A mechanism data double-driven flood forecasting method and system

Abstract

The application discloses a flood forecasting method and system driven by mechanism data, and belongs to the technical field of computers. The method is characterized in that a central routing intelligent agent dynamically allocates a calculation mode, combines a physical mechanism model prediction with a generative state error model analysis, and is supplemented by a physical consistency forced check, thereby effectively solving the problems of low calculation efficiency of the physical mechanism model and lack of physical constraints of the data-driven model in the traditional method. Thus, the method realizes deep cooperation of real-time physical check and dynamic error compensation in the forecasting process, thereby improving the physical credibility, calculation efficiency and adaptability to extreme flood events in flood forecasting.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a flood forecasting method and system driven by both mechanistic data. Background Technology

[0002] In the field of smart water management, flood forecasting, as a core support for flood control decision-making, has long relied on two technical approaches: physical mechanism models and data-driven models. Physical mechanism models, built upon hydrological physics equations, explicitly express fundamental laws such as mass and momentum conservation, reflecting the intrinsic mechanisms of watershed hydrological processes and possessing strong physical interpretability. On the other hand, pure data-driven models, represented by generative artificial intelligence, while efficiently mining nonlinear correlations in historical data for rapid forecasting, lack transparency in their internal operating mechanisms, resembling a "black box." Under rare scenarios outside the training data distribution (such as above-standard rainfall or sudden dam failure events), they often output forecast results that violate physical laws, such as non-physical oscillations in the flow process line or imbalances in water volume, severely damaging the reliability and credibility of the forecast.

[0003] To address the aforementioned issues, the relevant technologies urgently need improvement. Summary of the Invention

[0004] This application provides a flood forecasting method and system driven by both mechanistic data and practical principles. The technical solution is as follows: On the one hand, a flood forecasting method driven by both mechanistic data and other factors is provided, the method comprising: In response to the flood forecasting task for the target watershed, the real-time multi-source watershed characteristic data for the current time period is analyzed and routed by the central routing agent to obtain the model routing map of the target watershed. The model routing map is used to dynamically allocate computing modes to multiple computing units respectively. Based on the model routing diagram and the model state after correction in the previous time period, the computing units assigned to the coupled mode are used to perform forecast calculations through the physical mechanism model at the watershed scale to obtain the mechanism forecast flow results and the corresponding internal physical state sequence. For the internal physical state sequence and real-time observation data, a state error generation analysis is performed using a generative state error model to obtain the physical state correction amount. The physical state correction amount is then filtered and verified based on the physical conservation law using a physical consistency mandatory checker to obtain the effective state correction amount. The state variables of the physical mechanism model are corrected based on the effective state correction amount, and the forecast results of all computing units are fused and calculated based on the model routing diagram and the corrected state variables to obtain the flood forecast results of the target watershed.

[0005] On the one hand, a flood forecasting system driven by both mechanistic data and other factors is provided, the system comprising: The parsing module is used to respond to the flood forecasting task of the target watershed. It parses and makes routing decisions on the real-time multi-source watershed feature data of the current time period through the central routing agent to obtain the model routing map of the target watershed. The model routing map is used to dynamically allocate computing modes to multiple computing units respectively. The forecast calculation module is used to perform forecast calculations on the calculation units assigned to the coupled mode based on the model routing map and the model state after correction in the previous time period, and to obtain the mechanism forecast flow results and the corresponding internal physical state sequence. The error analysis module is used to perform state error generation analysis on the internal physical state sequence and real-time observation data through a generative state error model to obtain the physical state correction amount, and to perform verification and filtering of the physical state correction amount based on the physical conservation law through a physical consistency mandatory checker to obtain the effective state correction amount. The fusion calculation module is used to correct the state variables of the physical mechanism model based on the effective state correction amount, and to perform fusion calculation on the forecast results of all calculation units based on the model routing diagram and the corrected state variables to obtain the flood forecast results of the target watershed.

[0006] On one hand, a computer device is provided, the computer device including one or more processors and one or more memories, the one or more memories storing at least one computer program, the computer program being loaded and executed by the one or more processors to implement the mechanism data dual-driven flood forecasting method.

[0007] On the one hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, the computer program being loaded and executed by a processor to implement the mechanism data-driven flood forecasting method.

[0008] On the one hand, a computer program product or computer program is provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to perform the above-mentioned mechanism-data dual-driven flood forecasting method. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the implementation environment of a mechanism-data-driven flood forecasting method provided in this application embodiment; Figure 2 This is a flowchart of a mechanism-data-driven flood forecasting method provided in an embodiment of this application; Figure 3 This is a flowchart of another mechanism-data-driven flood forecasting method provided in this application embodiment; Figure 4 This is a flowchart of another mechanism-data-driven flood forecasting method provided in the embodiments of this application; Figure 5 This is a flowchart of another mechanism-data-driven flood forecasting method provided in the embodiments of this application; Figure 6 This is a flowchart of another mechanism-data-driven flood forecasting method provided in the embodiments of this application; Figure 7 This is a flowchart of another mechanism-data-driven flood forecasting method provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of a flood forecasting system driven by both mechanistic data, provided in an embodiment of this application. Figure 9 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0012] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0013] In order to illustrate the technical solutions provided in the embodiments of this application, some terms involved in the embodiments of this application will be introduced below.

[0014] Target basin: refers to the specific geographical area for flood forecasting, usually the catchment area of ​​one or more river systems, whose boundaries and internal hydrological characteristics form the basis of the forecasting task.

[0015] Real-time multi-source watershed characteristic data: refers to a dynamic data set that reflects the hydrological, meteorological, and geographical attributes of the target watershed, obtained from different sources during the current forecast period, such as rainfall, temperature, humidity, soil moisture content, topography, and land use type.

[0016] Central routing agent: refers to a computing module with intelligent decision-making capabilities. Its function is to receive and process real-time multi-source watershed characteristic data, and then intelligently allocate the most suitable computing mode to each computing unit in the watershed according to the dynamic changes and forecasting needs of the watershed.

[0017] Model routing map: refers to a type of structured data generated by the central routing agent. It represents the geographical location, interrelationships, and assigned computing modes of each computing unit within the target watershed in graphical or matrix form, providing guidance for subsequent forecast calculations.

[0018] Computational unit: refers to several sub-regions with independent computing capabilities into which the target watershed is divided. Each sub-region can be assigned different computing modes according to its characteristics and forecasting needs.

[0019] Computation mode: refers to the specific algorithm or model type assigned to the computing unit for performing forecast calculations, such as a purely physical mechanism model, a purely data-driven model, or a combination of both.

[0020] Coupled mode: refers to a special computing mode in which the physical mechanism model and the data-driven model work together to complete the forecast calculation, aiming to combine the advantages of both.

[0021] Physical mechanism model: refers to a mathematical model based on the laws of hydrophysical science (such as conservation of mass, energy and momentum) to simulate hydrological processes in a watershed, such as rainfall-runoff and river evolution.

[0022] Mechanism-based flow forecast results: These refer to the flow forecast values ​​calculated by a physical mechanism model under coupled mode based on its internal physical mechanisms.

[0023] Internal physical state sequence: refers to a series of snapshot data reflecting the internal hydrological and physical state of the watershed (such as soil moisture content, river water level, groundwater storage, etc.) recorded at different time steps during the forecast calculation process of the physical mechanism model.

[0024] Real-time observation data refers to measurement data that reflects the true hydrological state of the target watershed, obtained directly through sensors, remote sensing, and other means during the forecast period, such as actual rainfall, river level, and flow rate.

[0025] Generative state error model: refers to a model built based on generative artificial intelligence technology. Its function is to learn the deviation pattern between the internal physical state predicted by the physical mechanism model and the real-time observation data, and generate the corresponding physical state correction.

[0026] Physical state correction: refers to the value of physical state variables generated by the generative state error model, used to correct the value of physical state variables within the physical mechanism model in order to reduce prediction error.

[0027] Physical consistency enforced checker: This refers to a module specifically designed to check the physical rationality of physical state corrections. It filters and adjusts the corrections based on the laws of physical conservation to ensure that the corrections conform to physical laws.

[0028] Effective state correction: refers to the physical state correction that meets the physical conservation laws and rationality requirements after being verified and filtered by the physical consistency enforcer. It can be used to correct the state variables of the physical mechanism model.

[0029] State variables: These are key parameters in the physical mechanism model used to describe the hydrological and physical state within the watershed. Their numerical changes directly affect the simulation results of the model.

[0030] Flood forecast results: These refer to the predicted information about future flood events in the target watershed, which is output after calculation and fusion of the entire methodology. They are usually represented by the flow process curve at the outlet section.

[0031] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0032] Figure 1 This is a schematic diagram illustrating the implementation environment of a mechanism-data-driven flood forecasting method provided in this application embodiment. See also... Figure 1 This implementation environment may include node 110 and server 140.

[0033] Node 110 is connected to server 140 via a wireless or wired network. Optionally, node 110 can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. Node 110 has an application installed and running that supports flood forecasting driven by both mechanistic data and other data.

[0034] Server 140 is a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. Server 140 can provide background services for applications running on node 110.

[0035] Traditional flood forecasting methods primarily rely on physical mechanism models and data-driven models. Physical mechanism models have limitations in computational efficiency, handling complex underlying surface conditions, and simulating extreme rainfall responses, leading to insufficient forecast accuracy. Purely data-driven models, lacking physical constraints, often produce physically unreasonable or inexplicable forecasts under extreme scenarios outside the training data distribution, making reliability difficult to guarantee. Current attempts to combine these two types of models remain superficial, failing to fundamentally establish a deep collaborative mechanism for real-time physical verification and dynamic error compensation. This results in the inability to simultaneously achieve the three key objectives of physical reliability, computational efficiency, and adaptability to extreme flood events.

[0036] To address this, this application proposes a flood forecasting method driven by both mechanistic data and other relevant information. (See [link to relevant documentation]). Figure 2 Taking the server as the executing entity as an example, the method includes the following steps.

[0037] 201. In response to the flood forecasting task for the target watershed, the real-time multi-source watershed characteristic data for the current period is analyzed and routed through the central routing agent to obtain the model routing map of the target watershed. This model routing map is used to dynamically allocate computing modes to multiple computing units respectively.

[0038] 202. Based on the model routing diagram and the model state after correction in the previous time period, the computational units assigned to the coupled mode are used to perform forecast calculations through the physical mechanism model at the watershed scale to obtain the mechanism forecast flow results and the corresponding internal physical state sequence.

[0039] 203. For the internal physical state sequence and real-time observation data, perform state error generation analysis through a generative state error model to obtain the physical state correction amount, and perform verification and filtering based on the physical conservation law through a physical consistency mandatory checker to obtain the effective state correction amount.

[0040] 204. Based on the effective state correction, the state variables of the physical mechanism model are corrected, and based on the model routing diagram, the forecast results of all computing units are fused and calculated using the corrected state variables to obtain the flood forecast results of the target watershed.

[0041] It should be noted that the flood forecasting method provided in this application embodiment is an auxiliary method for professional technicians, and the specific results still need to be manually verified.

[0042] This embodiment provides a flood forecasting method driven by both mechanistic data and practical application. Responding to a flood forecasting task for a target watershed, this method processes real-time multi-source watershed characteristic data for the current time period. For example, data from different sources such as meteorological stations, hydrological stations, and remote sensing satellites can be collected manually and preliminarily converted and integrated. A central routing agent then parses this data and makes routing decisions. For instance, a set of rules based on expert experience can be preset, and the watershed can be simply divided into several regions based on indicators such as real-time rainfall and soil moisture. A computational mode is statically assigned to each region, thus obtaining a model routing map of the target watershed. This model routing map is used to dynamically assign computational modes to multiple computational units.

[0043] Furthermore, based on the aforementioned model routing diagram and the corrected model state from the previous time period, forecast calculations are performed on the computational units assigned to coupled modes using a watershed-scale physical mechanism model. For example, the observation data from the previous time period can be simply used as the initial conditions for the current time period to drive the physical mechanism model's calculations. Alternatively, the physical mechanism model can be run for forecast calculations on all computational units, regardless of their assigned mode. This yields the mechanistic forecast flow results and the corresponding internal physical state sequence.

[0044] Based on this, a generative state error model is used to generate and analyze the state error of the aforementioned internal physical state sequence and real-time observation data to obtain the physical state correction. For example, the internal physical state predicted by the mechanism can be subtracted point by point from the real-time observation data to obtain an original error value, which serves as the physical state correction. The physical state correction is then filtered and verified based on the law of physical conservation using a physical consistency enforced checker to obtain the effective state correction. For example, a fixed threshold can be set, and corrections exceeding this threshold can be directly discarded without further verification based on the law of physical conservation.

[0045] In a preferred implementation, the state variables of the physical mechanism model are corrected based on the aforementioned effective state correction. For example, the effective state correction can be directly and simply superimposed with the state variables of the physical mechanism model, without considering the physical mapping relationship or reasonable range between the two. Furthermore, based on the aforementioned model routing diagram, the forecast results of all computational units are fused using the corrected state variables to obtain the flood forecast result for the target watershed. For example, the forecast results of each computational unit can be simply averaged to obtain the flood forecast result, without considering the differences between different computational models or spatial topological relationships.

[0046] The mechanistic data-driven flood forecasting method proposed in this embodiment effectively solves the problems of low computational efficiency of physical mechanism models and lack of physical constraints in data-driven models in traditional methods by dynamically allocating computational modes through a central routing agent, combining physical mechanism model forecasting with generative state error model analysis, and supplementing it with mandatory physical consistency verification. This achieves deep synergy between real-time physical verification and dynamic error compensation during the forecasting process, thereby simultaneously improving physical reliability, computational efficiency, and adaptability to extreme flood events in flood forecasting.

[0047] In some of the embodiments described above in this application, a model routing map is obtained by parsing and routing decision-making of real-time multi-source watershed characteristic data, which is then used to dynamically allocate computational modes. However, in the implementation process, how to efficiently extract key feature sets to accurately characterize the urgency of hydrological response and the applicability of mechanism models in different regions, and based on this, to perform intelligent pattern matching and decision-making to ensure the accuracy and adaptability of the routing map, is a challenge.

[0048] To address this, this application further proposes a flood forecasting method driven by both mechanistic data and practical application. In this method, real-time multi-source watershed characteristic data for the current time period is analyzed and routed using a central routing agent to obtain a model routing map for the target watershed. (See [link to relevant documentation]). Figure 3 Taking the server as the executing entity as an example, the steps include the following: 301. Perform multi-dimensional feature analysis on the real-time multi-source watershed feature data and extract key feature sets. These key feature sets are used to characterize the urgency of hydrological response and the applicability of the mechanism model in different areas within the target watershed.

[0049] 302. Based on the key feature set, pattern matching and decision-making are performed through the routing decision network in the central routing agent to generate a corresponding pattern assignment label for each computing unit. The pattern assignment label is used to indicate the computing pattern.

[0050] 303. Assign labels based on the patterns of all computing units, generate and output the routing map of the model.

[0051] This real-time multi-source watershed characteristic data refers to a dataset describing the current state and characteristics of a watershed, acquired in real time from multiple sources. Its purpose is to provide comprehensive and real-time watershed information for flood forecasting, ensuring the accuracy and timeliness of forecasts. Specifically, this data can originate from satellite remote sensing (e.g., rainfall, soil moisture, land cover), ground sensor networks (e.g., rain gauges, water level stations, soil moisture probes), weather radar, numerical weather prediction model outputs, digital elevation models (DEMs), geological maps, and land use maps. Real-time data acquisition can be achieved through data streaming protocols (e.g., MQTT, Kafka), API integration with data providers, or automated data ingestion pipelines.

[0052] This central routing agent is an intelligent software component or system responsible for data parsing, feature extraction, and routing decision-making. Its core function is to act as a central coordinator, dynamically allocating computing modes to different computing units. It can be implemented as a microservices-based system, where different services are responsible for data ingestion, feature engineering, decision-making, and routing graph generation. This agent may include core components such as a data interface, a feature extraction module, a routing decision-making network, and a routing graph generator.

[0053] Multidimensional feature analysis refers to the process of in-depth analysis of real-time multi-source watershed feature data to extract meaningful information from multiple dimensions. Its purpose is to transform raw, heterogeneous data into a structured, information-rich set of key features to accurately reflect the hydrological dynamics and model applicability of the watershed, providing reliable input for subsequent decision-making. This analysis process can employ various techniques, such as statistical analysis (e.g., principal component analysis, correlation analysis), machine learning techniques (e.g., autoencoders for dimensionality reduction, feature importance ranking algorithms), or deriving parameters (e.g., runoff coefficient, infiltration rate) using domain-specific hydrological models. The dimensions of the analysis can cover time (e.g., rainfall intensity over time), space (e.g., soil type distribution), physical (e.g., elevation, slope), and hydrological (e.g., anterior soil moisture, river flow).

[0054] This key feature set is the most relevant and informative set of features selected after multi-dimensional feature analysis. These features are specifically used to characterize the urgency of hydrological response and the applicability of mechanistic models in different areas within the target watershed. The feature set may include dynamic features (such as current rainfall intensity, cumulative rainfall, soil moisture deficit, and river level) and static features (such as land cover type, soil permeability, slope, and watershed drainage density), as well as historical model performance indicators for specific sub-watersheds. Its structure can be a vector or tensor representation, where each element corresponds to a specific feature, and can be spatially organized (e.g., a feature map for each grid cell).

[0055] Hydrological response urgency is an indicator that measures the speed and intensity of a specific area's response to hydrological inputs (such as rainfall) within a watershed, indicating the potential for rapid runoff generation or flood formation. Its purpose is to identify areas that require immediate, more detailed (e.g., physically-based) modeling due to high flood risk or rapid response characteristics. This urgency can be quantified by combining dynamic characteristics (such as rainfall and soil moisture) and static characteristics (such as topography and soil type), for example, using hydrological indices or expert rules.

[0056] The suitability assessment of this mechanistic model evaluates the expected performance of a specific physical mechanism model in a particular region of a watershed, taking into account factors such as data availability, computational cost, and the dominant physical processes in that region. Its purpose is to guide the selection of the most appropriate computational model, ensuring that complex physical models are used where they have a clear advantage, while simpler or data-driven models are sufficient or more efficient where they are used. This suitability assessment can be based on criteria such as model calibration / driving data availability, computational resources, hydrological process complexity (e.g., karst topography, presence of urban areas), the historical performance of the mechanistic model in the specific region, and the model's inherent weakness, and can be conducted using a scoring system or machine learning classifier.

