Urban and rural rainstorm waterlogging early warning system and method fusing heat island and aerosol effects

Abstract

The present application relates to a kind of urban and rural rainstorm waterlogging early warning system and method of fusing heat island and aerosol effect, belong to urban rainstorm waterlogging early warning technical field.System includes multi-source data acquisition, dynamic environmental factor analysis, rainfall intelligent estimation, waterlogging intelligent prediction, model training optimization and early warning generation release module;Method is by real-time generation heat island intensity and aerosol optical thickness dynamic grid map, input convolution long short-term memory network of encoder-decoder architecture, short-term precipitation correction and probability forecast are carried out in combination with physical constraint, then by the bidirectional long short-term memory network of fusion attention mechanism Dynamic prediction of water depth, and further deduce traffic and power grid disaster damage situation.The present application realizes the rainfall-watering-disaster integration intelligent early warning driven by urban heat island and aerosol effect, significantly improves the local rainstorm forecast precision, prolongs early warning see period, enhances the waterlogging dynamic adaptation and multi-department linkage defense capability.

Description

Technical Field

[0001] This invention belongs to the field of urban rainstorm and waterlogging early warning technology, specifically relating to an urban and rural rainstorm and waterlogging early warning system and method that integrates the heat island and aerosol effects. Background Technology

[0002] With the acceleration of urbanization, the urban heat island effect leads to enhanced local thermal circulation. Atmospheric aerosol emissions significantly affect precipitation efficiency and distribution area through cloud microphysics processes. The coupling effect of these two factors results in frequent short-term heavy rainfall and sudden urban flooding disasters in urban and rural areas. However, existing rainstorm and flooding early warning systems mostly connect meteorological forecasts and hydrodynamic models offline. During the rainfall forecasting stage, they fail to assimilate highly dynamic urban environmental factors such as heat island intensity and aerosol optical thickness in real time, resulting in insufficient spatiotemporal accuracy and short lead time for forecasting local rainstorms in urban clusters. At the same time, flooding simulations often use static parameters such as fixed surface runoff coefficients and pipeline design flow capacity, which cannot reflect the key feedback of dynamic changes such as real-time evolution of soil saturation and pipeline siltation and blockage on the water accumulation process, leading to a large deviation between predicted water depth and actual conditions. In addition, traditional early warning outputs only stop at the level of water accumulation distribution maps, lacking quantitative extrapolation of water accumulation information into road congestion status, power grid failure probability, and other lifeline system disaster damage assessment links, making it difficult to support proactive and precise defense through multi-departmental collaboration. Therefore, there is an urgent need for an intelligent early warning solution that can integrate the heat island and aerosol effects into the entire chain of rainfall forecasting, waterlogging prediction, and disaster loss perception. Summary of the Invention

[0003] To address the aforementioned problems in the existing technology, this invention provides an urban and rural rainstorm flood early warning system and method that integrates the urban heat island and aerosol effects.

[0004] The objective of this invention can be achieved through the following technical solutions: In a first aspect, the present invention provides an urban and rural rainstorm flood early warning system that integrates the heat island and aerosol effects, comprising: The multi-source data acquisition module is used to collect and integrate real-time meteorological observation data, numerical weather prediction model output products, urban environmental dynamic monitoring data, underlying surface and infrastructure data, and water accumulation and disaster damage data. The urban environmental dynamic monitoring data is fused with near-real-time satellite remote sensing inversion products, ground monitoring station network data, and IoT sensor data. The multi-source data acquisition module is also used to standardize, interpolate, and assimilate the collected data with different spatiotemporal resolutions and different lag times to generate a time-consistent dynamic environmental factor field. The dynamic environmental factor analysis module is connected to the multi-source data acquisition module and is used to analyze and process the collected urban environmental dynamic monitoring data to generate a dynamic index grid map characterizing the heat island effect and aerosol effect. The intelligent rainfall prediction module is connected to the dynamic environmental factor analysis module. It is used to perform probability prediction or intensity correction of local short-term heavy rainfall based on meteorological forecast data and the dynamic index grid map, using a mechanism-data dual-driven deep learning model, and output rainfall prediction results. The intelligent prediction module for waterlogging is connected to the dynamic environmental factor analysis module and the intelligent rainfall prediction module, respectively. It is used to quickly predict the spatiotemporal evolution of surface water depth using a deep learning model based on the rainfall prediction results, at least a part of the dynamic index grid map, and static underlying surface data, and output the waterlogging prediction results. The model training and optimization module is used to train the deep learning models in the intelligent rainfall prediction module and the intelligent waterlogging prediction module using a historical database, and to optimize the hyperparameters using a Bayesian optimization algorithm. The early warning information generation and release module is connected to the intelligent prediction module for waterlogging and is used to generate multi-level early warning information based on the real-time rolling waterlogging prediction results and release it.

[0005] In a further embodiment of the present invention, the dynamic indicators of the heat island effect generated by the dynamic environmental factor analysis module include a spatial distribution map of heat island intensity, which is calculated based on the difference between the surface temperature and the near-surface air temperature obtained by satellite remote sensing, and is assimilated and corrected by combining anthropogenic heat emission intensity data; the generated dynamic indicators of the aerosol effect include a spatial distribution map of aerosol optical thickness, which integrates satellite remote sensing data and particulate matter concentration data from ground monitoring stations.

[0006] In a further embodiment of the present invention, the deep learning model built into the rainfall intelligent prediction module is a convolutional long short-term memory network with an encoder-decoder architecture. The encoder is used to extract the spatiotemporal coupling features of the rainfall field and the dynamic environmental factor field, and the decoder is used to output the rainfall prediction amount or probability of occurrence for a future preset period. During the training process of the intelligent rainfall prediction module, the loss function includes a data-driven mean squared error term and a physical constraint term, which is constructed based on the heat wave-convection triggering index and the aerosol-cloud microphysics index.

[0007] In a further embodiment of the present invention, the deep learning model built into the intelligent prediction module for urban flooding is a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks. The input feature vector of the intelligent prediction module for urban flooding explicitly includes a dynamic environmental factor feature map, which includes soil saturation, surface runoff coefficient, and pipeline flow capacity correction coefficient. The pipeline flow capacity correction coefficient is associated with a pipeline siltation level database.

