Smart water conservancy digital twin simulation system based on multi-source data

Abstract

The present invention provides a smart water conservancy digital twin simulation system based on multi-source data, which belongs to the field of digital monitoring technology. By constructing a three-dimensional perception network of air, space, and land, combined with edge intelligent preprocessing, minute-level data collection and transmission are realized, and millimeter-wave radar, lidar and other sensors are used to collect data and reduce noise, identify anomalies, and perform dynamic data fusion and intelligent calibration. With the help of time-space calibration algorithm, Kalman filtering and blockchain evidence storage technology, data alignment and trusted verification are achieved in seconds; a hybrid simulation and intelligent decision-making model is established, and physical models, machine learning architecture, and dynamic threshold decision trees are used to realize minute-level prediction and emergency response of flood evolution. The system improves the prediction accuracy of water conservancy monitoring and the efficiency of emergency response.

Description

Technical Field

[0001] The present invention belongs to the field of digital monitoring technology, and specifically to a smart water conservancy digital twin simulation system based on multi-source data. Background Art

[0002] In the field of smart water conservancy, the existing digital twin simulation system has many technical bottlenecks in practical applications.

[0003] On the one hand, traditional data acquisition networks rely on a single sensor. When faced with complex water conservancy scenarios, data acquisition efficiency is low and noise interference is severe. Burst pulse noise and Gaussian white noise cause data quality to deteriorate, and the ability to identify failed data and abnormal data is insufficient, making it difficult to achieve minute-level data collection and transmission.

[0004] On the other hand, the fusion of multi-source heterogeneous data has problems of temporal asynchrony and spatial heterogeneity. The clock deviation and coordinate system differences of lidar, millimeter-wave radar and camera equipment make it impossible to accurately align the data, affecting subsequent analysis.

[0005] In addition, physical models suffer from significant computational inefficiencies and robustness. Traditional models, such as the Saint-Venant equation, suffer from slow computational processes and are unable to meet the minute-by-minute forecasting needs of flood evolution. Furthermore, these models rely on parameters such as roughness and bottom slope, making them susceptible to environmental influences and errors.

[0006] At the decision-making level, fixed threshold strategies are difficult to adapt to complex flood scenarios, the visualization of flood evolution data is poor, and abstract data affects decision-making efficiency; at the same time, the system computing power scheduling lacks flexibility and cannot be dynamically adjusted according to flood evolution, the fault assessment and self-healing mechanisms are imperfect, and the data transmission continuity and trusted verification capabilities are insufficient in extreme weather conditions, which seriously restricts the emergency response and precise management of smart water conservancy.

[0007] To this end, the present invention provides a smart water conservancy digital twin simulation system based on multi-source data. Summary of the Invention

[0008] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0009] The technical solution adopted by the present invention to solve its technical problem is:

[0010] In a first aspect, the present invention provides a smart water conservancy digital twin simulation system based on multi-source data, comprising:

[0011] Edge intelligent preprocessing module: Builds a three-dimensional air-space-ground-air perception network and combines it with edge intelligent preprocessing to achieve minute-level data collection and transmission;

[0012] Trusted Verification Module: This module aligns the time consistency of the data collected by the space-air-ground-ground stereoscopic perception network, uses a spatiotemporal calibration algorithm to spatially calibrate the coordinates of the data collected by the space-air-ground-ground stereoscopic perception network, and uses a digital twin-driven Kalman filter to dynamically integrate the physical model with the spatially calibrated data, thereby improving the prediction accuracy of the physical model.

[0013] Intelligent Decision-Making Module: This module builds a hybrid model based on long-short-term memory networks and graph neural networks to address physical model prediction errors. It uses a lightweight architecture combining physical models and machine learning, a dynamic threshold decision tree, and digital twin visualization technology to achieve minute-by-minute predictions of flood evolution.

[0014] System self-healing module: Builds elastic computing power and self-healing mechanisms, and ensures dynamic scheduling of system computing power and fault self-healing through edge-cloud collaborative computing and Bayesian network fault assessment.

[0015] As a further improvement plan: the specific process of constructing the air-space-ground-stereoscopic perception network is as follows:

[0016] Millimeter-wave radars are deployed simultaneously at river sections and water level monitoring points, lidars are deployed within the flooded range of the river and in the basin terrain, and 4K high-definition cameras are deployed at the water's edge and overflow identification area.

[0017] As a further improvement, the specific process of combining edge intelligent preprocessing to achieve minute-level data collection and transmission is as follows:

[0018] Deploy AI edge boxes at sensor terminals to optimize data layer quality;

[0019] The Daubechies-4 wavelet is used to decompose the time series data collected by the space-ground-air stereo perception network into five layers, balancing the computational effort with the denoising effect. Soft threshold shrinkage is used to avoid signal mutations caused by hard thresholds. The soft threshold is adaptively calculated using the Stein unbiased risk estimation to identify and process invalid and abnormal data in the data collected by the space-ground-air stereo perception network.