[0057] This routing decision network is a component of the central routing agent, typically implemented using machine learning or deep learning techniques. Its aim is to learn the complex relationship between a set of key features and the optimal computational pattern for each computational unit. Its function is to intelligently match the characteristics of each computational unit (represented by its set of key features) with the most suitable computational pattern, thereby making dynamic and adaptive routing decisions. This network can employ various architectures, such as neural networks (e.g., multilayer perceptrons, convolutional neural networks for spatial features, graph neural networks for hydrological connectivity), decision trees, random forests, or reinforcement learning agents. Its training data typically includes watershed characteristics, the corresponding optimal model selection (potentially determined by expert knowledge or model performance retrospective analysis), and desired outcomes (e.g., accuracy, computational efficiency).

[0058] Pattern matching and decision-making refers to the process by which a routing decision network analyzes the key feature set of a given computational unit and determines, at a specific point in time, which computational mode (e.g., purely physical, purely data-driven, or a specific hybrid / coupled mode) is most suitable for that unit. Its aim is to assign an optimal computational strategy to each unit, balancing accuracy, computational cost, and physical consistency based on current watershed conditions and unit characteristics. This process can employ classification algorithms (e.g., the softmax output of a neural network, indicating the probability of each mode), rule-based expert systems, or multi-objective optimization algorithms considering multiple performance metrics. Its output can be a probability distribution of possible modes or a direct selection of a single mode.

[0059] This computational unit is a discrete spatial partition of the target watershed, representing a specific area for flood forecasting calculations. These units can vary in size and shape, such as grid cells, sub-watersheds, or hydrological response units. Their purpose is to enable spatially distributed and localized flood forecasting, allowing the application of different computational models based on the unique characteristics and real-time conditions of different parts of the watershed. The discretization of computational units can employ regular grid cells (such as square or hexagonal shapes), irregular sub-watersheds defined by hydrological boundaries, or hydrological response units (HRUs) based on similar land use / soil / slope characteristics. Each unit is associated with its geographic coordinates, its topological relationships with neighboring units, and a set of attributes (such as key features and allocation patterns).

[0060] The model assignment label is a specific identifier or flag assigned to each computational unit, indicating the specific computational model that should be used for that unit during the current forecast period. Its purpose is to explicitly instruct the forecasting system which model or model combination to deploy for each individual computational unit, thereby enabling dynamic routing decisions. This label can be an integer code (e.g., 0 for purely data-driven, 1 for purely physical, 2 for coupled models), a string identifier (e.g., "DataDriven", "PhysMech", "Hybrid"), or a one-hot encoded vector, directly corresponding to a predefined computational model in the system.

[0061] This model routing map is a structured data representation that encapsulates the spatial distribution of assigned computational patterns for all computational units in a target watershed within a given forecast period. Essentially, it provides a "map" indicating where to use which model. Its role is to serve as a comprehensive instruction set for subsequent flood forecast calculations, enabling the dynamic allocation of computational resources and the application of appropriate models to each part of the watershed. The map's structure can be a geospatial data layer (e.g., GeoJSON, Shapefile, NetCDF), where each feature (e.g., a polygon representing a computational unit) has attributes indicating its assigned pattern. Alternatively, it can be a grid-based raster, where each cell value represents a pattern. Its content includes the cell's spatial coordinates / boundaries, its unique identifier, and its corresponding pattern assignment label.

[0062] Through the aforementioned technical solution, this application effectively addresses the challenge of efficiently extracting key feature sets in flood forecasting to accurately characterize the urgency of hydrological responses and the applicability of mechanistic models in different regions, and then performing intelligent pattern matching and decision-making based on this to ensure the accuracy and adaptability of routing maps. Specifically, by performing multi-dimensional feature analysis on real-time multi-source watershed characteristic data, key information reflecting watershed hydrological dynamics and model applicability can be comprehensively and deeply extracted from complex heterogeneous data, avoiding feature omissions or biases that may exist in traditional methods, thus laying a solid data foundation for subsequent intelligent decision-making. On this basis, the routing decision network in the central routing agent can perform refined pattern matching and decision-making based on these key feature sets, generating customized pattern assignment labels for each computing unit. This dynamic and intelligent decision-making mechanism enables the flexible selection of the most suitable computing mode according to the urgency of real-time hydrological responses and the applicability of mechanistic models in different regions of the watershed, rather than adopting a static or uniform mode, thereby improving the local accuracy and overall efficiency of forecasting. The generated model routing diagram clearly indicates the calculation mode that each calculation unit should adopt, providing clear guidance for subsequent forecast calculations. This ensures that the entire flood forecasting system can achieve deep synergy of "physical reliability", "computational efficiency" and "adaptability to extreme conditions" under complex and variable hydrological conditions, effectively improving the reliability and adaptability of flood forecasts.

[0063] In some of the embodiments described above in this application, a multi-dimensional feature analysis of real-time multi-source watershed characteristic data is proposed to extract key feature sets to support routing decisions of a central routing agent. However, in its implementation, since multi-source data includes meteorological, hydrological, and underlying surface attributes, without detailed classification, analysis, and fusion, the key feature sets may not fully reflect key factors such as dynamic rainfall, water storage status, soil hydraulics, and topography, leading to inaccurate feature extraction, affecting the accuracy of subsequent routing decisions, and thus reducing the adaptability and reliability of flood forecasts.

[0064] To address this, this application further proposes a multi-dimensional feature analysis of the real-time multi-source watershed characteristic data to extract a key feature set, including: analyzing real-time meteorological and hydrological data from the real-time multi-source watershed characteristic data to extract dynamic rainfall intensity features and watershed water storage status features; analyzing underlying surface spatial attribute data from the real-time multi-source watershed characteristic data to extract soil hydraulic features and topographic features; and performing spatiotemporal fusion and importance assessment of the dynamic rainfall intensity features, watershed water storage status features, soil hydraulic features, and topographic features to obtain the key feature set.

[0065] Specifically, the analysis of real-time meteorological and hydrological data from real-time multi-source watershed characteristic data aims to capture the dynamic changes in hydrological processes within the watershed. Dynamic rainfall intensity characteristics reflect the instantaneous intensity and spatial distribution of rainfall events, serving as a key input for surface runoff formation. Watershed water storage status characteristics characterize the current water saturation level of the watershed, directly influencing rainfall infiltration, runoff generation, and confluence processes. Extraction of dynamic rainfall intensity characteristics can be achieved by fusing multi-source rainfall observation data (such as ground rain gauges, weather radar, and satellite remote sensing precipitation products) and combining them with data assimilation techniques (such as Kalman filtering and variational assimilation) to obtain a high spatiotemporal resolution rainfall field. Alternatively, deep learning models (such as convolutional neural networks and recurrent neural networks) can be used to train historical meteorological data to identify and predict rainfall intensity patterns. The extraction of watershed water storage characteristics can be based on simulation results from soil moisture sensor networks, remote sensing soil moisture products (such as SMAP and SMOS), and hydrological models (such as conceptual hydrological models and distributed hydrological models), characterized by calculating indicators such as soil moisture content, groundwater level, and surface runoff coefficient. Alternatively, it can be obtained by combining the simulation output of physical mechanism models and by real-time monitoring and correction of internal state variables (such as soil moisture content and groundwater storage).

[0066] Simultaneously, the spatial attribute data of the underlying surface in real-time multi-source watershed characteristic data are analyzed to identify the physical and geographical features affecting water flow and model applicability. Soil hydraulic characteristics (such as permeability, saturated hydraulic conductivity, and porosity) determine the capacity for rainfall infiltration and groundwater recharge. Topographic features (such as slope, aspect, elevation, and runoff path) control the direction, velocity, and runoff process of surface runoff. Soil hydraulic characteristics can be extracted based on soil type maps, land use / cover maps, and field soil sampling analysis data, using GIS spatial analysis functions combined with empirical formulas or hydraulic parameter databases. Alternatively, remote sensing technology (such as hyperspectral remote sensing) combined with machine learning methods can be used to invert soil texture, organic matter content, etc., and then calculate soil hydraulic parameters. Topographic features can be extracted using high-precision digital elevation models (DEMs) and topographic analysis algorithms (such as the D8 algorithm and multi-flow direction algorithms) to calculate slope, aspect, catchment area, river network density, etc. Alternatively, ultra-high resolution DEMs can be generated using LiDAR data and combined with hydrological connectivity analysis tools to accurately identify micro-topographic features and potential confluence paths.

[0067] Based on this, spatiotemporal fusion and importance assessment are performed on dynamic rainfall intensity characteristics, watershed water storage characteristics, soil hydraulic characteristics, and topographic features. The aim is to integrate feature data from different sources and at different spatiotemporal scales into a unified framework, ensuring information integrity and consistency, and identifying the features with the greatest impact on flood forecasting and routing decisions, thereby optimizing the expression efficiency and decision accuracy of the feature set. Spatiotemporal fusion can employ a multi-scale data assimilation framework, for example, using Kalman filtering or particle filtering to fuse real-time observation data with model simulation data, while considering the spatial correlation and temporal dynamics of different features. Alternatively, deep learning models such as graph neural networks (GNNs) or convolutional long short-term memory networks (ConvLSTMs) can be used for end-to-end learning and fusion of feature data with complex spatiotemporal dependencies. Importance assessment can quantify the impact of each feature on hydrological response and model performance using tree-based feature selection algorithms (such as random forests and gradient boosting trees) or mutual information methods. Alternatively, model interpretability techniques (such as SHAP values ​​and LIME) can be used to analyze the sensitivity of the forecast model to each feature, thereby determining its importance weight.

[0068] Through the aforementioned technical solutions, this application effectively addresses the problem of inaccurate key feature set extraction. Specifically, by classifying and analyzing real-time meteorological and hydrological data, dynamic rainfall intensity and watershed water storage status can be accurately captured, providing real-time and dynamic driving information for routing decisions. Simultaneously, the analysis of underlying surface spatial attribute data fully considers the watershed's physical and geographical background, such as soil hydraulic properties and topographic features, providing a solid foundation for evaluating the applicability of mechanistic models and the variability of hydrological responses. Based on this, by performing spatiotemporal fusion and importance assessment on these multi-dimensional features, information from different sources and scales can be integrated and weighted according to their importance to the flood forecasting task, thereby generating a comprehensive, accurate, and efficient key feature set. This refined key feature set can more accurately characterize the urgency of hydrological responses and the applicability of mechanistic models in different areas within the target watershed, improving the accuracy of pattern matching and decision-making by the central routing agent, and thus enhancing the adaptability and reliability of the entire flood forecasting method, especially in the face of complex underlying surface conditions and extreme rainfall events, providing more reliable forecast results.

[0069] In some of the above-mentioned schemes in this application, pattern matching and decision-making based on key feature sets are proposed, and pattern assignment labels are generated for each computing unit to dynamically allocate computing patterns. However, in this process, the decision may ignore spatial context association and multi-factor collaboration, resulting in pattern allocation being isolated from the influence of adjacent units or lacking consideration of key driving factors, which in turn leads to inconsistent allocation, violation of physical laws, and affects forecast accuracy and reliability.

[0070] To address this, this application further proposes a method for parsing and routing real-time multi-source watershed characteristic data for the current time period using a central routing agent to obtain a model routing map for the target watershed. This includes: for each computational unit, based on the key feature set of the computational unit and the key feature sets of its spatially adjacent computational units, spatial context association analysis is performed through the routing decision network to generate spatially enhanced decision features for the computational unit. Based on these spatially enhanced decision features, multi-factor collaborative decision-making is performed through the routing decision network to generate a preliminary model assignment probability for the computational unit. The factors in this multi-factor collaborative decision-making include at least a real-time rainfall intensity factor and a historical error trend factor from the mechanistic model. Based on the physical constraint rules of the target watershed, spatial consistency optimization is performed on the preliminary model assignment probability to obtain a model assignment label for the computational unit. These physical constraint rules are used to ensure that the model assignment between hydrologically connected units conforms to physical laws.

[0071] This spatial context association analysis aims to identify and integrate the mutual influences and dependencies between the target computational unit and its spatially adjacent computational units. Its core function is to break the isolation between units in traditional pattern allocation decisions, enabling the pattern selection of each computational unit to fully consider the integrity of its local hydrological environment, thereby more accurately reflecting the spatial continuity of the watershed's hydrological processes. Specifically, it can be implemented in one of the following ways: One approach is to construct a spatial adjacency matrix and perform weighted averaging or aggregation operations on the key feature sets of the target computational unit and its adjacent units. For example, different weights can be assigned based on distance or hydrological connectivity, integrating the key feature information of adjacent units into the feature representation of the target unit to form a more spatially representative feature vector. Another approach is to use graph neural network (GNN) technology. Computational units within the watershed are abstracted as nodes in a graph, with spatial adjacency or hydrological connectivity relationships between units as edges. Through graph convolution or graph attention mechanisms, nodes (computational units) can effectively aggregate the feature information of their neighboring nodes (adjacent computational units), thereby generating spatially enhanced decision features containing rich spatial context information.

[0072] This multi-factor collaborative decision-making refers to comprehensively considering and integrating multiple influencing factors when generating the initial model allocation probability of a computational unit, to ensure the comprehensiveness, dynamism, and adaptability of the decision. Its purpose is to avoid one-sided decisions caused by a single factor, enabling model allocation to simultaneously respond to current hydrological conditions (such as real-time rainfall intensity) and the model's own performance (such as the historical error trend of the mechanistic model). Specifically, it can be implemented in any of the following ways: One approach is to use a machine learning-based classification or regression model. For example, a support vector machine (SVM) or neural network model can be trained, taking spatial augmentation decision features, real-time rainfall intensity factors, and historical error trend factors of the mechanistic model as inputs, and outputting the probability of the computational unit being assigned to different models. The model learns complex patterns in historical data to capture the synergistic effects between these factors. Another approach is to construct a multi-criteria decision analysis (MCDA) framework. Weights and evaluation functions are assigned to each influencing factor, and then methods such as weighted summation, analytic hierarchy process (AHP), or fuzzy comprehensive evaluation are used to fuse the evaluation results of different factors, thereby obtaining a comprehensive initial model allocation probability.

[0073] These physical constraint rules are a series of restrictions formulated based on the principles of hydrophysics and the objective laws of watershed hydrological processes. Their function is to ensure that model allocation decisions conform to physical logic in space, such as the continuity of water flow and mass conservation, preventing physically unreasonable or contradictory model allocation results, thereby improving the reliability and credibility of flood forecasts. Specifically, these physical constraint rules can include any of the following types: One type is the hydrological connectivity rule, for example, if an upstream unit is assigned to a physical mechanism model, then its directly downstream units should also be preferentially assigned to a physical mechanism model under certain conditions to ensure the physical consistency of the water flow evolution process. Another type is the mass conservation-based rule, for example, in the model switching region, ensuring smooth water volume transitions between different models to avoid sudden changes or unreasonable losses in water volume due to model differences. This may be achieved by setting gradient constraints for model allocation or forcing similarity between adjacent unit models.

[0074] Spatial consistency optimization refers to using optimization algorithms or strategies to force the initial model assignment probabilities to meet preset physical constraints, thereby obtaining model assignment labels. Its purpose is to eliminate potential physical inconsistencies in the initial decision-making process, ensuring that the model assignments across the entire watershed are spatially coordinated and conform to physical laws. Specifically, it can be implemented in one of the following ways: One approach is to use an iterative optimization algorithm. In each iteration, it checks whether the initial model assignment probabilities of all computational units violate the physical constraints. For units that violate the rules, their probabilities are fine-tuned according to the physical constraints until the model assignments across the entire watershed achieve spatial consistency or converge to the preset optimization objective. Another approach is to construct a global optimization problem. The initial model assignment probabilities are used as optimization variables, the physical constraints as constraints, and an objective function is defined (e.g., minimizing the deviation from the initial probabilities). Then, linear programming, quadratic programming, or heuristic optimization algorithms (such as simulated annealing or genetic algorithms) are used to solve the optimization problem to obtain the optimal model assignment labels that satisfy the physical constraints.

[0075] Through the above technical solution, this application effectively solves the problem that model allocation decisions may ignore spatial context and multi-factor collaboration, leading to isolated model allocations or a lack of consideration for key driving factors, resulting in inconsistent allocations, violations of physical laws, and impacts on forecast accuracy and reliability. Specifically, by performing spatial context analysis on the key feature sets of each computational unit and its spatially adjacent computational units, spatially enhanced decision features are generated. This ensures that model allocation decisions are no longer isolated but fully consider the spatial dependence and connectivity of hydrological processes within the watershed, thereby more accurately capturing the overall nature of local hydrological responses. Based on this, through multi-factor collaborative decision-making, key dynamic factors such as real-time rainfall intensity factors and historical error trend factors of mechanistic models are integrated into the generation of preliminary model allocation probabilities, ensuring the comprehensiveness and dynamic adaptability of the decisions, enabling model selection to flexibly respond to current hydrological conditions and model performance. Furthermore, based on the physical constraints of the target watershed, spatial consistency optimization is performed on the preliminary model allocation probabilities, ensuring that model allocations between hydrologically connected units conform to physical laws, effectively avoiding physically unreasonable or contradictory allocation results, and improving the physical reliability and overall consistency of flood forecasts. This deeply collaborative decision-making mechanism makes the generation of model routing maps more accurate and reasonable, providing a solid foundation for subsequent forecast calculations, thereby improving the accuracy and reliability of flood forecasts.

[0076] In some of the embodiments described above in this application, a preliminary model allocation probability is generated based on spatially enhanced decision features through multi-factor collaborative decision-making to dynamically allocate calculation models. However, in its implementation, the lack of accurate quantitative evaluation of real-time rainfall intensity factors and historical error trend factors of mechanistic models leads to inaccurate preliminary model allocation probabilities, affecting the accuracy and reliability of flood forecasts.

[0077] To address this, this application further proposes a method based on the spatial augmentation decision-making characteristics, using a routing decision network to perform multi-factor collaborative decision-making to generate the preliminary mode allocation probability for the computing unit. This includes: performing state quantification evaluation of the real-time rainfall intensity factor based on the spatial augmentation decision-making characteristics to obtain a rainfall driving intensity index; performing dynamic performance evaluation of the historical error trend factor of the mechanism model based on the spatial augmentation decision-making characteristics to obtain a model reliability index; and calculating the preliminary mode allocation probability for the computing unit using a preset collaborative decision-making rule or probability mapping model based on the rainfall driving intensity index and the model reliability index.