[0008] In a further embodiment of the present invention, the system further includes: The multi-scale disaster damage situation awareness module is connected to the intelligent prediction module for urban flooding and waterlogging, as well as the external traffic management platform and power management platform. The multi-scale disaster situation awareness module incorporates an agent-based road traffic exposure prediction submodule and a multi-source data fusion-based power distribution network disaster situation awareness submodule. It is used to convert water accumulation prediction results into traffic congestion index prediction and power distribution network node failure probability prediction, and push customized early warning and handling suggestions to transportation and power management departments through the early warning information generation and release module.

[0009] In a second aspect embodiment, this system provides a method for early warning of urban and rural rainstorm flooding that integrates the urban heat island and aerosol effects, applied to the urban and rural rainstorm flooding early warning system integrating the urban heat island and aerosol effects in the first aspect embodiment of the present invention, comprising: Collect multi-source dynamic and static data within the study area to construct a standardized spatiotemporal database; the data includes at least meteorological forecast data, underlying surface and infrastructure data, water accumulation and disaster damage data, and urban environmental dynamic monitoring data; the urban environmental dynamic monitoring data includes dynamic indicators characterizing the urban heat island effect and dynamic indicators characterizing the atmospheric aerosol effect. A mechanism- and data-driven intelligent early warning model for rainstorms and urban flooding is constructed. The model includes a rainfall intelligent prediction submodule based on dynamic environmental factors, which uses real-time monitored urban environmental dynamic monitoring data as dynamic input and combines it with meteorological forecast data to make probability forecasts or intensity corrections for local short-term heavy rainfall. The intelligent prediction submodule for urban waterlogging considering environmental feedback is used to combine the rainfall prediction results of the intelligent rainfall prediction submodule and at least a portion of the dynamic indicators in the urban environmental dynamic monitoring data with static underlying surface data to quickly predict the spatiotemporal evolution of surface water depth. The early warning model is trained and its hyperparameters are optimized using the constructed spatiotemporal database. Based on real-time collected multi-source data, the trained early warning model is run continuously to dynamically update the rainfall and water accumulation prediction results, and multi-level early warning information is generated and released based on the prediction results.

[0010] In a further embodiment of the present invention, the dynamic indicators characterizing the urban heat island effect include at least one of the following: satellite remote sensing surface temperature, near-surface air temperature, anthropogenic heat emission intensity, impervious surface area, building density and average height, and population activity density; the anthropogenic heat emission intensity is estimated in real time based on industrial energy consumption, vehicle traffic flow, and building energy consumption data. The dynamic indicators characterizing the atmospheric aerosol effect include at least one of the following: satellite remote sensing aerosol optical thickness, particulate matter concentration at ground monitoring stations, black carbon mass concentration, and volatile organic compound concentration.

[0011] In a further embodiment of the present invention, the rainfall intelligent prediction submodule is constructed using a convolutional long short-term memory network or graph neural network with an encoder-decoder architecture; The input data for the intelligent rainfall prediction submodule includes time-series radar echo maps, numerically predicted rainfall fields, and synchronously gridded heat island intensity maps and aerosol optical thickness maps. The loss function of the intelligent rainfall prediction submodule incorporates physical constraint terms, which include heat wave-convection triggering index constraints and aerosol-cloud microphysical inhibition / enhancement index constraints.

[0012] In a further embodiment of the present invention, the intelligent prediction submodule for urban flooding adopts a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks; The input features of the intelligent prediction submodule for urban flooding include: the rainfall time series, static underlying surface features, and dynamic environmental factor feature map output by the intelligent rainfall prediction submodule; the dynamic environmental factor feature map includes real-time updated soil saturation, surface runoff coefficient, and pipeline flow capacity correction coefficient. The surface runoff coefficient is dynamically calculated based on the proportion of impermeable surface and rainfall intensity; the pipeline flow capacity correction coefficient is dynamically determined based on the pipeline siltation level.

[0013] In a further embodiment of the present invention, the early warning model further includes: The multi-scale disaster damage situation awareness submodule is used to input the spatiotemporal distribution data of surface water depth predicted by the intelligent waterlogging prediction submodule into the agent-based road traffic exposure prediction model and the power distribution network disaster damage situation awareness model based on multi-source data fusion, respectively, and output the traffic exposure prediction results of the road network and the damage probability prediction results of the power distribution network nodes.

[0014] The present invention has at least the following beneficial effects: 1. This invention improves the forecast accuracy and spatial precision of localized short-term heavy rainfall in cities by explicitly inputting dynamic environmental factors such as urban heat island intensity and aerosol optical thickness into the rainfall prediction model and applying physical constraint regularization, thereby reducing the underreporting and false alarms of extreme precipitation.

[0015] 2. The urban flooding prediction model dynamically couples the real-time soil saturation, the runoff coefficient driven by precipitation intensity, and the pipeline overflow correction coefficient associated with pipeline siltation level, so that the flooding simulation can truly respond to the actual working conditions of the drainage system and avoid prediction distortion caused by static assumptions.