[0020] The isolation forest anomaly detection method is used to calculate the anomaly score of the data collected by the air-space-ground-ground stereo perception network and identify sensor drift. The PCA feature compression method is used to perform principal component analysis on the data collected by the air-space-ground-ground stereo perception network to eliminate sudden abnormal data and retain valid principal components.

[0021] The AI ​​edge box runs a random forest model to conduct real-time monitoring of data collected by the air-space-ground-stereoscopic perception network after data layer quality optimization. When the water level exceeds the warning value, it triggers an audible and visual warning to prevent invalid data from being uploaded to the cloud.

[0022] As a further improvement scheme, the specific process of performing time consistency alignment on the data collected by the air-space-ground-space stereoscopic perception network is as follows:

[0023] Using the BeiDou-3 PPS pulse-per-second signal as the time reference, the deviation between the LiDAR, millimeter-wave radar, and 4K HD camera clocks and UTC is calculated. Sliding window averaging is used to suppress noise, and the LiDAR, millimeter-wave radar, and 4K HD camera are synchronized to UTC to ensure time consistency and achieve second-level precise alignment.

[0024] As a further improvement scheme: the specific process of using the time-space calibration algorithm to perform spatial calibration on the data coordinates collected by the space-ground-air stereoscopic perception network is as follows:

[0025] Convert polar coordinates to Cartesian coordinates. The specific conversion process is:

[0026] Transformation of horizontal axis: , vertical coordinate transformation: ,in, is the distance between the laser radar and the cross section, is the angle between the radar beam and the water flow direction; after conversion, it is the CGCS2000 coordinate system of the ground sensor;

[0027] Convert the pixel coordinates of 4K high-definition cameras into geodetic coordinates and perform the conversion by calibrating the camera's seven parameters of heat exchange Bursa: ,in, , , is the translation parameter, is the scaling factor, , , is the rotation angle, calibrated by field control points.

[0028] As a further improvement, the digital twin-driven Kalman filter is used to dynamically integrate the physical model with the spatially calibrated data to improve the prediction accuracy of the physical model. The specific process is as follows:

[0029] Based on the linearization and discretization of the Saint-Venant equation, the state vector and state equation are constructed respectively. Due to the errors in the physical model prediction, it is necessary to use the real-time data collected by the air-space-ground-air stereo perception network to correct them. Kalman filtering, as a fusion tool for model prediction and real-time observation, can use sensor data to calibrate state deviations in real time while retaining the physical model mechanism. Its core is achieved through the construction of observation equations and two-step iterative updates. The observation equations establish a mapping between the system state and the real-time data collected by the air-space-ground-air stereo perception network.

[0030] If the accuracy of the space-ground stereoscopic perception network is high, the observation data will dominate the correction; if the physical model is reliable, the model prediction will dominate the correction.

[0031] As a further improvement scheme: The specific process of constructing a hybrid model based on long short-term memory network and graph neural network to solve the prediction error of physical model is as follows:

[0032] A hybrid model of LSTM (long short-term memory) networks and GNN (graph neural network) networks was constructed. The LSTM network was used to process the time series of data collected by the air-space-ground-surface stereoscopic perception network, capturing the temporal dependencies of flood evolution. The GNN processed spatial topology and captured the spatial propagation of flood evolution.

[0033] The physical model simulation results of historical flood events and actual monitoring data are used as training sets to train the model. The loss function uses the mean absolute error, and the formula is: ,in, To monitor the water level, To predict water levels for the model, is the sample size;

[0034] When flood discharge from an upstream reservoir causes a sudden change in the downstream water level, the LSTM-GNN model can quickly identify changes in the relationship between flow and water level and correct the prediction errors of the physical model.

[0035] As a further improvement plan: the specific process of achieving minute-level prediction of flood evolution is as follows:

[0036] By extracting real-time features, including current water level, water level rise rate, flow, rainfall and historical features, the historical features are the evolution path of similar flood events;

[0037] The CART decision tree is used to train the relationship between real-time features and thresholds, and the adjusted dynamic thresholds are updated in real time. The three-dimensional terrain of the watershed is constructed using LiDAR point cloud and InSAR interferometry. The Shader shader is used to simulate the flow, reflection, and inundation effects of floods. Based on the water level data calculated based on the physical model, the water surface height and flow velocity vector are updated in real time to achieve minute-level predictions of flood evolution.

[0038] As a further improvement, the specific process of building elastic computing power is as follows:

[0039] By taking on real-time lightweight tasks at the edge and processing batch calculations in the cloud, flexible scheduling of computing power is achieved. The cloud is responsible for simulating flood evolution in the entire river basin, training machine learning models, and dynamically adjusting strategies. The edge is triggered by load thresholds to offload tasks to the cloud.