[0078] Specifically, when performing state quantification assessment on the real-time rainfall intensity factor, this factor characterizes the intensity, duration, and spatial distribution of rainfall within the target watershed at the current moment or over a recent period. Its role is to reflect the immediate driving force of the watershed's hydrological response. State quantification assessment aims to transform the raw real-time rainfall intensity factor into a unified and comparable numerical index. This can be achieved through: threshold-based grading, mapping the real-time rainfall intensity factor to different levels and assigning corresponding values ​​based on preset rainfall intensity thresholds (e.g., light rain, moderate rain, heavy rain, torrential rain, etc.); or statistical models, utilizing the statistical relationship between historical rainfall data and hydrological response to construct regression or classification models, inputting the real-time rainfall intensity factor into the model, and outputting a quantified intensity index. Through this assessment, a rainfall-driven intensity index can be obtained, which is a quantified numerical value used to intuitively reflect the driving intensity of current rainfall on the watershed's hydrological processes. A higher rainfall-driven intensity index usually indicates a faster and more intense hydrological response, thus potentially requiring more refined or responsive calculation models.

[0079] Meanwhile, when dynamically evaluating the historical error trend factor of the mechanistic model, this factor reflects the forecast error performance and its changing trend of the physical mechanism model under specific regional or hydrological conditions over a past period. Its role is to assess the reliability of the current physical mechanism model in a specific context. Dynamic performance evaluation aims to assess the applicability and reliability of the physical mechanism model in the current forecast context in real-time or near real-time, based on the historical error trend factor. This can be achieved through: time series analysis, using methods such as sliding window averaging and exponential smoothing to analyze historical error trends and predict the model's performance at the current moment; or, based on machine learning models, training a classification or regression model, inputting historical error trend factors and current watershed characteristics, and outputting a model performance level or reliability score. Through this evaluation, a model reliability index can be obtained, which is a quantifiable value representing the reliance of the physical mechanism model in the current forecast context. A higher model reliability index generally means that the model's output is closer to reality and can be given higher confidence.

[0080] Based on this, and using the rainfall-driven intensity index and the model reliability index, the preliminary mode assignment probability of the computing unit is calculated through a pre-defined collaborative decision rule or probabilistic mapping model. The pre-defined collaborative decision rule is a pre-defined set of logical judgments used to comprehensively consider the rainfall-driven intensity index and the model reliability index to determine the preliminary mode assignment probability of the computing unit; for example, it can take the form of a decision tree or rule set. The probabilistic mapping model is a mathematical or statistical model used to take the rainfall-driven intensity index and the model reliability index as input and output the probability of the computing unit being assigned to different computing modes; for example, it can take a logistic regression model or a neural network model. This preliminary mode assignment probability represents the likelihood that a computing unit will be assigned to a specific computing mode (e.g., a purely physical mechanism mode, a purely data-driven mode, or a coupled mode) in the current context, and it forms the basis for subsequent spatial consistency optimization.

[0081] Through the above technical solution, this application can accurately quantify and evaluate the real-time rainfall intensity factor and the historical error trend factor of the mechanistic model when making multi-factor collaborative decisions based on spatial augmentation decision features. Specifically, by performing state quantification evaluation on the real-time rainfall intensity factor, accurate rainfall driving intensity indicators can be obtained. This ensures that the dynamic changes in rainfall intensity are accurately captured, avoiding decision-making biases caused by coarse rainfall data and providing a reliable basis for subsequent decisions. Simultaneously, by performing dynamic performance evaluation on the historical error trend factor of the mechanistic model, accurate model reliability indicators can be obtained. This fully considers the model's historical performance and error trends, enhancing the adaptability and robustness of the decision-making process and effectively preventing the accumulation of model errors from affecting forecast reliability. Based on these two accurately quantified indicators, calculations can be performed using preset collaborative decision-making rules or probability mapping models to integrate the two key factors of rainfall driving and model reliability, generating a more comprehensive and coordinated preliminary model allocation probability. This solves the problem of inaccurate probability caused by single-factor evaluation, improves the accuracy of model allocation in the computational unit, and thus provides a more reliable dynamic model selection basis for the entire flood forecasting system, improving the overall accuracy and reliability of flood forecasts.

[0082] In some of the embodiments described above in this application, a method for pattern matching and decision-making based on a routing decision network is proposed, which generates a preliminary pattern allocation probability for each computing unit to dynamically allocate computing patterns. However, in its implementation, the preliminary decision may not fully consider hydrological connectivity, resulting in inconsistent pattern allocation in space, violating physical laws, and affecting the reliability and accuracy of flood forecasting.

[0083] To address this, this application further proposes a spatial consistency optimization process for the preliminary model assignment probabilities to obtain model assignment labels for computational units. Specifically, this includes: identifying upstream and downstream neighboring units of the computational unit based on hydrological connectivity; obtaining the preliminary model assignment probabilities of the upstream and downstream neighboring units; performing consistency constraint optimization on the preliminary model assignment probabilities of the computational unit, the upstream neighboring units, and the downstream neighboring units based on physical constraint rules to obtain optimized model assignment probabilities; and determining the model assignment label for the computational unit based on the optimized model assignment probabilities.

[0084] Among these methods, identifying upstream and downstream adjacent units of a computational unit based on hydrological connectivity aims to clarify the current computational unit's position within the entire hydrological network and its flow direction relationship with directly connected units. This is crucial for ensuring the physical consistency of model allocation in space. Specifically, based on Digital Elevation Model (DEM) data, the hydrological network topology within the watershed can be constructed through flow direction analysis and flow accumulation analysis, thereby determining the direct upstream and downstream adjacent units of each computational unit. For example, the D8 algorithm or multi-flow direction algorithm can be used to calculate the flow direction of each unit, and then the upstream catchment area and downstream runoff path can be traced based on the flow direction. Another approach is to utilize pre-constructed watershed river system maps or river network topology data. This data typically already contains the connectivity relationships between various hydrological units (such as sub-basins and river segments), and the upstream and downstream adjacent units of the target computational unit can be identified directly through queries or spatial relationship analysis (such as adjacency and inclusion relationships).

[0085] Obtaining the preliminary mode assignment probabilities of upstream and downstream neighboring units is crucial for making mode assignment decisions for the current computing unit while fully considering preliminary decision-making information about its surrounding hydrological environment. These preliminary probabilities reflect the likelihood of neighboring units being assigned to different computing modes in the initial stage. Specifically, after the central routing agent completes preliminary mode matching and decision-making, the preliminary mode assignment probabilities of all computing units can be stored in an accessible global data structure, such as a spatial index database or a distributed memory cache. When consistency constraint optimization is needed for a computing unit, the system can quickly retrieve the corresponding preliminary mode assignment probabilities from this data structure based on the identifiers of the identified upstream and downstream neighboring units. Alternatively, this can be achieved by defining an application programming interface (API) or messaging mechanism. When the optimization module of a computing unit starts, it can send a request to the central routing agent or other coordination modules to obtain the preliminary mode assignment probabilities of its upstream and downstream neighboring units.

[0086] Consistency constraint optimization is applied to the initial pattern allocation probabilities of the computing unit, its upstream neighboring units, and its downstream neighboring units to obtain optimized pattern allocation probabilities. The aim is to adjust these probabilities by introducing physical constraint rules, based on the initial pattern allocation probabilities of the current computing unit and its neighboring units, to ensure they spatially satisfy hydrophysical laws, thus obtaining more reasonable and reliable pattern allocation probabilities. In practice, an optimization model can be constructed, such as a graph-based optimization problem. The initial pattern allocation probabilities of the computing unit and its neighboring units are used as input, and physical constraint rules (such as mass conservation and flow continuity) are transformed into constraints for the optimization model. For example, a rule can be set: if an upstream unit is assigned to a purely data-driven mode (without outputting physical states), while its downstream unit is assigned to a physical mechanism mode (requiring physical input from the upstream), this assignment may violate physical continuity, and its probability needs to be adjusted. Optimization algorithms (such as the Lagrange multiplier method, sequential quadratic programming, etc.) will minimize the deviation from the initial probabilities under this constraint, thereby obtaining the optimized pattern allocation probabilities. Another approach is to employ an iterative adjustment mechanism based on rule-based reasoning. A series of physical constraint rules are defined, such as "the mode selection of upstream units should maintain a certain degree of compatibility with the mode selection of downstream units to ensure the physical continuity of the water flow." Then, the initial mode assignment probabilities of the current computational unit and its neighboring units are iteratively checked to see if they violate these rules. If a violation is found, the relevant probabilities are fine-tuned according to a pre-defined adjustment strategy (e.g., increasing or decreasing the probability of a specific mode until the rules are satisfied) until all rules are satisfied or the convergence condition is met.

[0087] Determining the mode assignment label for a computational unit based on the optimized mode assignment probabilities transforms the continuous optimized mode assignment probabilities into discrete, explicit computational mode instructions. This is a crucial decision-making step that directly impacts subsequent forecast calculations. In practice, a threshold-based method can be used. For example, for each computational unit, if its optimized "coupled mode" assignment probability exceeds a preset threshold (e.g., 0.6 or 0.7), the computational unit is determined to use a coupled mode. Otherwise, it is determined to use another mode (such as a pure data-driven mode). This threshold can be calibrated based on historical data or expert experience. Alternatively, a maximum probability selection method can be used. For each computational unit, its optimized mode assignment probability typically includes probability distributions of multiple modes (e.g., coupled mode probability, data-driven mode probability, etc.). The mode with the highest probability is directly selected as the mode assignment label for that computational unit.

[0088] Through the above technical solution, this application effectively solves the problem that preliminary model allocation decisions may not fully consider hydrological connectivity, leading to spatial inconsistencies in model allocation, violating physical laws, and affecting the reliability and accuracy of flood forecasts. Specifically, by identifying upstream and downstream adjacent units of a computational unit based on hydrological connectivity and obtaining their preliminary model allocation probabilities, necessary spatial context information is provided for subsequent optimization. Based on this, consistency constraint optimization is applied to the preliminary model allocation probabilities of the current computational unit and its adjacent units according to physical constraint rules, ensuring that model allocation follows physical conservation laws and the principle of flow continuity throughout the entire watershed. For example, it avoids the potential physical input mismatch problem when upstream units use a purely data-driven model while downstream units use a physical mechanism model. The optimized model allocation probabilities are used to determine the model allocation label of the computational unit, ensuring that the model selection of each computational unit not only considers its own characteristics but also incorporates the physical constraints of its hydrological connectivity environment, thereby improving the physical rationality, spatial consistency, and overall reliability and accuracy of flood forecast results. This deep collaboration mechanism ensures that model allocation can be physically verified in real time during forecast execution, providing a more robust and reliable foundation for subsequent mechanistic forecast calculations.

[0089] In some of the above-mentioned schemes in this application, a model routing map is proposed to dynamically allocate computational modes. However, in its implementation, directly generating the routing map based on discrete mode assignment labels may lead to spatial discontinuity or inconsistency, affecting the accuracy and efficiency of the forecast. This is because discrete labels cannot cover the continuous area of ​​the entire watershed and are prone to generating discontinuities at the boundaries or transition areas, thereby destroying the deep collaborative mechanism between the physical mechanism model and the data-driven model.

[0090] In response, this application further proposes a method for generating and outputting the model routing graph based on pattern assignment labels for all computing units, the method comprising: Establish a digital spatial reference system for the target watershed and map the mode assignment labels of all computing units to this digital spatial reference system.

[0091] Based on the pattern assignment label distribution in this digital spatial reference system, continuous pattern distribution data covering the target watershed is generated through spatial interpolation or topological connection methods.

[0092] The continuous pattern distribution data is structured and encoded to generate and output the model routing map, which is structured data containing spatial location and pattern allocation information.

[0093] Specifically, establishing a digital spatial reference system for the target watershed involves transforming its geospatial information into a unified, computable mathematical framework. This digital spatial reference system can employ a projection coordinate system commonly used in Geographic Information Systems (GIS), such as the Universal Transverse Mercator (UTM) projection or the Gauss-Kruger projection, converting the watershed's true geographic coordinates into planar coordinates, thus providing a standardized positioning benchmark for all spatial data. Alternatively, a digital spatial reference system based on a regular grid can be constructed, dividing the target watershed into a series of grid cells with fixed resolution, each with a unique spatial index. Mapping the schema assignment labels of all computational units to this digital spatial reference system means placing each discrete computational unit and its corresponding schema assignment label into this standardized spatial framework according to unified rules, such as associating the geometric center or boundary of the computational unit with specific coordinates or grid cells in the reference system. This process ensures a consistent spatial representation of all schema assignment information, laying the foundation for subsequent spatial analysis.

[0094] Based on the pattern assignment label distribution in this digital spatial reference system, continuous pattern distribution data covering the target watershed is generated through spatial interpolation or topological connectivity methods. The pattern assignment label distribution refers to the spatial arrangement and density of pattern assignment labels corresponding to each computational unit in the digital spatial reference system. Spatial interpolation methods aim to infer the attribute values ​​of unknown points (i.e., other locations within the watershed) based on the attribute values ​​(i.e., pattern assignment labels) of known discrete points (i.e., computational units), thereby generating a continuous spatial distribution field. For example, algorithms such as Kriging interpolation, inverse distance weighted (IDW) interpolation, or spline interpolation can be used to transform discrete pattern assignment labels (e.g., 0 representing data-driven patterns, 1 representing physical mechanism patterns) into a continuous numerical field. This numerical field can be further processed through thresholding or classification to represent the continuous pattern distribution. Topological connectivity methods utilize the spatial adjacency or hydrological connectivity between computational units to logically connect and extend discrete pattern assignment labels to form a continuous pattern distribution. For example, for sub-basins or river segments with well-defined hydrological topological relationships, model assignment labels can be extended from one unit to its neighboring units, based on their connectivity, through diffusion, aggregation, or propagation algorithms based on preset rules, until the entire basin is covered, while ensuring the physical consistency of model assignments. These methods effectively compensate for the spatial discontinuity of discrete labels, ensuring that model assignment information across the entire basin is complete and smoothly transitioned.

[0095] The continuous pattern distribution data is structured and encoded to generate and output the model routing map. This continuous pattern distribution data, after spatial interpolation or topological connection processing, is continuous data covering the entire target watershed and representing the spatial distribution of different computational models. Structured encoding aims to transform this continuous pattern distribution data into a dataset with a well-organized structure and data format, facilitating efficient storage, retrieval, parsing, and application by computer systems. For example, a raster data model can be used for encoding, discretizing the continuous pattern distribution data into a grid with a fixed resolution. Each grid cell stores its corresponding pattern type or pattern assignment probability, along with its spatial coordinate information. Another approach is to use a vector data model, vectorizing the boundaries of continuous regions with the same pattern to form polygonal objects. Each polygonal object stores its pattern type and geometric information, and can be supplemented with other spatial location attributes. The generated model routing map is a structured dataset containing spatial location information and pattern assignment information. It serves as the core scheduling basis, guiding subsequent forecast calculations, ensuring that each computational unit can perform calculations according to its assigned pattern, and supporting collaboration and fusion between different patterns.

[0096] Through the above technical solutions, this application effectively solves the problem of spatial discontinuity or inconsistency in forecasts caused by directly generating model routing maps based on discrete model assignment labels. Establishing a digital spatial reference system and mapping model assignment labels provides a unified and accurate spatial positioning basis for all model assignment information, eliminating potential errors caused by spatial reference inconsistencies. By using spatial interpolation or topological connectivity methods, discrete model assignment labels are transformed into continuous model distribution data covering the entire target watershed. This directly compensates for the deficiency of discrete labels in covering continuous areas, ensuring a smooth spatial transition of model assignments and avoiding discontinuities at boundaries or transition regions, thereby maintaining the integrity of the deep collaborative mechanism between the physical mechanism model and the data-driven model. Structured encoding of the continuous model distribution data ensures that the generated model routing map not only contains accurate spatial location information but also has clear model assignment instructions, making it easy for the forecast system to efficiently parse and execute. This structured, continuous routing map can more accurately guide the mode selection and switching of each computing unit, especially in complex underlying surfaces and hydrologically connected areas, ensuring the physical consistency and spatial continuity of forecast calculations, thereby improving the accuracy and efficiency of flood forecasts and strengthening the deep synergy between different computing modes.

[0097] In some of the above-mentioned schemes of this application, a forecast calculation is proposed to be performed on the computing units assigned as coupled modes to generate mechanistic forecast flow results and internal physical state sequences. However, in this process, how to efficiently select coupled mode units, accurately read the corrected initial state, and accurately perform forecast calculations to generate time-resolved internal state sequences are key challenges to ensure forecast accuracy and efficiency.

[0098] To address this, this application further proposes a flood forecasting method driven by both mechanistic data and computational data. Based on the model's routing diagram and the corrected model state from the previous time period, the computational units assigned to the coupled mode are used for forecasting calculations through a watershed-scale physical mechanism model. This yields the mechanistic forecasted flow results and the corresponding internal physical state sequence. (See [link to relevant documentation]). Figure 4 Taking the server as the executing entity as an example, the process includes: 401. Based on the routing diagram of this model, select the computing units that are assigned to the coupling mode from all computing units in the target watershed.

[0099] 402. For each selected computing unit, read the corresponding corrected model state from the physical verification and correction model state library, and use it as the initial forecast condition for the computing unit in this period.

[0100] 403. Based on the initial forecast conditions, drive the physical mechanism model at the watershed scale to perform forecast calculations, and output the mechanistic forecast flow results of the calculation unit and the internal physical state sequence generated at a preset time resolution.