[0016] 3. Through the multi-scale disaster situation awareness module, the prediction of water depth is automatically converted into an assessment of road traffic congestion exposure and a prediction of the failure probability of power distribution network nodes. Customized disposal suggestions are pushed to the transportation and power departments, realizing intelligent linkage early warning of cross-domain disasters. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of a method provided in one embodiment of the present invention; Figure 2 This is a dynamic indicator acquisition and processing flow provided in one embodiment of the present invention; Figure 3 The following is the operation flow of the intelligent rainfall prediction module provided in one embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects of the present invention in achieving the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0020] Please refer to the details. Figure 1-3 In a first aspect, the present invention provides an urban and rural rainstorm flooding early warning system that integrates the urban heat island and aerosol effects, comprising: The multi-source data acquisition module is used to collect and integrate real-time meteorological observation data, numerical weather prediction model output products, urban environmental dynamic monitoring data, underlying surface and infrastructure data, and water accumulation and disaster damage data. The urban environmental dynamic monitoring data is fused with near-real-time satellite remote sensing inversion products, ground monitoring station network data, and IoT sensor data. The multi-source data acquisition module is also used to standardize, interpolate, and assimilate the collected data with different spatiotemporal resolutions and different lag times to generate a time-consistent dynamic environmental factor field. The dynamic environmental factor analysis module is connected to the multi-source data acquisition module and is used to analyze and process the collected urban environmental dynamic monitoring data to generate a dynamic index grid map characterizing the heat island effect and aerosol effect. The intelligent rainfall prediction module is connected to the dynamic environmental factor analysis module. It is used to perform probability prediction or intensity correction of local short-term heavy rainfall based on meteorological forecast data and the dynamic index grid map, using a mechanism-data dual-driven deep learning model, and output rainfall prediction results. The intelligent prediction module for waterlogging is connected to the dynamic environmental factor analysis module and the intelligent rainfall prediction module, respectively. It is used to quickly predict the spatiotemporal evolution of surface water depth using a deep learning model based on the rainfall prediction results, at least a part of the dynamic index grid map, and static underlying surface data, and output the waterlogging prediction results. The model training and optimization module is used to train the deep learning models in the intelligent rainfall prediction module and the intelligent waterlogging prediction module using a historical database, and to optimize the hyperparameters using a Bayesian optimization algorithm. The early warning information generation and release module is connected to the intelligent prediction module for waterlogging and is used to generate multi-level early warning information based on the real-time rolling waterlogging prediction results and release it.

[0021] Traditional rainstorm and urban flooding early warning systems largely rely on a one-way coupling of numerical weather prediction and hydrodynamic models. Their rainfall forecasting often neglects the feedback from local urban thermal circulation and aerosol concentration changes on the precipitation process. The urban heat island effect enhances local convergence and upward motion, altering the precipitation location and intensity; high-concentration aerosols, acting as cloud condensation nuclei or ice nuclei, influence cloud microphysical processes, thereby affecting precipitation efficiency and total amount. However, existing systems fail to incorporate these highly dynamic urban environmental factors into rainfall forecasting and runoff calculations in real time, resulting in insufficient accuracy and timeliness in forecasting short-duration heavy rainfall and resulting urban flooding. Therefore, there is an urgent need for an early warning system that can deeply integrate the heat island and aerosol effects and achieve integrated intelligent prediction of rainfall and waterlogging.

[0022] In this embodiment, the early warning system consists of a multi-source data acquisition module, a dynamic environmental factor analysis module, a rainfall intelligent prediction module, an urban flooding intelligent prediction module, a model training and optimization module, and an early warning information generation and dissemination module. The multi-source data acquisition module is responsible for collecting and integrating multi-source heterogeneous data. Meteorological data sources include real-time meteorological observation data (such as radar reflectivity mosaics and observations from ground automatic weather stations) and numerical weather prediction model output products; urban environmental dynamic monitoring data integrates near-real-time satellite remote sensing inversion products, such as geostationary meteorological satellite surface temperature products, ground-based national control station particulate matter concentration data, and real-time traffic flow and energy consumption estimation data; underlying surface and infrastructure data, such as land use types, pipeline geographic information systems, and siltation detection records; and water accumulation and disaster damage data, such as historical water depth at flooding points and video surveillance water accumulation range. After receiving the above data, the dynamic environmental factor analysis module transforms discrete monitoring values ​​into a unified gridded dynamic index field through spatial interpolation, physical inversion, and data assimilation algorithms, specifically including a spatial distribution map of urban heat island intensity and a spatial distribution map of aerosol optical thickness. The intelligent rainfall prediction module receives dynamic index gridded maps and meteorological forecast gridded data, runs a pre-trained encoder-decoder architecture convolutional long short-term memory network, and outputs rainfall forecasts for each time step within a preset future period, as well as the distribution of rainfall exceeding probability thresholds. The intelligent urban flooding prediction module inputs rainfall prediction results, some dynamic index feature maps (such as updated soil saturation and runoff coefficients), and static underlying surface features into a bidirectional long short-term memory neural network to predict the spatiotemporal changes in surface water depth grid by grid. The model training and optimization module uses a spatiotemporal database composed of massive historical rainstorm cases to train the two deep learning models offline and automatically adjusts key hyperparameters such as the number of network layers and learning rate using a Bayesian optimization algorithm. The early warning information generation and dissemination module monitors the water depth prediction results in real time. Once the predicted value exceeds the set grading threshold, it automatically generates an early warning message containing the impact range, duration, and recommended measures, and disseminates it to relevant management departments and the public via dedicated lines, SMS, and application interfaces.

[0023] By constructing a closed-loop data flow that integrates the urban heat island effect and aerosol effects in real time, the barrier between meteorological drivers and urban environmental factors in traditional forecasting systems has been broken down. The intelligent rainfall prediction module, by incorporating heat island-triggered convection enhancement signals and cloud microphysical modulation information from aerosols, significantly improves the forecast accuracy and spatiotemporal precision of short-duration heavy rainfall in urban agglomerations. The intelligent urban flooding prediction module directly inherits the detailed results of rainfall prediction and dynamic environmental factor feedback, achieving seamless and rapid extrapolation from rainfall to flooding, significantly extending the effective warning time. The entire system forms a fully automated chain of data acquisition, environmental perception, intelligent prediction, optimization training, and warning issuance, greatly reducing the risk of human delays and missed reports in urban flooding warnings.

[0024] In a further embodiment of the present invention, the dynamic indicators of the heat island effect generated by the dynamic environmental factor analysis module include a spatial distribution map of heat island intensity, which is calculated based on the difference between the surface temperature and the near-surface air temperature obtained by satellite remote sensing, and is assimilated and corrected by combining anthropogenic heat emission intensity data; the generated dynamic indicators of the aerosol effect include a spatial distribution map of aerosol optical thickness, which integrates satellite remote sensing data and particulate matter concentration data from ground monitoring stations.