[0040] As a further improvement scheme: the specific process of the fault assessment is as follows:

[0041] A Bayesian network structure is constructed, nodes are defined as devices and functions, historical fault data is used to calculate the conditional probability table, and the edge end collects the three-dimensional air-space-ground perception network status in real time and inputs it into the Bayesian network as evidence. Through joint probability reasoning, the probability of fault occurrence is calculated.

[0042] In a second aspect, the present invention provides a smart water conservancy digital twin simulation method based on multi-source data, comprising:

[0043] S1: Build a three-dimensional air-space-ground-ground perception network, combined with edge intelligent preprocessing to achieve minute-level data collection and transmission;

[0044] S2: Align the data collected by the space-air-ground-ground stereoscopic perception network for temporal consistency, use a spatiotemporal calibration algorithm to spatially calibrate the coordinates of the data collected by the space-air-ground-ground stereoscopic perception network, and use a digital twin-driven Kalman filter to dynamically fuse the physical model with the spatially calibrated data to improve the prediction accuracy of the physical model.

[0045] S3: Build a hybrid model based on long short-term memory networks and graph neural networks to address physical model prediction errors. This model uses a lightweight architecture combining physical models and machine learning, dynamic threshold decision trees, and digital twin visualization technology to achieve minute-by-minute predictions of flood evolution.

[0046] S4: Build elastic computing power and self-healing mechanisms to ensure dynamic scheduling of system computing power and self-healing of faults through edge-cloud collaborative computing and Bayesian network fault assessment.

[0047] The beneficial effects of the present invention are as follows:

[0048] Through the three-dimensional air-space-ground perception network and edge intelligent preprocessing, minute-level data collection and transmission are achieved. Multi-source sensors such as millimeter-wave radar and lidar improve data comprehensiveness. Wavelet noise reduction and anomaly detection ensure data quality. Beidou / 5G dual links ensure the continuity of extreme weather data. The time-space calibration algorithm and Kalman filtering are used to solve the time-space asynchrony problem of multi-source data and achieve second-level alignment. The blockchain evidence storage is combined to ensure the credibility and traceability of data. The hybrid simulation model accelerates the calculation of the Saint-Venant equation and corrects the roughness error through the integration of physical models and machine learning. The dynamic threshold decision tree adapts to complex flood scenarios, and digital twin visualization improves decision-making efficiency. The elastic computing power mechanism realizes edge-cloud collaborative scheduling, and the Bayesian network fault assessment ensures system self-healing, which improves the overall water conservancy monitoring prediction accuracy and emergency response capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below with reference to the accompanying drawings.

[0050] Figure 1 is a system module diagram of embodiment 1 of the present invention;

[0051] Figure 2 This is a flowchart of the steps of Example 2 of the present invention. DETAILED DESCRIPTION

[0052] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0053] Example 1

[0054] like Figure 1 As shown, the smart water conservancy digital twin simulation system based on multi-source data according to the embodiment of the present invention includes:

[0055] Edge intelligent preprocessing module: Builds a three-dimensional air-space-ground-air perception network and combines it with edge intelligent preprocessing to achieve minute-level data collection and transmission;

[0056] Specifically, the specific process of constructing the air-space-ground-stereoscopic perception network is as follows:

[0057] In addition to traditional sensors, such as water level gauges and rain gauges, other types of sensors can be added. For example, millimeter-wave radars can be deployed simultaneously at river sections and water level monitoring points, lidars can be deployed within the flooded area of ​​the river and along the basin terrain, and 4K high-definition cameras can be deployed at the river's waterline and in overflow detection areas.

[0058] Among them, millimeter-wave radars at river sections and water level monitoring locations are used to measure the velocity of water flow, which is calculated using the Doppler frequency shift formula. The specific formula is: ,in, is the Doppler shift, is the water flow velocity, is the angle between the radar beam and the water flow direction, is the radar wavelength;

[0059] LiDAR is used to construct a three-dimensional model of the water surface. The three-dimensional model of the water surface is constructed through a point cloud fitting algorithm. The specific formula is: ,in, is the point cloud weight, is the i-th laser point cloud data, is the elevation;

[0060] A 4K high-definition camera is used to identify water edges and overtopping, and a U-Net image segmentation model is used to extract water boundaries;

[0061] A BeiDou-3 short message and 5G-MEC dual link was constructed to optimize transmission layer latency. Edge nodes were deployed within 5 kilometers of water conservancy facilities. Transmission was optimized using an end-to-end latency model. The specific formula is: ,in, is the transmission delay, For edge processing delay, is the queue delay;

[0062] When the 5G signal strength is less than -110dBm, it automatically switches to BeiDou-3 short message and uses the BeiDou-3 short message protocol to transmit key data, ensuring the continuity of data transmission in extreme weather conditions.

[0063] The LoRaWAN protocol is used to realize local networking between sensors and edge nodes. The edge nodes aggregate data and then upload it via 5G / Beidou, reducing air interface occupancy.

