same time, refer to the historical records of rainstorm and flood events, analyze the relationship between rainstorms and floods, find out the rainfall thresholds and conditions for the occurrence of rainstorms and floods, and combine the current rainfall data and topographic and geomorphic factors to predict the possible main rainstorm and flood periods in the future.

[0055] 2. Analyze and calculate urban rainstorm and flood resilience indicators

[0056] Considering the original resistance and recovery ability of the urban underlying surface, deeply analyze the maximum potential soil water holding capacity, land infiltration rate, total runoff and other rainstorm-related index parameters of land use change, and establish three levels of evaluation. One is the system performance, which measures the ability to resist rainstorms; the second is the recovery time, which measures the recovery ability after rainstorm erosion; the third is the recovery efficiency, which measures the ability to adapt to urban renewal. The evaluation concepts of system performance, recovery time and recovery efficiency are based on the direct impact of rainstorms on the urban natural environment. The specific evaluation methods are as follows:

[0057] 2.1 System Performance

[0058] System performance is used to estimate the ability of urban land to absorb specific disturbances or maintain its original state before the disturbance. The soil water storage and drainage capacity are the main factors determining whether the land can absorb precipitation and maintain its original state. The soil infiltration rate refers to the amount of water infiltrating into the soil per unit area of the ground surface per unit time, which measures the ability of the soil to absorb and transmit water in a given period. The maximum potential soil water holding capacity specifies the ability of the soil to hold water, which can be determined for various land covers and underlying surface environments. The formula for system performance is as follows:

[0059]

[0060] S is the maximum potential soil water holding capacity (mm), given by Equation (2); IR represents the infiltration rate of the land (mm / h). Taking Guangzhou City as an example according to the research of Guo Wenhao et al., the infiltration rates (IR) of ecological land, agricultural land and construction land in Guangzhou City are approximately 135.4 mm / h, 89.3 mm / h and 38.1 mm / h respectively.

[0061] CN is the runoff curve number (dimensionless), which is an important parameter for describing the rainfall-runoff relationship. The CN runoff curve number can be used as a quantitative description of the comprehensive characteristics of the underlying surface, including soil geology, topographic geology conditions, vegetation types, etc., and can reflect the runoff generation ability of different underlying surface types. The CN value theoretically ranges from 0 to 100, and the actual value range is between 30 and 98.

[0062] 2.2 Recovery Time

[0063] If the rainfall exceeds the infiltration rate of the soil, surface runoff will occur. Therefore, the recovery time can be used to measure the time required for the land to recover from the flood to its original state. The recovery time is calculated using the following equation:

[0064]

[0065] TR = R × FA (4)

[0066] OR = V × A (5)

[0067] where TR represents the total runoff (m 3 ), which is calculated by multiplying the runoff (R) (mm) by the flow accumulation (FA) (m 2 ) using ArcGIS software; OR represents the outflow rate (m 3 / s), which is calculated by multiplying the flow velocity (V) (m / s) by the flow cross-section area (A) (m 2 ).

[0068]

[0069] A = L × d (2 - 7) (7)

[0070]

[0071] where k is the constant velocity of overland flow (m / s), and the constant velocity values used are 0.48 m / s, 1.32 m / s, and 6.22 m / s, applicable to ecological land, agricultural land, and construction land; S0 represents the slope; L is the length (m) representing one side of the grid cell; d is the cumulative depth of runoff (m), calculated using formula (8); P is the cumulative rainfall depth (mm), with 250 mm of extreme heavy rain as the cumulative rainfall depth; S is as shown in the previous formula (2). When land use changes, taking the ecological space converted to construction land as an example, the recovery time is reduced from 10.8 hours (CN is 55, P is 15 mm, L is 1000 m, S0 is 5%, k is 0.48 m / s, TR is 10 9 cubic meters) to 9.4 h (CN is 93, k is 6.22 m / s, and the values of P, L, and S0 are the same). The change in the recovery time means that the recovery time of construction land is shorter than that of the ecological space, which implies the benefits of land use changing from the natural environment to the human construction environment.

