Intelligent algorithm for flood forecasting driven by fusion of multi-source data and physical mechanism

Abstract

This invention relates to the field of algorithm architecture technology, and discloses a dual-drive intelligent algorithm for flood forecasting that integrates multi-source data and physical mechanisms. The method includes: preprocessing and feature construction based on the multi-source hydrological data to generate a structured spatiotemporal sequence water source dataset, and simultaneously constructing a simplified distributed hydrological model; inputting the structured spatiotemporal sequence water source dataset into the simplified distributed hydrological model for simulation to generate a physical mechanism baseline forecast sequence; performing residual analysis on the structured spatiotemporal sequence water source dataset to obtain and correct historical residual sequences; fusing the physical mechanism baseline forecast sequence and the corrected residual sequences to obtain the dual-drive prediction result; and performing physical consistency verification on the dual-drive prediction result to generate the final flood forecast sequence. This invention can improve the reliability of dual-drive flood forecasting that integrates multi-source data and physical mechanisms.

Description

Technical Field

[0001] This invention relates to the field of algorithm architecture technology, and in particular to a dual-drive intelligent algorithm for flood forecasting that integrates multi-source data and physical mechanisms. Background Technology

[0002] Flood forecasting is a core technical support for flood control and disaster reduction. Existing flood forecasting methods are mainly divided into two categories: physical mechanism-driven models and data-driven models. Both have significant technical shortcomings when applied alone. Physical mechanism models are built based on hydrological and physical laws, and the forecast results have physical rationality. However, the model structure is complex, parameter calibration is difficult, and the simplification of underlying surface characteristics can easily lead to systematic biases in the forecast. Furthermore, they are difficult to capture the complex nonlinear relationships in hydrological processes. Data-driven models rely on big data and machine learning to mine data patterns, and the forecast accuracy is high. However, they lack physical mechanism support, are prone to forecast results that violate hydrological laws, have weak generalization ability, and their forecast reliability drops significantly under extreme hydrological scenarios.

[0003] To compensate for the shortcomings of a single model, existing technologies attempt to combine physical mechanism models with data-driven models. However, most of these methods employ simple serial or weighted fusion approaches, failing to achieve deep synergy between the two. This approach neither fully leverages the fundamental constraints of the physical model nor allows the data model to accurately correct errors in the physical model. Furthermore, the fusion process lacks an effective physical consistency verification step, which can easily lead to deviations from hydrological and physical laws in the fusion results. Consequently, it becomes difficult to balance accuracy and physical plausibility in forecasts, failing to meet the high accuracy and reliability requirements of actual flood control work. Therefore, improving forecast reliability has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a dual-drive intelligent flood forecasting algorithm that integrates multi-source data and physical mechanisms to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a dual-drive intelligent flood forecasting algorithm that integrates multi-source data and physical mechanisms, comprising: S1. Obtain multi-source hydrological data of the target watershed, preprocess and construct features of the multi-source hydrological data to generate a structured spatiotemporal sequence water source dataset, and construct a hydrological physical model to obtain a simplified distributed hydrological model. S2, input the structured spatiotemporal sequence water source dataset into the simplified distributed hydrological model for simulation to generate a physical mechanism baseline forecast sequence; S3, perform residual analysis on the structured spatiotemporal sequence water source dataset based on the simplified distributed hydrological model to obtain a historical residual sequence, and correct the historical residual sequence to obtain a corrected residual sequence; S4, the physical mechanism baseline forecast sequence and the corrected residual sequence are fused to obtain the dual-drive forecast result, and the physical consistency of the dual-drive forecast result is checked to generate the final flood forecast sequence.

[0006] In a preferred embodiment, the step of acquiring multi-source hydrological data of the target watershed, preprocessing the multi-source hydrological data, and constructing features to generate a structured spatiotemporal sequence water source dataset includes: Based on a pre-defined list of data sources for the target watershed, raw multi-source hydrological data are obtained. Based on a preset time scale and preset processing rules, the original multi-source hydrological data is preprocessed to obtain regularized multi-source hydrological data. Based on the physical mechanism knowledge base, the regularized multi-source hydrological data and associated static data are used to construct derived features and integrate static features to obtain a multi-source feature vector set. Based on the multi-source feature vector set, the structured spatiotemporal sequence water source dataset is constructed.

[0007] In a preferred embodiment, the simultaneous construction of a hydrophysical model to obtain a simplified distributed hydrological model includes: Based on the geographical boundary of the target watershed and the preset unit division rules, the target watershed is spatially discretized to determine at least one hydrological calculation unit. Based on historical hydrological observation datasets, a set of hydrophysical model parameters is determined for each hydrological calculation unit; Based on the preset runoff calculation equation and the parameters of the hydrophysical model, a runoff calculation module is constructed for each hydrological calculation unit; Based on the preset runoff calculation equation and the parameters of the hydrophysical model, the runoff generation calculation modules of each hydrological calculation unit are coupled in series to construct the runoff calculation module. The runoff calculation module and the confluence calculation module of all the hydrological calculation units are integrated to construct a simplified distributed hydrological model.

