A watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin

Abstract

This invention provides a watershed disaster chain early warning platform integrating dynamic coupling algorithms and digital twins, belonging to the field of disaster prediction technology. The platform includes a multi-dimensional data management module, a dynamic coupling algorithm module, a digital twin module, a comprehensive vulnerability assessment module, and a chain early warning module. Through the combined application of these modules, key risks in the disaster chain are identified, and the vulnerability levels of different regions are understood, making early warning reports more targeted and practical. This provides scientific guidance for relevant departments to carry out differentiated disaster prevention and emergency response work, effectively improving the overall efficiency of watershed disaster chain early warning and prevention.

Description

Technical Field

[0001] This invention relates to the field of disaster prediction technology, and more specifically, to a watershed disaster chain early warning platform that integrates dynamic coupling algorithms and digital twins. Background Technology

[0002] Watershed disaster early warning is a system platform engineering project that uses real-time monitoring of rainfall, water levels, and engineering conditions within a river basin, along with precise prediction and analysis using hydrological and meteorological models, to issue warning information to potentially affected areas before a disaster occurs. Watershed disaster early warning provides crucial decision-making basis for the precise scheduling of water conservancy projects and the scientific deployment of emergency resources, minimizing disaster losses to the greatest extent possible.

[0003] In related technologies, a single model is generally used for watershed disaster early warning. However, single models have problems such as lagging identification of risks in complex disaster chains, low model prediction accuracy, and disconnect from emergency response, making it difficult to achieve dynamic, accurate, and closed-loop prevention and control of watershed disasters. Summary of the Invention

[0004] The present invention aims to solve at least one of the above-mentioned problems.

[0005] To address the aforementioned issues, this invention provides a watershed disaster chain early warning platform that integrates a dynamic coupling algorithm and digital twin, comprising a multi-dimensional data management module, a dynamic coupling algorithm module, a digital twin module, a comprehensive vulnerability assessment module, and a chain early warning module;

[0006] The multidimensional data management module is used to store multi-source heterogeneous data;

[0007] The dynamic coupling algorithm module includes a forecasting layer, an analysis layer, and a scheduling layer. The forecasting layer is used to generate flood forecasts based on real-time rainfall or runoff data from the multi-source heterogeneous data, using a multi-head attention mechanism coupled with an LSTM model. The analysis layer is used to generate key risk point data of disaster chains based on historical disaster chain data from the multi-source heterogeneous data, using the Jaccard index and a Bayesian network. The scheduling layer is used to generate optimized reservoir scheduling schemes based on the flood forecasts, using a coupled particle swarm optimization algorithm.

[0008] The digital twin module includes a virtual inner-loop control unit and an actual outer-loop control unit. The virtual inner-loop control unit is used to generate inundation simulation data based on a one- or two-dimensional hydrodynamic coupling model and the reservoir optimization scheduling scheme. When the inundation simulation water depth in the inundation simulation data exceeds a preset water depth threshold, a scheduling re-optimization instruction is generated. The scheduling layer is also used to generate a new reservoir optimization scheduling scheme based on the scheduling re-optimization instruction. The actual outer-loop control unit is used to determine the similarity between the acquired on-site inundation image data and the inundation simulation data, and adjust the parameters of the one- or two-dimensional hydrodynamic coupling model based on the similarity.

[0009] The comprehensive vulnerability assessment module is used to determine the grid vulnerability based on the comprehensive vulnerability assessment model and according to the mobile phone signaling data and medical resource data in the multi-source heterogeneous data;

[0010] The chain-based early warning module is used to generate tiered early warning reports based on the key risk point data of the disaster chain and the vulnerability of the grid.

[0011] Optionally, the method for constructing the multidimensional data management module includes:

[0012] Acquire satellite remote sensing data, ground IoT sensor data, UAV oblique photography data, mobile phone signaling data, and socio-economic dynamic indicators to construct a multi-source heterogeneous data warehouse with a unified spatiotemporal benchmark;

[0013] Divide the data responsibility departments according to the water environment business chain and establish a data quality traceability list;

[0014] A lightweight algorithm with a quadratic error metric is used to compress BIM model data, and a database is compatible for storing heterogeneous data.

[0015] By performing protocol conversion through an edge computing gateway, multi-source data is encrypted and authenticated before being connected to the platform to build a multi-dimensional database.

[0016] Based on the multi-source heterogeneous data warehouse, the data quality traceability list, the heterogeneous data storage, and the multidimensional database, the multidimensional data management module is constructed. According to data attributes and business requirements, the multi-source heterogeneous data in the multidimensional data management module is classified into high-frequency real-time data, low-frequency business data, and historical disaster chain data.

[0017] Optionally, the method for constructing the forecast layer includes:

[0018] A multi-head attention mechanism is coupled with an LSTM to construct a coupled multi-head attention mechanism and an LSTM model, wherein the multi-head attention mechanism is used to extract the spatiotemporal features of the real-time rainfall or runoff data, and the LSTM is used to extract the stable long-sequence gradients of the spatiotemporal features.

[0019] The core parameters of the LSTM are searched using a Bayesian optimization algorithm to construct the initial prediction layer;

[0020] The historical real-time rainfall or historical runoff data in the multi-source heterogeneous data are globally normalized to obtain globally normalized data. The initial forecast layer is trained based on the globally normalized data and the Nash efficiency coefficient to generate the intermediate forecast layer.

[0021] The intermediate forecast layer uses a lightweight inference engine deployed at the edge of the AI ​​chip to update the input of the intermediate forecast layer in real time based on the second-level water level data of the edge nodes. Based on the attention weight matrix of the multi-head attention layer of the multi-head attention mechanism coupled with the LSTM model in the prediction process, the key period for the formation of the flood peak is determined, and the forecast layer is generated.

[0022] Optionally, the disaster chain key risk point data includes the overall risk value of the disaster chain and the critical path probability, and the analysis layer is specifically used for:

[0023] Based on the historical disaster chain data, a watershed disaster chain topology network is generated with water conservancy facilities as nodes and disaster transmission paths as edges;

[0024] The Jaccard index is used to screen nodes with high degree centrality and betweenness centrality in the watershed disaster chain topology network to determine the key nodes of the disaster chain;

[0025] The river siltation status identified by AI edge devices is mapped to Bayesian network evidence variables, and the risk probabilities of key nodes are dynamically updated based on the topology network of the watershed disaster chain, outputting the overall risk value of the disaster chain and the probability of the critical path.

[0026] Optionally, the scheduling layer is specifically used for:

[0027] Based on the flood hydrograph predicted by the flood forecast, and with maximizing the peak reduction rate as the optimization objective, an objective function for reservoir operation is established, which includes:

[0028] ;

[0029] Where maxF is the objective function value. For the peak inflow of floodwater, The outflow peak flow rate after scheduling;

[0030] Set flood control constraints, encode the discharge sequence into particle position vectors, and initialize the particle swarm. The flood control constraints are that the reservoir water level does not exceed the flood control limit water level and the water level fluctuation is less than the maximum water level fluctuation per unit time.

[0031] The objective function value of each particle is determined by the objective function, and the particle velocity and position are updated by dynamic inertial weights, which decrease linearly with the number of iterations.

[0032] The process is iteratively updated until the convergence condition is met, generating the optimized reservoir scheduling scheme.