[0025] In this embodiment, the dynamic environmental factor analysis module generates the spatial distribution map of heat island intensity using a process based on the difference between satellite remote sensing surface temperature and near-surface air temperature, combined with anthropogenic heat emission intensity data for assimilation and correction. First, surface temperature is acquired from real-time and near-real-time products from geostationary meteorological satellites, while simultaneously accessing standard Stevenson screen air temperature observations from the urban meteorological observation network. Spatial statistical interpolation is used to grid the air temperature to a resolution matching the satellite data. The difference between surface temperature and air temperature is calculated as the initial heat island intensity. Next, the module accesses anthropogenic heat emission intensity data, based on real-time monitoring of electricity load, industrial energy consumption, urban road network vehicle traffic flow, and building energy consumption sub-metering data. A bottom-up spatial distribution model is used to generate an hourly anthropogenic heat flux gridded field. This anthropogenic heat flux gridded field is then processed with the satellite-air temperature difference field using three-dimensional variational assimilation or similar data fusion techniques to correct local intensity distortions caused by advection effects and satellite inversion biases. Finally, a high-precision spatial distribution map of heat island intensity, incorporating anthropogenic heat contributions, is output. For the spatial distribution map of aerosol optical thickness, the module first acquires the aerosol optical thickness product retrieved by the Deep Blue or Dark Target algorithms from satellites. Then, it establishes a real-time conversion relationship between this product and the particulate matter mass concentration from ground environmental monitoring stations using geographic weighted regression or other statistical model methods. In areas with dense ground stations, the ground-corrected values ​​are used; in areas without ground stations, the gradient information retrieved from satellites is retained. Finally, an aerosol optical thickness distribution map matching the heat island grid is generated through interpolation.

[0026] In a further embodiment of the present invention, the deep learning model built into the rainfall intelligent prediction module is a convolutional long short-term memory network with an encoder-decoder architecture. The encoder is used to extract the spatiotemporal coupling features of the rainfall field and the dynamic environmental factor field, and the decoder is used to output the rainfall prediction amount or probability of occurrence for a future preset period. During the training process of the intelligent rainfall prediction module, the loss function includes a data-driven mean squared error term and a physical constraint term, which is constructed based on the heat wave-convection triggering index and the aerosol-cloud microphysics index.

[0027] In this embodiment, the deep learning model built into the intelligent rainfall prediction module specifically uses a convolutional long short-term memory network with an encoder-decoder architecture. Its encoder consists of several layers of ConvLSTM units. The input time-series data includes a gridded radar reflectivity map of past periods, a background field of future rainfall output from a regional numerical weather prediction model, and synchronously gridded heat island intensity and aerosol optical thickness maps. In each layer of the encoder, the rainfall field and two dynamic environmental factor fields are concatenated along the channel dimension through convolution operations to extract spatiotemporal coupling features. The encoder ultimately outputs a compact spatiotemporal hidden state tensor. The decoder consists of ConvLSTMs of the corresponding number of layers, using the encoder's hidden state for recursive prediction, outputting the future rainfall forecast gridded field step-by-step, and simultaneously outputting a branch representing the probability of precipitation. During training, the loss function consists of two parts: the first part is the mean squared error term, which measures the voxel-level difference between the predicted rainfall and the radar-estimated rainfall; the second part is the physical constraint term, specifically including two exponential constraints. The heat wave-convection triggering index constraint aims to penalize grid points where the model fails to predict corresponding upward motion or increasing precipitation trends even when the heat island intensity is extremely high. The aerosol-cloud microphysics index constraint, based on aerosol optical thickness and water vapor conditions, constrains the model to reasonably reflect the suppression or enhancement trend of precipitation efficiency in areas with high aerosol values ​​and sufficient water vapor. The model training is completed by jointly optimizing these two loss terms through backpropagation.

[0028] In a further embodiment of the present invention, the deep learning model built into the intelligent prediction module for urban flooding is a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks. The input feature vector of the intelligent prediction module for urban flooding explicitly includes a dynamic environmental factor feature map, which includes soil saturation, surface runoff coefficient, and pipeline flow capacity correction coefficient. The pipeline flow capacity correction coefficient is associated with a pipeline siltation level database.

[0029] In this embodiment, the built-in model of the intelligent prediction module for urban flooding adopts a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks. This model contains two parallel input branches. The first branch is the rainfall time series input, which extracts the time-step rainfall forecast sequence for future periods output by the intelligent rainfall prediction module through a one-dimensional convolutional layer. The second branch is the static and dynamic feature input, specifically including: static underlying surface features such as the proportion of impervious surface, topographic slope, and catchment area division extracted from the land use map; and explicitly included dynamic environmental factor feature maps. The dynamic environmental factor feature maps specifically contain several channels: the first channel is soil saturation, calculated in real-time from the previous accumulated rainfall using a water balance model; the second channel is the surface runoff coefficient, which is not a fixed value but dynamically calculated at the current time based on the proportion of impervious surface and rainfall intensity of each grid using a modified empirical formula for the runoff coefficient; the runoff coefficient is significantly different at the beginning of a rainstorm and after infiltration saturation; the third channel is the pipeline flow capacity correction coefficient, which is taken from an external pipeline siltation level database. This database assigns different siltation levels to each section of the pipeline network through regular closed-circuit television inspections and daily maintenance records. In the model, the corresponding manhole grid dynamically obtains an overflow correction coefficient within an effective range based on its siltation level. A bidirectional long short-term memory network receives the temporal features of the first branch and, through a fusion-associative perceptual attention mechanism, dynamically focuses on the static and dynamic features of different spatial locations in the second branch at each time step. Finally, a gridded predicted water depth sequence is generated by the convolutional output layer.

[0030] In a further embodiment of the present invention, the system further includes: The multi-scale disaster damage situation awareness module is connected to the intelligent prediction module for urban flooding and waterlogging, as well as the external traffic management platform and power management platform. The multi-scale disaster situation awareness module incorporates an agent-based road traffic exposure prediction submodule and a multi-source data fusion-based power distribution network disaster situation awareness submodule. It is used to convert water accumulation prediction results into traffic congestion index prediction and power distribution network node failure probability prediction, and push customized early warning and handling suggestions to transportation and power management departments through the early warning information generation and release module.