[0072] 2.3 Recovery Efficiency

[0073] Human intervention can help the land return to its original state. Even when the land exceeds its flood absorption capacity and cannot naturally recover, it can prevent flood damage to the land. Recovery efficiency is a method of evaluating the artificial ability of the land to return to its original state or resist typhoon and precipitation damage. Recovery efficiency is an estimate of its human recovery potential. The formula for recovery efficiency is:

[0074]

[0075] where Au represents "urban assets", calculated as the building area (km 2 ), that is, the construction land area within the unit grid. RR is the recovery execution speed (km 2 / h), the restoration efficiency is in hours (h), and the execution speeds for construction land and non-construction land are different. For example, based on the historical statistical data on the time taken for the Guangzhou Municipal Government to clear damages, the execution speeds for urban construction land and non-construction land are 0.3915 km 2 / h and 0.1566 km 2 / h, which means that compared with non-urban construction land, the government needs more than twice the effort to clear the damages caused by urban construction land rainwater floods.

[0076] 3. Conduct urban rainwater flood resilience assessment

[0077] To analyze the influencing factors of urban rainwater flood resilience, understand the spatial correlation between urban rainwater flood resilience and other datasets, and provide important support and reference for fields such as urban planning and natural resource management. Calculate the correlation between urban rainwater flood resilience and raster datasets using the Pearson correlation coefficient. The Pearson correlation coefficient is a statistic used to measure the degree of linear correlation between two variables, usually denoted by r. The range of correlation is from +1 to -1. A positive correlation indicates a positive relationship between two raster data, meaning that when the pixel value of one raster data increases, the pixel value of the other raster data may also increase. A negative correlation means that one raster data changes inversely with the other raster data. When the correlation is zero, it means there is no dependence relationship between the two raster data. The Pearson correlation coefficient is calculated as the ratio of the covariance between two raster data to the product of the standard deviations of the two raster data. Since the Pearson correlation coefficient is a ratio, it has no unit. The formula for calculating the Pearson correlation is as follows:

[0078]

[0079] where Cov ij represents the covariance between raster data i and raster data j, and δ i represents the standard deviation of raster data i. The covariance Cov between raster data i and raster data j ij :

[0080]

[0081] where: Z is the pixel value, k represents a specific pixel, i, j represent raster data, and Z ik represents the k-th pixel value of the i-th raster data; μ iIt represents the average value of the raster data; N represents the number of pixels. When performing a correlation analysis of two variables, the Pearson coefficient is calculated, and usually a p-value is also calculated to indicate whether the correlation coefficient is statistically significant. The p-value represents the probability of observing the correlation coefficient under the condition that the null hypothesis holds. The null hypothesis usually means that there is no correlation between the two variables. If the calculated P-value is less than the pre-set significance level (such as 0.05), the null hypothesis can be rejected, indicating that there is a significant correlation between the two variables. For example, if the calculated Pearson coefficient is 0.8 and the corresponding P-value is 0.01, it can be considered that there is a high positive correlation between these two variables, and this correlation is very significant. On the other hand, if the Pearson coefficient is 0.2 and the corresponding p-value is 0.6, it can be considered that there is a certain degree of positive correlation between the two variables, but this correlation is not significant. The Pearson correlation coefficient and p-value are calculated using the scientific computing foundation of python (numpy library) combined with the advanced scientific computing library (scipy library).

[0082] 4. Construct a simulation model for urban rainstorm flood disasters

[0083] The urban rainstorm flood process is a physical process that includes multiple processes and multi-scale fusions such as urban underlying surface runoff generation, slope runoff concentration, river channel runoff concentration, and pipe network runoff concentration. Urban rainstorm flood disasters are the responses to the risks of urban rainstorm floods, as shown in the appendix Figure 1 . Rainfall, surface water accumulation, urban pipe network and other data are the key input variables of the urban rainstorm flood model. Quantitative analysis of the spatio-temporal variation characteristics of rainfall is the basis for studying urban hydrological processes. The precipitation that falls on the urban underlying surface forms surface runoff and subsurface runoff after processes such as vegetation interception, soil infiltration, and land evaporation, which are the direct input sources in the urban rainstorm flood runoff concentration calculation process.

