An artificial intelligence-based urban flood disaster early warning method and system

Abstract

The application relates to the technical field of flood early warning, and discloses a city flood disaster early warning method and system based on artificial intelligence. The method comprises the following steps: collecting five types of information, i.e. meteorological sensing, hydrological monitoring, geographical space, city operation and social sensing, in real time to obtain multi-source data with accurate space-time coordinates; extracting rainfall accumulation and confluence evolution related features through preprocessing, grid space-time alignment and key feature screening; constructing a physically constrained space-time fusion deep learning model, combining a multi-head attention mechanism to output future waterlogging depth prediction results; adapting to environmental changes such as city topography and drainage facilities through incremental updating and transfer learning; and generating multi-level early warnings by fusing waterlogging depth, influence range and regional vulnerability features, and synchronously outputting spatial distribution maps, time evolution trends and affected object evaluation information. The application can effectively support city flood control and disaster reduction decision making and precise public risk avoidance.

Description

Technical Field

[0001] This application relates to the technical field of flood early warning, and in particular to an artificial intelligence-based method and system for urban flood disaster early warning. Background Technology

[0002] With the continuous advancement of urbanization, extreme weather events are becoming more frequent, and urban flooding has become a key risk factor restricting the safe operation of cities. Developing accurate and efficient flood disaster early warning technology is a core requirement for improving urban disaster emergency response capabilities and reducing loss of life and property.

[0003] Existing early warning technologies still suffer from prominent problems such as insufficient dynamic adaptability of models and low level of precision in early warning: On the one hand, environmental changes such as urban terrain modification and drainage facility upgrades can easily cause a mismatch between the model's preset parameters and the actual scenario. Traditional models mostly adopt static architectures and lack effective dynamic optimization mechanisms, requiring frequent manual calibration to maintain early warning performance, resulting in low adaptation efficiency. On the other hand, early warning level classifications are mostly based on a single water depth indicator, failing to deeply integrate vulnerability characteristics such as regional population density, economic activity intensity, and infrastructure importance. This makes it difficult to output differentiated risk avoidance guidance and decision support information for different functional areas, thus limiting the practical application effectiveness of early warning information.

[0004] As can be seen from the above, how to effectively support urban flood control and disaster reduction decision-making and the public's accurate risk avoidance still needs to be addressed. Summary of the Invention

[0005] To effectively support urban flood control and disaster reduction decision-making and enable the public to accurately avoid risks, this application provides an artificial intelligence-based urban flood disaster early warning method and system.

[0006] Firstly, this application provides an artificial intelligence-based method for early warning of urban flooding disasters, employing the following technical solution:

[0007] An artificial intelligence-based method for early warning of urban flooding disasters includes:

[0008] Real-time collection of meteorological sensing information, hydrological monitoring information, geospatial information, urban operation information, and social sensing information is performed to obtain corresponding real-time multi-source data. All real-time data are accompanied by precise spatiotemporal coordinates. Among them, meteorological sensing information includes rainfall content and precipitation forecast content, hydrological monitoring information includes water level data and pipeline flow data, geospatial information includes topography and drainage network distribution characteristics, urban operation information includes water accumulation status in flood-prone areas and drainage facility parameters, and social sensing information includes water accumulation-related content reported by the public.

[0009] The real-time multi-source data is sequentially subjected to outlier detection, missing value imputation, and standardization. Then, spatiotemporal alignment is achieved through gridding to filter out key spatiotemporal feature data reflecting rainfall accumulation and confluence evolution.

[0010] Based on the key spatiotemporal feature data, a spatiotemporal fusion deep learning early warning model with physical constraints is constructed. The early warning model integrates a spatial feature extraction module, a temporal evolution learning module, and a spatiotemporal gating mechanism. It transforms the discretized differential terms in the physical equation describing water flow motion into the initialization rules and dynamic adjustment mechanism of convolution kernel weights. Combined with a multi-head attention mechanism, it achieves adaptive fusion of multi-source features to output the prediction results of water depth in future time periods.

[0011] A dynamic optimization mechanism for the model is established. Based on the incremental updates of historical multi-source data and real-time multi-source data, transfer learning is used to transfer the model knowledge obtained from historical training to the new scene for areas with environmental changes. Through feature space mapping and parameter fine-tuning, the model can quickly adapt to environmental changes. The areas with environmental changes include changes in urban terrain and upgrades to drainage facilities.

[0012] Based on the water depth prediction results, the system integrates the impact range and regional vulnerability characteristics to generate multi-level refined early warnings. The regional vulnerability characteristics include population density, economic activity intensity, and infrastructure importance. The corresponding early warning level is determined by coupling the water depth threshold with the vulnerability level, and the system simultaneously outputs the spatial distribution map of the warning area, the temporal evolution trend, and specific assessment information of the affected objects.

[0013] Optionally, a multi-source data spatiotemporal correlation mechanism can be established during real-time data acquisition. Other methods include:

[0014] Dynamic timestamp calibration is performed on meteorological sensing information and hydrological monitoring information. The temporal resolution of meteorological data is adjusted according to the rainfall intensity level, and a high-frequency acquisition mode is used during periods of heavy rainfall.

[0015] At the same time, drainage network nodes in geospatial information are used as reference points to establish spatial mapping relationships between meteorological monitoring points, hydrological sensing equipment and network nodes, so that various types of data form a mutually verifying network in the spatiotemporal dimension.

[0016] Optionally, the method further includes:

[0017] The gridding process employs a topology-aware dynamic partitioning strategy, using natural water catchment units in geospatial information as the basic framework, and combining urban road red lines and land parcel boundaries to correct the grid shape, so that each grid unit contains topological attributes such as water catchment area, ground slope, drainage direction, and surrounding pipeline access points.

[0018] During spatiotemporal alignment, data is aggregated within grid cells, and the water flow exchange coefficient between adjacent grid cells is recorded synchronously. The water flow exchange coefficient is dynamically calculated based on the connectivity of the pipeline network and the terrain elevation difference.

[0019] Optionally, in constructing a spatiotemporal fusion deep learning-based early warning model with physical constraints, the method further includes:

[0020] The physical equations describing water flow motion are discretized into differential terms and transformed into a dynamic adjustment mechanism for the convolution kernel weights. The coefficient ranges corresponding to the inertial, gravity, and drag terms in the equations are extracted and used as the constraint boundaries for the convolution kernel weights.

[0021] The hydraulic gradient of the grid cell is calculated in real time. When the hydraulic gradient exceeds the set hydraulic gradient threshold, the emergency adjustment mode of the convolution kernel weight is triggered, which strengthens the rapid evolution feature of the water flow by amplifying the weight ratio of the gravity term.

[0022] When a sudden change in water level is detected at a pipeline node, the local weight distribution of the convolution kernel is dynamically corrected based on the resistance term coefficient to capture the congestion effect after the water flow is obstructed.

[0023] Optionally, the multi-head attention mechanism includes data type branching and spatial region branching, and the method further includes:

[0024] The data type branch sets up independent attention heads for data from different sources. The query vector generation rules for each attention head are related to the spatiotemporal sensitivity of the data. The different sources of data include meteorological, hydrological, and geospatial data.

[0025] The spatial region branch divides the attention scope according to urban functional zones. The key vector of each region's attention head incorporates the historical flood characteristics parameters of the corresponding region. Among them, urban functional zones include residential areas, commercial areas, and industrial areas.

[0026] Multi-source features are adaptively fused in both type and spatial dimensions by cross-product calculation of attention weights in two branches.

[0027] Optionally, in the model dynamic optimization mechanism, online learning and transfer learning form a collaborative triggering logic, and the method also includes:

[0028] When the urban environment changes gradually, the basic parameters of the model are first updated based on three consecutive months of incremental data through online learning, and then transfer learning is initiated to fine-tune the feature extraction module of the changed area.

[0029] When the environment changes abruptly, the core parameters related to global hydrological patterns in the model are first frozen. The physical constraint weight matrix of the changed area is reconstructed through transfer learning. Then, the output layer parameters of the model are optimized using new scenario data in conjunction with online learning.

[0030] Optionally, in the process of generating multi-level refined early warnings, the coupling between regional vulnerability characteristics and water depth adopts a hierarchical weighted mechanism, and the method also includes:

[0031] The basic layer uses water depth as the core indicator, accounting for 60% of the weight; the enhancement layer introduces the duration of water accumulation and the accessibility of surrounding emergency resources, accounting for 30% of the weight; the correction layer combines the actual impact deviation value of similar historical warnings, accounting for 10% of the weight; among them, surrounding emergency resources include drainage equipment and refuge areas.

[0032] The final warning level is determined by a three-level weighted calculation, and each warning level is bound to a corresponding list of emergency response actions. The list content is dynamically adapted to the regional functional attributes, which include school areas and transportation hubs.

[0033] Secondly, this application provides an artificial intelligence-based urban flood disaster early warning system, which adopts the following technical solution:

[0034] An artificial intelligence-based urban flood disaster early warning system includes:

[0035] The multi-source data acquisition and spatiotemporal calibration module collects meteorological sensing information, hydrological monitoring information, geospatial information, urban operation information, and social sensing information in real time to obtain corresponding real-time multi-source data. All real-time data are attached with precise spatiotemporal coordinates. Among them, meteorological sensing information includes rainfall content and precipitation forecast content, hydrological monitoring information includes water level data and pipeline flow data, geospatial information includes topography and drainage network distribution characteristics, urban operation information includes water accumulation status in flood-prone areas and drainage facility parameters, and social sensing information includes water accumulation-related content reported by the public.