[0101] The model routing diagram is a structured data format used to dynamically assign computational modes to multiple computational units. It contains a mode assignment label for each computational unit. This diagram typically exists in the form of a spatial indexed data structure (e.g., a raster map, vector polygon layer, or topology network diagram), where each computational unit is associated with a mode assignment label. In practical applications, computational units marked as "coupled modes" can be directly identified by parsing the mode assignment label field in the model routing diagram. For example, if the model routing diagram is a two-dimensional array, with each element storing a mode type code, the array can be traversed for filtering. Alternatively, a database or configuration file associated with the model routing diagram can be queried to retrieve a list of all matching computational unit IDs based on preset coupling mode identifiers. The computational unit represents the target watershed being discretized into multiple spatial units. Each unit is an independent computational entity, which can be a regular rectangular grid unit identified by its row and column indices, or an irregular sub-watershed unit identified by its unique ID or geographic coordinate range. This coupled mode refers to a computational approach in flood forecasting that requires the use of physical mechanism models for detailed hydrological process simulation. This typically occurs in regions with complex hydrological responses, where physical processes dominate, or where high-precision physical state descriptions are needed. For example, a coupled mode can refer to a mode that simulates hydrological processes entirely based on physical equations, such as calculations based on distributed hydrological models (e.g., SWAT, VIC). Alternatively, it can refer to a mode that deeply integrates physical mechanism models with data-driven models, where the data-driven model assists the physical mechanism model in parameter optimization or process correction.

[0102] The physically validated and corrected model state repository is a persistent storage system used to store the set of internal state variables of the physical mechanism model after physical validation and correction at the end of each forecast period. These state variables reflect the current hydrophysical conditions of the watershed, such as soil moisture content, groundwater level, and river storage. This state repository can be a relational database, where the state variables of each computational unit are stored in structured tables with timestamps and version information. Alternatively, it can be a distributed file system, storing the state variables of each computational unit at different time periods as independent binary files or NetCDF files. The corrected model state for the previous period is the set of state variables that best reflects the true physical conditions of the watershed, obtained by combining real-time observation data and physical conservation laws for error correction and consistency verification at the end of the previous forecast period. These states can be a set of values ​​for key hydrophysical variables such as soil moisture content, groundwater storage, river runoff, and snow depth. They can also be various internal storage quantities, energy quantities, etc., which evolve over time during model operation. The initial forecast conditions are the set of all internal state variables required by the physical mechanism model to begin simulation for the current forecast period. These conditions determine the starting point of the model simulation and are crucial to the accuracy of the forecast results. The initial forecast conditions can be directly derived from the corrected model states of the previous period without additional processing. Alternatively, after reading the corrected model states of the previous period, fine-tuning can be performed according to the specific needs of the current period (such as considering short-term evaporation).

[0103] The watershed-scale physical mechanism model is a mathematical model built upon hydrophysical laws (such as conservation of mass, energy, and momentum) to simulate hydrophysical processes within the watershed, including rainfall-runoff, evaporation, infiltration, groundwater movement, and river confluence. This model can be a distributed hydrological model, such as SWAT (Soil and Water Assessment Tool) or VIC (Variable Infiltration Capacity) models, which can simulate hydrological processes in different spatial units within the watershed. It can also be a semi-distributed or lumped hydrological model, such as the Xin'anjiang model or the SAC-SMA model, which consider the physical characteristics of the watershed to varying degrees. The predicted flow result is the runoff value at a specific time point or time period obtained by simulating hydrophysical processes under given input conditions and initial states. Examples include the outflow of a computational unit, such as grid runoff in a grid cell or the outlet flow of a sub-watershed. Alternatively, it can be a flow process curve of a river cross-section, representing the change in flow over time. The preset time resolution refers to the time interval during which internal physical state variables are recorded and output in the forecast calculation process performed by the physical mechanism model. This resolution can be set according to the needs of subsequent error analysis and state correction. For example, it can be an hour, half an hour, or shorter time interval to capture rapidly changing hydrological processes. It can also be a day or a longer time interval, suitable for simulating slower-changing hydrological processes. The internal physical state sequence is a set of snapshots of the changes over time of a series of key internal state variables (such as soil moisture content, groundwater level, river storage, etc.) recorded by the physical mechanism model according to the preset time resolution during the forecast calculation process. This sequence can include soil moisture content at various depths, groundwater level, river node flow and level, etc. It can also include process variables such as evapotranspiration, infiltration rate, and surface runoff. Although these variables are not "states" in the strict sense, they are of great significance for subsequent error analysis and correction.

[0104] Through the above technical solution, this application achieves efficient and accurate forecast calculations for computational units in the target watershed that require physical mechanism model forecasting. Utilizing the dynamic allocation information of the model routing diagram, units requiring coupled-mode calculations can be precisely selected, avoiding unnecessary complex physical calculations across the entire watershed and improving computational efficiency. Reading the corrected model state from the previous time period from the physically verified and corrected model state library as the initial forecast condition ensures the accuracy and physical consistency of the forecast starting point, effectively suppressing the accumulation and propagation of errors. Driving the watershed-scale physical mechanism model to perform forecast calculations and generating mechanistic forecast flow results and internal physical state sequences at a preset time resolution provides high-time-precision, physically meaningful input data for subsequent generative state error analysis and physical consistency verification. This lays a solid foundation for realizing a flood forecasting method driven by dual mechanistic data, improving the reliability and accuracy of the forecast.

[0105] In some of the solutions described above in this application, a forecast calculation based on a physical mechanism model driven by the initial forecast conditions is proposed to generate mechanistic forecast results. However, in its implementation, computational efficiency may be insufficient, and state capture may be inaccurate, leading to reduced reliability of forecast results under complex conditions. Specifically, the computation time step and output frequency are not dynamically optimized according to the coupling mode, resulting in resource waste or data loss. The iterative operation process lacks an efficient control mechanism, affecting the timeliness of simulation completion. Flow recording and state capture are not synchronized with the preset resolution, resulting in incomplete or distorted internal state sequences, which in turn affects the accuracy of subsequent error analysis and correction.

[0106] To address this, this application further proposes a method that, based on initial forecast conditions, drives a watershed-scale physical mechanism model to perform forecast calculations, outputting the mechanistic forecast flow results of the computational unit and an internal physical state sequence generated at a preset time resolution. Specifically, this includes: setting the computation time step and output frequency of the physical mechanism model within the current forecast period based on configuration parameters of the coupling mode; iteratively running the physical mechanism model at the computation time step based on the initial forecast conditions until the simulation of the current forecast period is completed; recording the flow value of the computational unit at each output frequency during the iterative process to form the mechanistic forecast flow results; and capturing snapshots of all key state variables of the physical mechanism model at the corresponding times to assemble an internal physical state sequence at a preset time resolution.

[0107] In this context, the configuration parameters of the coupling mode refer to the specific settings and instruction set used to guide the operation of the physical mechanism model in a flood forecasting method driven by both mechanistic data and physical mechanisms. These parameters define the model's behavior under a specific coupling mode, such as the model's complexity, the physical processes considered, and the selection of the numerical solver. Specifically, these configuration parameters can be predefined as a series of configuration files containing parameter combinations optimized for different hydrological scenarios or watershed characteristics. For example, for a rainstorm flood scenario, the configuration parameters might focus on surface runoff and runoff generation mechanisms. For a snowmelt flood scenario, they might focus on snowmelt and soil freeze-thaw processes. Furthermore, these configuration parameters can also be dynamically generated parameter sets, adjusted and optimized in real time based on real-time multi-source watershed characteristic data (such as rainfall intensity, soil moisture, and temperature) through machine learning models or expert systems to adapt to constantly changing watershed hydrological conditions.

[0108] The computation time step and output frequency are key parameters for the operation of a physical mechanism model. The computation time step refers to the time interval between each iteration of the numerical simulation, directly affecting the model's computational accuracy, stability, and efficiency. The output frequency refers to the time interval between recording and outputting forecast results (such as flow values) and snapshots of internal state variables during the simulation, determining the granularity of the output data and the temporal resolution of subsequent analysis. The computation time step and output frequency can be set according to the computational stability requirements and forecast accuracy needs of the coupled model. For example, for rapid-response flood processes, the computation time step might be set to minutes or even seconds to capture rapid changes. The output frequency might be set to hours or shorter to provide high temporal resolution forecast results. Alternatively, the computation time step and output frequency can be adaptive, dynamically adjusted based on the rate of change of internal state variables or the intensity of changes in external driving forces. For example, when rainfall intensity increases sharply or river levels rise rapidly, the computation time step is automatically reduced to ensure numerical stability. When the hydrological process is stable, the step size is appropriately increased to improve computational efficiency.

[0109] The iterative execution of the physical mechanism model refers to the process by which the physical mechanism model, starting from the initial forecast conditions, gradually advances the simulation time by repeatedly executing its internal physical equations and numerical algorithms according to a set computation time step, until it covers the entire forecast period. Specifically, explicit or implicit numerical solvers, such as the finite difference method, finite volume method, or finite element method, can be used to discretize and solve the hydrophysical equations (such as the Saint-Venant equation, Darcy's law, and energy conservation equations). Within each time step, the model calculates the state variables for the next time step based on the current state and external inputs (such as rainfall and evaporation). Furthermore, this iterative process can also be implemented using parallel or distributed computing frameworks, dividing the watershed into multiple sub-regions. The model calculations for each sub-region are performed on independent computing nodes, and information is exchanged through boundary conditions, thereby accelerating the simulation process for the entire watershed.

[0110] This method records the flow rate of the computational unit at each output frequency to form the mechanistic forecast flow rate result. It refers to extracting and storing the instantaneous flow rate values ​​of a specific computational unit (e.g., a watershed outlet section or key control point) from within the model according to a preset output frequency during the iterative operation of the physical mechanism model. These flow rate values ​​are arranged in chronological order, forming the mechanistic forecast flow rate process line for that unit. Specifically, a data acquisition module can be set up within the model to automatically extract the runoff output of the computational unit from the model's state variables each time the output frequency is reached, and write it to the forecast result database or file. Alternatively, it can be implemented through an event-triggered mechanism. When the simulation time reaches a preset output time, a callback function or script is triggered. This function is responsible for reading the flow rate data from the model's memory and performing necessary format conversions and storage.

[0111] Capturing a snapshot of all key state variables of the physical mechanism model at a given time point, and assembling it into an internal physical state sequence according to a preset time resolution, refers to recording all or selected key state variables (such as soil moisture content, groundwater level, river storage capacity, surface runoff depth, etc.) of the model at a specific time point during the iterative operation of the physical mechanism model, according to a preset time resolution (usually the same as or related to the output frequency), forming a "snapshot," and combining these snapshots into a sequence in chronological order. Specifically, the model can serialize the current values ​​of all key state variables at each preset time resolution point (e.g., store them as multidimensional arrays, matrices, or specific data structures), timestamp them, and then store them in memory or on disk. Alternatively, incremental storage or compression techniques can be used to record only the changes in state variables or to efficiently compress state snapshots to reduce storage space and improve I / O efficiency, while ensuring that the internal physical state sequence can be fully restored or reconstructed when needed.

[0112] Through the above technical solutions, this application effectively solves the problems of low computational efficiency and inaccurate state capture, ensuring greater reliability of the forecasting process under complex conditions. Specifically, by setting the computation time step and output frequency based on the configuration parameters of the coupled mode, the computation parameters can be dynamically adjusted according to the actual model requirements, avoiding resource waste or data redundancy caused by fixed step sizes, thereby improving computational efficiency. Iteratively running the model step-by-step based on the initial forecast conditions ensures that the simulation process progresses gradually from the initial state, efficiently completing the entire forecast period and preventing iteration interruptions or delays. Recording the flow value at each output frequency during the iterative process accurately captures flow change points, forming continuous mechanistic forecast flow results, providing an accurate data foundation for subsequent fusion. Capturing snapshots of all key state variables of the physical mechanism model at corresponding times and assembling them into an internal physical state sequence according to a preset time resolution ensures the integrity and consistency of the state data in the time dimension, facilitating subsequent error generation and physical verification, and improving the overall reliability of the forecast.

[0113] In some of the above-mentioned schemes in this application, a forecast calculation is proposed based on the forecast initial conditions to drive the physical mechanism model to perform forecast calculations in order to output the mechanism forecast flow results and the internal physical state sequence. However, in this process, the state capture may not be timely or accurate when iterating the model, resulting in an incomplete or inconsistent internal state sequence, which affects the subsequent error generation analysis and physical verification.

[0114] To address this, this application further proposes an iterative approach based on initial forecast conditions, iteratively running the physical mechanism model at computational time steps until the simulation of the current forecast period is completed. This includes: loading the initial forecast conditions into the physical mechanism model and initiating the simulation of the current forecast period. After the calculation at each computational time step is completed, the following process is executed: recording the flow value calculated by the physical mechanism model at the current moment. Determining whether the current moment is a state snapshot capture point based on a preset time resolution. If so, before updating the global state of the physical mechanism model, synchronously capturing all key state variables of the physical mechanism model at the current moment, forming a state variable snapshot and temporarily storing it. Based on the computational results of the computational time steps, updating the global state of the physical mechanism model and advancing the simulation time. Repeating the above process until the simulation time reaches the duration of the current forecast period.

[0115] Specifically, loading the initial forecast conditions into the physical mechanism model and initiating the simulation for the current forecast period aims to provide an accurate and calibrated starting state for the entire simulation process. This can be achieved in several ways. For example, the values ​​from the initial forecast conditions can be directly assigned to the corresponding state variables within the physical mechanism model, such as soil moisture content and groundwater level. Alternatively, the initial forecast conditions can be input as parameters by calling the physical mechanism model's initialization interface, allowing the model to automatically set the state variables and configure the simulation environment.

[0116] After the computation of each computational time step is completed, subsequent processes are executed, ensuring that necessary checks and data operations are performed at each discrete time point within the computational logic of the physical mechanism model. This can be achieved by immediately calling a post-processing function or module after each numerical integration or iterative calculation of a computational time step is completed in the main loop of the physical mechanism model. Alternatively, after the computation of the physical mechanism model completes the calculation of the current computational time step, an event or callback is triggered, which is captured by external control logic and subsequent steps are executed.

[0117] Recording the flow rate value calculated by the physical mechanism model for the current computational unit is to directly obtain and store the model's main output at each time step. This typically involves the physical mechanism model writing the instantaneous flow rate value (e.g., river cross-sectional flow, outflow) of its internally calculated computational unit to a pre-defined flow data buffer or file after completing the calculation for the current time step. Alternatively, the model can query and obtain the flow rate output variables of the computational unit through its provided API or data access interface and store them in a time series database.

[0118] This provides a flexible and efficient mechanism for acquiring state data by determining whether the current moment is a state snapshot capture point based on a preset time resolution. This can be achieved by maintaining a counter or timestamp, checking whether the current simulation time is aligned with an integer multiple of the preset time resolution after each calculation time step. For example, if the preset time resolution is 1 hour, check if the current time is an hour. Alternatively, a list of time points can be defined, containing all specific moments for which state snapshots need to be captured, and checking whether the current simulation time is in this list after each calculation time step.

[0119] If it is a state snapshot capture point, then before updating the global state of the physical mechanism model, simultaneously capturing all key state variables of the physical mechanism model at the current moment, forming a state variable snapshot and temporarily storing it, is a core step to ensure the integrity and consistency of the internal physical state sequence. This operation can be performed within the physical mechanism model. When a state snapshot capture point is determined, a dedicated state serialization function is immediately called. This function iterates through all key state variables of the physical mechanism model (such as soil moisture content, groundwater level, river water storage, evapotranspiration, etc.), packages their current values ​​into a data structure, and stores it in a temporary list in memory or a disk file. Alternatively, using memory snapshot technology or a deep copy mechanism, a complete copy of the entire state object or related data structure of the physical mechanism model is performed before the global state update instruction of the physical mechanism model is executed, and this copy is temporarily stored as a state variable snapshot. By capturing before updating the global state, it can be ensured that the captured state is the true instantaneous state at the end of the current calculation step, avoiding data inconsistency caused by subsequent update operations.

[0120] The standard advancement mechanism for physical mechanism model simulations is to update the global state of the physical mechanism model based on the computation results within the current time step and advance the simulation time. The physical mechanism model updates its internally stored global state variables based on the fluxes and changes of various physical processes (such as rainfall infiltration, surface runoff, evapotranspiration, and groundwater flow) calculated within the current time step, and advances the simulation clock forward by one computation time step. Alternatively, the physical mechanism model's internal state update function can be invoked. This function adjusts the global state of the physical mechanism model using appropriate physical equations and numerical methods based on the computation results of the current time step and automatically updates the internal time variables.

[0121] Repeat the above process until the simulation time reaches the duration of the current forecast period, ensuring a complete simulation of the entire forecast period. This is typically achieved through a loop structure that encapsulates the above steps, with the loop condition being that the current simulation time is less than or equal to the end time of the current forecast period. Alternatively, a predetermined simulation end time can be set, and the control module of the physical mechanism model will automatically manage the iteration process, terminating the simulation automatically when the internal simulation time reaches or exceeds this end time.

[0122] Through the above technical solution, this application effectively solves the problem that untimely or inaccurate state capture may occur during the iterative operation of the physical mechanism model, leading to incomplete or inconsistent internal state sequences. Specifically, by performing precise control after each computation time step and dynamically determining the state snapshot capture point according to a preset time resolution, unnecessary storage overhead is avoided, while ensuring the acquisition of state data at critical moments. Crucially, all key state variables are captured synchronously before updating the global state of the physical mechanism model, ensuring that the acquired state snapshot represents the true, consistent, and complete internal state of the model at the current moment, avoiding potential data deviations caused by the state update order. This precise and timely state capture mechanism provides a high-quality, highly reliable internal physical state sequence for subsequent generative state error model state error generation analysis, thereby improving the accuracy and robustness of the entire flood forecasting method. It also provides a solid data foundation for the physical consistency mandatory checker, enabling subsequent error correction and physical verification to be based on accurate internal states, thus improving the overall forecast reliability.

[0123] In some of the embodiments described above in this application, a state error generation analysis is proposed to obtain a physical state correction quantity through a generative state error model, which is used to correct the state variables of the physical mechanism model. However, in this process, the spatiotemporal inconsistency between the internal physical state sequence and the real-time observation data may lead to inaccurate bias pattern learning, and the generated initial correction quantity may not match the physical mechanism model in terms of dimension or physical meaning, affecting the effectiveness of the correction.

[0124] To address this, this application further proposes to perform state error generation analysis on the internal physical state sequence and real-time observation data using a generative state error model to obtain physical state correction quantities. See [link to relevant documentation]. Figure 5 Taking the server as the executing entity as an example, the method includes: 501. The internal physical state sequence is spatiotemporally aligned and feature-fused with the corresponding real-time observation data to form a fused data sample for error analysis.