[0008] In a preferred embodiment, the step of inputting the structured spatiotemporal sequence water source dataset into the simplified distributed hydrological model for simulation to generate a physical mechanism baseline forecast sequence includes: Dynamic driving features are extracted from the structured spatiotemporal sequence water source data to drive the simplified distributed hydrological model to perform continuous simulation calculations time by time period, so as to obtain the simulated flow sequence for the whole period. Based on the historical simulation portion of the full-period simulated flow sequence and the actual observed flow sequence, the residuals of the simplified distributed hydrological model in the historical period are calculated to obtain the historical residuals. From the full-period simulated flow sequence, subsequences corresponding to future forecast periods are extracted to obtain the physical mechanism baseline forecast sequence.

[0009] In a preferred embodiment, calculating the residuals of the simplified distributed hydrological model over historical periods to obtain historical residuals includes: The mathematical expression used to calculate the residuals of the simplified distributed hydrological model over historical periods is as follows: ; In the formula, Represents the historical residual at time t in the historical period. This represents the actual observed flow rate at time t. This represents the simulated flow value at time t in the historical period extracted from the full-period simulated flow sequence. For time indexing.

[0010] In a preferred embodiment, the residual analysis of the structured spatiotemporal sequence water source dataset based on the simplified distributed hydrological model to obtain the historical residual sequence includes: Obtain the actual observed flow sequence synchronized with the historical period of the structured spatiotemporal sequence water source dataset; Based on the simplified distributed hydrological model, historical data in the structured spatiotemporal sequence water source dataset are simulated to obtain the simulated historical flow sequence. The actual observed flow sequence and the historical simulated flow sequence are subtracted element-wise using algebraic subtraction at the same time index to obtain a standard historical residual sequence. At the same time, the historical residual at each time point is used to verify the standard historical residual sequence to obtain a historical residual sequence.

[0011] In a preferred embodiment, correcting the historical residual sequence to obtain a corrected residual sequence includes: Based on the historical residual sequence and the historical multi-source feature data in the structured spatiotemporal sequence water source dataset, a training sample set for deep learning model training is constructed. A residual prediction model is constructed, and the residual prediction model is trained and optimized using the training sample set to obtain a fully trained intelligent correction model. Based on the real-time updated structured spatiotemporal sequence water source dataset and the aforementioned intelligent correction model, multi-step rolling prediction is performed to generate corrected residual sequences corresponding to future forecast periods.

[0012] In a preferred embodiment, fusing the physical mechanism baseline prediction sequence with the corrected residual sequence to obtain the dual-drive prediction result includes: Based on the same time index, the physical mechanism baseline forecast sequence and the corrected residual sequence are time-series aligned to obtain aligned baseline forecast sequence and corrected residual sequence; The aligned baseline forecast sequence and the corrected residual sequence are subjected to time-by-time residual-driven fusion calculation to obtain the dual-drive prediction result.

[0013] In a preferred embodiment, the step of performing physical consistency verification on the dual-drive prediction results to generate the final flood forecast sequence includes: Based on a physical mechanism, at least one physical consistency constraint is established for the dual-drive prediction results. The dual-drive prediction results are verified based on the physical consistency constraints, and the time periods that do not meet the physical consistency requirements are identified. Based on the dynamic weight adjustment algorithm, the forecast values ​​of the identified time periods that do not meet the physical consistency are adjusted to obtain the adjusted forecast sequence; Based on the adjusted forecast sequence, the physically consistent final flood forecast sequence is generated.

[0014] In a preferred embodiment, the step of verifying the dual-drive prediction results based on the physical consistency constraints and identifying time periods that do not meet the physical consistency requirements includes: Based on the physical consistency constraints, the dual-drive prediction results are subjected to time-by-time verification calculations to obtain the verification results and quantization deviation values ​​for each time period. Based on the preset verification threshold, the verification results, and the quantization deviation value, a verification decision is made for each forecast period to identify periods that do not meet the physical consistency requirement.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention pioneers a physical-data dual-driven collaborative forecasting architecture, breaking through the limitations of existing technologies that simply connect, parallel, or weightedly fuse physical and data models. By constructing a collaborative framework for parallel interaction, it achieves deep integration and complementary advantages between the two. This architecture allows the physical model to fully leverage its strengths, outputting a baseline forecast sequence and corresponding historical residual sequence that conforms to hydrophysical laws, laying a physically sound foundation for flood forecasting. Simultaneously, it precisely focuses the learning objective of the data model on the residual sequence of the physical model, enabling the data model to be dedicated to uncovering complex nonlinear hydrological laws that the physical model has not captured and correcting its systematic and random errors. This allows the role of data intelligence to be played in a targeted manner. This non-obvious combination innovation effectively solves the problems of large deviations in single physical models and lack of physical support in single data models, significantly improving the basic accuracy and interpretability of flood forecasting.