[0036] The multi-source data preprocessing and feature screening module sequentially performs outlier detection, missing value imputation, and standardization on the real-time multi-source data, and then achieves spatiotemporal alignment through gridding processing to screen out key spatiotemporal feature data reflecting rainfall accumulation and confluence evolution.

[0037] The physical constraint spatiotemporal fusion early warning model construction module constructs a physical constraint spatiotemporal fusion deep learning early warning model based on the key spatiotemporal feature data. The early warning model integrates a spatial feature extraction module, a temporal evolution learning module, and a spatiotemporal gating mechanism. It transforms the discretized differential terms in the physical equations describing water flow motion into initialization rules and dynamic adjustment mechanisms for convolutional kernel weights. Combined with a multi-head attention mechanism, it achieves adaptive fusion of multi-source features to output the predicted water depth in future time periods.

[0038] The early warning model dynamic optimization and environment adaptation module establishes a model dynamic optimization mechanism. Based on the incremental updates of historical multi-source data and real-time multi-source data, it uses transfer learning to transfer the model knowledge obtained from historical training to new scenarios for areas with environmental changes. Through feature space mapping and parameter fine-tuning, it quickly adapts to environmental changes. The areas with environmental changes include changes in urban terrain and upgrades to drainage facilities.

[0039] The multi-level early warning generation and information output module generates multi-level refined early warnings based on the water depth prediction results and integrates the impact range and regional vulnerability characteristics. Among them, the regional vulnerability characteristics include population distribution density, economic activity intensity and infrastructure importance. The corresponding early warning level is determined by coupling the water depth threshold and vulnerability level, and the spatial distribution map of the early warning area, the time evolution trend and the specific assessment information of the affected objects are output simultaneously.

[0040] Thirdly, this application provides a computer device, which adopts the following technical solution:

[0041] A computer device includes a processor running a program for an artificial intelligence-based urban flood disaster early warning method as described in any one of the preceding claims.

[0042] Fourthly, this application provides a storage medium, which adopts the following technical solution:

[0043] A storage medium storing a program for an artificial intelligence-based urban flood disaster early warning method as described in any one of the above.

[0044] In summary, this application includes at least one of the following beneficial technical effects:

[0045] By deeply integrating multi-source data and employing precise modeling, the scientific rigor and reliability of early warning systems are enhanced from the outset, providing robust data and model support for urban flood control and disaster reduction decision-making. Real-time acquisition and spatiotemporal correlation mechanisms of multi-source heterogeneous data enable collaborative verification of information across all dimensions, including meteorology, hydrology, and geospatial data. Combined with dynamic gridding processing based on topological perception, this ensures the accuracy of data spatiotemporal alignment and the effectiveness of feature extraction. A deep fusion model of physical constraints and deep learning transforms the physical equations of water flow motion into convolutional kernel weight constraints and dynamic adjustment mechanisms. Coupled with a data type-spatial region dual-branch multi-head attention mechanism, it accurately captures the spatiotemporal evolution patterns of runoff generation and confluence under complex underlying surface conditions, significantly improving the accuracy and timeliness of water depth prediction. A collaborative optimization mechanism combining online learning and transfer learning dynamically adapts the model to both gradual and abrupt changes in the urban environment, ensuring stable long-term early warning performance. This allows decision-makers to formulate scientific flood control strategies in advance, such as drainage facility scheduling and risk area management, based on accurate and real-time prediction results.

[0046] Meanwhile, through refined early warning generation and differentiated information output, precise public risk avoidance and efficient emergency response are achieved. A layered weighted coupling mechanism based on regional vulnerability characteristics and water depth overcomes the limitations of single-indicator early warnings. Combined with historical impact bias corrections, this makes early warning levels more closely aligned with the actual risk levels of different regions. Each early warning level is linked to a dynamic list of emergency response actions, along with differentiated information packages for urban management departments, the public, and emergency rescue teams. This provides management departments with decision-making basis for drainage scheduling and resource allocation, while clearly presenting key guidance such as evacuation routes and refuge points to the public. The simultaneously outputted spatial distribution map and temporal evolution trend of the early warnings allow the public to intuitively grasp the scope and development of risks, effectively improving risk avoidance efficiency. A multi-channel, real-time updated information transmission model ensures timely delivery of early warning information, assisting in rapid public response and precise deployment of emergency rescue, minimizing loss of life and property caused by floods. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating an artificial intelligence-based urban flood disaster early warning method according to an exemplary embodiment.

[0048] Figure 2 This is a structural block diagram of an artificial intelligence-based urban flood disaster early warning system, illustrated according to an exemplary embodiment. Detailed Implementation

[0049] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0050] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0051] This application discloses an artificial intelligence-based urban flood disaster early warning method, referring to... Figure 1 ,include:

[0052] S100 collects meteorological sensing information, hydrological monitoring information, geospatial information, urban operation information, and social sensing information in real time, obtaining corresponding real-time multi-source data. All real-time data are accompanied by precise spatiotemporal coordinates. Among them, meteorological sensing information includes rainfall content and precipitation forecast content; hydrological monitoring information includes water level data and pipeline flow data; geospatial information includes topography and drainage network distribution characteristics; urban operation information includes water accumulation status in flood-prone areas and drainage facility parameters; and social sensing information includes water accumulation-related content reported by the public.

[0053] The core objective of S100 is to build a multi-dimensional, highly reliable, multi-source data acquisition system to achieve comprehensive capture of key information related to urban flooding and unification of spatiotemporal benchmarks. The specific execution process is as follows:

[0054] First, a multi-source heterogeneous data acquisition network was established. Based on the full-scenario needs of urban flood early warning, three types of acquisition channels—fixed monitoring, mobile sensing, and interface integration—were integrated to construct a comprehensive acquisition network covering meteorology, hydrology, geography, urban operations, and social sensing. The fixed monitoring channel deployed hardware facilities such as ground-based automatic weather stations, river water level sensors, pipeline flow monitoring terminals, and waterlogging monitoring equipment at flood-prone areas. The mobile sensing channel included a public mobile terminal location reporting portal and a drone inspection module. The interface integration channel connected with third-party systems such as numerical weather prediction systems, urban geographic information systems (GIS), drainage facility control systems, government feedback platforms, and social media monitoring platforms to ensure seamless access to multiple types of information.

[0055] Secondly, real-time data collection is implemented in a categorized manner. Differentiated collection strategies are adopted for five core types of information: For meteorological sensing information, real-time measured data such as rainfall intensity and amount are collected through automatic weather stations, and large-scale rainfall distribution data are obtained through weather radar. Simultaneously, the system interfaces with numerical weather prediction systems to acquire short- and medium-term precipitation forecast data, forming a combination of "measured + forecast" meteorological information; for hydrological monitoring information, sensors deployed along riverbanks and key nodes of the pipeline network are used to collect real-time data on water levels, instantaneous flow rates, and cumulative flow rates in the pipeline network, ensuring continuous capture of hydrological dynamics; for geospatial information, high-precision digital elevation models (DEMs) and urban drainage network GIS data are used as... The system is fundamentally based on regularly updated topographic and geomorphological change data through remote sensing imagery, and synchronized static and quasi-static characteristic data such as the distribution, diameter, and direction of drainage pipe networks through the pipeline network operation and maintenance system. For urban operation information, real-time status data such as water depth and extent are collected through flood-prone area monitoring equipment, and equipment operation data such as pump station operation parameters and gate opening and closing status are obtained through the drainage facility control system. For socially perceived information, water accumulation location, water accumulation degree, and other related content are collected from the public through multiple channels, including the water accumulation complaint portal of the government feedback platform, the water accumulation topic monitoring interface of social media, and the location reporting function of public mobile terminals, so as to achieve information complementarity between "professional monitoring and public participation".

[0056] Secondly, standardized and precise spatiotemporal coordinates are added. To address the issue of inconsistent spatiotemporal references for multi-source data, standardized spatiotemporal calibration rules are adopted: spatial coordinates uniformly adopt the geodetic coordinate system (such as CGCS2000), and precise latitude and longitude coordinates (accurate to the meter level) are added to each piece of collected data through methods such as GPS modules built into sensors and positioning associated with GIS systems; time coordinates uniformly adopt UTC+8 timestamps, accurate to the second level, where measured data (such as rainfall and water level) are based on the real-time clock of the data acquisition equipment, forecast data (such as precipitation forecasts) are based on the forecast release time, and public feedback data are based on the submission time, ensuring that all data have a unified reference that is correlated and comparable in the spatiotemporal dimensions.

[0057] Finally, preliminary data validation is performed. The collected real-time multi-source data undergoes preliminary screening, removing invalid data through preset rules: first, format validation, checking the completeness of data fields and the compliance of data types, removing data with format errors or missing fields; second, range validation, determining reasonable value ranges for various information types based on historical data, removing abnormal data exceeding normal thresholds (such as negative rainfall or water levels exceeding the designed upper limit of the river channel); and third, deduplication validation, removing duplicate data from the same source and at the same spatiotemporal coordinates, retaining the latest data record to ensure the validity and conciseness of the output data.