[0033] Optionally, the dynamic coupling algorithm module further includes a feedback optimization layer;

[0034] The feedback optimization layer is used to construct a Gaussian process proxy model by taking the reservoir optimization scheduling scheme as the initial solution and minimizing the deviation between the actual discharge process and the reservoir optimization scheduling scheme.

[0035] Using the Gaussian process proxy model, based on real-time collected downstream water level data, the Nash efficiency coefficient of the actual discharge process line and the reservoir optimization scheduling scheme is determined. When the Nash efficiency coefficient is lower than the preset coefficient threshold, the Bayesian optimizer is triggered.

[0036] The Bayesian optimizer iteratively searches for hyperparameter combinations of the Coupled Particle Swarm Optimization (CPSO) algorithm using its acquisition function, and then re-runs the CPSO algorithm based on these hyperparameter combinations to generate a corrected scheduling scheme, which is then fed back to the digital twin module.

[0037] Optionally, the flood simulation data includes the flood simulation water depth, flood simulation flow velocity, and flood simulation range, and the virtual inner loop control unit is specifically used for:

[0038] Based on the discharge process line in the reservoir optimization scheduling scheme, a one-dimensional hydrodynamic model is constructed to simulate the flood evolution of the main river channel.

[0039] Based on the acquired high-precision DEM data, the floodplain grid was divided, and a two-dimensional hydrodynamic model was constructed to simulate the surface runoff process.

[0040] The one-dimensional hydrodynamic model and the two-dimensional hydrodynamic model are coupled by lateral connection to obtain the one-dimensional and two-dimensional hydrodynamic coupled model. When the water level at the cross-section of the main channel exceeds the crest elevation during the flood evolution, an intermediate flow is transferred to the adjacent two-dimensional grid through a momentum exchange formula. The two-dimensional hydrodynamic model generates the inundation simulation velocity and the inundation simulation range based on the intermediate flow. The momentum exchange formula includes:

[0041] ;

[0042] Among them, Q overflow The intermediate flow rate is given by g, where g is the physical acceleration constant and C is the intermediate flow rate. d Here, H is the overflow coefficient, L is the cross-sectional length, and H is the overflow coefficient. 1D Z represents the water level at the river cross-section.crest The elevation of the top of the dike;

[0043] The simulated inundation depth in the flood evolution of the main river channel is extracted. When the simulated inundation depth exceeds the preset depth threshold, the scheduling re-optimization instruction is generated. The simulated inundation depth, the simulated inundation velocity, and the simulated inundation range are used for risk assessment.

[0044] Optionally, the actual outer loop control unit is specifically used for:

[0045] The video surveillance equipment deployed is used to collect real-time on-site flooding image data of the downstream flooded area, and the target detection algorithm is used to identify water body boundaries and flooding markers in the on-site flooding image data to generate a raster map of the actual flooding range.

[0046] Spatial overlay analysis is performed between the actual inundation range raster map and the simulated inundation range raster map in the inundation simulation data to determine the similarity. When the similarity is less than the similarity threshold, a loss function is constructed based on the inundation depth deviation.

[0047] The loss function is used to adjust the Manning roughness coefficient of the one-dimensional hydrodynamic model and the overflow coefficient of the two-dimensional hydrodynamic model in reverse, and the process is iteratively updated until the loss function value is less than the preset loss value.

[0048] Optionally, the river siltation status includes a siltation index, and the process of mapping the river siltation status identified by the AI ​​edge device to a Bayesian network evidence variable, and dynamically updating the risk probability of key nodes based on the watershed disaster chain topology network, includes:

[0049] The AI ​​edge device is deployed at key river sections and embankment hazard points in the watershed disaster chain topology network. The AI ​​edge device integrates a target detection algorithm to identify the amount of floating debris accumulation in the river and the seepage points in the embankment.

[0050] The siltation index is determined based on the amount of floating debris in the river channel. When the siltation index exceeds the preset siltation threshold, the conditional probability update of the corresponding node in the Bayesian network is triggered. When the levee seepage point is found to exist, the levee seepage point is added to the watershed disaster chain topology network to generate a new watershed disaster chain topology network.

[0051] The risk probability of the key nodes is updated based on the new watershed disaster chain topology network.

[0052] Optionally, the comprehensive vulnerability assessment module is specifically used for:

[0053] Based on the mobile phone signaling data, a dynamic population density distribution heat map is generated using a spatiotemporal clustering algorithm to obtain the dynamic population density.

[0054] Based on the aforementioned medical resource data, a medical accessibility index is determined;

[0055] The dynamic population density and the medical accessibility index are input into the vulnerability assessment formula of the comprehensive vulnerability assessment model to determine the grid vulnerability. The vulnerability assessment formula includes:

[0056] ;

[0057] Among them, V i For the aforementioned mesh fragility, and The contribution of population density and medical accessibility to disaster sensitivity are respectively obtained through weighting coefficients determined by the entropy weight method. The dynamic population density, This represents the minimum dynamic population density. This represents the maximum dynamic population density. The medical accessibility index is... This represents the minimum value of the healthcare accessibility index. This represents the maximum value of the healthcare accessibility index.

[0058] The beneficial effects of the watershed disaster chain early warning platform integrating the dynamic coupling algorithm and digital twin of this invention are:

[0059] The multidimensional data management module stores multi-source heterogeneous data, providing comprehensive and rich data support for the operation of subsequent modules. This avoids analytical and decision-making biases caused by missing or isolated data, ensuring the integrity and reliability of the entire early warning platform's data foundation. Based on this, in the dynamic coupling algorithm module, the forecasting layer relies on a multi-head attention mechanism and an LSTM coupling model to generate flood forecasts using real-time rainfall or runoff data from multi-source heterogeneous data. This model can accurately extract the spatiotemporal characteristics of the data, effectively improving the timeliness and accuracy of flood forecasts and laying a precise predictive foundation for subsequent disaster risk analysis. The analysis layer, based on the Jaccard index and Bayesian networks, combines historical disaster chain data from multi-source heterogeneous data to generate key risk point data for the disaster chain. The Jaccard index can efficiently locate key nodes, and the Bayesian network can scientifically calculate risk probabilities, thereby accurately identifying key risk points in the disaster chain and clarifying the key directions for risk prevention and control. The scheduling layer uses a coupled particle swarm optimization algorithm to generate optimized reservoir scheduling schemes based on flood forecasts. This algorithm aims to reduce peak flow rates and optimize reservoir scheduling while meeting flood control constraints, minimizing the impact of floods on downstream areas. The digital twin module receives the reservoir optimization scheduling scheme output by the scheduling layer. The virtual inner-loop control unit generates inundation simulation data based on a one-dimensional hydrodynamic coupling model. When the simulated inundation depth exceeds a preset threshold, a scheduling re-optimization instruction is generated in a timely manner. The scheduling layer updates the reservoir optimization scheduling scheme accordingly, forming a dynamic adjustment mechanism that can quickly respond to changes in flood inundation risk and avoid the expansion of disaster losses due to delays in the scheduling scheme. The actual outer-loop control unit adjusts the parameters of the one-dimensional hydrodynamic coupling model by determining the similarity between the on-site inundation image data and the inundation simulation data, making the model simulation results more consistent with the actual situation and further improving the accuracy of inundation range prediction, providing more reliable simulation data for risk assessment. The comprehensive vulnerability assessment module, based on a comprehensive vulnerability assessment model, uses mobile phone signaling data and medical resource data from multi-source heterogeneous data to determine grid vulnerability. Mobile phone signaling data reflects dynamic population distribution, while medical resource data reflects regional medical security capabilities. The combination of the two makes the grid vulnerability assessment more consistent with the actual disaster impact scenario and accurately measures the degree of disaster vulnerability of different grid areas. Ultimately, the chain-based early warning module generates tiered early warning reports based on key risk point data and grid vulnerability data in the disaster chain. This not only clarifies the key risks in the disaster chain but also reveals the vulnerability levels of different areas, making the early warning reports more targeted and practical. It provides scientific guidance for relevant departments to carry out differentiated disaster prevention and emergency response work, effectively improving the overall effectiveness of chain-based early warning and prevention of watershed disasters. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the structure of a watershed disaster chain early warning platform that integrates dynamic coupling algorithms and digital twins. Detailed Implementation