[0064] The specific process of combining edge intelligent preprocessing to achieve minute-level data collection and transmission is as follows:

[0065] Deploy AI edge boxes at sensor terminals to optimize data layer quality;

[0066] Since the noise of the data collected by the air-space-ground stereo perception network is mainly caused by burst impulse noise and Gaussian white noise, it is necessary to use wavelet basis and threshold strategy to reduce the noise of burst impulse noise and Gaussian white noise;

[0067] The specific noise reduction process is as follows:

[0068] The Daubechies-4 wavelet is used to decompose the time series data collected by the air-ground-space stereoscopic perception network into five layers, balancing the computational effort and denoising effect. Soft threshold shrinkage is used to avoid signal mutations caused by hard thresholds. The soft threshold is adaptively calculated using the Stein unbiased risk estimation: ,in, is the i-th original data, is the denoised data of the i-th data, is the noise variance, is the wavelet coefficient;

[0069] Furthermore, the quality of the data layer is optimized;

[0070] Identify and process invalid and abnormal data in the data collected by the air-space-ground-three-dimensional perception network;

[0071] Among them, the abnormal data types include: gradual abnormalities and sudden abnormalities. Gradual abnormalities are mainly sensor drift, and sudden abnormalities are mainly data jumps, such as electromagnetic interference causing invalid values ​​of the instantaneous output of the lidar;

[0072] The isolation forest anomaly detection method is used to calculate the anomaly score of the data collected by the air-ground-air stereo perception network and identify sensor drift. The specific calculation formula is: ,in, is the path length of the data collected by the air-space-ground-stereoscopic perception network in the tree, is the average path length, is the amount of data collected by the air-space-ground-stereoscopic perception network, Score for data anomaly; set the anomaly value of the data collected by the historical space-ground stereoscopic perception network, and set When it is greater than 0.6, it is an outlier;

[0073] Using the PCA feature compression method, principal component analysis is performed on the data collected by the air-space-ground-air stereoscopic perception network to eliminate abnormal mutation data and retain valid principal components.

[0074] The AI ​​edge box runs a random forest model to monitor data collected by the air-space-ground-air stereoscopic perception network, which has been optimized for data layer quality. When the water level exceeds the warning value, an audible and visual warning is triggered to prevent invalid data from being uploaded to the cloud and shorten the response chain.

[0075] Trusted Verification Module: This module aligns the time consistency of the data collected by the space-air-ground-ground stereoscopic perception network, uses a spatiotemporal calibration algorithm to spatially calibrate the coordinates of the data collected by the space-air-ground-ground stereoscopic perception network, and uses a digital twin-driven Kalman filter to dynamically integrate the physical model with the spatially calibrated data, thereby improving the prediction accuracy of the physical model.

[0076] Specifically, the spatiotemporal calibration algorithm solves the temporal asynchrony and spatial heterogeneity issues of LiDAR, millimeter-wave radar, and 4K high-definition cameras in the air-space-ground-surface stereoscopic perception network;

[0077] Using the BeiDou-3 PPS signal as the time reference, the deviation between the LiDAR, millimeter-wave radar, and 4K HD camera clocks and UTC is calculated. Sliding window averaging is used to suppress noise, synchronizing the LiDAR, millimeter-wave radar, and 4K HD camera to UTC to ensure time consistency and achieve second-level precision alignment.

[0078] Furthermore, the spatial-temporal calibration algorithm is used to perform spatial calibration on the data coordinates collected by the space-ground-air stereoscopic perception network;

[0079] Because the coordinate representation methods of millimeter-wave radar and 4K high-definition cameras in the air-space-ground stereoscopic perception network are different, LiDAR and millimeter-wave radar are represented by polar coordinates, while 4K high-definition cameras are represented by pixel coordinates. They cannot be directly superimposed, so coordinate system conversion and alignment are required;

[0080] Convert polar coordinates to Cartesian coordinates. The specific conversion process is:

[0081] Transformation of horizontal axis: , vertical coordinate transformation: ,in, is the distance between the laser radar and the cross section, is the angle between the radar beam and the water flow direction; after conversion, it is the CGCS2000 coordinate system of the ground sensor;

[0082] Convert the pixel coordinates of 4K high-definition cameras into geodetic coordinates and perform the conversion by calibrating the camera's seven parameters of heat exchange Bursa: ,in, , , is the translation parameter, is the scaling factor, , , is the rotation angle, calibrated by field control points;

[0083] For the data collected by the spatially calibrated air-space-ground stereoscopic perception network, a digital twin-driven Kalman filter is used to dynamically integrate the physical model with the spatially calibrated data, thereby improving the prediction accuracy of the physical model.