[0058] By constructing a multi-dimensional, multi-source data acquisition system and a standardized spatiotemporal calibration mechanism, we not only broke through the limitations of traditional early warning technologies with their single data source, achieving full-dimensional information coverage of meteorology, hydrology, geography, urban operation, and social perception, providing rich and comprehensive basic data for subsequent data preprocessing and model building; but also ensured the spatiotemporal consistency and data quality of multi-source data through unified spatiotemporal coordinates and preliminary verification, avoiding early warning deviations caused by inconsistent data benchmarks or invalid data, laying a solid data foundation for subsequent key spatiotemporal feature extraction, accurate calculation of early warning models, and refined early warning generation.

[0059] S200 sequentially performs outlier detection, missing value imputation, and standardization on real-time multi-source data, and then achieves spatiotemporal alignment through gridding to filter out key spatiotemporal feature data reflecting rainfall accumulation and confluence evolution.

[0060] S200 improves the quality of real-time multi-source data output by S100, eliminates data heterogeneity, achieves spatiotemporal dimension unification, and selects key features to support model building, providing high-quality input for subsequent accurate calculation of the early warning model. The specific execution process is as follows:

[0061] First, outlier detection and cleanup were conducted. Given the multi-source nature of the S100 data, a differentiated outlier detection strategy was adopted: For time-series data such as meteorological rainfall and hydrological water levels, a sliding window statistical method was used, using the mean and standard deviation within a fixed time window as a benchmark to identify abnormal fluctuations exceeding the range of "mean ± 3 times the standard deviation," while simultaneously verifying data rationality using domain interpolation. For static or quasi-static data such as geospatial and urban operation data, deviations from the normal range were identified by comparing with historical benchmark datasets (e.g., pipeline flow exceeding design capacity, severe discrepancies between topographic elevation data and remote sensing imagery). After detecting outliers, the data source and anomaly type (e.g., sudden fluctuation anomalies, systematic error anomalies) were marked. Isolated outliers were directly removed, while continuous outliers were marked and the anomaly period was recorded to provide a reference for weight adjustments in subsequent model calculations.

[0062] Secondly, precise missing value imputation is implemented. A hierarchical imputation strategy is adopted based on the type of missing data (random missing, continuous missing, structural missing): For short-term random missing data (such as missing single meteorological data), spatial interpolation is used, utilizing similar data from adjacent spatiotemporal nodes (such as monitoring stations in the same area, or adjacent grid data at the same time) for linear interpolation imputation; for long-term continuous missing data (such as missing hourly data at a certain monitoring point), a method of "historical similar scene mapping + associated data assistance" is used. First, datasets under similar meteorological and hydrological conditions from the same historical period are matched, and then associated data (such as referencing radar echo data when rainfall is missing, or referring to upstream and downstream node data when pipeline flow is missing) are combined for predictive imputation; for structurally missing data (such as areas lacking social perception feedback), zero-filling combined with missing markers is used to avoid data distortion caused by manual imputation, while weakening the weight of this type of data in subsequent feature selection. After imputation, consistency verification with adjacent time periods / regions ensures the rationality of the imputed data.

[0063] Next, data standardization is performed. To eliminate the differences in magnitude and dimensions between different types of data and to achieve the fusion calculation of multi-source data, a classification standardization strategy is adopted: For numerical continuous data such as meteorological rainfall, hydrological water level, and pipeline flow, the Z-score standardization method is used to convert the data into standardized data with a mean of 0 and a standard deviation of 1, preserving the distribution characteristics and relative differences of the data; for categorized data such as underlying surface type, drainage facility operation status, and public feedback level, one-hot coding or label coding is used to convert non-numerical information into a numerical format that the model can recognize; for data with fixed benchmarks such as coordinates and elevations in geospatial data, normalization is used to map the values ​​to the [0,1] interval to ensure uniform data scale. After standardization, a multi-source data matrix with a unified format is generated, providing a foundation for subsequent spatiotemporal alignment.

[0064] Then, spatiotemporal alignment is achieved through gridding. Based on the topography and drainage network distribution in geospatial information, a "dynamic grid division" strategy is adopted: combining the boundaries of natural catchment units, urban road red lines, and land parcel division, the warning area is divided into regular and topologically related grid units, ensuring that each grid contains complete topographic, pipeline, and monitoring point information. Standardized multi-source data is aggregated to the corresponding grid units according to the principle of "spatiotemporal coordinate matching": time-series data is aligned with the grid time granularity (e.g., hourly, half-hourly) according to the collection timestamp, and spatial data is accurately mapped to the grid according to latitude and longitude coordinates, achieving spatiotemporal unity of "one grid, one data set". At the same time, the topological attributes of each grid unit (e.g., upstream catchment grid, downstream drainage outlet grid) are recorded to provide a spatial correlation basis for subsequent extraction of confluence evolution features.

[0065] Finally, key spatiotemporal features were screened. Based on the core evolutionary pattern of flood disasters—"rainfall-runoff generation-confluence"—a dual strategy of "feature importance assessment + domain knowledge screening" was adopted: First, statistical methods (such as mutual information value, variance contribution, and random forest feature ranking) were used to quantify the influence weight of each data dimension on water depth, and candidate features with high weight rankings were screened out; then, combined with hydrological domain knowledge, features reflecting the core process were focused on, such as cumulative rainfall intensity, rainfall duration, confluence path length, pipeline load rate, topographic slope, and underlying surface runoff coefficient, etc., to eliminate redundant features unrelated to flood evolution (such as urban operation auxiliary data that are not directly related to water accumulation). Ultimately, a key spatiotemporal feature data set centered on "rainfall accumulation-runoff evolution-pipeline load" was formed, reducing the complexity of subsequent model calculations.

[0066] By employing differentiated preprocessing procedures and precise feature selection, this approach effectively addresses issues such as heterogeneity, missing data, and inconsistencies in spatiotemporal benchmarks among multi-source data. This significantly improves data quality and usability, preventing low-quality data from interfering with the early warning model. Furthermore, gridded spatiotemporal alignment enables the organic integration of multi-source data. By filtering key features to eliminate redundant information and focus on core patterns, the computational complexity of subsequent models is reduced, computational efficiency is improved, and targeted, reliable feature inputs are provided for the early warning model to accurately capture the spatiotemporal evolution patterns of floods. This serves as a crucial bridge connecting data acquisition and model building, directly determining the accuracy and stability of subsequent early warning model predictions and providing core data support for achieving precise early warning of urban floods.

[0067] S300 is a spatiotemporal fusion deep learning early warning model based on key spatiotemporal feature data and physical constraints. The early warning model integrates a spatial feature extraction module, a temporal evolution learning module, and a spatiotemporal gating mechanism. It transforms the discretized differential terms in the physical equations describing water flow into initialization rules and dynamic adjustment mechanisms for convolutional kernel weights. Combined with a multi-head attention mechanism, it achieves adaptive fusion of multi-source features to output the predicted water depth for future periods.

[0068] S300 is a spatiotemporal fusion deep learning model that combines the advantages of physical mechanism constraints and data-driven approaches. It accurately captures the spatiotemporal evolution of floods from rainfall to runoff to confluence, and outputs reliable predictions of future water depth. The specific execution process is as follows:

[0069] First, the overall architecture framework of the model is established. Following the logical thread of "feature extraction – constraint embedding – fusion computation – prediction output," a hierarchical collaborative architecture is constructed: the input layer receives key spatiotemporal feature data output by the S200, categorizing it into "spatial features, temporal features, and multi-source associated features" and inputting it into the corresponding modules; the intermediate layer integrates spatial feature extraction, temporal evolution learning, spatiotemporal gating mechanisms, physical constraint embedding, and multi-head attention fusion modules to achieve hierarchical feature extraction, constraint calibration, and adaptive fusion; the output layer is a fully connected layer that maps the fused high-dimensional features to grid-level water depth prediction values, forming a closed-loop computation process of "input – processing – output."

[0070] Secondly, the spatial feature extraction module performs computations. A three-dimensional convolutional neural network (3D-CNN) is used as the core architecture, with multi-scale convolutional kernel combinations (such as 3×3×1, 5×5×1, and 7×7×1) designed to extract features at different spatial scales: small-scale convolutional kernels focus on local features within the grid (such as terrain slope and the proportion of underlying surface types); medium-scale convolutional kernels capture short-distance correlation features between grids (such as connectivity of adjacent grid networks and local features of confluence paths); and large-scale convolutional kernels extract regional-level global features (such as large-scale catchment areas and the overall distribution density of drainage networks). Simultaneously, dilated convolution technology is introduced to expand the feature receptive field without increasing computational load, ensuring the model can capture cross-regional spatial correlation patterns and ultimately output a spatial feature vector with uniform dimensions.