[0061] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0062] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0063] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0064] It should be noted that the terms "one" and "more" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0065] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0066] To address the problems existing in the aforementioned related technologies, this embodiment provides a watershed disaster chain early warning platform that integrates dynamic coupling algorithm and digital twin.

[0067] like Figure 1 As shown in the figure, the watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin provided by the present invention includes a multi-dimensional data management module, a dynamic coupling algorithm module, a digital twin module, a comprehensive vulnerability assessment module, and a chain early warning module;

[0068] The multidimensional data management module is used to store multi-source heterogeneous data.

[0069] Specifically, the multi-dimensional data management module is used to store multi-source heterogeneous data, including data in the spatiotemporal, business, quality, and security dimensions. This provides comprehensive and rich data support for the operation of subsequent modules, avoids analysis and decision-making biases caused by missing or single data, and ensures the integrity and reliability of the data foundation of the entire early warning platform.

[0070] The dynamic coupling algorithm module includes a forecasting layer, an analysis layer, and a scheduling layer. The forecasting layer generates flood forecasts based on real-time rainfall or runoff data from the multi-source heterogeneous data, using a multi-head attention mechanism coupled with an LSTM model. The analysis layer generates critical risk point data of disaster chains based on historical disaster chain data from the multi-source heterogeneous data, using the Jaccard index and a Bayesian network. The scheduling layer generates optimized reservoir scheduling schemes based on the flood forecasts, using a coupled particle swarm optimization algorithm.

[0071] Specifically, the dynamic coupling algorithm module includes a forecasting layer, an analysis layer, and a scheduling layer. The forecasting layer uses a multi-head attention mechanism coupled with an LSTM model to generate flood forecasts based on real-time rainfall or runoff data from multi-source heterogeneous data. The multi-head attention mechanism extracts the spatiotemporal features of the real-time rainfall or runoff data, while the LSTM extracts stable long-sequence gradients. Historical real-time rainfall or runoff data can be used for training before forecasting. The analysis layer uses the Jaccard index and a Bayesian network to generate key risk point data for disaster chains based on historical disaster chain data from multi-source heterogeneous data. The Jaccard index (also known as intersection similarity) measures the degree of overlap between two sets, providing a basis for determining key risk point data in disaster chains by quantifying the similarity of data or networks. The Bayesian network performs the function of calculating and dynamically updating the risk probability of key nodes in the disaster chain, solving the problem that traditional static risk assessment cannot adapt to real-time changes in watershed conditions, and providing accurate risk quantification for subsequent scheduling optimization and graded early warning. The scheduling layer is used to generate optimal reservoir scheduling schemes based on flood forecasts using the Coupled Particle Swarm Optimization (CPSO) algorithm. The CPSO algorithm is an algorithm in which multiple particle swarms work together. It undertakes the key function of generating optimal reservoir scheduling schemes. By integrating the basin flood control constraints and the goal of maximizing peak reduction rate, it solves the problems of low accuracy and slow response of traditional experience-based scheduling schemes. At the same time, the linkage feedback optimization layer realizes the dynamic adjustment of hyperparameters, providing efficient scheduling support for the platform's "monitoring, simulation and decision-making" closed loop.

[0072] The digital twin module includes a virtual inner-loop control unit and an actual outer-loop control unit. The virtual inner-loop control unit is used to generate inundation simulation data based on a one- or two-dimensional hydrodynamic coupling model and the reservoir optimization scheduling scheme. When the inundation simulation water depth in the inundation simulation data exceeds a preset water depth threshold, a scheduling re-optimization instruction is generated. The scheduling layer is also used to generate a new reservoir optimization scheduling scheme based on the scheduling re-optimization instruction. The actual outer-loop control unit is used to determine the similarity between the acquired on-site inundation image data and the inundation simulation data, and adjust the parameters of the one- or two-dimensional hydrodynamic coupling model based on the similarity.

[0073] Specifically, the digital twin module combines real-world conditions with a virtual model to achieve a two-layer architecture of "virtual inner-loop control + actual outer-loop control." This enables closed-loop management of "simulation prediction, risk assessment, parameter correction, and scheduling optimization," addressing the problems of "disconnect between virtual and real systems, large simulation deviations, and delayed response" in traditional early warning platforms. It provides real-time verification and dynamic adjustment support for the scheduling scheme output by the dynamic coupling algorithm module. Its virtual inner-loop control unit, based on a one- or two-dimensional hydrodynamic coupling model, generates inundation simulation data according to the reservoir's optimized scheduling scheme. When the simulated inundation depth exceeds a preset threshold, a scheduling re-optimization instruction is generated, resolving the issue of "scheduling schemes not considering actual inundation risks." Its actual outer-loop control unit determines the similarity between acquired on-site inundation image data and inundation simulation data. Based on this similarity, it adjusts the parameters of the one- or two-dimensional hydrodynamic coupling model to address the problem of "static model parameters leading to large simulation deviations," thereby improving model accuracy.

[0074] The comprehensive vulnerability assessment module is used to determine the grid vulnerability based on the comprehensive vulnerability assessment model and the mobile signaling data and medical resource data in the multi-source heterogeneous data.

[0075] Specifically, the comprehensive vulnerability assessment module is used to determine grid vulnerability based on the comprehensive vulnerability assessment model and the mobile phone signaling data and medical resource data in the multi-source heterogeneous data. Grid vulnerability is a quantitative assessment of the vulnerability of socio-economic platforms in the basin to disaster threats. This allows for the development of differentiated early warning strategies for areas with different levels of vulnerability, such as increasing evacuation efforts and allocating more emergency resources in high-vulnerability areas in advance. This significantly improves the accuracy and effectiveness of early warning and provides strong support for basin disaster prevention and control.

[0076] The chain-based early warning module is used to generate tiered early warning reports based on the key risk point data of the disaster chain and the vulnerability of the grid.

[0077] Specifically, the chain-based early warning module generates tiered early warning reports based on key risk point data and grid vulnerability in the disaster chain. For example, when the product of key risk point data and grid vulnerability exceeds a preset early warning threshold, a tiered early warning report is generated and can be linked to the downstream response platform. The tiered early warning report refers to the different emergency response levels corresponding to disasters along different paths. The specific level can be determined according to the tiering mechanism set according to the actual situation. For example, risk level 1: matching plan A (extreme risk), the dispatching layer action is to open all floodgates + emergency evacuation of downstream residents; risk level 2: matching plan B (high risk), the dispatching layer action is to open 50% of the floodgates + pre-deployment of medical teams; risk level 3: matching plan C (medium risk), the dispatching layer action is to open 30% of the floodgates + reinforcement of dikes. The tiering mechanism in this embodiment is only an example, and the specific mechanism needs to be determined according to the actual situation, and is not limited here.