[0084] Dynamically fuse physical model predictions with spatially calibrated data to improve physical model prediction accuracy;

[0085] The state vector and state equation are constructed based on the linearization and discretization of Saint-Venant equation respectively;

[0086] State vector: ,in, for The water level at the moment, for The flow of time, is the time step;

[0087] Equation of state: ,in, is the state transition, the expression is , is the average water level, is the time step, is the space step, linearly approximating the Saint-Venant continuity equation, is the control matrix, corresponding to the gate opening control quantity , , is the flow rate opening conversion coefficient, is the process noise, for The system state vector at time , for The system state vector at time , for The state transition matrix at time , for Control input at each moment;

[0088] Physical model predictions contain errors, which require correction using real-time data collected by the air-space-ground-ground stereoscopic perception network. Kalman filtering, as a tool for fusing model predictions and real-time observations, can use sensor data to calibrate state deviations in real time while preserving the physical model mechanism. Its core is achieved through the construction of observation equations and a two-step iterative update:

[0089] The specific process of observation equation and Kalman update is:

[0090] The observation equation establishes a mapping between the system state and the real-time data collected by the air-space-ground-air stereoscopic perception network. For the core states of the water conservancy system, water level H, flow velocity V, and flow rate Q, the observation equation describes the direct relationship between the real-time data collected by the air-space-ground-air stereoscopic perception network and the system state:

[0091] The observation equation is updated as: ,in, ,for The physical model prediction object at the moment is the state vector corresponding to the real-time data collected by the air-space-ground stereoscopic perception network. is the observation noise, for Data collected by the time-space-ground-ground stereoscopic perception network;

[0092] Kalman filtering gradually optimizes state estimation and reduces model error through two steps: prior prediction and a posteriori correction.

[0093] The state transition relationship of physical models such as the Saint-Venant equation is used to describe the evolution of water level H, flow velocity V, and flow rate Q, and predict the state and uncertainty at the next moment:

[0094] The Kalman wave iterative update is:

[0095] Prediction step: , , is the state transition relationship, that is, the evolution law of water level H, flow velocity V, and flow rate Q, Predicted by physical model Momentary status, for The fusion of physical model and observation data at each moment, the corrected system state, for The uncertainty of the physical model prediction at that moment, yes The posterior state covariance matrix at time , State transition relationship The transpose of is the process noise covariance matrix, which is used to describe the uncertainty of the physical model prediction;

[0096] The state estimate is corrected by using the Kalman gain balance model to predict the weight of the real-time data collected by the air-space-ground stereoscopic perception network:

[0097] Update step: , , is the Kalman gain, for The uncertainty of the model prediction at that moment, is the observation matrix, is the observation noise covariance, for The posterior state estimate at time t, for The state predicted by the model at that moment, for Real-time data collected by the space-time-ground-space stereoscopic perception network;

[0098] If the accuracy of the three-dimensional perception network of the sky and earth is high, Approaching 1, the observational data dominates the correction;

[0099] If the physical model is reliable, Approaching 0, the model prediction dominates the revision;

[0100] The posterior state data corrected by the Kalman filter is the core basis for flood evolution simulation and emergency decision-making. This posterior state data corrected by the Kalman filter must be fully documented to ensure its credibility:

[0101] In existing technologies, uploading all data to the blockchain is inefficient and requires filtering core data to reduce the amount of data uploaded.

[0102] First, filter the posterior state data corrected by the Kalman filter. The specific rules for filtering are as follows:

[0103] The posterior state data after Kalman filter correction is referred to as data below;

[0104] Only upload extreme value data, fused key data, and abnormal events;

[0105] The SHA-256 algorithm is used to generate a unique hash value for the filtered data, and a water conservancy alliance chain is built based on Hyperledger Fabric to ensure the traceability and non-tamperability of the data on the chain;

[0106] In the water conservancy alliance chain, nodes include river basin management agencies, equipment manufacturers, and third-party testing agencies. Authorized devices are: only sensors on the whitelist can submit data to the chain, and data from other devices is automatically marked as unauthorized.

[0107] When you need to verify whether the data has been tampered with, query the hash value of the data from the blockchain browser and compare it with the hash calculated with the original data stored locally. If the comparison results are consistent, the data has not been tampered with;

[0108] Intelligent Decision-Making Module: This module builds a hybrid model based on long-short-term memory networks and graph neural networks to address physical model prediction errors. It uses a lightweight architecture combining physical models and machine learning, a dynamic threshold decision tree, and digital twin visualization technology to achieve minute-by-minute predictions of flood evolution.

[0109] In existing technologies, the calculation process of the Saint-Venant equation is slow and cannot support minute-level response. The Preissmann four-point implicit difference method is used to discretize the Saint-Venant equations to solve the problem of the slow calculation process of the Saint-Venant equation.