[0125] 502. Input the fused data sample into the generative state error model, learn the deviation pattern between the internal physical state sequence and the real-time observation data through the generative state error model, and generate the initial physical state correction.

[0126] 503. Standardize the initial physical state correction to obtain a physical state correction that matches the state variables of the physical mechanism model in both dimension and physical meaning.

[0127] Specifically, the internal physical state sequence is spatiotemporally aligned and feature-fused with the corresponding real-time observation data to form a fused data sample for error analysis. Spatiotemporal alignment aims to eliminate inconsistencies between different data sources in time and space. For example, temporal interpolation (such as linear interpolation or spline interpolation) can align the real-time observation data and the internal physical state sequence to a unified time step. Spatially, spatial overlay analysis, resampling, or gridding techniques from Geographic Information Systems (GIS) can unify data with different spatial resolutions or distributions onto the same computational unit or grid. This ensures the accuracy of subsequent error analysis and avoids misjudgments caused by data misalignment. Feature fusion aims to integrate the complementary information contained in the internal physical state sequence and the real-time observation data to provide a more comprehensive and robust foundation for error analysis. For example, simple data concatenation, weighted averaging, or deep learning models (such as convolutional neural networks or recurrent neural networks) can be used for feature extraction and fusion to integrate the key information of the two types of data into a unified feature vector or tensor. This allows the error model to understand the differences between the system state and the actual observations from a richer perspective. The fused data sample is a dataset that has undergone spatiotemporal alignment and feature fusion. Its structure and content are standardized, and it can be directly used as input to a generative state error model. This sample contains internal state information simulated by the physical model and corresponding real observation information, providing a direct basis for the model to learn the deviation pattern between the two.

[0128] The fused data sample is input into the generative state error model. This model learns the deviation patterns between the internal physical state sequence and the real-time observation data, and generates initial physical state corrections. The generative state error model is a machine learning model capable of learning data distribution and generating new samples. Here, it is used to capture complex, nonlinear deviation patterns between the simulation results of the physical mechanism model and the actual observation data. Its implementation can include, but is not limited to: an architecture based on Generative Adversarial Networks (GANs), where the generator learns to generate corrections to deceive the discriminator, making it unable to distinguish between real and generated deviations; an architecture based on Variational Autoencoders (VAEs), which learns a latent space to generate corrections that conform to the deviation distribution; or an architecture based on a Diffusion Model, which generates corrections through a progressive denoising process. This model can identify systematic error trends, random error fluctuations, and error characteristics under different scenarios from the fused data sample. Learning deviation patterns refers to the generative state error model automatically discovering and internalizing various discrepancies between the internal physical state sequence and the real-time observation data from a large number of fused data samples through the training process. These bias patterns may include systematic overestimation or underestimation, lag or lead in response to specific events (such as heavy rainfall), and spatial inconsistencies. By learning these patterns, the model can understand the sources and characteristics of the error. The initial physical state corrections are the raw correction values ​​output by the generative state error model for the current fused data sample based on the learned bias patterns. These corrections are typically high-dimensional and may not yet fully match the dimensions and spatial structure of the physical mechanism model's state variables, but they contain the model's preliminary estimate of the current state bias.

[0129] The initial physical state correction is standardized to obtain a correction value that matches the state variables of the physical mechanism model in both dimension and physical meaning. The standardization transformation aims to adjust the initial physical state correction output by the generative state error model to a form fully compatible with the internal state variables of the physical mechanism model. This includes unifying the dimensions, matching the spatial scale, and verifying physical rationality. For example, dimension unification can be achieved through linear scaling, nonlinear transformation, and unit conversion. Spatial scale matching can be achieved through spatial interpolation, aggregation, or decomposition. Matching the dimension and physical meaning means that the correction value is consistent with the state variables accepted by the physical mechanism model in terms of numerical magnitude, unit, physical meaning (e.g., whether it corrects soil moisture content, groundwater level, or river flow), and spatial distribution structure. This matching is crucial to ensuring that the correction value can be correctly understood and applied by the physical mechanism model, avoiding the introduction of new errors or model instability due to mismatch. The physical state correction value, after standardization, can be directly used to correct the state variables of the physical mechanism model. These corrections not only reflect the errors numerically, but are also physically reasonable, effectively improving the simulation accuracy and reliability of physical mechanism models.

[0130] Through the above technical solution, this application effectively solves the problems of inaccurate bias pattern learning caused by the spatiotemporal inconsistency between the internal physical state sequence and real-time observation data, and the potential mismatch between the dimension or physical meaning of the generated initial correction amount and the physical mechanism model, thus affecting the effectiveness of correction. Specifically, by accurately aligning and fusing the internal physical state sequence with the corresponding real-time observation data in a spatiotemporal manner, a high-quality fused data sample is constructed. This fundamentally eliminates the spatiotemporal misalignment between data sources, ensuring a high degree of consistency and integrity of the input data for error analysis, laying a solid foundation for subsequent bias pattern learning. This fused data sample is then input into a generative state error model, which can efficiently learn and capture the complex nonlinear bias patterns between the internal physical state sequence and real-time observation data, and generate preliminary correction amounts. This data-driven generative method makes the error compensation process more intelligent and adaptive, capable of handling error characteristics under various complex scenarios. The generated initial physical state correction amount is standardized to ensure that it fully matches the state variables of the physical mechanism model in terms of dimension and physical meaning. This crucial step ensures that the correction quantities can be accurately received and applied by the physical mechanism model, avoiding the introduction of new errors due to dimensional inconsistencies or spatial structure discrepancies, thereby improving the effectiveness and physical rationality of the state variable correction of the physical mechanism model. Overall, this application, through refined data preprocessing, intelligent error pattern learning, and rigorous correction quantity standardization, achieves accurate and physically consistent correction of the state variables of the physical mechanism model, greatly enhancing the accuracy and reliability of flood forecasting.

[0131] In some embodiments described above in this application, a generative state error model is proposed for generating state error generation analysis to produce physical state correction quantities. However, in its implementation, directly processing fused data samples may not effectively capture the spatiotemporal characteristics and physical correlation of the deviation pattern, resulting in inaccurate or physically inconsistent generated correction quantities, which in turn affects the accuracy of subsequent flood forecasts.

[0132] To address this, this application further proposes inputting fused data samples into a generative state error model. The generative state error model learns the deviation pattern between the internal physical state sequence and real-time observation data, and generates an initial physical state correction. Specifically, this process includes: constructing an explicit state deviation field between the internal physical state sequence and real-time observation data based on the fused data samples, serving as the conditional input to the generative state error model; inferring and generating a high-dimensional state correction potential field based on the spatiotemporal context of the explicit state deviation field and the internal physical state sequence, through the conditional generation mechanism of the generative state error model; and mapping the high-dimensional state correction potential field to the physical space and dimensions corresponding to the state variables of the physical mechanism model, forming the initial physical state correction.

[0133] This study constructs an explicit state deviation field between the internal physical state sequence and real-time observation data based on fused data samples. This field serves as the conditional input for the generative state error model, aiming to clarify the differences between the prediction results of the quantified physical mechanism model and actual observations. By spatiotemporally aligning the internal physical state sequence with the real-time observation data, the deviation can be calculated for the corresponding physical quantity (e.g., soil moisture content, groundwater level, runoff depth, etc.) at each spatiotemporal point. This deviation can be an absolute difference, a relative difference, or a standardized difference. For example, the simulated soil moisture content and the actual observed soil moisture content of each computational unit can be compared point-by-point in each forecast period. Based on these discrete deviation values, a continuous deviation field covering the entire target watershed can be generated through spatial interpolation (e.g., Kriging interpolation, inverse distance weighted interpolation) or gridding. This deviation field intuitively reflects the error distribution of the model in different regions and at different time points, providing a structured, high signal-to-noise ratio error signal for the generative state error model, avoiding the difficulty of the model extracting deviation patterns from the original, complex fused data samples.

[0134] Based on the explicit state bias field and the spatiotemporal context of the internal physical state sequence, a high-dimensional state correction potential field is inferred and generated through the conditional generation mechanism of a generative state error model. The aim is to utilize explicit error information and the dynamic evolution background of the physical model to generate an abstract yet comprehensive correction scheme. The explicit state bias field provides the error distribution at the current moment, while the spatiotemporal context of the internal physical state sequence includes the model's trajectory over a past period, the evolution trend of state variables, and the physical correlations between different regions. The generative state error model can employ architectures such as Conditional Generative Adversarial Networks (CGANs) or Conditional Variational Autoencoders (CVAEs). For example, in CGAN, the generator takes the explicit state bias field and spatiotemporal context as conditional inputs and learns to generate a high-dimensional potential field capable of correcting these biases, while the discriminator distinguishes the generated potential field from the true correction pattern. The introduction of the spatiotemporal context allows the generator to consider not only the current error but also historical evolution patterns and spatial correlations when generating the correction potential field, ensuring that the generated correction has physical rationality and temporal coherence.

[0135] Mapping a high-dimensional state correction latent field to the physical space and dimensions corresponding to the state variables of the physical mechanism model forms the initial physical state correction quantity. The aim is to transform the abstract correction latent field into a concrete physical quantity that can be directly applied to the physical mechanism model. The high-dimensional state correction latent field is typically an abstract representation learned by a generative model in the latent space, and its dimension and physical meaning may not directly correspond to the state variables of the physical mechanism model. Therefore, a mapping layer or decoder is needed to decode the information in the latent field into specific physical state correction quantities, such as correction values ​​for soil moisture content or groundwater level. This mapping process needs to ensure that the dimensions, units, and physical meaning of the correction quantity are completely consistent with the state variables of the physical mechanism model. For example, if a certain dimension in the latent field represents a "soil wettability adjustment factor," the mapping layer will convert it into a specific "soil moisture content increment (unit: millimeters)" and ensure that its spatial distribution matches the grid or cell structure of the physical mechanism model.

[0136] Through the above technical solution, this application effectively solves the problem of insufficient capture of deviation patterns when directly processing fused data samples. By constructing an explicit state deviation field, a clear and structured error signal is provided for the generative state error model, avoiding the ambiguity and uncertainty of the model extracting deviation patterns from the original data, thus ensuring accurate identification of deviation patterns. Based on the spatiotemporal context of the explicit state deviation field and the internal physical state sequence, a high-dimensional state correction potential field is inferred and generated through the conditional generation mechanism of the generative state error model. This allows the correction amount to dynamically adapt to the physical changes in the watershed, enhancing the accuracy and physical rationality of the correction by utilizing the spatial and temporal correlation characteristics. Mapping the high-dimensional state correction potential field to the physical space and dimensions corresponding to the state variables of the physical mechanism model forms the initial physical state correction amount. This ensures that the correction amount is consistent with the dimensions and spatial structure of the physical model, facilitating subsequent physical verification and integration, preventing the correction amount from deviating from physical constraints, improving the accuracy and physical consistency of the initial physical state correction amount, and thus improving the overall reliability of flood forecasting.

[0137] In some of the embodiments described above in this application, a spatiotemporal context based on explicit state deviation fields and internal physical state sequences is proposed. Through the conditional generation mechanism of a generative state error model, a high-dimensional state correction potential field is inferred and generated to improve the accuracy and physical rationality of state correction. However, in its implementation, due to the lack of sufficient consideration for the extraction and fusion of multi-scale spatiotemporal features and the lack of direct guidance from physical laws, the generated state correction potential field may fail to accurately capture complex deviation patterns or produce physically unreliable correction results under extreme scenarios, thereby affecting the reliability of subsequent flood forecasts.

[0138] To address this, this application further proposes a method for inferring and generating a high-dimensional state correction potential field based on the spatiotemporal context of an explicit state bias field and an internal physical state sequence, using a conditional generation mechanism of a generative state error model. This involves: extracting multi-scale spatiotemporal features from both the explicit state bias field and the internal physical state sequence to obtain multi-scale bias features and multi-scale state features; synergistically fusing the multi-scale bias features and multi-scale state features to construct fused conditional features for guiding generation; and iteratively sampling based on these fused conditional features using the conditional generation mechanism, introducing physical law-based generation guidance during the sampling process to output the high-dimensional state correction potential field.

[0139] Specifically, multi-scale spatiotemporal feature extraction is performed on the explicit state deviation field and the internal physical state sequence to capture key information from the data at different granularities. Multi-scale spatiotemporal feature extraction refers to analyzing data at different spatial resolutions and temporal granularities to reveal patterns and relationships that may be overlooked at a single scale. Its role is to comprehensively and meticulously acquire information on deviations and physical states, which is crucial for accurately modeling complex watershed hydrological processes and their errors. Various techniques can be employed for implementation. For example, in the spatial dimension, different sizes of convolutional kernels or multi-layer pooling operations in convolutional neural networks (CNNs) can be used to extract features at different spatial scales, or wavelet transform, multi-resolution analysis, and other methods can be applied to decompose spatial data into different frequency components. In the temporal dimension, recurrent neural networks (RNNs) such as Long Short-Term Memory (LSTM) networks or gated recurrent units (GRUs) can be used to capture short-term and long-term temporal dependencies by setting different sequence lengths or introducing attention mechanisms, or time series decomposition methods (such as STL decomposition) can be used to separate trends, seasonality, and residuals at different time scales.

[0140] Building upon this, the multi-scale bias features and multi-scale state features are collaboratively fused to construct fusion conditional features for guiding generation. Collaborative fusion refers to the effective integration of features extracted from different scales and sources (bias fields and physical state sequences) to form a more comprehensive and information-rich representation. Its role is to provide a rich and contextually informative conditional input for the generative state error model, ensuring that the subsequently generated state correction potential field can fully utilize the complementarity of bias information and physical states, thereby improving the accuracy of correction. In terms of implementation, one approach is to concatenate the multi-scale bias features and multi-scale state features along the feature dimension, learning their complex interactions through fully connected or convolutional layers and mapping them to a unified feature space. Another approach is to employ attention mechanisms, such as cross-attention or self-attention, to dynamically evaluate the importance of features of different scales and types (bias features and state features), assigning higher weights to more relevant features during the fusion process, thereby achieving adaptive feature fusion.

[0141] Furthermore, based on this fusion condition feature, iterative sampling is performed through this conditional generation mechanism, and physical law-based generation guidance is introduced during the sampling process to output the high-dimensional state correction potential field. The conditional generation mechanism refers to a model architecture capable of generating new data based on given conditional inputs, such as a Conditional Generative Adversarial Network (cGAN) or a Conditional Variational Autoencoder (cVAE). Iterative sampling refers to the generative model gradually generating outputs through multiple iterations or refinements, with each iteration potentially improving upon the previous result. Introducing physical law-based generation guidance means integrating known physical conservation laws, physical boundary conditions, or inherent constraints of physical processes into the training or sampling logic of the generative model during the generation process. Its purpose is to ensure that the generated high-dimensional state correction potential field not only accurately reflects observed biases but is also physically reasonable and consistent, avoiding correction results that violate physical laws, thereby improving the reliability of flood forecasts. Various strategies can be employed in its implementation. For example, physical constraints, such as mass conservation, energy conservation, or nonnegativity, can be added to the loss function of the generative model, causing the model to tend to generate correction quantities that satisfy these physical constraints during iterative sampling. Alternatively, after each iterative sampling, the generated potential correction quantities can be physically projected or post-processed to force them to be adjusted to a physically acceptable range. For example, if the correction quantity causes a physical variable to exceed its reasonable range, it can be truncated or redistributed. Furthermore, hybrid generative models can be designed to embed parts of the physical simulator or physical equations into the structure of the generative network, allowing it to directly follow physical laws during the generation process.

[0142] Through the aforementioned technical solutions, this application can capture subtle changes in explicit state deviation fields and internal physical state sequences from multiple scales and dimensions, ensuring that the information relied upon for generating high-dimensional state correction potential fields is comprehensive and detailed. By synergistically fusing these multi-scale features, richer and more contextually informative conditional inputs are provided to generative state error models, enabling the models to more accurately understand the sources of deviations and the evolution of physical states. More importantly, the introduction of physical law-based generation guidance during iterative sampling effectively solves the problem that traditional data-driven models may produce physically unreasonable correction results under extreme scenarios. This ensures that the generated state correction potential field not only accurately captures complex deviation patterns but also strictly follows physical conservation laws, thereby improving the accuracy, physical rationality, and reliability of state corrections in flood forecasting.

[0143] In some of the embodiments described above in this application, a standardization transformation of the initial physical state correction quantity is proposed to generate a correction quantity that matches the state variables of the physical mechanism model. However, in the implementation process, the initial correction quantity may have inconsistent dimensions, mismatched spatial scales, or contain physically unacceptable components due to the output of the generative state error model, which may cause the correction result to be unable to be effectively integrated into the physical mechanism model, affecting the accuracy and reliability of flood forecasting.

[0144] To address this, this application further proposes a standardization transformation of the initial physical state correction to obtain a physical state correction that matches the state variables of the physical mechanism model in terms of dimension and physical meaning. This standardization transformation includes: performing dimensional unification on the initial physical state correction, converting the dimensions output by the generative state error model into the standard physical dimensions used by the physical mechanism model. Based on the spatial distribution definition of the state variables of the physical mechanism model, an adaptive spatial scale transformation is performed on the dimensionally unified physical state correction to obtain an intermediate correction that matches the spatial organization structure of the state variables of the physical mechanism model. A physical permissibility check is performed on this intermediate correction, removing correction components that exceed the preset physical feasible region or violate the preset physical monotonicity relationship, and then outputting the physical state correction.

[0145] Specifically, dimensional unification is performed on the initial physical state correction to ensure data consistency across different models. Generative state error models may output dimensionless correction values ​​or dimensions incompatible with the physical mechanism model, such as percentage deviations instead of absolute physical quantities. Dimension unification aims to convert these correction values ​​into physical dimensions that the physical mechanism model can directly understand and apply, such as water depth, flow rate, and water content. This process can be achieved through multiplication or division using predefined dimensional conversion factors. For example, if the generative model outputs a relative deviation, it needs to be multiplied by the current or reference value of the corresponding state variable in the physical mechanism model and assigned the correct physical dimension. Alternatively, it can be implemented through lookup tables or rule-based mapping mechanisms. For instance, a mapping table can be created to map specific codes or ranges output by the generative model to the corresponding standard physical dimensions and units in the physical mechanism model.