[0078] In this embodiment, the multidimensional data management module stores multi-source heterogeneous data, providing comprehensive and rich data support for the operation of subsequent modules. This avoids analysis and decision-making biases caused by missing or single data, and ensures the integrity and reliability of the data foundation of the entire early warning platform. Based on this, in the dynamic coupling algorithm module, the forecasting layer relies on a multi-head attention mechanism and an LSTM coupling model to generate flood forecasts using real-time rainfall or runoff data from multi-source heterogeneous data. This model can accurately extract the spatiotemporal features of the data, effectively improving the timeliness and accuracy of flood forecasts and laying a precise predictive foundation for subsequent disaster risk analysis. The analysis layer, based on the Jaccard index and Bayesian networks, combines historical disaster chain data from multi-source heterogeneous data to generate key risk point data of the disaster chain. The Jaccard index can efficiently locate key nodes, and the Bayesian network can scientifically calculate risk probabilities, thereby accurately identifying key risk points in the disaster chain and clarifying the key directions for risk prevention and control. The scheduling layer uses a coupled particle swarm optimization algorithm to generate optimized reservoir scheduling schemes based on flood forecasts. This algorithm can target peak reduction rate to optimize reservoir scheduling under flood control constraints, minimizing the impact of floods on downstream areas. The digital twin module receives the reservoir optimization scheduling scheme output by the scheduling layer. The virtual inner-loop control unit generates inundation simulation data based on a one-dimensional hydrodynamic coupling model. When the simulated inundation depth exceeds a preset threshold, a scheduling re-optimization instruction is generated in a timely manner. The scheduling layer updates the reservoir optimization scheduling scheme accordingly, forming a dynamic adjustment mechanism that can quickly respond to changes in flood inundation risk and avoid the expansion of disaster losses due to delays in the scheduling scheme. The actual outer-loop control unit adjusts the parameters of the one-dimensional hydrodynamic coupling model by determining the similarity between the on-site inundation image data and the inundation simulation data, making the model simulation results more consistent with the actual situation and further improving the accuracy of inundation range prediction, providing more reliable simulation data for risk assessment. The comprehensive vulnerability assessment module, based on a comprehensive vulnerability assessment model, uses mobile phone signaling data and medical resource data from multi-source heterogeneous data to determine grid vulnerability. Mobile phone signaling data reflects dynamic population distribution, while medical resource data reflects regional medical security capabilities. The combination of the two makes the grid vulnerability assessment more consistent with the actual disaster impact scenario and accurately measures the degree of disaster vulnerability of different grid areas. Ultimately, the chain-based early warning module generates tiered early warning reports based on key risk point data and grid vulnerability data in the disaster chain. This not only clarifies the key risks in the disaster chain but also reveals the vulnerability levels of different areas, making the early warning reports more targeted and practical. It provides scientific guidance for relevant departments to carry out differentiated disaster prevention and emergency response work, effectively improving the overall effectiveness of chain-based early warning and prevention of watershed disasters.

[0079] Optionally, the method for constructing the multidimensional data management module includes:

[0080] Acquire satellite remote sensing data, ground IoT sensor data, UAV oblique photography data, mobile phone signaling data, and socio-economic dynamic indicators to construct a multi-source heterogeneous data warehouse with a unified spatiotemporal benchmark;

[0081] Divide the data responsibility departments according to the water environment business chain and establish a data quality traceability list;

[0082] A lightweight algorithm with a quadratic error metric is used to compress BIM model data, and a database is compatible for storing heterogeneous data.

[0083] By performing protocol conversion through an edge computing gateway, multi-source data is encrypted and authenticated before being connected to the platform to build a multi-dimensional database.

[0084] Based on the multi-source heterogeneous data warehouse, the data quality traceability list, the heterogeneous data storage, and the multidimensional database, the multidimensional data management module is constructed. According to data attributes and business requirements, the multi-source heterogeneous data in the multidimensional data management module is classified into high-frequency real-time data, low-frequency business data, and historical disaster chain data.

[0085] Specifically, data is acquired from satellite remote sensing, ground-based IoT sensor data, UAV oblique photography data, mobile phone signaling data, and socio-economic dynamic indicators. A unified WGS84 coordinate system spatial benchmark is established, and Gauss-Kruger projection is used to synchronize ground-based IoT sensors (water level gauges, rain gauges) and mobile phone signaling base stations at the millisecond level via the NTP protocol. Timestamps are unified in UTC format, and a unified time benchmark is established to construct a multi-source heterogeneous data warehouse. Data responsibility departments are assigned according to the "source-network-plant-river" water environment business chain. For example, "source" data is the responsibility of the local water resources bureau's reservoir management section, "network" data is the responsibility of the water group's pipeline department, "plant" data is the responsibility of the ecological environment bureau's pollution control section, and "river" data is the responsibility of the water resources bureau's river management department. Establish a data quality traceability list, clearly defining the responsible person for data collection, the model of the collection equipment, the calibration cycle, and the abnormal data handling process for each type of data. This records the collection equipment information, calibration cycle, and abnormal data handling process for each type of data. When data exceeds a reasonable range, an automatic manual verification mechanism is triggered to ensure data reliability and facilitate subsequent traceability. For example, when a verification rule triggers an alarm (such as a sudden water level exceeding a threshold), the responsible party is notified to conduct on-site verification. The source of the data problem is traced back using the identification code for accountability. For instance, HYDRO_5021_202308201200 can locate the data collection equipment of a hydrological station at 12:00 on August 20, 2023. A lightweight algorithm based on a secondary error metric is used to compress the BIM model within the watershed. After compression, the model file size is reduced, decreasing storage load. Spatial data can be stored using a PostgreSQL database (compatible with PostGIS spatial extensions), and non-spatial data can be stored using a MySQL database, achieving heterogeneous data classification and storage. An edge computing gateway is deployed, and multi-source data is encrypted and authenticated before being connected to the platform through protocols such as Modbus-RTU and MQTT. This constructs a multi-dimensional database, enabling secure edge access and reducing transmission latency for high-frequency data storage at the edge, meeting the requirement of second-level response. Based on the aforementioned multi-source heterogeneous data warehouse, data quality traceability list, heterogeneous data storage, and multi-dimensional database, a multi-dimensional data management module is constructed. Data is divided into three categories according to data attributes and business needs: high-frequency real-time data (second-level water level data) is stored on edge nodes, low-frequency business data (daily water consumption data) is stored in the central database, and historical disaster chain data is stored in a distributed graph database. The high-frequency real-time data (second-level water level) includes a spatiotemporal dimension (spatial data with timestamps) and a security dimension (encrypted storage on edge nodes). Its technical role is to support real-time input at the flood forecasting layer, meeting the requirement of minute-level early warning timeliness.Low-frequency business data (daily water consumption) is categorized into two dimensions: business dimension (attributed to water companies according to the "source-network-plant-river" hierarchy) and quality dimension (regular verification by the central database). Its technical role is to provide water consumption constraints at the scheduling layer, ensuring that reservoir scheduling plans conform to actual business rules. Historical disaster chain data (flood evolution paths) is categorized into two dimensions: spatiotemporal dimension (historical disaster spatiotemporal paths) and quality dimension (distributed storage integrity verification). Its technical role is to construct an analysis layer topology network to achieve dynamic positioning of key nodes in the disaster chain. The criteria for determining the three types of data are as follows: high-frequency real-time data has an update rate greater than or equal to 1 time / second, determined by the water level sensor acquisition frequency (e.g., IoT devices); low-frequency business data has an update rate of 1 time / day, determined by the business report generation cycle (e.g., water consumption statistics); and historical disaster chain data has an update rate less than or equal to 1 time / month or no update, determined by the static archiving characteristics of historical events.