[0084] 4.1 Simulate the rainfall-runoff process

[0085] The rainfall runoff process of the basin needs to be simulated using hydrodynamic modeling and high-resolution (or locally refined) grids. By coupling with the Green-Ampt model, the GPU acceleration and the finite volume shallow water model based on LTS (LTS: Local Time Step) are extended to rainfall runoff simulation. It is achieved by treating rainfall intensity and permeability as source terms in the control equation. In particular, this model is implemented on the GPU using the NVIDIA CUDA framework for parallel computing, and the LTS method is used for solution update. When applied to the basins of coastal cities in China, this model can also capture the main patterns of rainfall-induced floods.

[0086] Mathematical formula for the permeability of the water flow process

[0087] The control equation is written in vector form as follows

[0088]

[0089] wherein,

[0090] U is a vector of conserved physical variables; F, G are magnetic flux vectors containing advection terms and hydrostatic pressure; S s , S b and S f are source term vectors containing rainfall / infiltration, riverbed slope, and friction terms; t is time; x, y are horizontal directions; h is water depth; u, v are depth-averaged velocities in the x and y directions respectively; g = 9.8 m / s 2 ; and are riverbed slopes, where Z b is the riverbed elevation; and are friction slopes, where n is Manning's roughness; r is the rainfall intensity, specified according to the specific situation; i is the soil infiltration rate. The revised Green-Ampt model (Liu and Singh, 2004; Mein and Larson, 1973) is used. According to the Green-Ampt model, there is a wetting front extending downward in the soil as a dividing line. The soil is considered saturated above the wetting front and unsaturated below the wetting front. This model assumes that the wetting front can extend infinitely downward and ignores the horizontal diffusion and flow of infiltrated water. The calculation method of the infiltration rate is as follows:

[0091]

[0092] i is the infiltration rate; I is the cumulative infiltration; K s is the saturated hydraulic conductivity of the soil; Ψ is the suction of the wetting surface; θ s is the saturated volumetric water content (cm 3 / cm 3 ); is the initial volumetric water content (cm 3 / cm 3 ). When the rainfall intensity r is greater than the infiltration capacity, water begins to pond on the ground. When the infiltration rate i = r, ponding occurs. The cumulative infiltration I r can be obtained through the Green-Ampt model as follows:

[0093]

[0094] The ponding time is:

[0095] t r = I r / r (16)

[0096] Generally speaking, the permeability during the entire overland flow process can be expressed as:

[0097]

[0098] where t r is the time to reach ponding.

[0099] 4.2 Urban Rainfall-Runoff Concentration Calculation

[0100] Urban rainfall-runoff simulation models are generally classified into models mainly based on hydrological methods, models mainly based on hydrodynamic methods, and models mainly based on terrain analysis techniques according to the concentration calculation method. The runoff generation calculation in urban rainfall-runoff models generally does not distinguish by model type, and all runoff generation calculation methods can be used for various rainfall-runoff models. For impervious surfaces, the rainfall amount minus losses such as evaporation, interception, and depression storage is used for calculation. For pervious surfaces, the methods used by each model are different, mainly including the runoff coefficient method, SCS method, infiltration curve method, Ф-index method, and conceptual rainfall-runoff method, etc.

[0101] The calculation of urban rainfall-runoff generation and concentration is the basis for establishing urban rainfall-runoff models. Aiming at the characteristics of urban rainfall-runoff generation and concentration, scholars often summarize its calculation process as urban rainfall-runoff generation calculation, urban rainfall-runoff surface concentration calculation, and urban rainfall-runoff pipe network concentration calculation. Due to the complex diversity of the urban underlying surface, the urban underlying surface is usually simplified into pervious surfaces and impervious surfaces in the urban rainfall-runoff generation calculation. Currently, the common runoff generation calculation methods can be divided into statistical analysis methods, infiltration curve methods, and model methods. Among them, the SCS method in statistical analysis methods, the Green-Ampt infiltration curve and Horton infiltration curve in infiltration curve methods are widely used.