[0110] The specific calculation formula is: ,in, is the water flow area, For traffic, For friction resistance, is the bottom slope, is the spatial step length, is gravity, For the time layer, is a spatial node, For the Space nodes, For the Time layer;

[0111] Deploy the matrix operations of the differential equations to the GPU, use CUDA kernel functions for parallel computing, and compress the time steps of the operations;

[0112] In existing technologies, physical models rely on roughness and bottom slope parameters, which results in insufficient robustness and unavoidable errors.

[0113] The specific process of solving the error of the physical model is:

[0114] A hybrid model of LSTM (long short-term memory) networks and GNN (graph neural network) networks was constructed. The LSTM network was used to process the time series of data collected by the air-space-ground-surface stereoscopic perception network, capturing the temporal dependencies of flood evolution. The GNN processed spatial topology and captured the spatial propagation of flood evolution.

[0115] The physical model simulation results of historical flood events and actual monitoring data are used as training sets to train the model. The loss function uses the mean absolute error, and the formula is: ,in, To monitor the water level, To predict water levels for the model, is the sample size;

[0116] Using data collected by the latest air-space-ground-surface stereoscopic perception network, the model parameters are updated to correct the roughness error of the physical model. When the upstream reservoir releases floodwaters, causing a sudden change in the downstream water level, the LSTM-GNN model can quickly identify the changing relationship between flow and water level, correcting the prediction error results of the physical model.

[0117] In existing technologies, fixed thresholds cannot adapt to complex flood scenarios. Through feature engineering extraction and dynamic threshold training, the thresholds can be updated in real time.

[0118] Specifically, by extracting real-time features, including current water level, water level rise rate, flow, rainfall and historical features, the historical features are the evolution path of similar flood events;

[0119] Use CART decision tree to train the relationship between real-time features and thresholds, and update the adjusted dynamic thresholds in real time;

[0120] In existing technologies, abstract data on flood evolution is difficult to understand intuitively, which affects decision-making efficiency;

[0121] To address the problem that abstract data on flood evolution is difficult to intuitively understand and thus hinders decision-making efficiency, we used LiDAR point clouds and InSAR interferometry to construct the three-dimensional terrain of the watershed. Shaders were used to simulate flood flow, reflection, and inundation. Based on water level data calculated using physical models, water surface height and flow velocity vectors were updated in real time, enabling minute-by-minute predictions of flood evolution.

[0122] System self-healing module: Builds elastic computing power and self-healing mechanisms, ensuring dynamic scheduling of system computing power and fault self-healing through edge-cloud collaborative computing and Bayesian network fault assessment;

[0123] Existing technologies cannot dynamically adjust computing power in water conservancy scenarios based on demand, resulting in excessive computing load during periods of rapid flood growth. Flexible computing power scheduling can be achieved by enabling the edge to handle real-time lightweight tasks and the cloud to handle batch computing.

[0124] The edge is responsible for real-time, lightweight tasks; for example, it processes data collected by the air-space-ground-three-dimensional perception network;

[0125] The cloud handles batch and high-complexity tasks; for example, it handles basin-wide flood evolution simulation and machine learning model training.

[0126] Dynamically adjust the strategy, and the edge is triggered by the load threshold to offload tasks to the cloud;

[0127] For example, during a rainstorm, when water levels rise rapidly, the real-time computing workload at the edge is automatically offloaded to the cloud due to insufficient computing power. The cloud GPU is used to accelerate computing and reduce latency.

[0128] Construct a Bayesian network structure, defining nodes as devices and functions, and edges as associations. For example, a 5G module failure can cause data transmission delays, which in turn affects flood simulation services. Use historical failure data to calculate the conditional probability table;

[0129] The edge collects the network status of the air-space-ground-ground stereoscopic perception in real time and inputs it into the Bayesian network as evidence. Through joint probabilistic reasoning, the probability of failure is calculated.

[0130] The technical solution of the embodiment of the present invention is as follows: a multi-source sensor network is constructed, and millimeter-wave radar, lidar, and 4K high-definition cameras are added to traditional water level gauges and rain gauges; BeiDou-3 short message and 5G-MEC dual-link transmission are adopted, and edge nodes are deployed within 5 kilometers of water conservancy facilities. Transmission is optimized through an end-to-end delay model, and when the 5G signal is weak, the BeiDou link is automatically switched to. Local networking is achieved by combining the LoRaWAN protocol; an AI box is deployed on the edge side, and Daubechies-4 wavelet decomposition is used to reduce noise by 5 layers combined with soft threshold shrinkage. Anomalies are determined by the isolation forest algorithm, and invalid data is eliminated by PCA.

[0131] Using the Beidou PPS pulse-per-second (PPS) as the time reference, UTC time synchronization is achieved for the lidar, millimeter-wave radar, and camera through sliding window averaging. Polar coordinates are converted to Cartesian coordinates, and 4K camera pixel coordinates are converted to CGCS2000 geodetic coordinates using Bursa's seven parameters. A Kalman filter state model is constructed based on the Saint-Venant equation, integrating physical model predictions with real-time monitoring data through an observation equation and prediction-update two-step iteration. Core data is hashed with SHA-256 and uploaded to the Hyperledger Fabric Water Conservancy Alliance Chain. Only whitelisted sensors can submit data, and hash comparison verifies immutability.