[0086] Optionally, the method for constructing the forecast layer includes:

[0087] A multi-head attention mechanism is coupled with an LSTM to construct a coupled multi-head attention mechanism and an LSTM model, wherein the multi-head attention mechanism is used to extract the spatiotemporal features of the real-time rainfall or runoff data, and the LSTM is used to extract the stable long-sequence gradients of the spatiotemporal features.

[0088] The core parameters of the LSTM are searched using a Bayesian optimization algorithm to construct the initial prediction layer;

[0089] The historical real-time rainfall or historical runoff data in the multi-source heterogeneous data are globally normalized to obtain globally normalized data. The initial forecast layer is trained based on the globally normalized data and the Nash efficiency coefficient to generate the intermediate forecast layer.

[0090] The intermediate forecast layer uses a lightweight inference engine deployed at the edge of the AI ​​chip to update the input of the intermediate forecast layer in real time based on the second-level water level data of the edge nodes. Based on the attention weight matrix of the multi-head attention layer of the multi-head attention mechanism coupled with the LSTM model in the prediction process, the key period for the formation of the flood peak is determined, and the forecast layer is generated.

[0091] Specifically, a multi-head attention mechanism is coupled with LSTM to construct the forecast layer. The number of multi-head attention heads is set to extract spatiotemporal features from real-time rainfall or runoff data. Stable long-sequence gradients of these spatiotemporal features are extracted using LSTM. These stable long-sequence gradients are optimized through the multi-head attention mechanism, addressing the shortcomings of traditional flood forecasting models in processing long-term time-series data. Furthermore, a Bayesian optimization algorithm is used to search for the core parameters of the LSTM (such as the number of hidden layer neurons, learning rate, and regularization parameters) to solve the vanishing gradient problem and improve the model's ability to capture spatiotemporal features. Historical real-time rainfall and runoff data are selected, globally normalized, and then input into the initial forecast layer for training. The Nash efficiency coefficient is used as the accuracy evaluation index, and the model is continuously iterated and optimized until the Nash efficiency coefficient stably reaches the preset accuracy standard, generating an intermediate forecast layer. By deploying a lightweight inference engine on an edge computing gateway using an AI chip, real-time updates of water level data down to the second level are supported for edge deployment. This means that the model input is updated in real-time based on high-frequency real-time data (such as second-level water level data) from edge nodes, outputting runoff forecast results for a specified future period. Furthermore, the AI ​​chip reduces inference latency. By presenting stable long-sequence gradients and visualization tools to represent multi-head attention weights, the critical periods for flood peak formation can be identified, providing an intuitive basis for scheduling decisions.

[0092] Optionally, the disaster chain key risk point data includes the overall risk value of the disaster chain and the critical path probability, and the analysis layer is specifically used for:

[0093] Based on the historical disaster chain data, a watershed disaster chain topology network is generated with water conservancy facilities as nodes and disaster transmission paths as edges;

[0094] The Jaccard index is used to screen nodes with high degree centrality and betweenness centrality in the watershed disaster chain topology network to determine the key nodes of the disaster chain.

[0095] The river siltation status identified by AI edge devices is mapped to Bayesian network evidence variables, and the risk probabilities of key nodes are dynamically updated based on the topology network of the watershed disaster chain, outputting the overall risk value of the disaster chain and the probability of the critical path.

[0096] Specifically, based on historical disaster chain data, a watershed disaster chain topology network is constructed using existing software. Water conservancy facilities are used as nodes, and disaster transmission paths are used as edges. The weight of each edge is set to the probability of that path occurring in historical disasters, clearly presenting the disaster transmission relationship. Node similarity is calculated using the Jaccard index, and nodes with both high degree centrality and betweenness centrality are selected to identify key nodes in the disaster chain and clarify the key targets for disaster prevention and control. AI edge devices are deployed at key river sections and levee hazard points, integrating target detection algorithms to periodically collect video stream data and identify river siltation and levee seepage status. The identification results are mapped to Bayesian network evidence variables, dynamically updating the risk probabilities of key nodes, and outputting the overall disaster chain risk value and critical path probability, providing a quantitative basis for risk prevention and control. The critical path probability identifies the main disaster transmission chain and can be determined through topological network transmission. The overall disaster chain risk value assesses the platform-wide risk level of the entire watershed and can be determined based on the joint probability of all critical paths.

[0097] Optionally, the scheduling layer is specifically used for:

[0098] Based on the flood hydrograph predicted by the flood forecast, and with maximizing the peak reduction rate as the optimization objective, an objective function for reservoir operation is established, which includes:

[0099] ;

[0100] Where maxF is the objective function value. For the peak inflow of floodwater, The outflow peak flow rate after scheduling;

[0101] Set flood control constraints, encode the discharge sequence into particle position vectors, and initialize the particle swarm. The flood control constraints are that the reservoir water level does not exceed the flood control limit water level and the water level fluctuation is less than the maximum water level fluctuation per unit time.

[0102] The objective function value of each particle is determined by the objective function, and the particle velocity and position are updated by dynamic inertial weights, which decrease linearly with the number of iterations.

[0103] The process is iteratively updated until the convergence condition is met, generating the optimized reservoir scheduling scheme.

[0104] Specifically, based on the flood hydrograph output by the forecast layer, a reservoir scheduling objective function is established with maximizing the peak reduction rate as the optimization objective. Simultaneously, flood control constraints are set (such as the reservoir water level not exceeding the flood control limit level and the water level fluctuation not exceeding the maximum allowable value per unit time) to ensure the safety and feasibility of the scheduling scheme. The discharge sequence is encoded as a particle position vector, i.e. , where X iLet Q be a set of vectors. T For the position of the T-th particle, initialize the particle swarm size and iteration count, and set dynamic inertia weights (decreasing linearly with the iteration count), learning factors, and random number ranges. Then, determine the objective function value, i.e., the particle fitness value, for each particle using the objective function. Iterate through the particle velocity and position update formula (existing PSO formula) to search for the optimal solution. When the iteration meets the convergence condition (e.g., the change in the swarm's optimal fitness value is less than a preset threshold after multiple consecutive iterations), output the global optimal solution as the reservoir optimization scheduling scheme, achieving scientific and refined reservoir scheduling. Compared to traditional static optimization, the coupled particle swarm algorithm balances global and local search capabilities through dynamic inertia weights, improving peak reduction efficiency. Simultaneously, the rigid constraints of flood control prevent the scheduling scheme from violating engineering safety rules, ensuring the feasibility of the scheme.