[0071] Next, the temporal evolution learning module is implemented. A bidirectional long short-term memory (Bi-LSTM) network architecture is adopted to mine temporal features from two directions for time-series data with key spatiotemporal characteristics (such as rainfall intensity time series and water level change time series): the forward LSTM branch follows the "history → current" time series direction to learn the gradual evolution patterns of rainfall accumulation and water level increase; the backward LSTM branch follows the "current → history" backtracking direction to mine the abrupt change patterns of key time nodes (such as the start of heavy rainfall and the moment when the pipeline network is at full load). Through the complementary fusion of bidirectional temporal information, a temporal feature vector containing historical dependencies and trend evolution characteristics is generated, solving the problem that traditional unidirectional time series models are insufficient in capturing sudden changes.

[0072] Then, a physical constraint mechanism is embedded and the convolution kernel weights are dynamically adjusted. Based on the Saint-Venant equations (continuity and momentum equations), the discretized differential terms (inertia, gravity, and drag terms) are first mathematically decomposed to extract reasonable ranges for each coefficient. These ranges are then used as the initial constraint boundaries for the convolution kernel weights, ensuring that the initial model parameters conform to the basic physical laws of water flow and avoiding potential "physical paradoxes" (such as backflow due to water accumulation or runoff without rainfall) that may occur in data-driven models. Based on this, a dynamic weight adjustment rule is constructed: the hydraulic gradient of each grid cell is calculated in real time. When the hydraulic gradient exceeds a preset threshold (reflecting the rapid evolution of water flow), an emergency adjustment mode is triggered. This mode amplifies the weight proportion of the convolution kernel corresponding to the gravity term, strengthening the model's ability to capture rapid water flow convergence. When a sudden change in water level at a network node is detected (reflecting water flow obstruction), the local weight distribution of the convolution kernel in the corresponding region is dynamically corrected based on the drag term coefficients, highlighting the characteristic expression of water flow congestion effects and ensuring that model predictions always conform to physical laws.

[0073] Next, a spatiotemporal feature fusion is achieved through a spatiotemporal gating mechanism. A dual-gating unit containing a "spatial gate" and a "temporal gate" is designed. The spatial gate calculates the weight coefficients of spatial features using a sigmoid activation function, filtering out key spatial features from different regions (such as spatial features of flood-prone areas and areas with weak pipe networks). Similarly, the temporal gate calculates the weight coefficients of temporal features using a sigmoid activation function, focusing on temporal features at key time-series nodes (such as periods of heavy rainfall and periods of rapid water accumulation). The gating unit dynamically controls the transmission ratio of spatial and temporal features, suppressing interference from redundant spatiotemporal features (such as temporal features during periods without rainfall and spatial features from non-critical areas), and outputs a filtered and calibrated fused feature vector.

[0074] Then, a multi-head attention mechanism is used to achieve adaptive fusion of multi-source features. A two-branch multi-head attention structure of "data type branch + spatial region branch" is adopted: the data type branch configures independent attention heads for features from different sources such as meteorology, hydrology, and geospatial data. The query vector generation rules of each attention head are related to the spatiotemporal sensitivity of the corresponding data (e.g., meteorological data has high time sensitivity, so the query vector strengthens the weight of the time-series dimension; geospatial data has high spatial sensitivity, so the query vector strengthens the weight of the spatial dimension). The spatial region branch divides the attention range according to urban functional zones (residential areas, commercial areas, industrial areas, etc.), and the key vector of each region's attention head incorporates the historical flood feature parameters of the corresponding region (e.g., historical waterlogging frequency, maximum waterlogging depth). By calculating the cross-product of the attention weights of the two branches, the multi-source features are achieved through dual adaptive weighting in the two-dimensional dimension of "data type - spatial region", allowing the model to automatically focus on feature combinations that play a key role in waterlogging prediction (e.g., the feature combination of heavy rainfall periods + flood-prone residential areas).

[0075] Finally, model training and water depth prediction are performed. Historical key spatiotemporal feature data and corresponding measured water depth data are used as training samples. Mean squared error (MSE) is used as the loss function, and the model parameters are iteratively trained using the Adaptive Moment Estimation (Adam) optimization algorithm until the model converges (the loss function value stabilizes below a preset threshold). After training, real-time key spatiotemporal feature data is input into the model. Through the collaborative computation of the above modules, the predicted water depth for each grid cell in the future time period is output. The prediction results are correlated with grid coordinates and timestamps, forming a precise mapping relationship between "spatiotemporal grid - water depth".

[0076] By constructing an early warning model that deeply integrates physical constraints and spatiotemporal fusion, this approach overcomes the limitations of traditional hydrological and hydrodynamic models, which suffer from computational complexity and poor real-time performance. It also addresses the issues of purely data-driven models lacking physical plausibility and having weak generalization capabilities. The physical constraint mechanism ensures that prediction results conform to the basic laws of water flow, avoiding unrealistic prediction biases. The spatiotemporal fusion architecture, in collaboration with bidirectional LSTM and multi-scale 3D-CNN, accurately captures fine-grained features of flood spatiotemporal evolution. A multi-head attention mechanism enables adaptive focusing of multi-source features, enhancing the model's ability to capture features from key scenarios (heavy rainfall, flood-prone areas). Ultimately, the module outputs grid-level, future-period water depth predictions with high accuracy, reliability, and timeliness, providing core computational support for subsequent dynamic model optimization and refined early warning generation. This is the core technology for achieving accurate urban flood early warning.

[0077] S400 establishes a dynamic model optimization mechanism, based on incremental updates of historical multi-source data and real-time multi-source data. For areas with environmental changes, it uses transfer learning to transfer the model knowledge obtained from historical training to new scenarios. Through feature space mapping and parameter fine-tuning, it quickly adapts to environmental changes. These areas with environmental changes include changes in urban terrain and upgrades to drainage facilities.

[0078] S400 is a dynamic optimization mechanism that combines incremental updates and transfer learning to address the performance degradation of models caused by changes in the urban environment. This ensures that the model remains adapted to the real-world scenario and maintains high-precision early warning capabilities. The specific execution process is as follows:

[0079] First, a basic framework for dynamic model optimization is established. An environmental change identification and classification module is built first. By continuously monitoring updated geospatial information data (such as changes in topographical remote sensing images) and facility operation and maintenance records in urban operational information (such as pump station renovation and pipeline laying logs), the module automatically identifies the type and scope of urban environmental changes, classifying them into two categories: gradual changes (such as pipeline aging and local topographical fine-tuning) and abrupt changes (such as the construction of large-scale drainage pump stations and large-scale topographical modifications like lake reclamation). Simultaneously, a historical data and model knowledge repository is constructed, archiving historical multi-source data, model training parameters, feature extraction weights, and other core information along the time dimension. This provides data support for incremental updates and reserves reusable model knowledge (such as parameters for the general hydrological law extraction module and physical constraint rules) for transfer learning.

[0080] Secondly, an incremental update execution process was designed. A dual-mode approach of "regular periodic update + triggered update" was adopted: Under normal conditions, real-time multi-source data and historical similar scene data for the current day are automatically extracted on a fixed daily cycle to construct an incremental training dataset. The dataset size is controlled to be 10%–20% of the total training data to avoid resource waste. Mini-batch SGD is used to locally update model parameters, focusing on updating parameters strongly correlated with real-time data, such as fully connected layers and attention mechanism weights. Parameters of core physical constraint modules and general feature extraction modules are frozen to ensure the stability of the model's basic performance. When the fluctuation of key data (such as rainfall intensity or pipeline flow) exceeds a preset threshold for multiple consecutive collection cycles (e.g., fluctuation exceeding 50% within 24 hours), or when slight, gradual changes occur in the environment, an immediate incremental update is triggered, expanding the incremental dataset size to 30%–40% of the total data and simultaneously updating the model's feature mapping layer parameters to quickly adapt to data fluctuations or slight environmental changes.

[0081] Secondly, implement a differentiated transfer learning strategy. For different types of areas experiencing environmental changes, develop a customized solution of "feature layer transfer + decision layer fine-tuning": For areas with gradual environmental changes (such as localized pipe network aging or small-scale green space renovation), first optimize the basic parameters of the model by incrementally updating based on three consecutive months of incremental data, and then start transfer learning—retain the core module parameters in the model that capture global hydrological patterns (such as general rainfall-runoff relationships), and only fine-tune the spatial feature extraction sub-module that is strongly correlated with environmental changes. Through feature space mapping, transform the feature extraction rules in the historical model that were adapted to the original environment into rules that are adapted to the changed environment, ensuring a smooth transition of the model. For areas experiencing sudden environmental changes (such as new drainage pumping stations or terrain changes caused by large-scale road reconstruction), the core parameters related to global hydrological patterns in the model (such as the basic structure of the physical constraint weight matrix) are first frozen to avoid interference from new scene data on the ability to extract general patterns. Then, the physical constraint weight matrix of the changed area is reconstructed through transfer learning, transferring the physical rules of water flow from the historical model to the new scene. The constraint boundaries are adjusted in combination with the drainage facility parameters and terrain data of the new scene. Finally, incremental updates are initiated, using real-time multi-source data collected in the new scene to optimize the model output layer parameters and quickly improve the model's adaptation accuracy to the new scene.