[0102] Urban rainfall-runoff surface concentration is the process in which the net rainfall in each drainage sub-basin converges to the outlet control section or directly discharges into the river. The calculation of urban surface concentration uses hydrological methods and hydrodynamic methods. The hydrological method is based on the idea of the system, establishing the relationship between input and output to simulate surface concentration, and using a nonlinear reservoir for simulation. The hydrodynamic method is based on microscopic physical laws. By solving the Saint-Venant equations or their simplified forms, a more detailed surface concentration process can be obtained. When the measured data cannot meet the solution requirements of the hydrodynamic method, the hydrological-hydrodynamic method can be used as an alternative.

[0103] Compared with runoff calculation and surface runoff calculation, the confluence calculation of urban rainwater pipe networks is relatively mature. Common methods include simple hydrological methods and complex hydrodynamic methods. Hydrological methods include the instantaneous unit hydrograph method and the Muskingum method. Among them, the Muskingum method is relatively simple in calculation, has fewer parameters, requires less data, has higher calculation accuracy, and is widely used. The hydrodynamic method is based on the Saint-Venant equation and adopts its simplified forms, including kinematic wave, diffusion wave, and dynamic wave. Among them, the kinematic wave calculation is relatively simple, but this method is only applicable to the situation where the downstream backwater effect is small and the pipe slope is large; the calculation accuracy of the diffusion wave is similar to that of the kinematic wave, but it is not applicable to the water flow calculation of looped pipe networks with various flow states coexisting; the dynamic wave calculation has higher accuracy, and peak attenuation and backwater effects are considered during the calculation, which is applicable to various inflow conditions and pipe slopes. However, this method requires high data requirements and the calculation is also more complex. Existing research results show that when the data conditions are good and the accuracy requirements are high, the diffusion wave or dynamic wave can be selected for simulation calculation according to the pipe shape and water flow state. In other cases, the Muskingum method is considered a better choice.

[0104] 4.3 Construction of Urban Rainstorm Disaster Simulation Model

[0105] The rainstorm disaster simulation technology is the basis for rainstorm disaster early warning. The urban rainstorm model is an important means for rainstorm disaster simulation, and data such as rainfall, surface water accumulation, and urban pipe networks are the key input variables of the urban rainstorm model. The rainstorm disaster simulation technology and the rainstorm disaster early warning technology are closely related. The comprehensive sorting of the rainstorm disaster simulation and early warning technology is as Figure 1 shown. In terms of data collection, real-time rainfall data and predicted rainfall data can be obtained by comprehensively analyzing multi-source rainfall data. By combining the rainfall data and the real-time water accumulation data measured by the water accumulation observation equipment, the predicted water accumulation data can be calculated. Inputting the rainfall data, surface water accumulation data, etc. into the urban rainstorm model can perform relevant simulations. According to the model simulation output data, the spatio-temporal distribution characteristics of rainfall and water accumulation can be obtained, the urban river channel pipe network can be evaluated to guide the emergency rescue work, cooperate with the management level for intelligent scheduling, and formulate a visual emergency plan. The model output data can also be used for the assessment of waterlogging risks and the classification of levels. The assessment results are sent to the public through platforms such as radio, television, and mobile phones, which can minimize casualties and property losses to the greatest extent. The management level can not only issue early warning information and carry out emergency rescue work based on this system, but also carry out post-disaster reconstruction, risk area governance, sponge city construction and evaluation according to the analysis results, and control disasters from the source.