[0105] Optionally, such as Figure 1 As shown, the dynamic coupling algorithm module also includes a feedback optimization layer;

[0106] The feedback optimization layer is used to construct a Gaussian process proxy model by taking the reservoir optimization scheduling scheme as the initial solution and minimizing the deviation between the actual discharge process and the reservoir optimization scheduling scheme.

[0107] Using the Gaussian process proxy model, based on real-time collected downstream water level data, the Nash efficiency coefficient of the actual discharge process line and the reservoir optimization scheduling scheme is determined. When the Nash efficiency coefficient is lower than the preset coefficient threshold, the Bayesian optimizer is triggered.

[0108] The Bayesian optimizer iteratively searches for hyperparameter combinations of the Coupled Particle Swarm Optimization (CPSO) algorithm using its acquisition function, and then re-runs the CPSO algorithm based on these hyperparameter combinations to generate a corrected scheduling scheme, which is then fed back to the digital twin module.

[0109] Specifically, the optimized reservoir scheduling scheme output by the scheduling layer is used as the initial solution. With the objective of minimizing the deviation between the actual discharge process and the optimized scheme, a Gaussian process surrogate model is constructed to fit the relationship between the deviation and hyperparameters. Based on downstream water level data collected in real time from edge nodes, the Nash efficiency coefficient (NSE) of the actual discharge process line and the optimized scheme is calculated. When the coefficient is lower than a preset threshold, for example, below 0.8, the Bayesian optimizer is triggered, initiating the scheme correction process. The hyperparameter combination of the particle swarm optimization algorithm is iteratively searched through the Bayesian optimizer's acquisition function. The updated hyperparameters are re-input into the scheduling layer, the coupled particle swarm optimization algorithm is run, and a corrected scheduling scheme is generated and fed back to the digital twin module. This achieves dynamic optimization of the scheduling scheme, solving the problem of low efficiency in traditional trial-and-error methods and reducing scheduling scheme deviation. Simultaneously, the NSE triggering mechanism ensures that optimization is initiated only when necessary, reducing computational resource waste. The Gaussian process surrogate model abstracts the actual particle swarm calculation into a differentiable mathematical function, while the Bayesian optimizer is a probability-guided intelligent search mechanism. Together, they address the problems of long computation time and blind parameter optimization in traditional methods, supporting the overall technical solution and achieving minute-level scheduling correction. The Gaussian process surrogate model takes historical hyperparameter combinations as input and outputs a bias prediction function and its confidence interval. Its purpose is to replace computationally expensive particle swarm optimization simulations, abstracting actual optimization biases into differentiable mathematical functions. The Bayesian optimizer takes bias observations as input and outputs hyperparameter optimization suggestions. Its purpose is to explore and utilize the acquisition function balance to search for hyperparameter combinations that minimize the bias.

[0110] Optionally, the flood simulation data includes the flood simulation water depth, flood simulation flow velocity, and flood simulation range, and the virtual inner loop control unit is specifically used for:

[0111] Based on the discharge process line in the reservoir optimization scheduling scheme, a one-dimensional hydrodynamic model is constructed to simulate the flood evolution of the main river channel.

[0112] Based on the acquired high-precision DEM data, the floodplain grid was divided, and a two-dimensional hydrodynamic model was constructed to simulate the surface runoff process.

[0113] The one-dimensional hydrodynamic model and the two-dimensional hydrodynamic model are coupled by lateral connection to obtain the one-dimensional and two-dimensional hydrodynamic coupled model. When the water level at the cross-section of the main channel exceeds the crest elevation during the flood evolution, an intermediate flow is transferred to the adjacent two-dimensional grid through a momentum exchange formula. The two-dimensional hydrodynamic model generates the inundation simulation velocity and the inundation simulation range based on the intermediate flow. The momentum exchange formula includes:

[0114] ;

[0115] Among them, Q overflow The intermediate flow rate is given by g, where g is the physical acceleration constant and C is the intermediate flow rate.d Here, H is the overflow coefficient, L is the cross-sectional length, and H is the overflow coefficient. 1D Z represents the water level at the river cross-section. crest The elevation of the top of the dike;

[0116] The simulated inundation depth in the flood evolution of the main river channel is extracted. When the simulated inundation depth exceeds the preset depth threshold, the scheduling re-optimization instruction is generated. The simulated inundation depth, the simulated inundation velocity, and the simulated inundation range are used for risk assessment.

[0117] Specifically, based on the discharge process line output by the scheduling layer, a one-dimensional hydrodynamic model of the main channel is constructed using professional hydrodynamic simulation software. This model only considers one-dimensional spatial variations along the flow direction, sets channel cross-sectional parameters and the Manning roughness coefficient, simulates the flood evolution process of the main channel, and outputs water level and velocity data for each cross-section. A two-dimensional hydrodynamic model is constructed by dividing the floodplain into grids based on high-precision DEM data. This model considers two-dimensional spatial variations in the plane, sets an overflow coefficient, and simulates the surface overflow process of floodwater in the floodplain, outputting inundation range and inundation depth data. The one-dimensional and two-dimensional models are coupled via lateral connections. When the water level at the one-dimensional model channel cross-section exceeds the levee crest elevation, the flow rate is transferred to the adjacent two-dimensional grid using a momentum exchange formula. When the maximum inundation depth output by the coupled model exceeds a preset threshold, such as 3m, a scheduling re-optimization instruction is generated. The scheduling layer re-runs the particle swarm optimization algorithm, adjusts the discharge sequence, and optimizes the scheduling scheme. This allows the one-dimensional coupled model to solve the problem of traditional single models failing to accurately simulate overflow, reducing the prediction error of the inundation range. Simultaneously, the water depth threshold triggering mechanism enables dynamic closed-loop control of risk, improving response speed. The inundation simulation flow velocity is a grid velocity vector output by the two-dimensional hydrodynamic model, used to assess the risk of hydraulic impact. For example, the one-dimensional and two-dimensional hydrodynamic models can be obtained by setting the above parameters using existing models.

[0118] Optionally, the actual outer loop control unit is specifically used for:

[0119] The video surveillance equipment deployed is used to collect real-time on-site flooding image data of the downstream flooded area, and the target detection algorithm is used to identify water body boundaries and flooding markers in the on-site flooding image data to generate a raster map of the actual flooding range.

[0120] Spatial overlay analysis is performed between the actual inundation range raster map and the simulated inundation range raster map in the inundation simulation data to determine the similarity. When the similarity is less than the similarity threshold, a loss function is constructed based on the inundation depth deviation.

[0121] The loss function is used to adjust the Manning roughness coefficient of the one-dimensional hydrodynamic model and the overflow coefficient of the two-dimensional hydrodynamic model in reverse, and the process is iteratively updated until the loss function value is less than the preset loss value.

[0122] Specifically, video surveillance equipment is deployed in the downstream inundation area to collect real-time inundation image data. Object detection algorithms are used to identify water boundaries and inundation markers in the images, generating a raster map of the actual inundation area. The simulated inundation area raster map and the actual inundation area raster map are then spatially overlaid to calculate the similarity, i.e., the Jaccard similarity coefficient.

[0123] ;

[0124] Among them, S sim To simulate the inundation range raster map, S obs This is a raster map showing the actual flooding area.