[0082] Then, a model optimization effect verification mechanism is established. After each incremental update or transfer learning, a verification dataset (composed of historical measured data not used in training and real-time verification data collected for new scenarios) is automatically invoked. The performance of the optimized model is evaluated by calculating the mean squared error (MSE), mean absolute error (MAE), and early warning accuracy of the predicted water depth and measured data. If the indicators meet the preset threshold (e.g., MSE is lower than 120% of the historical best value), the updated model parameters are fixed as the operating parameters for subsequent early warnings. If the indicators do not meet the threshold, the optimization process is backtracked, the size of the incremental dataset, the fine-tuning range of transfer learning parameters, or the learning rate are adjusted, and the optimization process is re-executed until the model performance meets the requirements, forming a closed-loop mechanism of "optimization-verification-fixation".

[0083] Finally, a model optimization log and iteration mechanism are constructed. This automatically records the triggering conditions for each optimization (such as environmental change type and data fluctuation), optimization parameters (such as incremental dataset size and transfer learning fine-tuning layers), performance validation results, and other information, forming a complete optimization log to provide a reference for subsequent optimization strategy adjustments. Simultaneously, a regular comprehensive optimization cycle (e.g., every six months) is set. Combining the accumulated optimization log with long-term environmental change trends, the overall architecture and core parameters of the model are systematically calibrated to ensure the model adapts to the cumulative changes in the urban environment over the long term and avoids performance deviations caused by localized optimizations.

[0084] By constructing a dynamic optimization mechanism that combines incremental updates and transfer learning, this module effectively addresses the core issues of insufficient environmental adaptability and performance degradation caused by the static nature of traditional early warning models. The incremental update mode avoids the high resource consumption of full retraining, enabling efficient iteration of model parameters. The differentiated transfer learning strategy fully reuses historical model knowledge, significantly reducing model training costs in new scenarios and enabling rapid adaptation to gradual and abrupt environmental changes. This module ensures that the early warning model maintains high accuracy and reliability even under continuously changing environmental factors such as urban terrain and drainage facilities, preventing the expansion of early warning deviations due to environmental changes. It provides stable model support for the continued effectiveness of subsequent multi-level refined early warnings and is a crucial guarantee for achieving long-term accurate early warnings of urban flood disasters.

[0085] S500 generates multi-level refined early warnings based on water depth prediction results and integrates the impact range and regional vulnerability characteristics. Among them, regional vulnerability characteristics include population density, economic activity intensity and infrastructure importance. The corresponding early warning level is determined by coupling the water depth threshold with the vulnerability level, and the spatial distribution map, temporal evolution trend and specific assessment information of the affected objects are output simultaneously.

[0086] Among them, S500 transforms the water depth prediction results output by S300 into multi-level refined early warning information that fits the actual application scenario, providing direct guidance for urban flood control decisions and public evacuation. The specific implementation process is as follows:

[0087] First, we conducted data integration and standardization of early warning systems. We first connected with the grid-level water depth prediction results output by S300, splitting the prediction sequence by time dimension (future 1 hour, 3 hours, 6 hours, 12 hours, 24 hours) to clarify the peak water depth and duration of each grid in each time period. Then, through Geographic Information System (GIS) overlay analysis, we automatically delineated the impact range of each time period based on water depth thresholds (such as the depth standards corresponding to potential water accumulation, significant water accumulation, and severe water accumulation), marking the boundary of the range, the core affected area, and the diffusion trend. Simultaneously, we retrieved regional vulnerability characteristic data and standardized three types of indicators: population density, economic activity intensity, and infrastructure importance. Population density was converted into a quantitative value of 0-10 points based on the proportion of the region's population to its area; economic activity intensity was assigned a value based on GDP density and industrial agglomeration; and infrastructure importance was divided into levels of 5-10 points according to categories such as transportation hubs, medical facilities, schools, and ordinary buildings, ensuring that the three types of characteristics could directly participate in the coupled calculation.

[0088] Secondly, a regional vulnerability characteristic classification and weighting system was implemented. Based on standardized vulnerability data, the K-means clustering algorithm was used to classify regional vulnerability into three levels: low, medium, and high. Low-vulnerability areas are sparsely populated, with moderate economic activity and no core infrastructure; medium-vulnerability areas are areas with moderate population density and general commercial facilities or secondary infrastructure; high-vulnerability areas are areas with dense population (such as residential areas and commercial districts), concentrated economic activity, and core infrastructure such as transportation hubs, hospitals, and schools. Simultaneously, based on urban flood control and emergency management needs, differentiated weights were assigned to the three vulnerability characteristics: population density accounted for 40%, infrastructure importance accounted for 35%, and economic activity intensity accounted for 25%. A weighted summation was used to obtain the comprehensive vulnerability level score for each region.

[0089] Secondly, the warning level is determined through a "water depth-vulnerability" coupling mechanism. A two-factor coupling judgment matrix is ​​constructed, with water depth as the vertical dimension (divided into four gradients: mild, moderate, severe, and extremely severe) and comprehensive vulnerability level as the horizontal dimension (low, medium, and high). A one-to-one mapping relationship for warning levels is established: mild water depth + low vulnerability corresponds to Level 1 warning; mild water depth + moderate vulnerability / moderate water depth + low vulnerability corresponds to Level 2 warning; mild water depth + high vulnerability / moderate water depth + moderate vulnerability / severe water depth + low vulnerability corresponds to Level 3 warning; moderate water depth + high vulnerability / severe water depth + moderate vulnerability corresponds to Level 4 warning; and extremely severe water depth + any vulnerability / severe water depth + high vulnerability corresponds to Level 5 warning. A dynamic correction coefficient is introduced during the coupling process, and the judgment results are fine-tuned by combining the warning verification results of similar scenarios in the same period in history, avoiding the distortion of warning levels due to single data deviations.

[0090] Then, multi-dimensional early warning auxiliary information is generated. First, a spatial distribution map of the warning area is drawn. Based on the GIS platform, the warning level, impact range, and core risk points (such as flood-prone areas, severely flooded road sections, and the location of core infrastructure) are visualized. Key information such as the distribution of refuge sites, emergency rescue channels, and drainage facilities are marked. Different colors are used to distinguish the warning level to ensure intuitiveness and ease of understanding. Second, a time evolution trend chart is generated. The chart shows the changing trend of water depth, peak occurrence time, and receding cycle in the form of a time-series curve. Key time nodes are marked (such as the moment when water begins to affect traffic and the moment when water reaches the critical value for safe evacuation), providing decision-makers and the public with a time-dimensional basis for prediction. Third, specific assessment information of affected objects is compiled. By overlaying population, economic, and infrastructure data, the number and distribution of the affected population in each warning area, the list of key protected infrastructure (including hospitals, schools, substations, etc.), and the estimated range of potential economic losses are calculated. The groups that need to be prioritized for relocation, the facilities that need to be prioritized for protection, and the key points of emergency resource allocation are identified.

[0091] Next, a differentiated information output system was constructed. Information packages were customized to meet the needs of different users: a "Decision Support Package" for urban management departments, including explanations of warning levels, details of the affected area, suggestions for drainage facility scheduling, emergency resource (such as rescue teams and supplies) allocation plans, and suggestions for key area control; a "Risk Avoidance Guidance Package" for the public, including warning levels for their area, flood risk warnings, safe evacuation routes (avoiding flooded areas and dangerous road sections), locations of nearby shelters and ways to get there, and emergency contact numbers; and a "Rescue Deployment Package" for emergency rescue departments, including the distribution of high-risk points, location of trapped at-risk individuals, accessibility of rescue channels, a list of key infrastructure requiring rescue, and suggestions for deploying rescue forces.

[0092] Finally, a real-time update and multi-channel push mechanism for early warning information was established. Using an hourly cycle, combining the S300's updated water depth prediction results with real-time multi-source data (such as actual rainfall intensity and drainage facility operation status), the warning level, impact range, and auxiliary information were dynamically adjusted to ensure timeliness. Simultaneously, a multi-channel push mechanism was established, using government apps, social media platforms, SMS, emergency broadcasts, road electronic displays, and community notices to accurately push warning information to management departments, the public, and rescue teams in affected areas. High-level warnings were pushed simultaneously through multiple channels including pop-ups, SMS, and broadcasts to ensure full coverage and no omissions.

[0093] Through a comprehensive design encompassing "data integration – level determination – information output," the system transforms flood depth prediction into practical early warning information. On one hand, by coupling flood depth with regional vulnerability, it overcomes the limitations of traditional single-indicator early warnings, making warning levels more aligned with the actual risk levels of different areas, thus enhancing the precision and targeting of early warnings. On the other hand, the multi-dimensional output of early warning information (spatial maps, temporal trends, and assessment details) and the design of differentiated information packages provide urban management departments with a scientific basis for decision-making, facilitating the precise implementation of flood control measures such as drainage scheduling, resource allocation, and regional control. It also provides the public with clear and easy-to-understand evacuation guidelines, helping them quickly grasp the risk situation and choose safe routes, while simultaneously clarifying the focus and deployment direction for emergency rescue departments. This module effectively connects model prediction with practical application, transforming technical prediction results into practical information supporting urban flood control and disaster reduction decisions and enabling precise public evacuation, maximizing the practical value of early warning technology.