[0106] The principles of each model are summarized as follows:

[0107] (1) SWMM Model

[0108] SWMM (Storm Water Management Model) can be used to conduct urban sub - rainfall or long - term continuous water quantity and water quality simulations. Among them, the runoff module mainly deals with rainfall - runoff generation in each sub - catchment area, and the confluence module conducts water quantity transmission through pipe networks, channels, water storage and treatment facilities, regulating gates, etc.

[0109] The simulation steps of the SWMM model are as follows: sub - basin generalization; surface runoff and confluence calculation; pipe network confluence calculation. The nonlinear reservoir model is adopted in SWMM to describe the surface confluence process. Each sub - catchment area is generalized as a non - linear reservoir with a very shallow water depth. Rainfall is the input, and soil infiltration and surface runoff are the outflows, which are jointly solved by the continuity equation and the Manning formula. The data required for SWMM simulation include catchment area data (land use type, pipe network data, digital elevation data), meteorological data, and hydrological observation data for calibrating the model, etc.

[0110] (2) SWAT model

[0111] The SWAT (Soil and Water Assessment Tool) model is used to simulate surface water and groundwater quality and quantity, and predict the impacts of land management measures on hydrology, sediment, and agricultural chemical yields in large - scale complex basins with different soil types, land use patterns, and management conditions. The SWAT model uses a daily time step and can conduct long - term continuous simulations. Its calculation results include the water yield, sediment yield, and nitrogen and phosphorus contents in each sub - basin during each simulation period, as well as the water yield, sediment yield, and nitrogen and phosphorus contents at the outlets of each main river channel during each simulation period.

[0112] (3) DHI MIKE11 model

[0113] This model can be used to simulate river flow and water level, and can also simulate the diffusion and attenuation of pollutants in the river channel. It is mainly used for hydrology, hydraulics, water quality, and sediment transport simulations of estuaries, rivers, irrigation systems, and other inland waters, and can be widely applied in flood control and flood forecasting, water resources quantity and quality management, and water conservancy project planning and design demonstration.

[0114] MIKE11 includes the following basic modules: Hydrodynamic module (HD), Advection - Diffusion module (AD), Rainfall - Runoff module (RR), Water Quality module (WQ), Sediment Transport module (ST), etc. In the comprehensive water quality model, MIKE11 selects the Hydrodynamic module (HD) to simulate river channel flow and water level, and on this basis, selects the Advection - Diffusion module (AD) to simulate the advection and diffusion process of pollutants in the water body.

[0115] 5. Evaluation and prediction learning methods based on machine learning

[0116] 5.1 Data processing

[0117] Apply NLP technology to extract information from text and construct a database, extract entity relationships through deep learning, construct a knowledge graph, train and optimize the NLP model with a labeled dataset. Design a database, import data, store and verify it for practical applications, and continuously optimize the NLP model and database based on feedback. This series of steps can efficiently extract text information and construct a practical database. Subsequently, input this information into the determined model. For example, attached Figure 2 shows the deep learning input, output, and prediction processes of a machine learning-based urban stormwater resilience assessment method. This assessment prediction framework model diagram details the entire process from data collection to model training and then to the final result output. Collect urban stormwater-related data through various data sources, including historical rainfall, urban drainage system design parameters, topographic and geomorphic information, etc. In the model training stage, use machine learning algorithms to analyze the data and identify the key factors affecting urban stormwater resilience. Then, preprocess these data through repeated iteration and optimization to ensure their quality and consistency and prepare for model training.

[0118] 5.2 Model evaluation and determination

[0119] Use the stormwater resilience index as the input variable of the model, and adopt four machine learning methods: random forest, support vector machine, backpropagation neural network, and extreme gradient boosting to construct an estimation model of urban stormwater resilience. All algorithms are implemented through Matlab R 2016b. Use the coefficient of determination (R 2 ) and root mean square error (RMSE) to comprehensively evaluate the model accuracy. The closer R 2 is to 1 and the smaller the RMSE, the higher the model accuracy and the better the modeling effect. The calculation formulas are as follows:

[0120]

[0121] where y i , represent the estimated value, measured value, and average value of the urban stormwater resilience index respectively; n is the number of samples.