[0031] In this embodiment, the early warning system is additionally configured with a multi-scale disaster situation awareness module. This module, as a parallel processing unit, connects simultaneously with the intelligent flooding prediction module, external traffic management platform, and power management platform via a standard data interface. The module contains two core sub-modules. The first is an agent-based road traffic exposure prediction sub-module. It abstracts the road network as a graph structure composed of road segments and intersection nodes, treating each vehicle as a mobile agent and loading the city's real-time traffic origin-destination demand. Upon receiving spatialized flood depth prediction time-series data, the agent dynamically avoids road segments with flooding exceeding a safe threshold during route planning. The model calculates the travel delays and queue lengths of all vehicles in the road network through microscopic simulation, ultimately aggregating and outputting a traffic congestion index prediction and a list of high-risk traffic routes at a set time step, using road segments as units. The second is a distribution network disaster situation awareness sub-module based on multi-source data fusion. This sub-module integrates the distribution network topology, equipment locations and their historical fault records, and flood resistance standards. The model overlays the predicted water depth distribution with the spatial location of equipment: for nodes where the water level exceeds the equipment foundation's waterproofing platform, the failure probability is calculated using a pre-trained Bayesian network or vulnerability curve. The model also considers the topological impact of load transfer, predicting the potential load loss and outage area. The prediction results from both submodules are pushed to the early warning information generation and dissemination module. This module uses these results to generate instructions for traffic management departments, including suggested detour routes and intersection control timings, and to generate customized notifications for power departments, including early warnings for critical node reinforcement and advance recommendations for emergency repair resources.

[0032] In a second aspect embodiment of the present invention, a method for early warning of urban and rural rainstorm flooding that integrates the effects of urban heat island and aerosols is provided, applied to the urban and rural rainstorm flooding early warning system integrating the effects of urban heat island and aerosols in the first aspect embodiment, comprising: Collect multi-source dynamic and static data within the study area to construct a standardized spatiotemporal database; the data includes at least meteorological forecast data, underlying surface and infrastructure data, water accumulation and disaster damage data, and urban environmental dynamic monitoring data; the urban environmental dynamic monitoring data includes dynamic indicators characterizing the urban heat island effect and dynamic indicators characterizing the atmospheric aerosol effect. A mechanism- and data-driven intelligent early warning model for rainstorms and urban flooding is constructed. The model includes a rainfall intelligent prediction submodule based on dynamic environmental factors, which uses real-time monitored urban environmental dynamic monitoring data as dynamic input and combines it with meteorological forecast data to make probability forecasts or intensity corrections for local short-term heavy rainfall. The intelligent prediction submodule for urban waterlogging considering environmental feedback is used to combine the rainfall prediction results of the intelligent rainfall prediction submodule and at least a portion of the dynamic indicators in the urban environmental dynamic monitoring data with static underlying surface data to quickly predict the spatiotemporal evolution of surface water depth. The early warning model is trained and its hyperparameters are optimized using the constructed spatiotemporal database. Based on real-time collected multi-source data, the trained early warning model is run continuously to dynamically update the rainfall and water accumulation prediction results, and multi-level early warning information is generated and released based on the prediction results.

[0033] Existing methods for early warning of urban rainstorms and flooding often separate rainfall forecasting and flooding simulation into two independent stages, handled separately by meteorological and municipal departments. This results in data exchange gaps and hinders rolling updates at the minute-by-minute response level. Furthermore, conventional early warning methods rarely consider the urbanization effects, such as the urban heat island effect exacerbating localized rainfall and aerosols altering the precipitation particle spectrum. Additionally, runoff generation models lag in recognizing dynamic information such as pipe network siltation and pre-existing soil moisture, leading to insufficient accuracy in early warnings of the urban rainstorm-flooding chain. Therefore, a fully automated method is needed, encompassing data collection, environmental sensing, integrated rainfall and waterlogging forecasting, and early warning dissemination, integrating disturbances such as the urban heat island and aerosols throughout the entire forecasting chain.

[0034] In this embodiment, the first step is to collect multi-source dynamic and static data within the study area to construct a standardized spatiotemporal database. Weather forecast data comes from hourly grid products output by numerical models; underlying surface and infrastructure data includes the latest land survey data, above-ground building outlines and heights, underground pipe network geographic information systems, and siltation monitoring records; waterlogging and disaster damage data comes from historical flooding point markers and historical water levels from street video sensors; urban environmental dynamic monitoring data includes dynamic indicators characterizing the urban heat island effect and dynamic indicators characterizing the atmospheric aerosol effect. All data are stored in the spatiotemporal database with a unified spatiotemporal reference and format. The second step is to construct a mechanism-data-driven intelligent early warning model for rainstorm-induced urban flooding. The model comprises two cascaded sub-modules: a rainfall intelligent prediction sub-module based on dynamic environmental factors, which transforms real-time monitored dynamic environmental factors such as the urban heat island and aerosols into a gridded map, inputting it along with radar data and numerical weather prediction background fields into an encoder-decoder deep learning network, outputting a gridded correction field of precipitation magnitude and probability; and a waterlogging intelligent prediction sub-module considering environmental feedback, which inputs the rainfall forecast sequence, dynamically updated soil saturation and pipe network flow capacity correction coefficients, and static underlying surface features into a bidirectional long short-term memory model, outputting a grid-by-grid waterlogging depth prediction sequence. The third step involves jointly fine-tuning the two sub-modules using a large number of historical rainstorm event records in a constructed spatiotemporal database, and using Bayesian optimization to find the optimal hyperparameter combination. The fourth step involves daily rolling operation after the system goes live: real-time collection of the latest multi-source data, driving the early warning model periodically, and dynamically updating rainfall and waterlogging predictions for the next few hours. The early warning information generation service continuously monitors the prediction results. Once the water depth triggers the preset threshold, it automatically extracts information such as the scope and severity of the disaster, generates multi-level early warnings, and releases them through multiple channels.