[0146] Based on this, and according to the spatial distribution definition of the state variables of the physical mechanism model, an adaptive spatial scale transformation is performed on the dimensionally unified physical state correction quantity to address the differences in spatial resolution or discretization methods between different models. Generative state error models may generate correction quantities at different spatial grids, watershed partitions, or discrete points, while the physical mechanism model has its fixed computational units and state variable definitions. The adaptive spatial scale transformation aims to convert the correction quantity from its original spatial representation to a spatial organization structure that precisely matches the state variables of the physical mechanism model, ensuring that the correction can accurately act on the corresponding spatial location of the physical mechanism model. This transformation can employ spatial interpolation techniques, such as Kriging interpolation, inverse distance weighted interpolation, or spline interpolation, to interpolate the dimensionally unified correction quantity based on the discrete points or grid centers of the state variables of the physical mechanism model in space, generating a continuous or discrete correction field at the spatial resolution of the physical mechanism model. Simultaneously, it can also be achieved through aggregation or decomposition operations. For example, if the resolution of the correction quantity output by the generative model is higher than that of the physical mechanism model, multiple correction quantities can be spatially averaged or weighted aggregated. If the value is lower than that of the physical mechanism model, the correction amount can be spatially decomposed or replicated and refined by combining local features.

[0147] Furthermore, a physical permissibility test is performed on this intermediate correction to ensure that the correction is physically reasonable. While generative models are powerful, their outputs may not always conform to the constraints of physical laws or actual physical processes. The physical permissibility test aims to identify, correct, or remove correction components that cause the state variables of the physical mechanism model to exceed reasonable ranges (e.g., water content cannot be negative, flow rate cannot exceed the maximum carrying capacity of the watershed) or violate physical monotonicity (e.g., soil moisture content increases instead of decreasing after rainfall), thereby ensuring the physical consistency of the corrected physical state variables. This test can be implemented by setting hard thresholds or range checks. For example, for soil moisture correction, ensure that the corrected moisture content is not lower than 0 and not higher than the saturation moisture content. For flow rate correction, ensure that the corrected flow rate is not negative. Correction components that exceed the range can be directly truncated to boundary values ​​or marked as invalid. Alternatively, physical monotonicity or trend constraints can be introduced. For example, when correcting soil moisture content, if the moisture content in the upstream area increases, the moisture content correction in the downstream area should also tend to increase or remain unchanged, rather than decrease significantly. For correction components that violate these physical trends, they can be smoothed, weighted averaged, or removed directly.

[0148] Through the above technical solution, this application can effectively solve the mismatch problem in terms of dimensions, spatial scale, and physical feasibility of the initial physical state correction, thereby ensuring that the correction can be seamlessly integrated into the physical mechanism model and improving the physical consistency and accuracy of flood forecasts. Specifically, by performing dimension unification processing on the initial physical state correction, the dimensions output by the generative state error model are converted into the standard physical dimensions used by the physical mechanism model. This solves the problem that the correction cannot be directly applied due to dimensional differences between different models, laying the foundation for the effectiveness of subsequent corrections. On this basis, based on the spatial distribution definition of the state variables of the physical mechanism model, an adaptive spatial scale transformation is performed on the dimension-unified physical state correction, solving the defect of mismatch between the spatial organization structure of the initial correction and the physical model. Through adaptive adjustment, spatial scale alignment is achieved, avoiding forecast bias caused by scale inconsistency. Furthermore, a physical permissibility check is performed on intermediate corrections, removing correction components that exceed the preset physical feasible region or violate the preset physical monotonicity relationship. This effectively addresses the risk of physically unreasonable components that may arise from generative models. Forced verification ensures that the corrections remain within the range of physical laws, thus outputting physical state corrections that perfectly match the state variables of the physical mechanism model. This multi-stage standardization conversion mechanism enables deep coupling between the generative state error model and the physical mechanism model, ensuring that the corrected model state reflects both data-driven error correction and strictly adheres to physical conservation laws, thereby improving the reliability and accuracy of flood forecasting.

[0149] In some embodiments described above, a method is proposed to verify and filter the physical state correction amount to obtain an effective state correction amount. However, during implementation, the physical state correction amount may violate physical conservation laws, leading to unreliable prediction results. Therefore, this application further proposes a method to obtain an effective state correction amount by performing verification and filtering based on physical conservation laws on the physical state correction amount using a physical consistency enforcement checker. See [link to relevant documentation]. Figure 6 The method includes the following steps: 601. Input the physical state correction amount and the current state variable of the physical mechanism model into the physical consistency enforcement checker.

[0150] The physical state correction, output by the generative state error model, is used to correct the state variables of the physical mechanism model, such as adjustments to soil moisture content, groundwater level, or river flow. The current state variables of the physical mechanism model refer to the internal physical state of the model at the start of the current forecast period or at a specific moment, such as soil moisture distribution, groundwater storage, and river water level within the watershed. These variables collectively describe the instantaneous physical condition of the watershed hydrological system. The physical consistency enforcement checker is a module or algorithm specifically designed to evaluate and enforce the correction against physical laws. Its core function is to ensure that any modification to the model state does not violate fundamental physical conservation principles. Inputting the physical state correction and the current state variables of the physical mechanism model into the checker provides a complete context, allowing the checker to evaluate the reasonableness of the correction under the current physical state, rather than simply examining the correction itself in isolation. For example, the checker needs to know the current soil moisture content to determine whether a soil moisture content correction would cause the moisture content to exceed saturation or fall below zero.

[0151] 602. In this physical consistency mandatory checker, a set of check equations for the correction amount of the physical state is constructed based on the laws of conservation of mass and conservation of momentum.

[0152] In hydrological processes, the law of conservation of mass means that the input and output of water bodies within a watershed, as well as changes in internal storage, must remain in balance. For example, changes in soil moisture content in a region must correspond to changes in net infiltration, evaporation, and runoff. The law of conservation of momentum describes the relationship between changes in the velocity and direction of water flow and the forces acting upon it. For instance, the acceleration or deceleration of river flow must conform to factors such as hydraulic gradient and roughness. This set of verification equations is a series of mathematical equations derived from these physical laws. They correlate physical state corrections with the current state variables of the physical mechanism model, quantifying whether applying a certain correction would cause the physical system to violate conservation laws.

[0153] 603. Solve the set of verification equations to obtain the theoretical conservation deviation corresponding to the physical state correction.

[0154] Solving the verification equations involves using numerical or analytical methods to calculate the degree of theoretical deviation of the physical system from the conservation of mass and momentum after applying a given physical state correction. This theoretical conservation deviation is an indicator of the degree to which the physical state correction violates the physical conservation laws. If the deviation is zero or within an acceptable range, the correction is physically consistent. If the deviation is too large, it indicates that the correction may lead to physically unreasonable results.

[0155] 604. Compare the theoretical conservation deviation with the preset tolerance threshold, and based on the comparison result, filter or correct the physical state correction amount, and output the portion that meets the tolerance threshold as the effective state correction amount.

[0156] The preset tolerance threshold is a configurable parameter that defines the maximum acceptable range of theoretical conservation deviations. It allows the system to tolerate minor inconsistencies arising from numerical calculations or model simplifications to a certain extent, while rejecting corrections that violate physical laws. This threshold can be adjusted based on watershed characteristics, forecast accuracy requirements, and model sensitivity. Based on the comparison results, physical state corrections with theoretical conservation deviations exceeding the tolerance threshold can be directly discarded (filtered) or adjusted (corrected) to be as close as possible to the original correction while satisfying conservation laws. This may involve optimization algorithms, such as least squares or projection methods, to project the correction into a physically feasible space that satisfies conservation laws. After verification and filtering, physical state corrections that satisfy physical conservation laws (i.e., theoretical conservation deviations within the tolerance threshold) are determined as the effective state corrections.

[0157] Through the above technical solution, this application can provide a complete model state context for verification, avoiding isolated evaluation of correction quantities and thus more accurately judging the physical rationality of correction quantities. By constructing a set of verification equations based on the laws of mass and momentum conservation, strict physical constraints are directly imposed on the correction quantities, ensuring that they conform to the basic physical laws of hydrological processes. By solving the set of verification equations and quantifying the theoretical conservation deviation, an objective basis is provided for subsequent decision-making. Furthermore, by comparing the theoretical conservation deviation with a preset tolerance threshold and filtering or correcting the physical state correction quantities based on the comparison results, dynamic adjustments can be made to eliminate non-conservative parts, while introducing adjustable tolerance to prevent overly strict or lenient filtering. This ensures that the state variables used to correct the physical mechanism model are physically reasonable and consistent, improving the reliability and interpretability of flood forecast results and effectively solving the problem that physical state correction quantities may violate physical conservation laws.

[0158] In some of the embodiments described above in this application, a method for screening or correcting physical state correction quantities is proposed to ensure the effectiveness of the correction based on the law of physical conservation. However, in its implementation, screening or correcting directly based on comparison results may not be able to accurately process each component in the correction quantity, resulting in some components with excessive deviations not being fully adjusted or omitted, thereby affecting the physical consistency of the effective correction quantity and the accuracy of flood forecasting.

[0159] To address this, this application further proposes a method for filtering or modifying the physical state correction based on comparison results, outputting the portion that satisfies the tolerance threshold as the effective state correction. Specifically, this involves: decomposing the physical state correction into multiple independent correction components; comparing the theoretical conservation deviation corresponding to each correction component with the tolerance threshold to identify correction components with excessive deviations and those that meet the requirements; applying constraints based on the conservation law to the correction components with excessive deviations until the adjusted theoretical conservation deviation corresponding to the correction component with excessive deviations falls within the tolerance threshold; and recombining the constrained correction components with the compliant correction components to generate the effective state correction.

[0160] Specifically, the physical state correction is decomposed into multiple independent correction components, aiming to refine the overall correction into more manageable and analyzable parts. This decomposition can be based on various dimensions. For example, it can be based on the type of different state variables in the physical mechanism model (such as soil moisture content, groundwater level, surface runoff, etc.), so that each component corresponds to a specific physical process or state. Alternatively, it can be based on the spatial discretization units of the target watershed (such as grid cells, sub-watershed cells), so that each component represents the correction value for a specific spatial region. Furthermore, it can be decomposed based on the time step or different time points within the forecast period. This decomposition enables refined management of the correction, avoiding coarse processing of the overall correction, thus laying the foundation for subsequent accurate verification and adjustment.

[0161] By comparing the theoretical conservation deviation corresponding to each correction component with the tolerance threshold, the system identifies correction components with deviations exceeding the limit and those that meet the requirements. The aim is to precisely determine which correction components exhibit physical inconsistencies and the degree of these inconsistencies. Specifically, direct numerical comparison can be used, determining whether the absolute value of the theoretical conservation deviation caused by each correction component exceeds the preset tolerance threshold. For example, if the absolute value of the deviation is greater than the threshold, the component is marked as "deviation exceeding the limit," and vice versa. Alternatively, statistical methods can be employed, such as calculating the confidence interval of the deviation to determine whether it falls within an acceptable range. This step-by-step comparison and identification mechanism ensures a comprehensive and thorough review of all correction components, providing a clear target for subsequent targeted adjustments.

[0162] Furthermore, constraints based on the conservation law are applied to the correction components that exceed the limit until the adjusted theoretical conservation deviation corresponding to the correction component that exceeds the limit falls within the tolerance threshold. This aims to correct correction components that do not conform to the physical conservation law, ensuring they meet physical consistency requirements. One implementation method is to use a scaling method, adjusting the value of the correction component proportionally according to the magnitude of the deviation to reduce its violation of the conservation law. Another implementation method is to use an optimization algorithm, for example, constructing an optimization problem with the objective of minimizing the adjustment magnitude of the correction amount while using the conservation law as a constraint, and obtaining the adjusted correction component by solving this problem. An iterative correction method can also be used, where the component that exceeds the limit is slightly adjusted in each iteration, and its theoretical conservation deviation is recalculated until the deviation meets the tolerance threshold. Alternatively, the excess portion can be redistributed to other related, non-exceeding components to maintain overall conservation.

[0163] The adjusted correction component, after constraint adjustment, is recombine with the compliant correction component to generate the effective state correction quantity. The purpose is to integrate the physically verified and adjusted correction components into a complete correction quantity that can be used for model state correction. Specifically, if the correction component is decomposed based on the type of physical state variable, the adjusted state variable correction values ​​can be recombine into a multidimensional vector or data structure. If the correction component is decomposed based on spatial units, the adjusted spatial unit correction values ​​can be reassembled into a spatial distribution field. This recombination ensures that the generated effective state correction quantity is physically consistent overall and accurately reflects the correction requirements for the state variables of the physical mechanism model.

[0164] Through the above technical solution, this application overcomes the limitations that may exist when directly screening or correcting physical state correction quantities based on comparison results. By decomposing the physical state correction quantity into multiple independent correction components, this application achieves fine-grained management and analysis of the correction quantity, avoiding omissions or ambiguities that may occur during overall processing. Each correction component undergoes precise comparison with theoretical conservation deviations and tolerance thresholds, ensuring accurate identification of all inconsistent components. For correction components with deviations exceeding limits, this application applies constraint adjustments based on conservation laws, ensuring that each component strictly conforms to physical laws, thereby effectively eliminating physically unreasonable deviations. This targeted adjustment process, until the adjusted theoretical conservation deviation falls within the tolerance threshold, guarantees the thoroughness and effectiveness of the correction. By recombinating the constrained components with the compliant components, this application generates physically highly consistent and reliable effective state correction quantities. This improves the accuracy and reliability of state variable correction in physical mechanism models, thereby improving the accuracy and credibility of flood forecasts, especially in dealing with complex hydrological scenarios and extreme events, providing more physically plausible forecast results.

[0165] In some embodiments described above, this application proposes to correct the state variables of a physical mechanism model based on an effective state correction amount to improve flood forecast accuracy. However, in its implementation, directly adding the correction amount may cause the state variables to exceed the physically reasonable range, thereby affecting the reliability and accuracy of the forecast. To address this, this application further proposes a method for correcting the state variables of a physical mechanism model based on an effective state correction amount. This method includes: determining the mapping relationship between the state variable to be corrected in the physical mechanism model and the effective state correction amount; based on the mapping relationship, adding the effective state correction amount to the state variable to be corrected using a preset weighting method to obtain a corrected candidate state variable; performing a physical reasonableness test on the corrected candidate state variable to verify whether the candidate state variable is within the reasonable state space defined by the physical mechanism model; and determining the candidate state variable that passes the verification as the corrected state variable.

[0166] Specifically, when determining the mapping relationship between the state variables to be corrected in the physical mechanism model and the effective state correction quantity, the aim is to establish a precise correspondence between the effective state correction quantity and the state variables within the physical mechanism model. This ensures that the correction quantity can accurately act on the target physical quantity, avoiding ineffective corrections due to variable mismatch. For example, this can be achieved through a predefined lookup table or configuration file, which explicitly lists which state variables(s) in the physical mechanism model each effective state correction quantity component should act on. Alternatively, it can be dynamically established using an intelligent matching algorithm based on variable names or physical dimensions. For instance, the portion of the correction quantity representing "change in soil moisture content" can be mapped to the corresponding "soil moisture content" state variable in the physical mechanism model.

[0167] Based on this mapping relationship, the effective state correction is superimposed onto the state variable to be corrected using a preset weighting method to obtain the corrected candidate state variable. This step aims to apply the effective state correction, filtered by the physical consistency enforcer, to the corresponding state variable of the physical mechanism model in a controlled manner, thereby controlling the correction magnitude and preventing over-correction or state abrupt changes. For example, the weighting method can be a simple linear superposition, where the corrected candidate state variable equals the original state variable plus the weight coefficient multiplied by the effective state correction. The weight coefficient can be set according to factors such as the reliability of the correction, the model's sensitivity to the state variable, or the forecast timeliness. Alternatively, a nonlinear weighting function can be used, for example, using a larger weight when the correction is small and a smaller weight when the correction is large, to avoid over-correction. Another approach is a dynamic weighting method based on model uncertainty estimation.

[0168] To ensure that the corrected state variables conform to physical laws, this application further conducts a physical rationality test on the corrected candidate state variables to verify whether they fall within the reasonable state space defined by the physical mechanism model. This test is a crucial step in ensuring the physical reliability of the correction results, preventing the introduction of unreasonable physical states due to the correction amount. For example, the physical rationality test may include checking the range of values ​​for the state variables, such as soil moisture content not being negative or exceeding saturation moisture content, river flow not being negative, and temperature not falling below absolute zero. Furthermore, it may include checking the physical relationships between state variables, such as the relationship between water depth and water level, and the conservation relationship between runoff and runoff, to ensure that the corrected combination of variables still satisfies these fundamental physical laws.

[0169] The candidate state variable that passes the verification is identified as the corrected state variable. This step aims to formally adopt those candidate state variables that pass the physical plausibility test as the correction results, providing a physically consistent and reliable initial state for subsequent forecast calculations. For example, if a candidate state variable passes all physical plausibility tests, it is directly adopted as the corrected state variable. If a candidate state variable fails the test, a fallback strategy can be adopted, such as using the original state variable, iteratively adjusting the correction amount until it passes the test, or retesting with a conservative correction amount (such as half the correction amount) until a physically plausible correction value is found.

[0170] Through the above technical solution, this application introduces a rigorous physical rationality verification mechanism when correcting the state variables of the physical mechanism model using effective state correction quantities. By determining the mapping relationship, it ensures that the correction quantity can accurately act on the target state variable. By using a preset weighting method to superimpose the correction quantity, the magnitude of the correction is effectively controlled, preventing over-correction or state abrupt changes, thereby maintaining the stability of the model state. More importantly, physical rationality verification of the corrected candidate state variables can promptly identify and eliminate any correction results that exceed the physical feasible region or violate physical laws, ensuring that the corrected state variables always remain within the reasonable state space defined by the physical mechanism model. Only the verified candidate state variables are determined as the corrected state variables, thereby ensuring that subsequent flood forecast calculations can be based on physically credible and reliable data. This not only improves the reliability and accuracy of flood forecasts but also effectively solves the problem that purely data-driven models may produce physically unreasonable forecast results under extreme scenarios. This makes the mechanism-data dual-driven flood forecasting method maintain computational efficiency while enhancing its physical interpretability and robustness.