[0016] 2. The residual-driven physical constraint fusion method proposed in this invention achieves a dual improvement in forecast accuracy and physical rationality compared to the traditional weighted average fusion method. On the one hand, by directly superimposing the corrected residuals predicted by the data model with the basic physical quantities of the physical model, it achieves precise residual-driven fusion, allowing the baseline forecast of the physical model to be refined, further reducing the deviation between the forecast and the actual values, and significantly improving the numerical accuracy of flood forecasts. On the other hand, a dedicated physical consistency verification step is designed. By establishing physical constraints such as water balance and non-negativity, the fused dual-drive forecast results are verified time-by-time. Combined with a dynamic weight adjustment algorithm, forecast values ​​that do not meet physical laws are corrected, effectively avoiding situations where the fusion results violate hydrological and physical principles. This ensures that the final flood forecast sequence has both the high accuracy brought by data-driven methods and solid physical rationality and reliability, making the forecast results more consistent with the actual hydrological process and providing more scientific and effective data support for practical applications such as flood control decision-making and flood warning. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a dual-drive intelligent flood forecasting algorithm that integrates multi-source data and physical mechanisms, provided in an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a dual-drive intelligent flood forecasting algorithm that integrates multi-source data and physical mechanisms. The executing entity of this algorithm includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the dual-drive intelligent flood forecasting algorithm can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server 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 communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1The diagram shown is a flowchart illustrating a dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms, provided in an embodiment of the present invention. In this embodiment, the dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms includes: S1. Obtain multi-source hydrological data of the target watershed, preprocess and construct features of the multi-source hydrological data to generate a structured spatiotemporal sequence water source dataset, and construct a hydrological physical model to obtain a simplified distributed hydrological model. In this embodiment of the invention, the step of acquiring multi-source hydrological data of the target watershed, preprocessing the multi-source hydrological data, and constructing features to generate a structured spatiotemporal sequence water source dataset includes: Based on a pre-defined list of data sources for the target watershed, raw multi-source hydrological data are obtained. Based on a preset time scale and preset processing rules, the original multi-source hydrological data is preprocessed to obtain regularized multi-source hydrological data. Based on the physical mechanism knowledge base, the regularized multi-source hydrological data and associated static data are used to construct derived features and integrate static features to obtain a multi-source feature vector set. Based on the multi-source feature vector set, the structured spatiotemporal sequence water source dataset is constructed.

[0021] It should be noted that the raw multi-source hydrological data refers to unprocessed raw observation data sequences that are synchronously acquired from heterogeneous data sources such as meteorological stations, hydrological monitoring stations, and remote sensing satellite databases, and stored at different temporal and spatial resolutions, including but not limited to rainfall intensity sequences, river water level sequences, and river flow sequences.

[0022] It should be noted that the preprocessing includes time alignment, missing value imputation, and outlier correction. Time alignment unifies all data sequences to a preset forecast base time scale. Missing value imputation fills in missing time periods in the data sequence using observation data from time- or spatially adjacent stations, through preset statistical relationships or spatial interpolation algorithms, to form a continuous sequence. Outlier correction identifies and corrects non-physical data spikes or abrupt changes caused by sensor malfunctions or extreme interference.

[0023] Furthermore, one specific implementation of outlier correction is to statistically identify the data sequence based on the Raida criterion: calculate the mean μ and standard deviation σ of the entire data sequence, and if the value x of a certain data point satisfies |x - μ|>3σ, it is determined to be an outlier, and it is smoothly replaced by the moving average or median value of the data in the adjacent time period.

[0024] It should be noted that the regularized multi-source hydrological data is a basic hydrological data sequence with a unified time step, continuous data, and reliable quality, after time alignment, missing data filling, and anomaly correction, providing a clean and consistent input for subsequent feature engineering.

[0025] It should be noted that the construction of the derived features is based on hydrological and meteorological physical laws. Indirect features that can reflect changes in the hydrological state of the watershed are calculated from the regularized multi-source hydrological data. For example, the amount of rainfall that affected the previous period is calculated based on the cumulative rainfall intensity sequence to characterize the soil moisture status; or the rate of change of rainfall intensity is calculated to reflect the intensity change trend of the rainfall process.

[0026] Furthermore, the initial impact on rainfall The calculation formula is as follows: ; In the formula, This represents the amount of rainfall occurring before time t. The attenuation coefficient is... This indicates the amount of rainfall affecting the area before time t-1. This represents the observed rainfall at time t. For time indexing.

[0027] It should be noted that the static feature integration refers to extracting features that do not change or change slowly from the static geographic information data of the target watershed, and spatially matching and splicing them with derived features. The associated static data includes digital elevation models and land use type data.

[0028] Furthermore, the average slope of watershed units is extracted from the digital elevation model. The algorithm expression is as follows: ; In the formula, The average slope of the watershed unit. This is the difference between the maximum and minimum elevations within a watershed unit, calculated based on DEM data. This represents the horizontal projected distance between the corresponding maximum and minimum elevation points. The greater the slope value, the faster the surface runoff converges.

[0029] It should be noted that the multi-source feature vector set refers to a multi-dimensional vector formed by splicing preprocessed regular hydrological data, derived features, and integrated static features for each unified time sample point. This vector set is continuous in the time dimension and fully represents the comprehensive state of the watershed system at each moment.