[0122] 5.2.1 LSTM model

[0123] Long short-term memory network (LSTM) is a specially designed time-recurrent neural network model, whose main purpose is to overcome the limitations of traditional recurrent neural networks (RNN) in dealing with long-term dependency problems. The basic structure of all RNNs is composed of a series of identical neural network modules connected in series to form a chain-like structure. Specifically, as attached Figure 3As shown, LSTM is an improvement of RNN, which introduces three gating mechanisms: the forget gate, the input gate, and the output gate. Among them, the forget gate is responsible for determining which information in the cell state memory should be discarded; the input gate determines which new information should be written into the cell state memory; and the output gate controls how the information stored in the cell state memory is passed to the next step in the time series. The cell memory unit in the LSTM network endows it with good memory ability and is widely used in fields such as time series prediction. The LSTM network model can be trained for parameters by using the time backpropagation algorithm to precisely adjust and optimize the performance of the entire LSTM network.

[0124] Its specific calculation formula is as follows:

[0125] f t = σ(W f [h t-1 , x t ) (20)

[0126] i t = σ(W i [h t-1 , x t + b i ) (21)

[0127]

[0128] O t = σ(W o [h t-1 , x t + b o ) (24)

[0129]

[0130] 5.2.2 CNN-LSTM Model

[0131] CNN can achieve good performance in predicting urban rain floods. In addition, CNN requires a large training dataset and computing power and often searches for patterns from high-dimensional data. The CNN-LSTM model is a coupled model that effectively reduces the number of training parameters (as shown in Appendix Figure 4 ). It can extract high-dimensional features and time series features from the data, and use activation functions to establish a non-linear mapping relationship between the optimal features and the rain flood resilience index, with strong generalization ability, effectively improving the accuracy and stability of urban rain flood resilience prediction.

[0132] 5.2.3 CNN and LSTM Hybrid Model

[0133] A hybrid network model is constructed through CNN and LSTM technologies to perform in-depth optimization processing on the input data, effectively removing abnormal components such as noise and irregular jumps contained therein. The CNN algorithm is used to mine the relationships between the data in the parameters, remove the noise, make the sequence features of each parameter more obvious, so as to obtain the data sequence features of the parameters related to the unit operation. Then, the LSTM network is used to mine the temporal features of the reconstructed data, and a regularization method is added after the LSTM network to reduce the overfitting phenomenon. The output data of the LSTM network is completed with prediction after inverse normalization processing.

[0134] 5.2.4 Model Verification and Evaluation

[0135] Based on historical detection data, the present invention verifies the performance of the CNN-LSTM model in the inversion of urban stormwater resilience assessment. Through correlation statistical analysis with the measured stormwater resilience indicators, statistical regression models are respectively constructed to achieve urban stormwater resilience assessment. An urban stormwater resilience inversion model is constructed using the measured data, and the flowchart of urban stormwater resilience prediction based on machine learning is as shown in the attached figure. In terms of the modeling and verification accuracy, the R of the CNN-LSTM model 2 is 0.90, and the RMSE is 1.75 mg / L. CNN-LSTM is a high-performance spatio-temporal deep learning model with powerful feature extraction and model expression capabilities. The results show that the model has a good effect on the prediction of urban stormwater. Establishing an applicable and highly accurate assessment inversion model in a complex urban stormwater system is still a challenge. By using the CNN-LSTM model, multi-scale features can be learned and extracted in the spatio-temporal dimension, and the non-linear relationships in the spatio-temporal data can be effectively captured. This spatio-temporal modeling method plays a key role in achieving accurate prediction, verifying the good performance of the CNN-LSTM model in terms of inversion accuracy and model stability. Therefore, the model can effectively utilize the time series information, extract meaningful features from the spatio-temporal data, and make accurate predictions. This is of great significance for fields such as stormwater disasters and urban resilience, which can better understand the urban stormwater situation, take corresponding measures in a timely manner, and minimize the urban losses caused by stormwater.