[0094] In this embodiment, a case study illustrates the situation: A coastal city in southern China experienced heavy rainfall during the summer, with the meteorological department forecasting localized downpours reaching torrential levels within 24 hours. After activating this early warning system, the system's S100 collected real-time multi-source data, including rainfall, water levels, and pipeline flow, through ground weather stations, radar, pipeline sensors, and public reporting channels, and performed spatiotemporal calibration. The S200 system preprocessed the data and filtered features to extract key characteristics such as cumulative rainfall intensity and pipeline load rate. The S300 system used a physically constrained deep learning model to accurately predict varying degrees of flooding in three areas—the old city, the riverside area, and others—with the core commercial district of the old city expected to experience flooding depths of 30-40 cm within the next 12 hours. Based on this prediction, the city management department proactively deployed three surrounding pumping stations to full capacity, closed low-lying roads in the old city, and allocated emergency supplies such as sandbags to flood-prone areas, gaining crucial time for flood prevention and disaster reduction.

[0095] Following the heavy rainfall, the S400 system, based on real-time rainfall data and pipeline operation status, dynamically optimized model parameters through incremental updates. The revised forecast indicated that the peak flooding in the old city area would arrive two hours earlier. The S500 system then generated a Level IV warning, sending a "guidance package" to residents of the old city area containing a map of the flooded area and evacuation routes, explicitly advising against three severely flooded sections and recommending two nearby shelters. A "deployment package" was also sent to emergency rescue departments, marking three high-risk areas for residents to be trapped (including one kindergarten and two older residential areas). Based on the warning information, residents either evacuated to safe areas in advance or chose alternative routes, and no one was trapped due to flooding. The public's need for precise evacuation was effectively met.

[0096] During the rainfall, the plan continuously updated early warning information hourly and adjusted emergency response strategies based on the receding trend of floodwaters. Thanks to the decision-makers' reliance on accurate forecasts, drainage facilities were scientifically scheduled and resources were precisely allocated. The floodwaters in the old city area completely receded within three hours after the rainfall stopped, preventing prolonged disruption to economic activities in the core business district. Emergency rescue departments, based on the location of trapped risk points provided by the early warnings, quickly completed inspections of key areas to ensure no hidden dangers were overlooked. Throughout the entire process, the plan, through a closed-loop process of "data collection – model prediction – dynamic optimization – accurate early warning," provided scientific support for urban flood control decision-making and clear guidance for public evacuation, effectively reducing the loss of life and property caused by floods.

[0097] In this embodiment of the application, a multi-source data spatiotemporal correlation mechanism is established during real-time acquisition, and the specific execution process further includes:

[0098] First, dynamic timestamp calibration of meteorological and hydrological data is implemented. The original acquisition timestamps of meteorological sensing information (such as rainfall intensity and amount) and hydrological monitoring information (such as water level and pipeline flow) are extracted. A time calibration model is established based on the transmission delay characteristics of these two types of data (such as weather radar data transmission delay and sensor data upload interval), and the timestamps are synchronously corrected to eliminate timing misalignments caused by clock deviations or transmission delays in the acquisition equipment. Simultaneously, acquisition levels are categorized according to rainfall intensity: For non-rainfall or light rain periods (daily rainfall < 10 mm), a conventional time resolution (e.g., 10 minutes / time) is used; for moderate to heavy rain periods (10 mm ≤ daily rainfall < 50 mm), the resolution is increased to 5 minutes / time; and for heavy rain and above (daily rainfall ≥ 50 mm), a high-frequency acquisition mode (1 minute / time) is automatically switched to ensure sufficient data density during key rainfall periods to support confluence process analysis.

[0099] Secondly, a spatial mapping relationship is constructed based on drainage network nodes. Key nodes of the drainage network (such as network junctions, pump station inlets and outlets, inspection wells, etc.) are extracted from geospatial information, and the precise latitude and longitude coordinates of each node (accurate to 0.1 meters) are recorded as spatial reference points. Through spatial matching algorithms, the coordinates of meteorological monitoring points (such as automatic weather stations) and hydrological sensing devices (such as water level gauges and flow meters) are bound to the nearest network nodes, establishing a one-to-one correspondence between "monitoring devices and network nodes," forming a spatial association index table. For example, meteorological station data in a certain area is associated with network nodes within 500 meters of it, and these nodes are also associated with three surrounding hydrological sensors, enabling meteorological data (such as rainfall in the area) and hydrological data (such as the network flow of the corresponding node) to achieve spatial cross-verification through the reference node (such as whether the trends of rainfall and network flow match).

[0100] Dynamic timestamp calibration enhances the timeliness and temporal consistency of data during critical periods (heavy rainfall). Spatial correlation and mutual verification of multi-source data are achieved through spatial mapping of pipeline node benchmarks, further improving the spatiotemporal matching accuracy and reliability of multi-source data, and providing a more accurate correlation data foundation for subsequent data fusion and model prediction.

[0101] In this embodiment of the application, the method further includes:

[0102] First, topology-aware dynamic grid morphology is used for grid division. Based on natural water catchment units (such as small watersheds and gully catchment areas) in geospatial information, the core boundaries of the grid division are determined. Then, urban road red lines and land parcel planning boundaries are overlaid to correct the basic grid morphology, preventing grid splitting across roads or land parcels and ensuring the grid matches the actual urban spatial management units. After division, each grid unit is automatically associated with topological attributes, including catchment area size, average ground slope, natural drainage direction, and the location and diameter of surrounding connected pipe network nodes, forming an integrated grid structure of "spatial morphology + topological attributes".

[0103] Secondly, spatiotemporal alignment and water exchange coefficient calculation are performed. During spatiotemporal alignment, standardized multi-source data (meteorological, hydrological, urban operation, etc.) are precisely aggregated to the corresponding grid cells according to coordinates, achieving spatiotemporal unification of multiple types of data within the same grid. At the same time, for adjacent grid cells, geospatial analysis is used to extract data on pipeline connectivity (such as whether they are directly connected and the diameter of the connecting pipes) and topographic elevation differences. A hydraulic calculation model is used to dynamically calculate the water exchange coefficient—the better the pipeline connectivity and the greater the topographic elevation difference, the higher the coefficient value, and vice versa. This coefficient directly reflects the water exchange capacity between adjacent grids and is synchronously recorded in the grid attributes.

[0104] By using topology-aware dynamic partitioning and water flow exchange coefficient calculation, the grid cells not only conform to the actual geographical and urban management boundaries, but also contain complete confluence topology relationships, which improves the accuracy and relevance of data spatiotemporal alignment, and provides grid-level data support that is more in line with actual water flow movement for subsequent models to capture the confluence evolution law.

[0105] In this embodiment of the application, the method for constructing a spatiotemporal fusion deep learning-based early warning model with physical constraints further includes:

[0106] First, the differential terms of the physical equations are transformed into a dynamic adjustment mechanism for the convolution kernel weights. Based on the Saint-Venant equations describing water flow, the discretized inertial terms (reflecting the inertial effect of water flow), gravity terms (reflecting gravity-driven water flow), and resistance terms (reflecting resistance such as pipe network friction) are mathematically decomposed. Through numerical simulation and fitting with historical data, reasonable ranges for the coefficients of the three terms are determined (e.g., the resistance term coefficient is limited to 0.02-0.08 due to different pipe network materials). These ranges are used as the constraint boundaries for the model's convolution kernel weights, ensuring that the weight initialization and dynamic adjustment do not exceed the range allowed by physical laws.

[0107] Secondly, the gravity term weight adjustment is triggered based on the hydraulic gradient. The hydraulic gradient (i.e., the water level difference per unit distance) of each grid cell is calculated in real time and compared with a preset threshold (e.g., 0.05 / m, corresponding to a rapid water flow state). When the hydraulic gradient exceeds the threshold, it is determined to be a rapid water flow evolution scenario, and the emergency adjustment mode of the convolution kernel weight is immediately triggered. The weight ratio of the convolution kernel corresponding to the gravity term is increased by the algorithm (e.g., from the basic ratio of 30% to 50%), which strengthens the model's capture of the feature of water flow accelerating and converging due to gravity.

[0108] Finally, the weight distribution of the resistance term is corrected based on sudden changes in the water level in the pipeline network. Water level changes at pipeline nodes are monitored in real time using hydrological monitoring data. When a sudden change in water level exceeding a threshold (e.g., a rise of 0.5m) is detected within a short period (e.g., within 5 minutes), it is determined that the water flow is obstructed (e.g., pipeline blockage, gate closure). At this point, based on the dynamic change rules of the resistance term coefficient, the weight distribution of the convolution kernel in the corresponding region is locally corrected (e.g., increasing the proportion of the resistance term weight in the grid surrounding the pipeline node), enabling the model to accurately capture the characteristics of blockage and high water level after the water flow is obstructed.

[0109] By transforming the differential terms of the physical equations into executable dynamic adjustment rules for convolutional kernel weights, the model parameters are constrained by physical laws to avoid unreasonable predictions. At the same time, the model can be specifically enhanced to capture features of key scenarios such as rapid water flow evolution and obstruction, thereby improving the adaptability and prediction accuracy of the early warning model to complex water flow movements.

[0110] In this embodiment of the application, the multi-head attention mechanism includes data type branching and spatial region branching, and the method further includes:

[0111] First, we construct a data type branch attention structure. For three different data sources—meteorological, hydrological, and geospatial—we configure independent attention heads (three in total). For each attention head, we customize query vector generation rules, strongly binding them to the spatiotemporal sensitivity of the data. Meteorological data (e.g., rainfall) has high temporal sensitivity, so the query vector emphasizes the weight allocation logic of the temporal dimension; hydrological data (e.g., water level) has balanced spatiotemporal sensitivity, so the query vector considers both temporal changes and spatial relationships; geospatial data (e.g., topography) has high spatial sensitivity, so the query vector focuses on the weight proportion of spatial topological relationships, ensuring that the core features of different data types are accurately captured.