[0125] When the similarity coefficient is less than the similarity threshold, a loss function is constructed based on the submerged water depth deviation. The Manning roughness coefficient of the one-dimensional model and the overflow coefficient of the two-dimensional model are then adjusted in reverse using the gradient descent method. This process is iteratively updated until the loss function value is less than a preset value. The corrected parameters are then fed back to the module to update the hydrodynamic coupling model parameter library, thereby quantitatively evaluating the model accuracy. This method is more efficient than traditional visual inspection, and the dynamic parameter correction ensures the model continuously approximates reality, improving long-term prediction stability. The similarity threshold is determined experimentally based on historical data.

[0126] Optionally, the river siltation status includes a siltation index, and the process of mapping the river siltation status identified by the AI ​​edge device to a Bayesian network evidence variable, and dynamically updating the risk probability of key nodes based on the watershed disaster chain topology network, includes:

[0127] The AI ​​edge device is deployed at key river sections and embankment hazard points in the watershed disaster chain topology network. The AI ​​edge device integrates a target detection algorithm to identify the amount of floating debris accumulation in the river and the seepage points in the embankment.

[0128] The siltation index is determined based on the amount of floating debris in the river channel. When the siltation index exceeds the preset siltation threshold, the conditional probability update of the corresponding node in the Bayesian network is triggered. When the levee seepage point is found to exist, the levee seepage point is added to the watershed disaster chain topology network to generate a new watershed disaster chain topology network.

[0129] The risk probability of the key nodes is updated based on the new watershed disaster chain topology network.

[0130] Specifically, AI edge devices are deployed at key sections of the watershed disaster chain topology network and at potential levee locations. These AI edge devices are data acquisition devices that integrate target detection algorithms, such as impact monitoring equipment. Every 5 minutes, the AI ​​edge devices collect video stream data to identify river siltation and levee seepage points. During their use, the identification results can be transmitted to the analysis layer via the MQTT protocol. When the siltation index exceeds a preset siltation threshold, the conditional probabilities of the Bayesian network nodes are updated. When a seepage point is detected, a new topology network path is added, and the levee seepage point is added to the watershed disaster chain topology network, generating a new watershed disaster chain topology network. After the update, the risk probability of key nodes is recalculated, and initial screening is performed using degree centrality. The top 3 high-risk nodes are pushed to the scheduling layer based on degree centrality, and betweenness centrality is calculated asynchronously in the background for secondary verification. If the betweenness centrality of the top 3 high-risk nodes does not rank in the top 5 overall, they are replaced with nodes ranked by a combination of dual indicators. Degree centrality enables rapid initial screening with minute-level response, while betweenness centrality background verification ensures that no key nodes are missed. This enables AI edge devices to monitor river conditions at the minute level, improving efficiency compared to manual inspections. Meanwhile, dynamic updates to the topology network improve the timeliness of disaster chain path prediction, avoiding the omission of key risk nodes.

[0131] Optionally, the comprehensive vulnerability assessment module is specifically used for:

[0132] Based on the mobile phone signaling data, a dynamic population density distribution heat map is generated using a spatiotemporal clustering algorithm to obtain the dynamic population density.

[0133] Based on the aforementioned medical resource data, a medical accessibility index is determined;

[0134] The dynamic population density and the medical accessibility index are input into the vulnerability assessment formula of the comprehensive vulnerability assessment model to determine the grid vulnerability. The vulnerability assessment formula includes:

[0135] ;

[0136] Among them, V i For the aforementioned mesh fragility, and The contribution of population density and medical accessibility to disaster sensitivity are respectively obtained through weighting coefficients determined by the entropy weight method. The dynamic population density, This represents the minimum dynamic population density. This represents the maximum dynamic population density. The medical accessibility index is... This represents the minimum value of the healthcare accessibility index. This represents the maximum value of the healthcare accessibility index.

[0137] Specifically, mobile signaling acquisition gateways deployed at edge nodes acquire base station location data and generate dynamic population density distribution heatmaps using spatiotemporal clustering algorithms to obtain the dynamic population density of each grid. By connecting to the health department's API interface, spatial data of medical resources is obtained, and a network analysis method is used to calculate the medical accessibility index, quantifying the regional medical security capacity. The dynamic population density and medical accessibility index are input into a comprehensive vulnerability assessment model, and the entropy weight method is used to calculate their contribution to disaster sensitivity (weight coefficients). The vulnerability of each grid is then calculated using a vulnerability assessment formula, quantifying the degree of disaster vulnerability in different grid areas and providing a basis for differentiated disaster prevention and control.

[0138] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0139] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin, characterized in that, It includes a multi-dimensional data management module, a dynamic coupling algorithm module, a digital twin module, a comprehensive vulnerability assessment module, and a chain-based early warning module; The multidimensional data management module is used to store multi-source heterogeneous data; The dynamic coupling algorithm module includes a forecasting layer, an analysis layer, and a scheduling layer. The forecasting layer is used to generate flood forecasts based on real-time rainfall or runoff data from the multi-source heterogeneous data, using a multi-head attention mechanism coupled with an LSTM model. The analysis layer is used to generate key risk point data of disaster chains based on historical disaster chain data from the multi-source heterogeneous data, using the Jaccard index and a Bayesian network. The scheduling layer is used to generate optimized reservoir scheduling schemes based on the flood forecasts, using a coupled particle swarm optimization algorithm. The digital twin module includes a virtual inner-loop control unit and an actual outer-loop control unit. The virtual inner-loop control unit is used to generate inundation simulation data based on a one- or two-dimensional hydrodynamic coupling model and the reservoir optimization scheduling scheme. When the inundation simulation water depth in the inundation simulation data exceeds a preset water depth threshold, a scheduling re-optimization instruction is generated. The scheduling layer is also used to generate a new reservoir optimization scheduling scheme based on the scheduling re-optimization instruction. The actual outer-loop control unit is used to determine the similarity between the acquired on-site inundation image data and the inundation simulation data, and adjust the parameters of the one- or two-dimensional hydrodynamic coupling model based on the similarity. The comprehensive vulnerability assessment module is used to determine the grid vulnerability based on the comprehensive vulnerability assessment model and according to the mobile phone signaling data and medical resource data in the multi-source heterogeneous data; The chain-based early warning module is used to generate tiered early warning reports based on the key risk point data of the disaster chain and the vulnerability of the grid.

2. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 1, characterized in that, The construction method of the multidimensional data management module includes: Acquire satellite remote sensing data, ground IoT sensor data, UAV oblique photography data, mobile phone signaling data, and socio-economic dynamic indicators to construct a multi-source heterogeneous data warehouse with a unified spatiotemporal benchmark; Divide the data responsibility departments according to the water environment business chain and establish a data quality traceability list; A lightweight algorithm with a quadratic error metric is used to compress BIM model data, and a database is compatible for storing heterogeneous data. By performing protocol conversion through an edge computing gateway, multi-source data is encrypted and authenticated before being connected to the platform to build a multi-dimensional database. Based on the multi-source heterogeneous data warehouse, the data quality traceability list, the heterogeneous data storage, and the multidimensional database, the multidimensional data management module is constructed, and according to data attributes and business requirements, the multi-source heterogeneous data in the multidimensional data management module is classified into high-frequency real-time data, low-frequency business data, and the historical disaster chain data.

3. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 1, characterized in that, The method for constructing the forecast layer includes: A multi-head attention mechanism is coupled with an LSTM to construct a coupled multi-head attention mechanism and an LSTM model, wherein the multi-head attention mechanism is used to extract the spatiotemporal features of the real-time rainfall or runoff data, and the LSTM is used to extract the stable long-sequence gradients of the spatiotemporal features. The core parameters of the LSTM are searched using a Bayesian optimization algorithm to construct the initial prediction layer; The historical real-time rainfall or historical runoff data in the multi-source heterogeneous data are globally normalized to obtain globally normalized data. The initial forecast layer is trained based on the globally normalized data and the Nash efficiency coefficient to generate the intermediate forecast layer. The intermediate forecast layer uses a lightweight inference engine deployed at the edge of the AI ​​chip to update the input of the intermediate forecast layer in real time based on the second-level water level data of the edge nodes. Based on the attention weight matrix of the multi-head attention layer of the multi-head attention mechanism coupled with the LSTM model in the prediction process, the key period for the formation of the flood peak is determined, and the forecast layer is generated.

4. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 1, characterized in that, The disaster chain key risk point data includes the overall risk value of the disaster chain and the critical path probability. The analysis layer is specifically used for: Based on the historical disaster chain data, a watershed disaster chain topology network is generated with water conservancy facilities as nodes and disaster transmission paths as edges; The Jaccard index is used to screen nodes with high degree centrality and betweenness centrality in the watershed disaster chain topology network to determine the key nodes of the disaster chain; The river siltation status identified by AI edge devices is mapped to Bayesian network evidence variables, and the risk probabilities of key nodes are dynamically updated based on the topology network of the watershed disaster chain, outputting the overall risk value of the disaster chain and the probability of the critical path.

5. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 1, characterized in that, The scheduling layer is specifically used for: Based on the flood hydrograph predicted by the flood forecast, and with maximizing the peak reduction rate as the optimization objective, an objective function for reservoir operation is established, which includes: ； Where maxF is the objective function value. For the peak inflow of floodwater, The outflow peak flow rate after scheduling; Set flood control constraints, encode the discharge sequence into particle position vectors, and initialize the particle swarm. The flood control constraints are that the reservoir water level does not exceed the flood control limit water level and the water level fluctuation is less than the maximum water level fluctuation per unit time. The objective function value of each particle is determined by the objective function, and the particle velocity and position are updated by dynamic inertial weights, which decrease linearly with the number of iterations. The process is iteratively updated until the convergence condition is met, generating the optimized reservoir scheduling scheme.

6. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 1, characterized in that, The dynamic coupling algorithm module also includes a feedback optimization layer; The feedback optimization layer is used to construct a Gaussian process proxy model by taking the reservoir optimization scheduling scheme as the initial solution and minimizing the deviation between the actual discharge process and the reservoir optimization scheduling scheme. Using the Gaussian process proxy model, based on real-time collected downstream water level data, the Nash efficiency coefficient of the actual discharge process line and the reservoir optimization scheduling scheme is determined. When the Nash efficiency coefficient is lower than the preset coefficient threshold, the Bayesian optimizer is triggered. The Bayesian optimizer iteratively searches for the hyperparameter combination of the Coupled Particle Swarm Optimization (CPSO) algorithm using its acquisition function, and then re-runs the CPSO algorithm based on the hyperparameter combination to generate a corrected scheduling scheme, which is then fed back to the digital twin module.

7. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 1, characterized in that, The flood simulation data includes the flood simulation water depth, flood simulation flow velocity, and flood simulation range. The virtual inner loop control unit is specifically used for: Based on the discharge process line in the reservoir optimization scheduling scheme, a one-dimensional hydrodynamic model is constructed to simulate the flood evolution of the main river channel. Based on the acquired high-precision DEM data, the floodplain grid was divided, and a two-dimensional hydrodynamic model was constructed to simulate the surface runoff process. The one-dimensional hydrodynamic model and the two-dimensional hydrodynamic model are coupled by lateral connection to obtain the one-dimensional and two-dimensional hydrodynamic coupled model. When the water level at the cross-section of the main channel exceeds the crest elevation during the flood evolution, an intermediate flow is transferred to the adjacent two-dimensional grid through a momentum exchange formula. The two-dimensional hydrodynamic model generates the inundation simulation velocity and the inundation simulation range based on the intermediate flow. The momentum exchange formula includes: ； Among them, Q overflow The intermediate flow rate is given by g, where g is the physical acceleration constant and C is the intermediate flow rate. d Here, H is the overflow coefficient, L is the cross-sectional length, and H is the overflow coefficient. 1D Z represents the water level at the river cross-section. crest The elevation of the top of the dike; The simulated inundation depth in the flood evolution of the main river channel is extracted. When the simulated inundation depth exceeds the preset depth threshold, the scheduling re-optimization instruction is generated. The simulated inundation depth, the simulated inundation velocity, and the simulated inundation range are used for risk assessment.

8. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 7, characterized in that, The actual outer ring control unit is specifically used for: The video surveillance equipment deployed is used to collect real-time flood image data of the downstream flooded area, and the water body boundary and flooded markers in the flood image data are identified by the target detection algorithm to generate a raster map of the actual flooded area. Spatial overlay analysis is performed between the actual inundation range raster map and the simulated inundation range raster map in the inundation simulation data to determine the similarity. When the similarity is less than the similarity threshold, a loss function is constructed based on the inundation depth deviation. The loss function is used to adjust the Manning roughness coefficient of the one-dimensional hydrodynamic model and the overflow coefficient of the two-dimensional hydrodynamic model in reverse, and the process is iteratively updated until the loss function value is less than the preset loss value.

9. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 4, characterized in that, The river channel siltation status includes a siltation index. Mapping the river channel siltation status identified by AI edge devices to Bayesian network evidence variables, and dynamically updating the risk probability of key nodes based on the watershed disaster chain topology network, includes: The AI ​​edge device is deployed at key river sections and embankment hazard points in the watershed disaster chain topology network. The AI ​​edge device integrates a target detection algorithm to identify the amount of floating debris accumulation in the river and the seepage points in the embankment. The siltation index is determined based on the amount of floating debris in the river channel. When the siltation index exceeds the preset siltation threshold, the conditional probability update of the corresponding node in the Bayesian network is triggered. When the levee seepage point is found to exist, the levee seepage point is added to the watershed disaster chain topology network to generate a new watershed disaster chain topology network. The risk probability of the key nodes is updated based on the new watershed disaster chain topology network.

10. The watershed disaster chain early warning platform integrating dynamic coupling algorithm and digital twin as described in claim 1, characterized in that, The comprehensive vulnerability assessment module is specifically used for: Based on the mobile phone signaling data, a dynamic population density distribution heat map is generated using a spatiotemporal clustering algorithm to obtain the dynamic population density. Based on the aforementioned medical resource data, a medical accessibility index is determined; The dynamic population density and the medical accessibility index are input into the vulnerability assessment formula of the comprehensive vulnerability assessment model to determine the grid vulnerability. The vulnerability assessment formula includes: ； Among them, V i For the aforementioned mesh fragility, and The contribution of population density and medical accessibility to disaster sensitivity are respectively obtained through weighting coefficients determined by the entropy weight method. The dynamic population density, This represents the minimum dynamic population density. This represents the maximum dynamic population density. The medical accessibility index is... This represents the minimum value of the healthcare accessibility index. This represents the maximum value of the healthcare accessibility index.

Images