[0035] By explicitly incorporating heat island and aerosol factors into the complete data stream from precipitation forecasting to urban flooding simulation, this method overcomes the limitations of traditional methods that rely on unidirectional static environmental forcing. Through the rolling operation of the integrated "rainfall prediction-flooding prediction" model, near real-time responses to environmental changes and initial disturbances are achieved, significantly extending the system's lead time for sudden, localized rainstorm-induced flooding. The combination of a spatiotemporal database and Bayesian optimization ensures the model's generalization and adaptability across different climate zones and urban characteristics. The resulting multi-level early warning information includes both flood depth and spatiotemporal evolution trends, providing decision-makers with a scientific, accurate, and timely digital basis for tiered responses.

[0036] In a further embodiment of the present invention, the dynamic indicators characterizing the urban heat island effect include at least one of the following: satellite remote sensing surface temperature, near-surface air temperature, anthropogenic heat emission intensity, impervious surface area, building density and average height, and population activity density; the anthropogenic heat emission intensity is estimated in real time based on industrial energy consumption, vehicle traffic flow, and building energy consumption data. The dynamic indicators characterizing the atmospheric aerosol effect include at least one of the following: satellite remote sensing aerosol optical thickness, particulate matter concentration at ground monitoring stations, black carbon mass concentration, and volatile organic compound concentration.

[0037] In this embodiment, the dynamic indicators characterizing the urban heat island effect include at least the following six: Satellite remote sensing surface temperature, obtained using thermal infrared channel data from polar-orbiting and geostationary meteorological satellites through a split-window algorithm or other inversion methods, is used to capture the thermal substrate of the urban surface. Near-surface air temperature, obtained from a dense network of urban automatic weather stations, is used to calculate the actual near-surface temperature gradient after quality control. Anthropogenic heat emission intensity, a dynamic estimation indicator, is generated by the system using industrial electricity load data from the power dispatch center, real-time vehicle traffic flow and vehicle type distribution data from the traffic management bureau, and sub-metered energy consumption data from large public buildings. A bottom-up heat flux model is applied to sum the traffic heat emission, building heat emission, and industrial heat emission of each grid to generate real-time anthropogenic heat flux tiles. The proportion of impermeable surface, based on annual products generated from high-resolution remote sensing image interpretation, serves as a stable parameter characterizing heat capacity and evaporation characteristics. Building density and average height, extracted from a three-dimensional urban reality model, affect the radiation interception and ventilation capacity of street canyons. Population activity density, a real-time estimate of population thermal distribution based on anonymous mobile signaling data, reflects the metabolic heat production of human activities and the attraction effect of traffic. Dynamic indicators characterizing atmospheric aerosol effects include at least the following four: Satellite-sensed aerosol optical thickness, acquired from advanced sensors, providing a wide-area aerosol load; Particulate matter concentration from ground monitoring stations, serving as a ground-based true value for calibrating and correcting satellite inversion; Black carbon mass concentration, derived from real-time monitoring and mobile observations by black carbon meters, characterizing the heating effect of absorbing aerosols on atmospheric stability; Volatile organic compound (VOC) concentration, obtained from photochemical reaction observations at environmental monitoring superstations, serving as a tracer of secondary aerosol generation and influencing the activation spectrum of cloud condensation nuclei.

[0038] In a further embodiment of the present invention, the rainfall intelligent prediction submodule is constructed using a convolutional long short-term memory network or graph neural network with an encoder-decoder architecture; The input data for the intelligent rainfall prediction submodule includes time-series radar echo maps, numerically predicted rainfall fields, and synchronously gridded heat island intensity maps and aerosol optical thickness maps. The loss function of the intelligent rainfall prediction submodule incorporates physical constraint terms, which include heat wave-convection triggering index constraints and aerosol-cloud microphysical inhibition / enhancement index constraints.

[0039] In this embodiment, the intelligent rainfall prediction submodule is constructed using a convolutional long short-term memory network with an encoder-decoder architecture. The time series of input data includes: a mixed-scan reflectivity grid map retrieved from Doppler weather radar, a near-term forecast field output from a high-resolution numerical weather prediction model, a synchronously gridded heat island intensity map, and an aerosol optical thickness map. Before the data enters the encoder, all variables are uniformly interpolated to a set fine grid. The encoder uses several layers of ConvLSTM to progressively compress the temporal and spatial dimensions, extracting the coupled evolution features of "rainfall-heat island-aerosol". The decoder then maps these features back to the original space, outputting a time-step rainfall forecast sequence for a preset future period. The key technical point lies in the model's loss function: it consists of a standard data-driven mean squared error term and two specially designed physical constraint terms. The heat wave-convection triggering index constraint is implemented as follows: first, at each grid point, regions where the heat island intensity exceeds a given threshold and simultaneously possesses certain water vapor flux convergence conditions are marked as high-potential triggering zones. If the model predicts an abnormally low initial probability of convection or future rainfall at these grid points, a penalty is imposed. The aerosol-cloud microphysical suppression / enhancement index constraint establishes a piecewise precipitation response curve based on the magnitude of aerosol optical thickness and relative humidity. When aerosol levels are within a suitable range and relative humidity is high, the model will be penalized for weaker-than-expected rainfall output; conversely, when aerosol levels are too high and water vapor is insufficient, the model will be penalized for excessive rainfall output. These two constraints are regularized and added to the total loss, forcing the network parameters to conform to cloud physics constraints.

[0040] In a further embodiment of the present invention, the intelligent prediction submodule for urban flooding adopts a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks; The input features of the intelligent prediction submodule for urban flooding include: the rainfall time series, static underlying surface features, and dynamic environmental factor feature map output by the intelligent rainfall prediction submodule; the dynamic environmental factor feature map includes real-time updated soil saturation, surface runoff coefficient, and pipeline flow capacity correction coefficient. The surface runoff coefficient is dynamically calculated based on the proportion of impermeable surface and rainfall intensity; the pipeline flow capacity correction coefficient is dynamically determined based on the pipeline siltation level.