[0132] The Saint-Venant equations were discretized using Preissmann four-point implicit differencing, and GPUs were deployed to accelerate parallel computing using CUDA. An LSTM-GNN hybrid model was constructed, with LSTM processing time series and GNN capturing spatial topology. Historical data was combined to correct roughness errors. Dynamic thresholds were trained using a CART decision tree, integrating real-time features such as current water level and rise rate with historical flood paths. LiDAR point clouds and InSAR were used to construct 3D terrain, and shaders were used to simulate flood flow effects, updating water surface height and flow velocity vectors in real time.

[0133] The edge is responsible for lightweight tasks such as data preprocessing, while the cloud is responsible for full-basin simulation and model training. When the edge computing power is insufficient during heavy rain, the task is automatically offloaded to the cloud GPU; a Bayesian network is constructed, device association nodes are defined, and conditional probability tables are calculated based on historical fault data. The device status is collected in real time to infer the fault probability and achieve system self-healing.

[0134] Example 2

[0135] Based on Example 1, the present invention provides a smart water conservancy digital twin simulation method based on multi-source data, including:

[0136] S1: Build a three-dimensional air-space-ground-ground perception network, combined with edge intelligent preprocessing to achieve minute-level data collection and transmission;

[0137] S2: Align the data collected by the space-air-ground-ground stereoscopic perception network for temporal consistency, use a spatiotemporal calibration algorithm to spatially calibrate the coordinates of the data collected by the space-air-ground-ground stereoscopic perception network, and use a digital twin-driven Kalman filter to dynamically fuse the physical model with the spatially calibrated data to improve the prediction accuracy of the physical model.

[0138] S3: Build a hybrid model based on long short-term memory networks and graph neural networks to address physical model prediction errors. This model uses a lightweight architecture combining physical models and machine learning, dynamic threshold decision trees, and digital twin visualization technology to achieve minute-by-minute predictions of flood evolution.

[0139] S4: Build elastic computing power and self-healing mechanisms to ensure dynamic scheduling of system computing power and self-healing of faults through edge-cloud collaborative computing and Bayesian network fault assessment.

[0140] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The intelligent water conservancy digital twin simulation system based on multi-source data is characterized by: include: Edge intelligent preprocessing module: Builds a three-dimensional air-space-ground-air perception network and combines it with edge intelligent preprocessing to achieve minute-level data collection and transmission; The specific process of constructing the air-space-ground-stereoscopic perception network is as follows: Millimeter-wave radars are deployed simultaneously at river sections and water level monitoring points, lidars are deployed within the flooded areas of the river and along the basin terrain, and 4K high-definition cameras are deployed along the river's waterline and in overflow identification areas. The specific process of combining edge intelligent preprocessing to achieve minute-level data collection and transmission is as follows: Deploy AI edge boxes at sensor terminals to optimize data layer quality; The Daubechies-4 wavelet is used to decompose the time series data collected by the space-ground-air stereo perception network into five layers, balancing the computational effort with the denoising effect. Soft threshold shrinkage is used to avoid signal mutations caused by hard thresholds. The soft threshold is adaptively calculated using the Stein unbiased risk estimation to identify and process invalid and abnormal data in the data collected by the space-ground-air stereo perception network. The isolation forest anomaly detection method is used to calculate the anomaly score of the data collected by the air-space-ground-ground stereo perception network and identify sensor drift. The PCA feature compression method is used to perform principal component analysis on the data collected by the air-space-ground-ground stereo perception network to eliminate sudden abnormal data and retain valid principal components. The AI ​​edge box runs a random forest model to monitor data collected by the air-space-ground-air stereoscopic perception network, which has been optimized for data layer quality. When the water level exceeds the warning value, an audible and visual warning is triggered to prevent invalid data from being uploaded to the cloud. Trusted Verification Module: This module aligns the time consistency of the data collected by the space-air-ground-ground stereoscopic perception network, uses a spatiotemporal calibration algorithm to spatially calibrate the coordinates of the data collected by the space-air-ground-ground stereoscopic perception network, and uses a digital twin-driven Kalman filter to dynamically integrate the physical model with the spatially calibrated data, thereby improving the prediction accuracy of the physical model. Intelligent Decision-Making Module: This module builds a hybrid model based on long-short-term memory networks and graph neural networks to address physical model prediction errors. It uses a lightweight architecture combining physical models and machine learning, a dynamic threshold decision tree, and digital twin visualization technology to achieve minute-by-minute predictions of flood evolution. System self-healing module: Builds elastic computing power and self-healing mechanisms, and ensures dynamic scheduling of system computing power and fault self-healing through edge-cloud collaborative computing and Bayesian network fault assessment.