[0171] In some of the schemes mentioned above in this application, the forecast results of all computing units are fused using the corrected state variables. However, in this process, due to the differences in the computing modes of different computing units and the complexity of spatial topological relationships, directly fusing the forecast results may lead to spatial discontinuities and boundary inconsistencies. For example, the forecast results generated by the coupled model unit based on the physical mechanism model may jump or mismatch with the data-driven results of other model units in the transition region, which in turn affects the overall consistency of the watershed state field. This makes it impossible for the hydrological evolution calculation to accurately simulate the actual flood process, reducing the reliability and physical credibility of the flood forecast results.

[0172] To address this issue, this application proposes a fusion calculation method. Through a systematic fusion calculation mechanism, it effectively resolves the spatial discontinuity problem in integrating forecast results from different model calculation units, ensuring the physical consistency and accuracy of flood forecasts. (See [link to relevant documentation]). Figure 7 Taking the server as the executing entity as an example, the method includes: 701. Based on the routing diagram of the model, identify all computing units in the target watershed that adopt different computing modes and the first spatial topology relationship between the computing units.

[0173] In detail, the model routing map is a type of structured data that records the dynamically assigned computational modes of each computational unit within the target watershed, as well as the spatial locations and potential connections of these units. In this step, it serves as an index and guide, clarifying which computational units employ which forecasting mode (e.g., physical mechanism model, data-driven model, or coupled model) and assisting in identifying the spatial relationships between these units. Specifically, by parsing the model routing map, traversing all computational units, classifying them according to their mode assignment labels, and calculating the distances, shared boundaries, or hydrological connectivity paths between adjacent units based on predefined watershed grids or Geographic Information System (GIS) data, a first spatial topological relationship can be established. For example, spatial indexing techniques (such as R-trees and quadtrees) can be used to quickly query adjacent computational units, and predefined connectivity rules (e.g., shared edges, flow direction) can be combined to determine their topological relationships. This step aims to provide accurate spatial context information for subsequent forecast result fusion, ensuring that boundaries and connections between different model units are correctly handled during the fusion process.

[0174] 702. For each computing unit, obtain the prediction result corresponding to that computing unit. For the computing unit of the coupled mode, the prediction result is calculated based on the corrected state variable.

[0175] Specifically, this step involves collecting all forecast data involved in the fusion. For computational units of uncoupled models, their forecast results can be directly obtained from the corresponding data-driven or simplified models. For computational units of coupled models, their forecast results are calculated by correcting the state variables of the physical mechanism model using effective state corrections processed by a physical consistency enforcer. This ensures that these results not only have a physical basis but also undergo error correction, thereby enhancing the reliability of the data.

[0176] 703. Based on the first spatial topological relationship and the computing mode adopted by each computing unit, the forecast results of all computing units are integrated into a spatially continuous watershed state field through a spatial fusion strategy.

[0177] In detail, spatial fusion strategies are a series of rules and algorithms used to integrate forecast results from different computational units, which may have different spatial resolutions or physical meanings, into a unified, continuous, and physically consistent watershed state field. It addresses the spatial discontinuities and inconsistencies in forecast results from different models, ensuring that the fused watershed state field accurately reflects the actual physical processes. For example, interpolation methods based on physical conservation laws can be used to coordinate flow or water level at the boundaries of different model units through mass or energy conservation principles. Alternatively, multi-scale fusion techniques can be employed to fuse high-resolution coupled model results with low-resolution data-driven model results using weighted averaging, kriging interpolation, or physical model-based downscaling methods. This step involves boundary coordination processing of forecast results from different model units, such as introducing transition regions between adjacent units and using smoothing functions or physical constraints to eliminate discontinuities, thereby generating a data representation with physical meaning and spatial continuity throughout the entire target watershed.

[0178] 704. Perform hydrological evolution calculations on the state field of the watershed, obtain and output the outlet section flow process of the target watershed as the flood forecast result.

[0179] Specifically, hydrological evolution calculation refers to simulating the process of water flow propagation, convergence, and attenuation from upstream to downstream within a watershed, typically based on hydrodynamic equations or simplified hydrological models. It uses a fused watershed state field as input to simulate the propagation of floodwater within the watershed, obtaining the flow process at the watershed outlet section. For example, one-dimensional or two-dimensional hydrodynamic models (such as the Saint-Venant equations) can be used for simulation, considering the physical characteristics of flow paths such as channels and floodplains. Alternatively, hydrological confluence models (such as the unit hydrograph method or Muskingan method) can be used, dividing the watershed into multiple sub-watersheds or confluence units, and performing flow superposition and evolution based on their hydrological response characteristics. After the hydrological evolution calculation is completed, the time-varying sequence data of the flow at the specified outlet section is extracted from the model and output as graphs, numerical tables, or via an API interface, serving as the output of the entire flood forecasting method.

[0180] Through the above technical solution, this application effectively solves the spatial discontinuity problem in integrating forecast results from different model computational units by employing a systematic fusion computation mechanism, ensuring the physical consistency and accuracy of flood forecasts. Specifically, based on the model routing map, all computational units and their first spatial topological relationships are identified. This step fully utilizes the dynamic allocation information of the routing map to accurately capture the spatial connectivity between units within the watershed, laying a structured foundation for subsequent fusion. Forecast results are obtained for each computational unit, with special emphasis on coupled model units based on corrected state variables. This ensures that the forecast results have passed physical verification and error correction, enhancing data reliability. Based on the first spatial topological relationship and the computational model adopted by each computational unit, the forecast results are integrated into a spatially continuous watershed state field through a spatial fusion strategy. This strategy adaptively handles the boundary transitions of different model units, achieving smooth connection of the state field and avoiding local inconsistencies caused by model differences. Hydrological evolution calculations are performed on the watershed state field, and the outlet section flow process is output. Combined with physical mechanisms, the flood evolution path is simulated to generate forecast results, thereby improving the overall accuracy and reliability of flood forecasts.

[0181] In some of the embodiments described above in this application, a method is proposed to fuse the forecast results of all computing units based on the model routing diagram and the corrected state variables. However, in its implementation, since different computing units may adopt different computing modes, direct integration may lead to discontinuous or uncoordinated state fields in the spatial transition area, affecting the accuracy of subsequent hydrological evolution calculations, especially the physical inconsistency problem that may occur at the boundary.

[0182] To address this, this application further proposes a spatial fusion strategy to integrate the forecast results of all computational units into a spatially continuous watershed state field. This strategy is based on a first spatial topology and the computational mode adopted by each computational unit. Specifically, it includes: for adjacent computational units using different computational modes, based on the model routing map and the first spatial topology, performing boundary coordination processing in the spatial transition region between adjacent computational units to achieve a smooth transition of the state field. For all computational units under each computational mode, an adaptive spatial interpolation method matching the characteristics of that computational mode is used to generate continuously distributed sub-state fields within the mode from the discrete unit values ​​of the forecast results corresponding to that computational unit. All sub-state fields are then globally coordinated and fused based on physical laws to obtain the watershed state field covering the target watershed.

[0183] The aforementioned spatial fusion strategy aims to integrate discrete forecast results from different computational models into a unified, continuous, and physically consistent watershed state field. This strategy systematically addresses the potential spatial discontinuities and physical inconsistencies between different models through a series of steps.

[0184] In this process, boundary coordination processing is performed in the spatial transition region between adjacent computational units. This aims to address potential discontinuities or physical inconsistencies that may arise at the boundaries between different computational models, ensuring a smooth spatial transition of the state field. Specifically, interpolation smoothing methods can be employed. In the boundary region between adjacent computational units, weighted averaging, linear interpolation, or spline interpolation of the forecast results on both sides can generate boundary values ​​for a smooth transition. For example, different weights can be assigned based on the distance from the boundary, allowing the forecast results in the transition region to gradually and smoothly transition from one model to another. Furthermore, flux conservation correction methods can be used. For mass or energy fluxes commonly found in hydrological models, flux conservation constraints are introduced at the boundary. By adjusting the forecast results on both sides of the boundary, it is ensured that the flux across the boundary is physically continuous and conserved. For example, the water exchange on both sides of the boundary can be calculated, and the forecast flow can be fine-tuned to satisfy the law of mass conservation.

[0185] For all computational units under each computational mode, an adaptive spatial interpolation method matching the characteristics of that computational mode is employed to transform the prediction results of discrete computational units into a continuously distributed sub-state field within the mode. This method also considers the characteristics of different computational modes to improve interpolation accuracy and physical plausibility. Specifically, interpolation methods based on the model's physical mechanisms can be used. For example, for physical mechanism models, interpolation methods based on hydrological processes (such as surface flow and groundwater flow) can be used, utilizing digital elevation models (DEMs) and flow path information to ensure the interpolation results conform to hydrological connectivity. For data-driven models, machine learning-based interpolation methods, such as Kriging interpolation, radial basis function (RBF) interpolation, or deep learning models (such as U-Net), can be used for spatial prediction. These methods can learn nonlinear spatial relationships in the data. Furthermore, dynamically parameter-adjusted interpolation methods can be used, dynamically selecting or adjusting the parameters of the interpolation algorithm based on the type of computational mode (e.g., distributed hydrological models, conceptual hydrological models) and its output characteristics (e.g., flow rate, water level, soil moisture). For example, for patterns with drastic output changes, a more local interpolation method can be used. For patterns with gradual output changes, a global interpolation method can be used.

[0186] The goal of globally coordinating and fusing all sub-state fields based on physical laws is to ensure that all sub-state fields satisfy unified physical laws (such as mass conservation and energy conservation) throughout the entire target watershed, thereby generating a physically consistent and continuous watershed state field. Specifically, an optimization adjustment method can be employed, establishing a global optimization problem. All sub-state fields are treated as variables, with physical conservation laws (such as the conservation of total watershed water volume) as constraints. The sub-state fields are adjusted by minimizing a certain error function (e.g., the deviation from the original sub-state fields) to achieve global consistency. This can be achieved using the Lagrange multiplier method or iterative optimization algorithms. Alternatively, a multi-scale data assimilation method can be used. Sub-state fields generated by different models are treated as observation data from different sources. Combined with a unified physical model (e.g., a simplified watershed hydrological model), data assimilation techniques (such as Kalman filtering, ensemble Kalman filtering, or variational assimilation) are used to fuse these sub-state fields into an optimal watershed state field, ensuring that the fusion result conforms to physical laws and is as close as possible to the information of each sub-state field.

[0187] Through the above technical solutions, this application effectively addresses the spatial discontinuities and physical inconsistencies caused by model differences when fusing forecast results from different computational models. By performing boundary coordination processing in the spatial transition region between adjacent computational units, a smooth transition of the state field at the boundary between different models can be ensured, avoiding abrupt changes and physical inconsistencies. Simultaneously, by employing an adaptive spatial interpolation method matching the characteristics of each computational model, discrete unit forecast results can be transformed into continuously distributed sub-state fields within the model, guaranteeing the continuity and consistency of results within each model. Based on this, global coordination and fusion of all sub-state fields based on physical laws further ensures the physical consistency and continuity of the entire watershed state field, such as satisfying fundamental physical laws like mass conservation. This comprehensive spatial fusion strategy provides a high-quality, highly reliable spatially continuous watershed state field for subsequent hydrological evolution calculations, improving the accuracy and stability of flood forecasts. Especially in complex watersheds and multi-model coupled forecasting scenarios, it can effectively address the heterogeneity of different hydrological response regions, thereby obtaining more reliable and interpretable flood forecast results.

[0188] In some of the above-mentioned schemes in this application, it is proposed to perform hydrological evolution calculations on the watershed state field after fusing the forecast results of all computing units based on the model routing diagram to obtain the outlet section flow process. However, in this process, due to the differences in the calculation modes adopted by different computing units, the evolution parameters may lack adaptive adjustment, resulting in low efficiency and insufficient accuracy of the hydrological calculation process, especially in extreme flood events where the forecast results are unreliable or physically inconsistent.

[0189] To address this, this application further proposes hydrological evolution calculations of the watershed state field to obtain and output the outlet section flow process of the target watershed as the flood forecast result. This calculation process includes: extracting the hydrological state and second spatial topological relationship of each hydrological response unit within the target watershed based on the model routing diagram and the watershed state field; employing a physics-based hydrological evolution method based on the second spatial topological relationship, using the output of the upstream calculation unit as the input of the downstream calculation unit for step-by-step calculation; dynamically adjusting the evolution parameters in conjunction with the calculation mode adopted by the upstream calculation unit until the calculation reaches the outlet section of the target watershed, obtaining a continuous flow process line as the flood forecast result.

[0190] Specifically, the step of "extracting the hydrological status and second spatial topological relationships of each hydrological response unit within the target watershed based on the model routing map and the watershed state field" aims to provide accurate input data and spatial connectivity information for subsequent hydrological evolution calculations. The model routing map provides pattern allocation information for each calculation unit, while the watershed state field provides continuous hydrological status data. Combining the two can identify the hydrological characteristics (such as runoff and water storage) of different areas within the watershed and the physical connections between them. One implementation method is to overlay and analyze the model routing map (containing pattern allocation information) and the watershed state field (containing the spatial distribution of hydrological variables) through a Geographic Information System (GIS) platform. Using predefined watershed divisions (such as sub-watersheds and grid units), the average or representative hydrological status values ​​of each hydrological response unit are extracted from the watershed state field. Simultaneously, based on the digital elevation model (DEM) and river network data in GIS, the flow paths and confluence relationships between each hydrological response unit are analyzed and constructed, thereby obtaining the second spatial topological relationships. Another approach involves establishing a dedicated data parsing module that receives structured data from the model routing diagram and raster or vector data from the watershed state field. Using spatial query algorithms, hydrological elements (such as soil moisture, surface runoff, and groundwater storage) from the watershed state field are aggregated into the boundaries of the computational units defined by the model routing diagram, forming the hydrological state of each unit. Simultaneously, using a pre-stored watershed connectivity database or a real-time confluence path algorithm, the upstream and downstream relationships between each computational unit—that is, the second spatial topology—are determined.

[0191] The core of hydrological evolution lies in the step of "using a physics-based hydrological evolution method based on second-space topology, and using the output of upstream computing units as input for downstream computing units for step-by-step calculations," ensuring accurate simulation of the physical transmission process of water flow within the watershed. By utilizing the established upstream-downstream connections (second-space topology), the runoff or flow calculated by upstream units is used as input for downstream units to simulate the convergence and propagation of water flow. The physics-based approach ensures that the calculation process conforms to hydrophysical laws, improving forecast reliability. One implementation method is to use the confluence module in a distributed hydrological model, such as methods based on kinematic wave equations or diffusion wave equations. These methods can simulate the propagation process of water flow on river channels and slopes, considering physical parameters such as water depth, velocity, and roughness. Step-by-step calculation means starting from the upstream unit of the watershed and sequentially transmitting water flow information to downstream units until the watershed outlet. Another implementation method is to use a series of reservoirs or a linear reservoir model in a conceptual hydrological model, combined with hydrodynamic principles for calculations. For example, each computing unit is treated as one or more reservoirs, and water balance calculations are performed based on its inflow (from upstream units and local runoff) and outflow (to downstream units).

[0192] The step of "dynamically adjusting evolution parameters in conjunction with the calculation mode adopted by the upstream calculation unit during the step-by-step calculation process until the outlet section of the target watershed is calculated, obtaining a continuous flow process line as the flood forecast result" aims to solve the problem of mismatched evolution parameters between different calculation mode units. By dynamically adjusting the evolution parameters, the hydrological evolution process can adapt to the actual calculation mode of the upstream unit (e.g., whether it is a physical mechanism mode or a data-driven mode), thereby improving the accuracy and physical consistency of the hydrological evolution of the entire watershed, especially in the mode switching area. One implementation method is to establish a parameter adjustment rule base or a machine learning model. When the upstream calculation unit adopts a coupled mode, the evolution parameters (such as channel roughness, slope roughness, hydraulic conductivity, etc.) can be finely adjusted according to the internal state of the physical mechanism model or its output physical quantities. When the upstream calculation unit adopts a pure data-driven mode, the evolution parameters can be adjusted according to the characteristics output by the data-driven model (such as the rate of change of flow, water depth change trend) or through pre-trained mapping relationships to ensure the connection with the downstream physical evolution. This adjustment can be based on real-time data or historical experience. Another approach is to introduce an adaptive parameter optimization algorithm. At each step of the progressive calculation, the impact of the current evolution parameters on water flow propagation is dynamically evaluated based on the mode type of the upstream computing unit. For example, if the upstream unit is in a data-driven mode, its output may be smoother or more turbulent. In this case, the confluence parameters of the downstream unit (such as propagation velocity and attenuation coefficient) can be adjusted to better match the characteristics of the upstream output while maintaining physical rationality. This adjustment can be based on error feedback or optimization of the objective function.

[0193] The above technical solution extracts hydrological status and second spatial topological relationships based on the model routing diagram and the watershed state field. Utilizing the computational model information provided by the model routing diagram, it accurately identifies the hydrological attributes and spatial connections of each unit, laying a data foundation for the calculation and avoiding input data confusion caused by model differences. Based on the second spatial topological relationships, a physical mechanism-based hydrological evolution method is used for step-by-step calculations. This physical mechanism ensures that the calculations conform to natural laws, and upstream-downstream dependencies are handled based on topological relationships, improving the accuracy of the calculations. During the step-by-step calculation process, the evolution parameters are dynamically adjusted according to the computational model adopted by the upstream computational unit. Parameters are adjusted in real time according to the actual model of the upstream unit (such as coupled or pure data model), solving the parameter mismatch problem caused by model switching and enhancing adaptability in extreme events. The calculation continues until a continuous flow process line is obtained at the outlet section, completing the calculation output for the entire watershed and providing complete and reliable flood forecast results. This effectively solves the problem of unreliable forecasts caused by parameter inconsistencies in fusion calculations, ensuring the efficiency, accuracy, and physical consistency of the calculation process.

[0194] The following example will provide a more detailed explanation of the above technical solution: If flood forecasting is required for a target watershed with complex topography and variable hydrological characteristics, the watershed, which comprises multiple sub-watersheds and river sections, is divided into several computational units.

[0195] When a flood forecast task for the target watershed is received, multi-source watershed characteristic data for the current time period is collected in real time. This data may include real-time rainfall data from meteorological radar, flow and water level observation data from surface hydrological stations, soil moisture data acquired by satellite remote sensing, and underlying surface spatial attribute data such as the watershed's digital elevation model (DEM) and land use type. This real-time multi-source watershed characteristic data is input into the central routing agent for analysis and routing decisions.