[0030] It should be noted that the structured spatiotemporal sequence water source dataset is a two-dimensional data matrix formed by organizing the multi-source feature vector set in chronological order.

[0031] In this embodiment of the invention, the simultaneous construction of a hydrophysical model to obtain a simplified distributed hydrological model includes: Based on the geographical boundary of the target watershed and the preset unit division rules, the target watershed is spatially discretized to determine at least one hydrological calculation unit. Based on historical hydrological observation datasets, a set of hydrophysical model parameters is determined for each hydrological calculation unit; Based on the preset runoff calculation equation and the parameters of the hydrophysical model, a runoff calculation module is constructed for each hydrological calculation unit; Based on the preset runoff calculation equation and the parameters of the hydrophysical model, the runoff generation calculation modules of each hydrological calculation unit are coupled in series to construct the runoff calculation module. The runoff calculation module and the confluence calculation module of all the hydrological calculation units are integrated to construct a simplified distributed hydrological model.

[0032] It should be noted that the spatial discretization refers to the process of dividing a continuous geographical watershed into several relatively independent basic computational units with consistent hydrological response based on its topography and river network characteristics, wherein each hydrological computational unit is regarded as an independent runoff generation and confluence system.

[0033] It should be noted that the hydrological calculation unit is a basic hydrological response unit obtained after spatial discretization, and its internal structure is considered to have uniform hydrological characteristics; each unit receives its own rainfall input, performs runoff calculation independently, and its output net rainfall is used as the input of the runoff system.

[0034] It should be noted that the parameters of the hydrophysical model are a set of values ​​used to quantify the physical characteristics of the hydrological process. They determine the model’s response mode and intensity to the conversion of rainfall input into river flow. Key parameters include, but are not limited to, the number of runoff curves CN, which reflects the runoff generation capacity of the underlying surface, and the linear reservoir receding coefficient K, which reflects the residence time of water flow in the watershed.

[0035] Furthermore, parameter calibration refers to the process of using historical rainfall data and corresponding actual observed flow data as input and reference, and repeatedly adjusting the parameters of the hydrophysical model through the SCE-UA algorithm to minimize the error between the flow sequence simulated by the model and the observed flow sequence.

[0036] Furthermore, when the target watershed lacks sufficient historical observation data, a parameter regionalization transplantation method can be adopted. This involves transplanting the hydrophysical model parameters determined by a neighboring calibrated watershed with similar hydrological and meteorological conditions, either directly or after correction by factors such as topography and soil, to the corresponding hydrological calculation unit of the target watershed.

[0037] It should be noted that the runoff calculation module is an algorithm unit based on hydrophysical mechanisms used to simulate and calculate the net surface rainfall process after deducting losses from rainfall; its function is to quantify the portion of each rainfall event that can form direct runoff.

[0038] Furthermore, the preset production flow calculation equation used in constructing the production flow calculation module is the SCS-CN equation, and its mathematical expression is as follows: ; ; in, For direct runoff, This represents the cumulative rainfall. This is the initial loss amount. The potential maximum retention is where the initial loss is 0.2 times the maximum retention.

[0039] Furthermore, the relationship between maximum retention and runoff curve number CN is as follows: .

[0040] Furthermore, the preset flow calculation equation used in constructing the flow calculation module is a linear reservoir equation, the mathematical expression of which is as follows: ; In the formula, Let t be the outflow rate of the reservoir. Let t be the water storage capacity of the reservoir. The recession coefficient of the reservoir. Let t be the total inflow rate into the reservoir. For time indexing.

[0041] It should be noted that the simplified distributed hydrological model is a complete computational framework formed by integrating the runoff calculation modules of all spatial discrete units and coupling them in series through the confluence calculation module according to the river network topology.

[0042] S2, input the structured spatiotemporal sequence water source dataset into the simplified distributed hydrological model for simulation to generate a physical mechanism baseline forecast sequence; In this embodiment of the invention, the step of inputting the structured spatiotemporal sequence water source dataset into the simplified distributed hydrological model for simulation to generate a physical mechanism baseline forecast sequence includes: Dynamic driving features are extracted from the structured spatiotemporal sequence water source data to drive the simplified distributed hydrological model to perform continuous simulation calculations time by time period, so as to obtain the simulated flow sequence for the whole period. Based on the historical simulation portion of the full-period simulated flow sequence and the actual observed flow sequence, the residuals of the simplified distributed hydrological model in the historical period are calculated to obtain the historical residuals. From the full-period simulated flow sequence, subsequences corresponding to future forecast periods are extracted to obtain the physical mechanism baseline forecast sequence.

[0043] It should be noted that the dynamic driving features refer to the core input data selected from the structured spatiotemporal sequence water source dataset that can directly drive the operation of the simplified distributed hydrological model, mainly the rainfall of each hydrological calculation unit at each time step.

[0044] It should be noted that the continuous simulation calculation by time period refers to the calculation process of the simplified distributed hydrological model, which is executed sequentially from the first moment of the historical period to the last moment of the future forecast period, using the time step of the structured spatiotemporal sequence water source dataset as the unit and following the time order. At each moment, the model receives the rainfall input from each unit at that moment. Based on the hydrological state at the previous moment, it calculates the simulated flow value of the watershed outlet section at the current moment through the runoff calculation module and the confluence calculation module.