2. The smart water conservancy digital twin simulation system based on multi-source data according to claim 1 is characterized by: The specific process of performing time consistency alignment on the data collected by the air-space-ground-space stereoscopic perception network is as follows: Using the BeiDou-3 PPS pulse-per-second signal as the time reference, the deviation between the LiDAR, millimeter-wave radar, and 4K HD camera clocks and UTC is calculated. Sliding window averaging is used to suppress noise, and the LiDAR, millimeter-wave radar, and 4K HD camera are synchronized to UTC to ensure time consistency and achieve second-level precise alignment.

3. The smart water conservancy digital twin simulation system based on multi-source data according to claim 1 is characterized by: The specific process of using the time-space calibration algorithm to perform spatial calibration on the data coordinates collected by the space-ground stereoscopic perception network is as follows: Convert polar coordinates to Cartesian coordinates. The specific conversion process is: Transformation of horizontal axis: , vertical coordinate transformation: ,in, is the distance between the laser radar and the cross section, is the angle between the radar beam and the water flow direction; after conversion, it is the CGCS2000 coordinate system of the ground sensor; Convert the pixel coordinates of 4K high-definition cameras into geodetic coordinates and perform the conversion by calibrating the camera's seven parameters of heat exchange Bursa: ,in, , , is the translation parameter, is the scaling factor, , , is the rotation angle, calibrated by field control points.

4. The smart water conservancy digital twin simulation system based on multi-source data according to claim 1 is characterized by: The specific process of using digital twin-driven Kalman filtering to achieve dynamic fusion of physical models and spatially calibrated data and improve the prediction accuracy of physical models is as follows: Based on the linearization and discretization of the Saint-Venant equation, the state vector and state equation are constructed respectively. Due to the errors in the physical model prediction, it is necessary to use the real-time data collected by the air-space-ground-air stereo perception network to correct them. Kalman filtering, as a fusion tool for model prediction and real-time observation, can use sensor data to calibrate state deviations in real time while retaining the physical model mechanism. Its core is achieved through the construction of observation equations and two-step iterative updates. The observation equations establish a mapping between the system state and the real-time data collected by the air-space-ground-air stereo perception network. If the accuracy of the space-ground stereoscopic perception network is high, the observation data will dominate the correction; if the physical model is reliable, the model prediction will dominate the correction.

5. The smart water conservancy digital twin simulation system based on multi-source data according to claim 1 is characterized by: The specific process of constructing a hybrid model based on long short-term memory network and graph neural network to solve the prediction error of physical model is as follows: A hybrid model of LSTM (long short-term memory) networks and GNN (graph neural network) networks was constructed. The LSTM network was used to process the time series of data collected by the air-space-ground-surface stereoscopic perception network, capturing the temporal dependencies of flood evolution. The GNN processed spatial topology and captured the spatial propagation of flood evolution. The physical model simulation results of historical flood events and actual monitoring data are used as training sets to train the model. The loss function uses the mean absolute error, and the formula is: ,in, To monitor the water level, To predict water levels for the model, is the sample size; When flood discharge from an upstream reservoir causes a sudden change in the downstream water level, the LSTM-GNN model can quickly identify changes in the relationship between flow and water level and correct the prediction errors of the physical model.

6. The smart water conservancy digital twin simulation system based on multi-source data according to claim 1 is characterized by: The specific process of achieving minute-level prediction of flood evolution is as follows: By extracting real-time features, including current water level, water level rise rate, flow, rainfall and historical features, the historical features are the evolution path of similar flood events; The CART decision tree is used to train the relationship between real-time features and thresholds, and the adjusted dynamic thresholds are updated in real time. The three-dimensional terrain of the watershed is constructed using LiDAR point cloud and InSAR interferometry. The Shader shader is used to simulate the flow, reflection, and inundation effects of floods. Based on the water level data calculated based on the physical model, the water surface height and flow velocity vector are updated in real time to achieve minute-level predictions of flood evolution.

7. The smart water conservancy digital twin simulation system based on multi-source data according to claim 1 is characterized by: The specific process of building elastic computing power is as follows: By taking on real-time lightweight tasks at the edge and processing batch calculations in the cloud, flexible scheduling of computing power is achieved. The cloud is responsible for simulating flood evolution in the entire river basin, training machine learning models, and dynamically adjusting strategies. The edge is triggered by load thresholds to offload tasks to the cloud.

8. The smart water conservancy digital twin simulation system based on multi-source data according to claim 1 is characterized by: The specific process of the fault assessment is as follows: A Bayesian network structure is constructed, nodes are defined as devices and functions, historical fault data is used to calculate the conditional probability table, and the edge end collects the three-dimensional air-space-ground perception network status in real time and inputs it into the Bayesian network as evidence. Through joint probability reasoning, the probability of fault occurrence is calculated.

Images