[0041] In this embodiment, the intelligent prediction submodule for urban flooding employs a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks. The model's input features are carefully organized into three tensor sets: the first is a rainfall time-series tensor, where each grid stores the time-step sequence of rainfall intensity for the future period output by the intelligent rainfall prediction submodule. The second is a static underlying surface feature tensor, including elevation, slope, flow direction, runoff accumulation, and impervious area ratio. The third is a dynamic environmental factor feature map sequence, which changes with each real-time data update. This dynamic feature map specifically includes multiple channels: the first channel is soil saturation, using a distributed antecedent rainfall index model, continuously updating the saturation factor of each grid based on past actual rainfall and evapotranspiration. The second channel is the surface runoff coefficient, which is dynamically calculated. Each grid runs a lookup and interpolation function in real time, taking the current impervious surface ratio and the first-time-term rainfall intensity of the future forecast for that grid as input, and outputting the dynamic runoff coefficient. When rainfall intensity exceeds a high value, even permeable areas will output runoff due to excessive infiltration, and the coefficient approaches its maximum value. The third channel is the pipeline flow capacity correction coefficient. The system maintains a pipeline siltation level database, which is updated after each pipeline dredging or inspection. Each manhole in this database is assigned a siltation reduction factor to the corresponding pipeline segment. When the waterlogging prediction module extrapolates, it reads these coefficients to ensure that the spatial distribution of drainage capacity reflects the latest health status of the pipeline network in real time. The first stage of the model uses one-dimensional convolution to extract multi-scale temporal patterns from the rainfall sequence, and then feeds them into a bidirectional long short-term memory network. At the same time, the hidden state of each time step of the bidirectional long short-term memory network uses an association-aware attention mechanism to query the spatial regions and feature channels more relevant to the current water accumulation evolution in the third set of dynamic environmental factor feature maps. Finally, the deconvolution output layer generates a water accumulation depth map.

[0042] In a further embodiment of the present invention, the early warning model further includes: The multi-scale disaster damage situation awareness submodule is used to input the spatiotemporal distribution data of surface water depth predicted by the intelligent waterlogging prediction submodule into the agent-based road traffic exposure prediction model and the power distribution network disaster damage situation awareness model based on multi-source data fusion, respectively, and output the traffic exposure prediction results of the road network and the damage probability prediction results of the power distribution network nodes.

[0043] In this embodiment, the early warning model further includes a multi-scale disaster situation awareness submodule. After receiving the spatiotemporal distribution data of surface water depth from the intelligent flooding prediction submodule, this submodule initiates two parallel simulation threads. Thread one enters an agent-based road traffic exposure prediction model. This model initializes a road network topology mapped to the real world and deploys a large number of vehicle agents on the road network based on a real-time traffic demand matrix generated from mobile phone signaling and floating car data. As the simulation clock advances, each agent plans its route based on current road conditions and destination. When the simulation kernel detects that the predicted water depth of a road segment exceeds the maximum wading depth threshold for vehicles, that road segment is dynamically marked as unsafe. The model outputs quantitative indicators such as the future average speed reduction ratio, congestion index, and exposure rating for each road segment by simulating the detours and congestion propagation process of all agents. Thread two enters a distribution network disaster situation awareness model based on multi-source data fusion. The model first resolves the distribution network topology into a node-branch connection matrix, assigning latitude and longitude and equipment waterproof reference elevation to each distribution transformer and switchgear. The predicted water depth is superimposed with the equipment elevation to calculate the relative submersion depth at each piece of equipment. For submerged equipment, a vulnerability curve calibrated based on historical flooding failure events is used to calculate its individual outage probability. Furthermore, a Monte Carlo or DC power flow calculation model is used to sample a large number of possible node failure scenarios to assess the risk of cascading overloads, ultimately outputting the failure probability, estimated load loss, and number of affected users for each distribution node. The prediction outputs from both threads—traffic exposure prediction results and distribution network node damage probabilities—are uniformly transmitted to the early warning generation and distribution stage, serving as the core basis for forming customized professional early warning instructions.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A city / township rainstorm flood early warning system integrating heat island and aerosol effects, characterized in that, include: The multi-source data acquisition module is used to collect and integrate real-time meteorological observation data, numerical weather prediction model output products, urban environmental dynamic monitoring data, underlying surface and infrastructure data, and water accumulation and disaster damage data. The urban environmental dynamic monitoring data is fused with near-real-time satellite remote sensing inversion products, ground monitoring station network data, and IoT sensor data. The multi-source data acquisition module is also used to standardize, interpolate, and assimilate the collected data with different spatiotemporal resolutions and different lag times to generate a time-consistent dynamic environmental factor field. The dynamic environmental factor analysis module is connected to the multi-source data acquisition module and is used to analyze and process the collected urban environmental dynamic monitoring data to generate a dynamic index grid map characterizing the heat island effect and aerosol effect. The intelligent rainfall prediction module is connected to the dynamic environmental factor analysis module. It is used to perform probability prediction or intensity correction of local short-term heavy rainfall based on meteorological forecast data and the dynamic index grid map, using a mechanism-data dual-driven deep learning model, and output rainfall prediction results. The intelligent prediction module for waterlogging is connected to the dynamic environmental factor analysis module and the intelligent rainfall prediction module, respectively. It is used to quickly predict the spatiotemporal evolution of surface water depth using a deep learning model based on the rainfall prediction results, at least a part of the dynamic index grid map, and static underlying surface data, and output the waterlogging prediction results. The model training and optimization module is used to train the deep learning models in the intelligent rainfall prediction module and the intelligent waterlogging prediction module using a historical database, and to optimize the hyperparameters using a Bayesian optimization algorithm. The early warning information generation and release module is connected to the intelligent prediction module for waterlogging and is used to generate multi-level early warning information based on the real-time rolling waterlogging prediction results and release it.

2. The urban and rural rainstorm flood early warning system integrating heat island and aerosol effects according to claim 1, characterized in that, The dynamic environmental factor analysis module generates dynamic indicators of the heat island effect, including a spatial distribution map of heat island intensity, which is calculated based on the difference between satellite remote sensing surface temperature and near-surface air temperature, and is assimilated and corrected by combining anthropogenic heat emission intensity data; the generated dynamic indicators of the aerosol effect include a spatial distribution map of aerosol optical thickness, which integrates satellite remote sensing data and particulate matter concentration data from ground monitoring stations.