[0045] Furthermore, for a simplified distributed hydrological model containing N hydrological calculation units, its simulation calculation process at time t can be abstractly described as follows: First, for the i-th unit (i=1,2,...,N), its direct runoff is calculated using its rainfall input and runoff calculation module; then, according to the watershed river network topology, the direct runoff of all units and the confluence flow from upstream units are used as inputs to the confluence module. By solving the confluence equations, the total simulated flow at the watershed outlet at time t is finally obtained. This process iteratively progresses over time.

[0046] It should be noted that the full-period simulated flow sequence is obtained through continuous simulation calculations over time periods. It is a sequence of continuous river flow values ​​arranged in chronological order, covering historical periods and future forecast periods. This sequence fully records the simulation results of the simplified distributed hydrological model on the hydrological response of the basin throughout the entire study period.

[0047] It should be noted that the physical mechanism baseline forecast sequence is a subsequence extracted from the full-period simulated flow sequence, corresponding to a specific time period that needs to be forecasted in the future, such as the next 24 hours or 72 hours. This sequence is generated entirely by a simplified distributed hydrological model based on physical mechanisms, without any data-driven model correction. It represents the physical model's "first principles" prediction of future hydrological conditions, providing a stable and interpretable baseline anchor for subsequent intelligent corrections.

[0048] It should be noted that the actual observed flow sequence is a sequence of real river flow observations obtained from hydrological monitoring stations, corresponding to the historical simulation period of the model, and arranged in chronological order after quality control.

[0049] It should be noted that the historical residual values ​​are the core target variables for subsequent training and prediction of the deep learning model.

[0050] In this embodiment of the invention, calculating the residuals of the simplified distributed hydrological model over a historical period to obtain historical residuals includes: The mathematical expression used to calculate the residuals of the simplified distributed hydrological model over historical periods is as follows: ; In the formula, Represents the historical residual at time t in the historical period. This represents the actual observed flow rate at time t. This represents the simulated flow value at time t in the historical period extracted from the full-period simulated flow sequence. For time indexing.

[0051] S3, perform residual analysis on the structured spatiotemporal sequence water source dataset based on the simplified distributed hydrological model to obtain a historical residual sequence, and correct the historical residual sequence to obtain a corrected residual sequence; In this embodiment of the invention, the step of performing residual analysis on the structured spatiotemporal sequence water source dataset based on the simplified distributed hydrological model to obtain a historical residual sequence includes: Obtain the actual observed flow sequence synchronized with the historical period of the structured spatiotemporal sequence water source dataset; Based on the simplified distributed hydrological model, historical data in the structured spatiotemporal sequence water source dataset are simulated to obtain the simulated historical flow sequence. The actual observed flow sequence and the historical simulated flow sequence are subtracted element-wise using algebraic subtraction at the same time index to obtain a standard historical residual sequence. At the same time, the historical residual at each time point is used to verify the standard historical residual sequence to obtain a historical residual sequence.

[0052] It should be noted that the actual observed flow sequence is a continuous flow observation sequence obtained from the hydrological monitoring station located at the outlet section of the target watershed. It has undergone quality control and is arranged in chronological order. The time range of this sequence completely corresponds to the historical period portion of the structured spatiotemporal sequence water source dataset, and the time resolution has also been aligned.

[0053] It should be noted that the historical simulated flow sequence is a sequence of calculated river flow values ​​that corresponds one-to-one with the actual observation time points in the historical period, obtained by re-running the model using the simplified distributed hydrological model that has been constructed and calibrated, with the dynamic driving features such as rainfall in the historical period of the structured spatiotemporal sequence water source dataset as input.

[0054] Furthermore, the simulation process drives the simplified distributed hydrological model to sequentially calculate the flow at the watershed outlet at each time step, starting from the beginning of the historical period. At each time step, the model receives the rainfall input from each hydrological calculation unit at that time. Based on the internal state of the model at the previous time step, it outputs the simulated flow at that time step through runoff generation and runoff calculation. This process is repeated iteratively until the end of the historical period.

[0055] It should be noted that the purpose of verifying the standard historical residual sequence with the historical residual at each time step is to detect whether the data of the standard historical residual sequence has uncontrollable deviations in the process. The two data are compared. If they are the same, it means that no deviation has occurred. When a deviation occurs, the historical residual is used to replace the corresponding data in the standard historical residual sequence.

[0056] It should be noted that the historical residual sequence is a time series that systematically records the systematic deviations, random errors, and hydrological process information that the physical model failed to capture due to model simplification or parameter uncertainty during the entire historical simulation period relative to actual observations.

[0057] In this embodiment of the invention, correcting the historical residual sequence to obtain a corrected residual sequence includes: Based on the historical residual sequence and the historical multi-source feature data in the structured spatiotemporal sequence water source dataset, a training sample set for deep learning model training is constructed. A residual prediction model is constructed, and the residual prediction model is trained and optimized using the training sample set to obtain a fully trained intelligent correction model. Based on the real-time updated structured spatiotemporal sequence water source dataset and the aforementioned intelligent correction model, multi-step rolling prediction is performed to generate corrected residual sequences corresponding to future forecast periods.