[0196] Specifically, the central routing agent performs multi-dimensional feature analysis on these real-time multi-source watershed characteristic data to extract key feature sets. For example, it extracts dynamic rainfall intensity features and watershed water storage status features from real-time meteorological and hydrological data. It extracts soil hydraulic features and topographic features from underlying surface spatial attribute data. These features are then spatiotemporally fused and their importance assessed to form a comprehensive key feature set. This key feature set can characterize the urgency of hydrological response and the applicability of physical mechanism models in different areas within the target watershed.

[0197] Based on this key feature set, the routing decision network in the central routing agent performs pattern matching and decision-making. For each computing unit, the network comprehensively considers its own key feature set and the key feature sets of its spatially adjacent computing units, performs spatial context association analysis, and generates spatially enhanced decision features for that computing unit. For example, if a computing unit currently experiences extremely high rainfall intensity and its upstream unit is also in a high-intensity rainfall area, its spatially enhanced decision features will reflect a higher urgency of hydrological response. Based on these spatially enhanced decision features, the routing decision network performs multi-factor collaborative decision-making. For example, it combines real-time rainfall intensity factors (such as rainfall-driven intensity indicators) and historical error trend factors of physical mechanism models (such as model reliability indicators), and generates a preliminary pattern allocation probability for that computing unit through preset collaborative decision-making rules. Based on the physical constraint rules of the target watershed (e.g., upstream and downstream units with hydrological connectivity should adopt consistent or compatible computing modes as much as possible), the preliminary pattern allocation probability is spatially consistent to obtain a pattern allocation label for each computing unit, indicating the computing mode it should adopt (e.g., pure physical mechanism mode, data-driven mode, or coupled mode).

[0198] After the model assignment labels for all computational units are determined, a digital spatial reference frame for the target watershed is established, and these model assignment labels are mapped into this reference frame. Continuous model distribution data covering the target watershed is generated using spatial interpolation or topological connectivity methods, and this data is then structured and encoded to generate and output a model routing map. This model routing map is structured data containing spatial location and model assignment information, dynamically assigning computational models to multiple computational units within the watershed.

[0199] After obtaining the model routing map, forecast calculations are performed on the computational units assigned to coupled mode based on the map and the model state corrected in the previous time period. For example, all computational units assigned to coupled mode are selected from the model routing map. For each selected computational unit, the corresponding model state corrected in the previous time period is read from the physically verified and corrected model state library as the initial forecast condition for the current time period. Based on these initial forecast conditions, the watershed-scale physical mechanism model is driven to perform forecast calculations. During the calculation process, the model iterates stepwise at a preset calculation time step and records the flow value of the computational unit at a preset time resolution to form the mechanistic forecast flow result. At the same time, it captures snapshots of all key state variables of the model at the corresponding time and assembles them into an internal physical state sequence.

[0200] Next, to correct potential errors in the physical mechanism model, these internal physical state sequences and real-time observation data are processed. The internal physical state sequences and corresponding real-time observation data are spatiotemporally aligned and feature-fused to form a fused data sample for error analysis. This fused data sample is then input into a generative state error model. This model generates initial physical state corrections by learning the deviation patterns between the internal physical state sequences and the real-time observation data. For example, the model constructs an explicit state deviation field based on the fused data sample and, combined with the spatiotemporal context of the internal physical state sequences, infers and generates a high-dimensional state correction potential field through a conditional generation mechanism, which is then mapped to the physical space to form the initial physical state correction. The initial physical state correction is then standardized, including dimensional unification, adaptive spatial scale transformation, and physical admissibility testing, to obtain a physical state correction that matches the state variables of the physical mechanism model in both dimension and physical meaning.

[0201] To ensure the physical rationality of the corrections, these physical state corrections are filtered and verified based on physical conservation laws using a physical consistency enforcement checker. The checker combines the current state variables of the physical mechanism model with a set of verification equations based on the laws of mass and momentum conservation, solving for the theoretical conservation deviation corresponding to the physical state correction. For example, if the correction causes a non-conservation of water in a certain calculation unit, a significant deviation will occur. This theoretical conservation deviation is compared with a preset tolerance threshold. If the deviation exceeds the limit, the physical state correction is filtered or corrected, for example, by decomposing it into multiple correction components. Constraints based on conservation laws are applied to the components with excessive deviations until the adjusted theoretical conservation deviation falls within the tolerance threshold. The portion that meets the tolerance threshold is then output as the valid state correction.

[0202] Unlike data-driven models in related technologies that may produce physically unreasonable results, this solution uses a generative state error model combined with a physical consistency enforced verifier to ensure the physical rationality of the correction amount, thus avoiding the problem of unreliable prediction results in extreme scenarios caused by "black box" models.

[0203] Based on these effective state corrections, the state variables of the physical mechanism model are corrected. This includes determining the mapping relationship between the state variables to be corrected and the effective state corrections, and then superimposing the effective state corrections onto the state variables to be corrected using a preset weighting method to obtain corrected candidate state variables. The candidate state variables are then subjected to a physical rationality test to verify whether they fall within the reasonable state space defined by the physical mechanism model, and are thus determined as corrected state variables.

[0204] Based on the corrected state variables, and using the model routing diagram, the forecast results of all computational units are fused and calculated. All computational units employing different computational modes in the target watershed and their spatial topological relationships are identified. For each computational unit, its corresponding forecast result is obtained (for computational units in coupled modes, the forecast result is calculated based on the corrected state variables). Through spatial fusion strategies, such as performing boundary coordination processing in the spatial transition region for adjacent computational units employing different computational modes, a smooth transition of the state field is achieved. For all computational units under each computational mode, an adaptive spatial interpolation method is used to generate continuously distributed sub-state fields within the mode. Then, all sub-state fields are globally coordinated and fused based on physical laws, integrating the forecast results of all computational units into a spatially continuous watershed state field.

[0205] Hydrological evolution calculations are performed on the watershed state field to obtain and output the outlet section flow process of the target watershed as the flood forecast result. This includes extracting the hydrological state and spatial topological relationships of each hydrological response unit based on the model routing map and watershed state field, and using a physical mechanism-based hydrological evolution method, taking the output of the upstream calculation unit as the input of the downstream calculation unit for step-by-step calculation. During the calculation process, the evolution parameters are dynamically adjusted according to the calculation mode adopted by the upstream calculation unit until the calculation reaches the outlet section of the target watershed, obtaining a continuous flow process line.

[0206] Through the above process, this scheme achieves deep collaboration between the physical mechanism model and the data-driven model. The dynamic allocation of computation by the central routing agent solves the problem of low computational efficiency in traditional physical mechanism models. The generative state error model combined with the physical consistency enforcement checker compensates for the lack of physical constraints in pure data-driven models, ensuring the physical reliability of the forecast results. The fusion calculation of the corrected state variables enables real-time interaction and error compensation between models in the core computation process, overcoming the limitation of model integration in related technologies that only focuses on result fusion. This allows for more accurate and reliable flood forecast results in sudden flood events.

[0207] Figure 8 This is a schematic diagram of the structure of a mechanism-data-driven flood forecasting system provided in an embodiment of this application. See also... Figure 8 The system includes: The parsing module 801 is used to respond to the flood forecasting task for the target watershed by parsing and making routing decisions on the real-time multi-source watershed feature data of the current time period through the central routing agent to obtain the model routing map of the target watershed. This model routing map is used to dynamically allocate computing modes to multiple computing units respectively.

[0208] The forecast calculation module 802 is used to perform forecast calculations on the calculation units assigned to the coupled mode based on the model routing map and the model state after correction in the previous time period, and to obtain the mechanism forecast flow results and the corresponding internal physical state sequence.

[0209] Error analysis module 803 is used to perform state error generation analysis on the internal physical state sequence and real-time observation data through a generative state error model to obtain the physical state correction amount, and to perform verification and filtering on the physical state correction amount based on the physical conservation law through a physical consistency mandatory checker to obtain the effective state correction amount.

[0210] The fusion calculation module 804 is used to correct the state variables of the physical mechanism model based on the effective state correction amount, and based on the model routing diagram, use the corrected state variables to perform fusion calculation on the forecast results of all calculation units to obtain the flood forecast results of the target watershed.

[0211] It should be noted that the mechanism-data-driven flood forecasting system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer equipment can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the mechanism-data-driven flood forecasting system and the mechanism-data-driven flood forecasting method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0212] Figure 9 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 900 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 901 and one or more memories 902. The one or more memories 902 store at least one computer program, which is loaded and executed by the one or more processors 901 to implement the methods provided in the various method embodiments described above. Of course, the server 900 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server 900 may also include other components for implementing device functions, which will not be elaborated upon here.

[0213] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program that can be executed by a processor to perform the mechanism data-driven flood forecasting method in the above embodiments. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0214] In an exemplary embodiment, a computer program product or computer program is also provided, which includes program code stored in a computer-readable storage medium. A processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to perform the above-described mechanism-data dual-driven flood forecasting method.

[0215] In some embodiments, the computer program involved in the present application embodiments may be deployed and executed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0216] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0217] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A flood forecasting method driven by both mechanistic data and theoretical data, characterized in that, The method includes: In response to the flood forecasting task for the target watershed, multi-dimensional feature analysis is performed on the real-time multi-source watershed characteristic data for the current time period to extract key feature sets. These key feature sets characterize the urgency of hydrological response and the applicability of mechanistic models in different areas within the target watershed. For each computational unit, based on the key feature sets of the computational unit and the key feature sets of its spatially adjacent computational units, spatial context association analysis is performed through the routing decision network in the central routing agent to generate spatially enhanced decision features for the computational unit. Based on these spatially enhanced decision features, multi-factor collaborative decision-making is performed through the routing decision network to generate... The preliminary model allocation probability of the computing unit is obtained, and the factors of the multi-factor collaborative decision-making include at least the real-time rainfall intensity factor and the historical error trend factor of the mechanism model. Based on the physical constraint rules of the target watershed, the preliminary model allocation probability is subjected to spatial consistency optimization processing to obtain the model allocation label of the computing unit. The physical constraint rules are used to ensure that the model allocation among hydrologically connected units conforms to physical laws, and the model allocation label is used to indicate the computing mode. Based on the model allocation labels of all computing units, a model routing map of the target watershed is generated and output. The model routing map is used to dynamically allocate computing modes to multiple computing units respectively. Based on the model routing diagram and the model state after correction in the previous time period, the computing units assigned to the coupled mode are used to perform forecast calculations through the physical mechanism model at the watershed scale to obtain the mechanism forecast flow results and the corresponding internal physical state sequence. The internal physical state sequence and real-time observation data are analyzed by generating state error model to obtain physical state correction amount. The physical state correction amount is then filtered by physical consistency mandatory checker based on physical conservation law to obtain effective state correction amount. The state variables of the physical mechanism model are corrected based on the effective state correction amount, and the forecast results of all computing units are fused and calculated based on the model routing diagram and the corrected state variables to obtain the flood forecast results of the target watershed.

2. The method according to claim 1, characterized in that, The step of generating the preliminary pattern allocation probability of the computing unit by performing multi-factor collaborative decision-making through the routing decision network based on the spatial augmentation decision features includes: Based on the spatial augmentation decision features, the real-time rainfall intensity factor is evaluated by state quantification to obtain the rainfall driving intensity index; Based on the spatial augmentation decision characteristics, the historical error trend factor of the mechanism model is dynamically evaluated to obtain the model reliability index. Based on the rainfall driving intensity index and the model reliability index, the preliminary mode allocation probability of the calculation unit is obtained by calculation through preset collaborative decision-making rules or probability mapping models.

3. The method according to claim 1, characterized in that, The physical constraint rules based on the target watershed are used to perform spatial consistency optimization on the preliminary mode assignment probability to obtain the mode assignment label of the computing unit, including: Based on hydrological connectivity, the upstream and downstream adjacent units of the computing unit are identified; Obtain the preliminary mode allocation probability of the upstream adjacent unit and the preliminary mode allocation probability of the downstream adjacent unit; Based on the physical constraint rules, consistency constraint optimization is performed on the initial mode allocation probability of the computing unit, the initial mode allocation probability of the upstream adjacent unit, and the initial mode allocation probability of the downstream adjacent unit to obtain the optimized mode allocation probability. Based on the optimized pattern allocation probability, the pattern allocation label of the computing unit is determined.

4. The method according to claim 1, characterized in that, Based on the model routing diagram and the model state corrected in the previous time period, the computational units assigned to the coupled mode are used to perform forecast calculations through a watershed-scale physical mechanism model to obtain the mechanism-based forecast flow results and the corresponding internal physical state sequence, including: Based on the model routing diagram, computing units assigned to the coupling mode are selected from all computing units in the target watershed; For each selected computing unit, the corresponding corrected model state from the previous time period is read from the model state library after physical verification and correction, and used as the initial forecast condition for the computing unit in this time period. Based on the initial forecast conditions, the physical mechanism model at the watershed scale is driven to perform forecast calculations, and the mechanistic forecast flow results of the calculation unit and the internal physical state sequence generated at a preset time resolution are output.

5. The method according to claim 4, characterized in that, The process of driving the watershed-scale physical mechanism model to perform forecast calculations based on the initial forecast conditions, and outputting the mechanistic forecast flow results of the calculation unit and the internal physical state sequence generated at a preset time resolution includes: Based on the configuration parameters of the coupling mode, the calculation time step and output frequency of the physical mechanism model are set during the current forecast period; Based on the initial forecast conditions, the physical mechanism model is iteratively run step by step with the calculated time step until the simulation of the current forecast period is completed; During the iterative operation, the flow rate value of the computing unit at each output frequency is recorded to form the mechanism-predicted flow rate result; and snapshots of all key state variables of the physical mechanism model at the corresponding time are captured to assemble the internal physical state sequence according to the preset time resolution.

6. The method according to claim 1, characterized in that, The process of generating and analyzing state errors using a generative state error model on the internal physical state sequence and real-time observation data yields physical state corrections, including: The internal physical state sequence is spatiotemporally aligned and feature-fused with the corresponding real-time observation data to form a fused data sample for error analysis. The fused data samples are input into the generative state error model, which learns the deviation pattern between the internal physical state sequence and the real-time observation data, and generates an initial physical state correction. The initial physical state correction is standardized to obtain a physical state correction that matches the state variables of the physical mechanism model in both dimension and physical meaning.

7. The method according to claim 6, characterized in that, The step of inputting the fused data samples into the generative state error model, learning the deviation pattern between the internal physical state sequence and the real-time observation data through the generative state error model, and generating an initial physical state correction includes: Based on the fused data samples, an explicit state deviation field is constructed between the internal physical state sequence and the real-time observation data, which serves as the conditional input to the generative state error model. Based on the spatiotemporal context of the explicit state deviation field and the internal physical state sequence, a high-dimensional state correction potential field is inferred and generated through the conditional generation mechanism of the generative state error model. The high-dimensional state correction potential field is mapped to the physical space and dimensions corresponding to the state variables of the physical mechanism model to form the initial physical state correction quantity.

8. The method according to claim 1, characterized in that, The physical state correction amount is filtered and verified based on the physical conservation law using a physical consistency enforcer to obtain an effective state correction amount, including: The physical state correction amount and the current state variables of the physical mechanism model are input into the physical consistency enforcement checker; In the physical consistency enforcement checker, a set of check equations for the physical state correction quantity is constructed based on the laws of conservation of mass and momentum. Solve the set of verification equations to obtain the theoretical conservation deviation corresponding to the physical state correction; The theoretical conservation deviation is compared with a preset tolerance threshold, and the physical state correction amount is filtered or corrected based on the comparison result. The portion that satisfies the tolerance threshold is output as the effective state correction amount.

9. The method according to claim 1, characterized in that, Based on the model routing diagram, and using the corrected state variables, the forecast results of all computational units are fused and calculated to obtain the flood forecast results for the target watershed, including: Based on the model routing diagram, identify all computing units in the target watershed that adopt different computing modes and the first spatial topological relationship between the computing units. For each computing unit, the prediction result corresponding to the computing unit is obtained, wherein, for the computing unit in coupled mode, the prediction result is calculated based on the corrected state variables; Based on the first spatial topological relationship and the computing mode adopted by each computing unit, the prediction results of all computing units are integrated into a spatially continuous watershed state field through a spatial fusion strategy. Hydrological evolution calculations are performed on the watershed state field to obtain and output the outlet section flow process of the target watershed as the flood forecast result.

10. A flood forecasting system driven by both mechanistic data and [other factors], characterized in that, The system includes: The parsing module, in response to flood forecasting tasks for the target watershed, performs multi-dimensional feature analysis on real-time multi-source watershed characteristic data for the current time period, extracting a key feature set. This key feature set characterizes the urgency of hydrological response and the applicability of mechanistic models in different areas within the target watershed. For each computational unit, based on its key feature set and the key feature sets of its spatially adjacent computational units, spatial context association analysis is performed through the routing decision network in the central routing agent to generate spatially enhanced decision features for that computational unit. Based on these spatially enhanced decision features, multi-factor collaborative decision-making is performed through the routing decision network. A preliminary model allocation probability is generated for the computing units. The factors in the multi-factor collaborative decision-making include at least a real-time rainfall intensity factor and a historical error trend factor of the mechanistic model. Based on the physical constraint rules of the target watershed, the preliminary model allocation probability is spatially consistent to obtain the model allocation label of the computing units. The physical constraint rules are used to ensure that the model allocation among hydrologically connected units conforms to physical laws, and the model allocation label is used to indicate the computing mode. Based on the model allocation labels of all computing units, a model routing map of the target watershed is generated and output. The model routing map is used to dynamically allocate computing modes to multiple computing units respectively. The forecast calculation module is used to perform forecast calculations on the calculation units assigned to the coupled mode based on the model routing map and the model state after correction in the previous time period, and to obtain the mechanism forecast flow results and the corresponding internal physical state sequence. The error analysis module is used to perform state error generation analysis on the internal physical state sequence and real-time observation data through a generative state error model to obtain the physical state correction amount, and to perform verification and filtering of the physical state correction amount based on the physical conservation law through a physical consistency mandatory checker to obtain the effective state correction amount. The fusion calculation module is used to correct the state variables of the physical mechanism model based on the effective state correction amount, and to perform fusion calculation on the forecast results of all calculation units based on the model routing diagram and the corrected state variables to obtain the flood forecast results of the target watershed.

Images