[0136] The above technical solutions only reflect the preferred technical solutions of the technical solutions of the present invention. Some changes that may be made to some parts by those skilled in the art of this technology field all reflect the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A deep learning assessment method for urban flood resilience based on machine learning, characterized in that: The following steps are involved: Step 1: Analyze the temporal and spatial characteristics of urban rain and flood; Step 2: Analyze and calculate urban stormwater resilience indicators; Step 3: Conduct urban stormwater resilience assessment; Step 4: Construct an urban rain and flood disaster simulation model; Step 5: Evaluation and prediction learning method based on machine learning.

2. According to the deep learning assessment method of urban flood resilience based on machine learning in claim 1, it is characterized in that: The step 1 comprises the following working steps: Step 1): Investigate the current status of urban stormwater resilience sources; Step 2): Determine the main rainy and flooding periods.

3. The urban flood resilience deep learning assessment method based on machine learning according to claim 2 is characterized in that: The SCS runoff curve number method was used to estimate the relationship between rainfall and runoff during the main rainy and flood periods, and the SWMM model was used to simulate the urban rainfall-runoff process.

4. The urban flood resilience deep learning assessment method based on machine learning according to claim 1 is characterized in that: The step 2 includes the following working steps: Step 1): Calculate the maximum potential soil water holding capacity and infiltration rate. Soil water storage and drainage capacity are the main factors that determine whether the land can absorb precipitation and maintain its original state; Step 2): Calculate the precipitation runoff, total runoff and outflow rate. If the rainfall exceeds the infiltration rate of the soil, surface runoff will occur. Step 3): Calculate city assets and execution speed.

5. The urban flood resilience deep learning assessment method based on machine learning according to claim 1 is characterized in that: The step 3 includes the following working steps: Step 1): Use ArcGIS spatial analysis tools to couple flood information with social engineering information through the ArcGIS platform to analyze urban stormwater resilience; Step 2): By evaluating land use changes and their impact on parameters such as maximum potential water holding capacity and infiltration rate, the urban stormwater resilience is calculated using the Pearson correlation coefficient, thereby reflecting the system's performance, recovery time and recovery efficiency, and evaluating the spatiotemporal evolution of urban stormwater resilience.

6. The urban flood resilience deep learning assessment method based on machine learning according to claim 1 is characterized in that: The step 4 includes the following working steps: Step 1): Collect urban rainfall data, surface water accumulation data and underground pipe network monitoring data; Step 2): Use the coupling of SWMM and InfoworksICM-2D model to establish an urban rainwater flood disaster model.

7. The urban flood resilience deep learning assessment method based on machine learning according to claim 2 is characterized in that: The step 1) includes rainfall period, surface water monitoring equipment and underground pipe network detection equipment.

8. The urban flood resilience deep learning assessment method based on machine learning according to claim 7 is characterized in that: The step 5 comprises the following working steps: Step 1): Use the NLP text processing method in the machine learning algorithm to extract data and establish a database; Step 2): Input the extracted data into the improved neural network model based on deep learning; Step 3): Introduce the MIV algorithm to filter and purify the extracted text data; Step 4): Use a neural network model based on deep learning to predict and evaluate urban stormwater resilience.

9. The deep learning assessment method for urban flood resilience based on machine learning according to claim 5 is characterized in that: In the step three, the urban stormwater resilience prediction and assessment uses the random forest-entropy weight method to determine the weights, and then proposes a feasible construction strategy based on the prediction and assessment of the neural network model of deep learning under machine learning.

10. The urban stormwater resilience deep learning assessment method based on machine learning according to claim 9 is characterized in that: Set urban stormwater targets and spatiotemporal differentiation characteristics, verify and estimate urban stormwater runoff processes, underlying surface resilience and stormwater disaster simulation, and finally use machine learning methods to predict and evaluate urban stormwater resilience.

Images