[0058] It should be noted that the construction of the training sample set involves establishing a temporal mapping relationship between historical multidimensional hydrological data and historical residual sequences. Specifically, for the residual at each historical moment, the structured spatiotemporal sequence data of its past consecutive moments is used as an input sample, and the residual at that historical moment is used as the target label corresponding to the input sample. By traversing the entire historical period through a sliding time window, a large number of sample pairs in the form of (input sample, target label) are generated, which constitute the training sample set.

[0059] It should be noted that the deep temporal neural network architecture selected during the construction process is preferably a long short-term memory network. Its core advantage lies in its internal gating mechanism, which can effectively capture and remember long-term dependencies and complex nonlinear dynamics in time series.

[0060] Furthermore, the training and optimization of the residual prediction model are achieved by using the backpropagation algorithm and gradient descent optimizer to minimize the gap between the model's predicted value and the actual historical residual label. The training process iteratively adjusts tens of thousands of connection weight parameters within the network until the model's prediction performance on the independent validation set tends to stabilize. At this point, the model is considered to have learned the complex patterns hidden in historical data that cause errors in the physical model, thus obtaining the fully trained intelligent correction model.

[0061] It should be noted that the intelligent correction model is the final product of the training process. It is a function that has fixed the optimal parameters and has the ability to predict future residuals from real-time multi-source feature data. The model encapsulates data-driven intelligence, and its predictive ability comes from the mining of historical residual generation patterns, rather than explicit physical equations.

[0062] It should be noted that the multi-step rolling prediction is a key operation for generating the complete corrected residual sequence for the future. At the real-time forecast time, the latest continuous observation structured spatiotemporal sequence data is first acquired to form the initial input window. This window data is then input into the intelligent correction model, and the model output is the residual prediction value for the first forecast period. Subsequently, the window is rolled: the new observation data of the actual first forecast period or the forecast-driven data are used to update the input window, the oldest time data is removed, and the latest time data is added to form a new input window, which is then input into the model again to predict the residual for the second period. This process is iterated until all future forecast periods are predicted, thus obtaining an ordered corrected residual sequence.

[0063] S4, the physical mechanism baseline forecast sequence and the corrected residual sequence are fused to obtain the dual-drive forecast result, and the physical consistency of the dual-drive forecast result is checked to generate the final flood forecast sequence.

[0064] In this embodiment of the invention, fusing the physical mechanism baseline prediction sequence with the corrected residual sequence to obtain a dual-drive prediction result includes: Based on the same time index, the physical mechanism baseline forecast sequence and the corrected residual sequence are time-series aligned to obtain aligned baseline forecast sequence and corrected residual sequence; The aligned baseline forecast sequence and the corrected residual sequence are subjected to time-by-time residual-driven fusion calculation to obtain the dual-drive prediction result.

[0065] It should be noted that the time alignment refers to ensuring that the physical mechanism baseline forecast sequence and the corrected residual sequence are completely corresponding in the time dimension. Since both are predictions made for the same future forecast period, the baseline forecast value and the corrected residual value of the corresponding period are matched one by one by matching their time step indices.

[0066] It should be noted that the time-by-time residual-driven fusion calculation is an algebraic addition operation performed on each future forecast period, which involves adding the physical mechanism baseline forecast value to the corrected residual value of the intelligent prediction for that period.

[0067] In this embodiment of the invention, the step of performing physical consistency verification on the dual-drive prediction results to generate the final flood forecast sequence includes: Based on a physical mechanism, at least one physical consistency constraint is established for the dual-drive prediction results. The dual-drive prediction results are verified based on the physical consistency constraints, and the time periods that do not meet the physical consistency requirements are identified. Based on the dynamic weight adjustment algorithm, the forecast values ​​of the identified time periods that do not meet the physical consistency are adjusted to obtain the adjusted forecast sequence; Based on the adjusted forecast sequence, the physically consistent final flood forecast sequence is generated.

[0068] It should be noted that the physical consistency constraint is a reasonable boundary set for the forecast results based on the physical laws that are universally followed in nature. Its core purpose is to prevent the fused prediction sequence from violating basic physical principles, such as "water volume appearing out of nowhere" or "flow velocity being negative". At least one constraint includes: water balance constraint, that is, the change in river storage volume during the forecast period must be consistent with the difference between inflow and outflow; and non-negativity constraint, that is, the calculated flow velocity or flow rate should not have a physically impossible negative value.

[0069] It should be noted that the dynamic weight adjustment algorithm is a weighted smoothing calculation method used for periods where verification fails, in order to achieve a balance between trusting the basic framework of the physical model and absorbing the reasonable trend of the fusion sequence. Its core idea is that the greater the degree of physical inconsistency, the more it tends to trust the baseline forecast value of the physical mechanism; conversely, it refers more to the shape of the fusion sequence.