[0112] Secondly, a spatial regional branch attention structure is constructed. The attention range is divided according to urban functional zones (residential areas, commercial areas, and industrial areas), and a dedicated attention head is configured for each zone. Historical flood characteristic parameters (such as historical water accumulation frequency, maximum water accumulation depth, and water receding time) of each zone are extracted, standardized, and integrated into the key vector of the corresponding regional attention head. This ensures that the key vector contains both real-time feature information and historical scenario experience, improving the adaptability to flood patterns in different functional areas.

[0113] Finally, the attention weights of the two branches are cross-fused and calculated. The attention weight matrices of the two branches are calculated separately: the data type branch outputs a weight matrix of "data type - feature importance", and the spatial region branch outputs a weight matrix of "region - feature importance". Through matrix cross-product operation, a three-dimensional weight matrix of "data type - spatial region - feature importance" is obtained, realizing dual adaptive weighting of multi-source features in the type dimension (prioritizing key data types) and spatial dimension (prioritizing high-risk regions). Finally, the fused high-dimensional feature vector is output.

[0114] By customizing the generation of dual-branch vectors and fusing them with cross-products, the attention mechanism can accurately match the spatiotemporal characteristics of different data and conform to the historical flood patterns of different regions. This enables dual focusing and adaptive fusion of multi-source features, significantly improving the model's ability to capture features of key data and high-risk areas, and further optimizing the prediction accuracy of the early warning model.

[0115] In this embodiment of the application, in the model dynamic optimization mechanism, online learning and transfer learning form a collaborative triggering logic, and the method further includes:

[0116] First, collaborative optimization is implemented to handle gradual environmental changes. When the environmental change identification module determines that a change is gradual (such as slow changes like local pipeline aging or small-scale green space renovation), an online learning mechanism is initiated: incremental data for three consecutive months (including comparative data before and after the change) is extracted from historical and real-time data to construct a targeted training set. The basic parameters of the model (such as feature mapping layers and attention weights) are updated using mini-batch gradient descent to allow the model to initially adapt to the slow environmental changes. After online learning is completed, transfer learning is initiated, focusing on areas strongly correlated with environmental changes (such as the grid around aging pipelines). Only the parameters of the feature extraction sub-modules (such as local convolutional kernels in spatial feature extraction) in this area are fine-tuned. Through feature space mapping, the feature rules adapted to the original environment in the historical model are transferred to the changed scene to achieve a smooth transition.

[0117] Secondly, collaborative optimization is used to handle abrupt environmental changes. When an abrupt change is identified (such as the construction of a large drainage pumping station or drastic changes like lake reclamation), the core parameters in the model that capture global hydrological patterns (such as the basic structure of the physical constraint weight matrix and parameters of the general rainfall-runoff generation module) are first frozen to prevent new scenario data from interfering with the model's learning of general patterns. Then, transfer learning is initiated. Based on the new topographic data and drainage facility parameters (such as pumping station head and new pipe diameter) of the changed area, the physical constraint weight matrix of the area is reconstructed, and the physical rules of water flow movement in the historical model are transferred to the new scenario and the constraint boundaries are adjusted. After the transfer learning is completed, online learning is initiated. Using real-time multi-source data collected in the new scenario (such as the operation data of the newly built pumping station and the water level monitoring data of the changed area), the output layer parameters of the model are optimized in a focused manner to quickly improve the prediction accuracy of water depth in the new scenario.

[0118] By leveraging the collaborative triggering logic of online learning and transfer learning, a layered processing approach of "basic adaptation - knowledge transfer - precise optimization" is implemented for different types of environmental changes. This approach not only reuses historical model knowledge to reduce optimization costs but also rapidly adapts to new scenarios through incremental data, significantly improving the model's response efficiency and adaptation accuracy to environmental changes and ensuring the long-term stable operation of the early warning model.

[0119] In this embodiment of the application, during the generation of multi-level refined early warning, the coupling between regional vulnerability characteristics and water depth adopts a hierarchical weighted mechanism, and the method further includes:

[0120] First, a weighted calculation of the base layer is constructed. Using the grid-level water depth prediction results output by S300 as the core indicator, the water depth is quantified into a score of 0-10 according to the gradient (slight, moderate, heavy, and extremely heavy) (e.g., extremely heavy water depth corresponds to 10 points), and then multiplied by 60% weight to obtain the base layer score, which reflects the physical risk level of the water itself.

[0121] Secondly, a weighted calculation for the enhancement layer is implemented. The enhancement layer includes two indicators: one is the duration of water accumulation, which is quantified into 0-10 points based on the predicted duration (e.g., <1 hour, 1-3 hours, >3 hours), with a weight of 15%; the other is the accessibility of surrounding emergency resources, which is quantified into 0-10 points based on the straight-line distance between drainage equipment, refuge sites and grids, and road traffic conditions (e.g., smooth, congested) (the closer the distance and the better the traffic, the higher the score), with a weight of 15%. The scores of the two indicators are added together and multiplied by 30% to obtain the enhancement layer score, which supplements the reflection of the time impact of risks and emergency support capabilities.

[0122] Next, a weighted calculation of the correction layer is performed. The deviation values ​​between the warning results and the actual impact of similar scenarios (such as similar water depths or areas) in the same historical period are extracted (such as the difference rate between the predicted water depth and the measured value). The deviation values ​​are standardized to -2 to 2 points (negative deviation indicates that the warning is too light, and positive deviation indicates that the warning is too heavy). The score is multiplied by 10% to obtain the correction layer score, which is used to calibrate the calculation deviation between the base layer and the enhancement layer.

[0123] Then, the final warning level is determined. The scores of the basic layer, the enhanced layer, and the correction layer are added together to obtain a comprehensive score, which is then mapped to the five-level warning system according to the score range (e.g., 0-2 points corresponds to Level 1, 2-4 points corresponds to Level 2, etc.). At the same time, an emergency response action list is bound to each warning level, and the content of the list is dynamically adjusted according to the functional attributes of the area—for example, the list for school areas focuses on "class suspension notices and evacuation routes for teachers and students," while the list for transportation hubs focuses on "traffic control and vehicle detour guidance."

[0124] By comprehensively considering the physical risks of water accumulation, temporal impact, emergency response capabilities, and historical biases through a tiered weighted mechanism, the accuracy of early warning levels has been improved. By dynamically adapting the emergency list to regional functional attributes, precise coordination between early warning and emergency actions has been achieved, enhancing the guiding value of early warning for actual flood control and disaster reduction work.

[0125] This application discloses an artificial intelligence-based urban flood disaster early warning system, referring to... Figure 2 ,include:

[0126] The multi-source data acquisition and spatiotemporal calibration module 001 collects meteorological sensing information, hydrological monitoring information, geospatial information, urban operation information, and social sensing information in real time to obtain corresponding real-time multi-source data. All real-time data are attached with precise spatiotemporal coordinates. Among them, meteorological sensing information includes rainfall content and precipitation forecast content, hydrological monitoring information includes water level data and pipeline flow data, geospatial information includes topography and drainage network distribution characteristics, urban operation information includes water accumulation status in flood-prone areas and drainage facility parameters, and social sensing information includes water accumulation-related content reported by the public.

[0127] The multi-source data preprocessing and feature screening module 002 performs outlier detection, missing value imputation, and standardization on real-time multi-source data in sequence, and then achieves spatiotemporal alignment through gridding processing to screen out key spatiotemporal feature data reflecting rainfall accumulation and confluence evolution.

[0128] The physical constraint spatiotemporal fusion early warning model construction module 003 constructs a physical constraint spatiotemporal fusion deep learning early warning model based on key spatiotemporal feature data. The early warning model integrates a spatial feature extraction module, a temporal evolution learning module, and a spatiotemporal gating mechanism. It transforms the discretized differential terms in the physical equations describing water flow motion into initialization rules and dynamic adjustment mechanisms for convolutional kernel weights. Combined with a multi-head attention mechanism, it achieves adaptive fusion of multi-source features to output the predicted water depth in future time periods.

[0129] The early warning model dynamic optimization and environment adaptation module 004 establishes a model dynamic optimization mechanism. Based on the incremental updates of historical multi-source data and real-time multi-source data, it adopts transfer learning to transfer the model knowledge obtained from historical training to the new scene for areas with environmental changes. It quickly adapts to environmental changes through feature space mapping and parameter fine-tuning. Among them, areas with environmental changes include changes in urban terrain and upgrades to drainage facilities.

[0130] The multi-level early warning generation and information output module 005 generates multi-level refined early warnings based on the water depth prediction results and integrates the impact range and regional vulnerability characteristics. Among them, the regional vulnerability characteristics include population distribution density, economic activity intensity and infrastructure importance. The corresponding early warning level is determined by coupling the water depth threshold and vulnerability level, and the spatial distribution map, time evolution trend and specific assessment information of the affected objects of the early warning area are output simultaneously.