3. The urban and rural rainstorm flood early warning system integrating heat island and aerosol effects according to claim 1, characterized in that, The deep learning model built into the intelligent rainfall prediction module is a convolutional long short-term memory network with an encoder-decoder architecture. Its encoder is used to extract the spatiotemporal coupling features of the rainfall field and the dynamic environmental factor field, and its decoder is used to output the rainfall prediction amount or probability of occurrence for a future preset period. During the training process of the intelligent rainfall prediction module, the loss function includes a data-driven mean squared error term and a physical constraint term, which is constructed based on the heat wave-convection triggering index and the aerosol-cloud microphysics index.

4. The urban and rural rainstorm flood early warning system integrating heat island and aerosol effects according to claim 1, characterized in that, The deep learning model built into the intelligent prediction module for urban flooding is a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks. The input feature vector of the intelligent prediction module for urban flooding explicitly includes a dynamic environmental factor feature map, which includes soil saturation, surface runoff coefficient, and pipeline flow capacity correction coefficient. The pipeline flow capacity correction coefficient is associated with a pipeline siltation level database.

5. The urban and rural rainstorm flood early warning system integrating heat island and aerosol effects according to claim 1, characterized in that, The system also includes: The multi-scale disaster damage situation awareness module is connected to the intelligent prediction module for urban flooding and waterlogging, as well as the external traffic management platform and power management platform. The multi-scale disaster situation awareness module incorporates an agent-based road traffic exposure prediction submodule and a multi-source data fusion-based power distribution network disaster situation awareness submodule. It is used to convert water accumulation prediction results into traffic congestion index prediction and power distribution network node failure probability prediction, and push customized early warning and handling suggestions to transportation and power management departments through the early warning information generation and release module.

6. A method for early warning of urban and rural rainstorm flooding that integrates the urban heat island and aerosol effects, applied to the urban and rural rainstorm flooding early warning system that integrates the urban heat island and aerosol effects as described in any one of claims 1 to 5, characterized in that, include: Collect multi-source dynamic and static data within the study area to construct a standardized spatiotemporal database; the data includes at least meteorological forecast data, underlying surface and infrastructure data, water accumulation and disaster damage data, and urban environmental dynamic monitoring data; the urban environmental dynamic monitoring data includes dynamic indicators characterizing the urban heat island effect and dynamic indicators characterizing the atmospheric aerosol effect. A mechanism- and data-driven intelligent early warning model for rainstorms and urban flooding is constructed. The model includes a rainfall intelligent prediction submodule based on dynamic environmental factors, which uses real-time monitored urban environmental dynamic monitoring data as dynamic input and combines it with meteorological forecast data to make probability forecasts or intensity corrections for local short-term heavy rainfall. The intelligent prediction submodule for urban waterlogging considering environmental feedback is used to combine the rainfall prediction results of the intelligent rainfall prediction submodule and at least a portion of the dynamic indicators in the urban environmental dynamic monitoring data with static underlying surface data to quickly predict the spatiotemporal evolution of surface water depth. The early warning model is trained and its hyperparameters are optimized using the constructed spatiotemporal database. Based on real-time collected multi-source data, the trained early warning model is run continuously to dynamically update the rainfall and water accumulation prediction results, and multi-level early warning information is generated and released based on the prediction results.

7. The urban and rural rainstorm flood early warning method integrating heat island and aerosol effects according to claim 6, characterized in that, The dynamic indicators characterizing the urban heat island effect include at least one of the following: satellite remote sensing surface temperature, near-surface air temperature, anthropogenic heat emission intensity, percentage of impervious surface, building density and average height, and population activity density; the anthropogenic heat emission intensity is estimated in real time based on industrial energy consumption, vehicle traffic flow, and building energy consumption data. The dynamic indicators characterizing the atmospheric aerosol effect include at least one of the following: satellite remote sensing aerosol optical thickness, particulate matter concentration at ground monitoring stations, black carbon mass concentration, and volatile organic compound concentration.

8. The urban and rural rainstorm flood early warning method integrating heat island and aerosol effects according to claim 6, characterized in that, The intelligent rainfall prediction submodule is constructed using a convolutional long short-term memory network or graph neural network with an encoder-decoder architecture. The input data for the intelligent rainfall prediction submodule includes time-series radar echo maps, numerically predicted rainfall fields, and synchronously gridded heat island intensity maps and aerosol optical thickness maps. The loss function of the intelligent rainfall prediction submodule incorporates physical constraint terms, which include heat wave-convection triggering index constraints and aerosol-cloud microphysical inhibition / enhancement index constraints.

9. The urban and rural rainstorm flood early warning method integrating heat island and aerosol effects according to claim 6, characterized in that, The intelligent prediction submodule for urban flooding adopts a bidirectional long short-term memory neural network model that integrates associative perception attention and convolutional neural networks. The input features of the intelligent prediction submodule for urban flooding include: the rainfall time series, static underlying surface features, and dynamic environmental factor feature map output by the intelligent rainfall prediction submodule; the dynamic environmental factor feature map includes real-time updated soil saturation, surface runoff coefficient, and pipeline flow capacity correction coefficient. The surface runoff coefficient is dynamically calculated based on the proportion of impermeable surface and rainfall intensity; the pipeline flow capacity correction coefficient is dynamically determined based on the pipeline siltation level.

10. The urban and rural rainstorm flood early warning method integrating heat island and aerosol effects according to claim 6, characterized in that, The early warning model also includes: The multi-scale disaster damage situation awareness submodule is used to input the spatiotemporal distribution data of surface water depth predicted by the intelligent waterlogging prediction submodule into the agent-based road traffic exposure prediction model and the power distribution network disaster damage situation awareness model based on multi-source data fusion, respectively, and output the traffic exposure prediction results of the road network and the damage probability prediction results of the power distribution network nodes.

Images