[0070] Furthermore, the mathematical expression of the dynamic weight adjustment algorithm is as follows: ;

[0071] In the formula, This represents the final forecast flow value after adjustment for time period t. This represents the dynamically adjusted weight for time period t. This represents the physical mechanism baseline forecast flow value for time period t. This represents a sliding time window centered on time period t. The median value of the sequence. The water imbalance difference during time period t. This is the preset verification threshold.

[0072] It should be noted that the physical meaning of this dynamic weight adjustment algorithm is to achieve a conservative yet flexible correction. First, the dynamic weight is calculated based on the relationship between the current water imbalance difference and the threshold. Greater than or equal to hour, A value of 1 means that the baseline prediction value of the physical mechanism is fully trusted; when exist and In between, It varies linearly between 0.5 and 1; the greater the deviation, the higher the weight of the physical reference. when Less than This algorithm usually won't trigger at this time because the verification has already passed, but theoretically it should. The smaller the value, the more it relies on the trend value of the fusion sequence. Then, the physical baseline value and the local trend value of the fusion sequence are weighted and averaged to obtain the adjusted forecast value for that period. This adjustment method not only corrects serious physical inconsistencies but also avoids abrupt changes that may be caused by the failure of a single data source, thus maintaining the smoothness of the forecast sequence.

[0073] It should be noted that the final flood forecast sequence is the adjusted forecast sequence, which is the final output of the entire dual-drive flood forecasting intelligent algorithm. It not only integrates the dual advantages of physical mechanisms and data intelligence, but also passes rigorous physical law verification, ensuring that the forecast results have high accuracy while possessing solid physical rationality and reliability, and can be directly used for flood control decision-making and early warning issuance.

[0074] In this embodiment of the invention, the step of verifying the dual-drive prediction result based on the physical consistency constraint and identifying time periods that do not meet the physical consistency requirement includes: Based on the physical consistency constraints, the dual-drive prediction results are subjected to time-by-time verification calculations to obtain the verification results and quantization deviation values ​​for each time period. Based on the preset verification threshold, the verification results, and the quantization deviation value, a verification decision is made for each forecast period to identify periods that do not meet the physical consistency requirement.

[0075] It should be noted that the time-period verification calculation is a process of quantitative evaluation by applying the physical consistency constraints for each forecast period in the future. For the water balance constraint, the calculation involves converting the flow forecast value into the river storage capacity and checking the relationship between the storage capacity change and the net inflow.

[0076] Furthermore, the mathematical expression for the verification calculation based on water balance constraints is as follows: ; In the formula, This represents the water imbalance difference during time period t. Represents the forecast flow based on time period t. The water storage capacity of the river section at the end of time period t is calculated from the river cross-sectional morphology. This represents the water storage volume of the river section at the end of time period t-1; Indicates the time step of the forecast period. This represents the total inflow into this river segment during time period t. This represents the predicted flow rate for the dual-drive system during time period t.

[0077] It should be noted that the verification result and the quantization deviation value are two key pieces of information output for each time period. The verification result is a Boolean value, derived from the quantization deviation value. With the preset verification threshold The comparison shows that the quantization deviation value is... It objectively measures the degree to which the forecast results for that period deviate from the laws of physical conservation.

[0078] It should be noted that the verification decision is a logical process for judging the physical rationality of the forecast results for each time period. The rule is: the quantization deviation value of the current time period t is... With a reasonable threshold pre-set according to the characteristics of the watershed and the length of the time period If a comparison is made, If the water balance constraint is met during this period, the verification result is "passed," and the preliminary fusion forecast value for this period is... It is physically acceptable, if If the water conservation rule is violated during that period, the verification result will be "failed," and an adjustment mechanism needs to be initiated.

[0079] Furthermore, the preset verification threshold ,in, and These are the mean and standard deviation of the historical water imbalance series, respectively.

[0080] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.

[0081] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0082] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dual-drive intelligent algorithm for flood forecasting that integrates multi-source data and physical mechanisms, characterized in that: The method includes: S1. Obtain multi-source hydrological data of the target watershed, preprocess and construct features of the multi-source hydrological data to generate a structured spatiotemporal sequence water source dataset, and construct a hydrological physical model to obtain a simplified distributed hydrological model. S2, input the structured spatiotemporal sequence water source dataset into the simplified distributed hydrological model for simulation to generate a physical mechanism baseline forecast sequence; S3, perform residual analysis on the structured spatiotemporal sequence water source dataset based on the simplified distributed hydrological model to obtain a historical residual sequence, and correct the historical residual sequence to obtain a corrected residual sequence; S4, the physical mechanism baseline forecast sequence and the corrected residual sequence are fused to obtain the dual-drive forecast result, and the physical consistency of the dual-drive forecast result is checked to generate the final flood forecast sequence.

2. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 1, characterized in that, The process of acquiring multi-source hydrological data of the target watershed, preprocessing the multi-source hydrological data, and constructing features to generate a structured spatiotemporal sequence water source dataset includes: Based on a pre-defined list of data sources for the target watershed, raw multi-source hydrological data are obtained. Based on a preset time scale and preset processing rules, the original multi-source hydrological data is preprocessed to obtain regularized multi-source hydrological data. Based on the physical mechanism knowledge base, the regularized multi-source hydrological data and associated static data are used to construct derived features and integrate static features to obtain a multi-source feature vector set. Based on the multi-source feature vector set, the structured spatiotemporal sequence water source dataset is constructed.

3. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 1, characterized in that, The simultaneous construction of a hydrophysical model to obtain a simplified distributed hydrological model includes: Based on the geographical boundary of the target watershed and the preset unit division rules, the target watershed is spatially discretized to determine at least one hydrological calculation unit. Based on historical hydrological observation datasets, a set of hydrophysical model parameters is determined for each hydrological calculation unit; Based on the preset runoff calculation equation and the parameters of the hydrophysical model, a runoff calculation module is constructed for each hydrological calculation unit; Based on the preset runoff calculation equation and the parameters of the hydrophysical model, the runoff generation calculation modules of each hydrological calculation unit are coupled in series to construct the runoff calculation module. The runoff calculation module and the confluence calculation module of all the hydrological calculation units are integrated to construct a simplified distributed hydrological model.

4. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 1, characterized in that, The step of inputting the structured spatiotemporal sequence water source dataset into the simplified distributed hydrological model for simulation to generate a physical mechanism baseline forecast sequence includes: Dynamic driving features are extracted from the structured spatiotemporal sequence water source data to drive the simplified distributed hydrological model to perform continuous simulation calculations time by time period, so as to obtain the simulated flow sequence for the whole period. Based on the historical simulation portion of the full-period simulated flow sequence and the actual observed flow sequence, the residuals of the simplified distributed hydrological model in the historical period are calculated to obtain the historical residuals. From the full-period simulated flow sequence, subsequences corresponding to future forecast periods are extracted to obtain the physical mechanism baseline forecast sequence.

5. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 4, characterized in that, The calculation of the residuals of the simplified distributed hydrological model over historical periods to obtain historical residuals includes: The mathematical expression used to calculate the residuals of the simplified distributed hydrological model over historical periods is as follows: ； In the formula, Represents the historical residual at time t in the historical period. This represents the actual observed flow rate at time t. This represents the simulated flow value at time t in the historical period extracted from the full-period simulated flow sequence. For time indexing.

6. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 5, characterized in that, The residual analysis of the structured spatiotemporal sequence water source dataset based on the simplified distributed hydrological model yields a historical residual sequence, including: Obtain the actual observed flow sequence synchronized with the historical period of the structured spatiotemporal sequence water source dataset; Based on the simplified distributed hydrological model, historical data in the structured spatiotemporal sequence water source dataset are simulated to obtain the simulated historical flow sequence. The actual observed flow sequence and the historical simulated flow sequence are subtracted element-wise using algebraic subtraction at the same time index to obtain a standard historical residual sequence. At the same time, the historical residual at each time point is used to verify the standard historical residual sequence to obtain a historical residual sequence.

7. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 1, characterized in that, The process of correcting the historical residual sequence to obtain the corrected residual sequence includes: Based on the historical residual sequence and the historical multi-source feature data in the structured spatiotemporal sequence water source dataset, a training sample set for deep learning model training is constructed. A residual prediction model is constructed, and the residual prediction model is trained and optimized using the training sample set to obtain a fully trained intelligent correction model. Based on the real-time updated structured spatiotemporal sequence water source dataset and the aforementioned intelligent correction model, multi-step rolling prediction is performed to generate corrected residual sequences corresponding to future forecast periods.

8. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 1, characterized in that, The process of fusing the physical mechanism baseline prediction sequence with the corrected residual sequence to obtain the dual-drive prediction result includes: Based on the same time index, the physical mechanism baseline forecast sequence and the corrected residual sequence are time-series aligned to obtain aligned baseline forecast sequence and corrected residual sequence; The aligned baseline forecast sequence and the corrected residual sequence are subjected to time-by-time residual-driven fusion calculation to obtain the dual-drive prediction result.

9. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 1, characterized in that, The step of performing physical consistency verification on the dual-drive prediction results to generate the final flood forecast sequence includes: Based on a physical mechanism, at least one physical consistency constraint is established for the dual-drive prediction results. The dual-drive prediction results are verified based on the physical consistency constraints, and the time periods that do not meet the physical consistency requirements are identified. Based on the dynamic weight adjustment algorithm, the forecast values ​​of the identified time periods that do not meet the physical consistency are adjusted to obtain the adjusted forecast sequence; Based on the adjusted forecast sequence, the physically consistent final flood forecast sequence is generated.

10. The dual-drive intelligent flood forecasting algorithm integrating multi-source data and physical mechanisms as described in claim 9, characterized in that, The step of verifying the dual-drive prediction results based on the physical consistency constraints and identifying time periods that do not meet the physical consistency requirements includes: Based on the physical consistency constraints, the dual-drive prediction results are subjected to time-by-time verification calculations to obtain the verification results and quantization deviation values ​​for each time period. Based on the preset verification threshold, the verification results, and the quantization deviation value, a verification decision is made for each forecast period to identify periods that do not meet the physical consistency requirement.

Images