[0131] This application also discloses a computer device, including a processor, wherein the processor runs a program for the artificial intelligence-based urban flood disaster early warning method described in any one of the above embodiments.

[0132] This application also discloses a storage medium storing a program for the artificial intelligence-based urban flood disaster early warning method described in any one of the above embodiments.

[0133] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An artificial intelligence-based method for early warning of urban flooding disasters, characterized in that, include: Real-time collection of meteorological sensing information, hydrological monitoring information, geospatial information, urban operation information, and social sensing information is performed to obtain corresponding real-time multi-source data. All real-time data are accompanied by precise spatiotemporal coordinates. Among them, meteorological sensing information includes rainfall content and precipitation forecast content, hydrological monitoring information includes water level data and pipeline flow data, geospatial information includes topography and drainage network distribution characteristics, urban operation information includes water accumulation status in flood-prone areas and drainage facility parameters, and social sensing information includes water accumulation-related content reported by the public. The real-time multi-source data is sequentially subjected to outlier detection, missing value imputation, and standardization. Then, spatiotemporal alignment is achieved through gridding to filter out key spatiotemporal feature data reflecting rainfall accumulation and confluence evolution. Based on the key spatiotemporal feature data, a spatiotemporal fusion deep learning early warning model with physical constraints is constructed. The early warning model integrates a spatial feature extraction module, a temporal evolution learning module, and a spatiotemporal gating mechanism. It transforms the discretized differential terms in the physical equation describing water flow motion into the initialization rules and dynamic adjustment mechanism of convolution kernel weights. Combined with a multi-head attention mechanism, it achieves adaptive fusion of multi-source features to output the prediction results of water depth in future time periods. A dynamic optimization mechanism for the model is established. Based on the incremental updates of historical multi-source data and real-time multi-source data, transfer learning is used to transfer the model knowledge obtained from historical training to the new scene for areas with environmental changes. Through feature space mapping and parameter fine-tuning, the model can quickly adapt to environmental changes. The areas with environmental changes include changes in urban terrain and upgrades to drainage facilities. Based on the water depth prediction results, the system integrates the impact range and regional vulnerability characteristics to generate multi-level refined early warnings. The regional vulnerability characteristics include population density, economic activity intensity, and infrastructure importance. The corresponding early warning level is determined by coupling the water depth threshold with the vulnerability level, and the system simultaneously outputs the spatial distribution map of the warning area, the temporal evolution trend, and specific assessment information of the affected objects.

2. The urban flood disaster early warning method based on artificial intelligence according to claim 1, characterized in that, Establishing a multi-source data spatiotemporal correlation mechanism during real-time data acquisition also includes the following methods: Dynamic timestamp calibration is performed on meteorological sensing information and hydrological monitoring information. The temporal resolution of meteorological data is adjusted according to the rainfall intensity level, and a high-frequency acquisition mode is used during periods of heavy rainfall. At the same time, drainage network nodes in geospatial information are used as reference points to establish spatial mapping relationships between meteorological monitoring points, hydrological sensing equipment and network nodes, so that various types of data form a mutually verifying network in the spatiotemporal dimension.

3. The urban flood disaster early warning method based on artificial intelligence according to claim 1, characterized in that, The method also includes: The gridding process employs a topology-aware dynamic partitioning strategy, using natural water catchment units in geospatial information as the basic framework. It combines urban road red lines and land parcel boundaries to correct the grid shape, so that each grid unit contains topological attributes such as catchment area, ground slope, drainage direction, and surrounding pipe network access points. During spatiotemporal alignment, data is aggregated within grid cells, and the water flow exchange coefficient between adjacent grid cells is recorded synchronously. The water flow exchange coefficient is dynamically calculated based on the connectivity of the pipeline network and the terrain elevation difference.

4. The urban flood disaster early warning method based on artificial intelligence according to claim 1, characterized in that, In constructing a spatiotemporal fusion deep learning-based early warning model with physical constraints, the methods also include: The physical equations describing water flow motion are discretized into differential terms and transformed into a dynamic adjustment mechanism for the convolution kernel weights. The coefficient ranges corresponding to the inertial, gravity, and drag terms in the equations are extracted and used as the constraint boundaries for the convolution kernel weights. The hydraulic gradient of the grid cell is calculated in real time. When the hydraulic gradient exceeds the set hydraulic gradient threshold, the emergency adjustment mode of the convolution kernel weight is triggered, which strengthens the rapid evolution feature of the water flow by amplifying the weight ratio of the gravity term. When a sudden change in water level is detected at a pipeline node, the local weight distribution of the convolution kernel is dynamically corrected based on the resistance term coefficient to capture the congestion effect after the water flow is obstructed.

5. The urban flood disaster early warning method based on artificial intelligence according to claim 1, characterized in that, The multi-head attention mechanism includes data type branching and spatial region branching, and the method further includes: The data type branch sets up independent attention heads for data from different sources. The query vector generation rules for each attention head are related to the spatiotemporal sensitivity of the data. The different sources of data include meteorological, hydrological, and geospatial data. The spatial region branch divides the attention scope according to urban functional zones. The key vector of each region's attention head incorporates the historical flood characteristics parameters of the corresponding region. Among them, urban functional zones include residential areas, commercial areas, and industrial areas. Multi-source features are adaptively fused in both type and spatial dimensions by cross-product calculation of attention weights in two branches.

6. The urban flood disaster early warning method based on artificial intelligence according to claim 1, characterized in that, In the model dynamic optimization mechanism, online learning and transfer learning form a collaborative triggering logic, and the methods also include: When the urban environment changes gradually, the basic parameters of the model are first updated based on three consecutive months of incremental data through online learning, and then transfer learning is initiated to fine-tune the feature extraction module of the changed area. When the environment changes abruptly, the core parameters related to global hydrological patterns in the model are first frozen. The physical constraint weight matrix of the changed area is reconstructed through transfer learning. Then, the output layer parameters of the model are optimized using new scenario data in conjunction with online learning.

7. The urban flood disaster early warning method based on artificial intelligence according to claim 1, characterized in that, In the process of generating multi-level refined early warnings, the coupling between regional vulnerability characteristics and water depth adopts a hierarchical weighted mechanism. The method also includes: The basic layer uses water depth as the core indicator, accounting for 60% of the weight; the enhancement layer introduces the duration of water accumulation and the accessibility of surrounding emergency resources, accounting for 30% of the weight; the correction layer combines the actual impact deviation value of similar historical warnings, accounting for 10% of the weight; among them, surrounding emergency resources include drainage equipment and refuge areas. The final warning level is determined by a three-level weighted calculation, and each warning level is bound to a corresponding list of emergency response actions. The list content is dynamically adapted to the regional functional attributes, which include school areas and transportation hubs.

8. An artificial intelligence-based urban flood disaster early warning system, characterized in that, include: The multi-source data acquisition and spatiotemporal calibration module collects meteorological sensing information, hydrological monitoring information, geospatial information, urban operation information, and social sensing information in real time to obtain corresponding real-time multi-source data. All real-time data are attached with precise spatiotemporal coordinates. Among them, meteorological sensing information includes rainfall content and precipitation forecast content, hydrological monitoring information includes water level data and pipeline flow data, geospatial information includes topography and drainage network distribution characteristics, urban operation information includes water accumulation status in flood-prone areas and drainage facility parameters, and social sensing information includes water accumulation-related content reported by the public. The multi-source data preprocessing and feature filtering module sequentially performs outlier detection, missing value imputation, and standardization on the real-time multi-source data, and then achieves spatiotemporal alignment through gridding processing to filter out key spatiotemporal feature data reflecting rainfall accumulation and confluence evolution. The physical constraint spatiotemporal fusion early warning model construction module constructs a physical constraint spatiotemporal fusion deep learning early warning model based on the key spatiotemporal feature data. The early warning model integrates a spatial feature extraction module, a temporal evolution learning module, and a spatiotemporal gating mechanism. It transforms the discretized differential terms in the physical equations describing water flow motion into initialization rules and dynamic adjustment mechanisms for convolutional kernel weights. Combined with a multi-head attention mechanism, it achieves adaptive fusion of multi-source features to output the predicted water depth in future time periods. The early warning model dynamic optimization and environment adaptation module establishes a model dynamic optimization mechanism. Based on the incremental updates of historical multi-source data and real-time multi-source data, it uses transfer learning to transfer the model knowledge obtained from historical training to new scenarios for areas with environmental changes. Through feature space mapping and parameter fine-tuning, it quickly adapts to environmental changes. The areas with environmental changes include changes in urban terrain and upgrades to drainage facilities. The multi-level early warning generation and information output module generates multi-level refined early warnings based on the water depth prediction results and integrates the impact range and regional vulnerability characteristics. Among them, the regional vulnerability characteristics include population distribution density, economic activity intensity and infrastructure importance. The corresponding early warning level is determined by coupling the water depth threshold and vulnerability level, and the spatial distribution map of the early warning area, the time evolution trend and the specific assessment information of the affected objects are output simultaneously.

9. A computer device, characterized in that, Includes a processor, wherein the processor runs a program for an artificial intelligence-based urban flood disaster early warning method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The program stores the AI-based urban flood disaster early warning method as described in any one of claims 1-7.

Images