A water conservancy project operation monitoring method based on digital twinning
The evolutionary graph neural operator model constructed using digital twin technology solves the problems of multi-source data processing and model adaptive adjustment in water conservancy project monitoring, realizing high-precision, stable, and rapid-response water conservancy project operation monitoring, and possessing intelligent early warning capabilities.
Patent Information
- Application Number
- CN202610306974.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-13
- Publication Date
- 2026-06-16
AI Technical Summary
Existing water conservancy engineering monitoring technologies have shortcomings in multi-field information fusion, model self-learning, and system stability control. They are difficult to achieve unified processing of multi-source heterogeneous data and extraction of spatiotemporal coupling features, resulting in unstable prediction results and difficulty in adaptive adjustment of models.
Adopting the concept of digital twins, an evolutionary graph neural operator model is constructed. By combining multi-source monitoring data processing, cognitive stabilization double-loop assimilation and counterfactual inverse update mechanism, synchronous evolution prediction of hydraulic field, seepage field and structural field is achieved. The model parameters and topology are dynamically adjusted through spatiotemporal evolution memory kernel and multi-field coupling calculation.
It has achieved high-precision monitoring of the operational status of water conservancy projects, with strong model stability, fast dynamic response, and intelligent early warning capabilities. It has reduced human intervention and operation and maintenance costs, and improved the robustness and adaptability of the monitoring system.
Smart Images
Figure CN122220764A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of information technology and intelligent monitoring of water conservancy projects, and in particular to a method for monitoring the operation of water conservancy projects based on digital twins. Background Technology
[0002] With the continuous expansion of water conservancy projects and the increasing complexity of their operating environments, traditional monitoring and safety assessment methods are no longer sufficient to meet the demands for refined and real-time management. Current monitoring systems largely rely on manual inspections, fixed-point instrument observations, and single-field data analysis, primarily using static models or empirical algorithms to independently assess monitored parameters such as water level, seepage pressure, flow rate, and strain. While traditional methods can provide basic operational status information, their high degree of model simplification and low data utilization often fail to accurately reflect the dynamic response characteristics of engineering structures under complex operating conditions.
[0003] In existing technologies, some studies have attempted to integrate machine learning and numerical simulation to improve the intelligence level of monitoring. However, traditional methods generally suffer from problems such as difficulty in uniformly processing multi-source heterogeneous data, difficulty in extracting spatiotemporal coupling features, and lack of physical constraints in the model, leading to unstable and uninterpretable prediction results. Furthermore, the parameter update and model correction mechanisms are lagging; when monitoring data shows anomalies or boundary conditions change, the model struggles to adapt and adjust in a timely manner, easily causing misjudgments or delayed warnings.
[0004] Existing water conservancy project monitoring technologies still have significant shortcomings in multi-field information fusion, model self-learning, and system stability control.
[0005] Therefore, how to provide a method for monitoring the operation of water conservancy projects based on digital twins is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose a digital twin-based method for monitoring the operation of water conservancy projects. This invention integrates technologies such as multi-source monitoring data processing, evolutionary graph neural operator modeling, cognitive stabilization dual-loop assimilation, and counterfactual inverse update mechanisms to construct a digital twin that can dynamically map the actual operational state of water conservancy projects. By introducing a spatiotemporal evolution memory kernel and a multi-field coupled computation module into the model, synchronous evolution prediction of the hydraulic field, seepage field, and structural field is achieved, and the model parameters and topology can be adaptively corrected based on monitoring data deviations. This invention has advantages such as high monitoring accuracy, strong model stability, fast dynamic response, and superior intelligent early warning capabilities, and can be widely applied to the safety monitoring and operation management of water conservancy projects such as dams, canals, and pumping stations.
[0007] A method for monitoring the operation of a water conservancy project based on digital twins according to an embodiment of the present invention includes: Acquire multi-source monitoring data collected by sensors deployed at key locations in water conservancy projects, preprocess the multi-source monitoring data, and form a monitoring dataset; Based on monitoring datasets and hydraulic engineering structural information, an evolutionary graph neural operator model is constructed. A spatiotemporal evolution memory kernel is embedded in the evolutionary graph neural operator model to complete multi-field coupling calculations of hydraulic field, seepage field and structural field, and obtain the first twin prediction result. Using the monitoring dataset and the first twin prediction results, a cognitive stabilization double-loop assimilation is performed. The inner loop updates the material parameters, permeability coefficient, and boundary condition parameters based on the deviation between the monitoring data and the first twin prediction results, while the outer loop performs topological correction on the edge connectivity of the engineering component diagram based on the deviation distribution, thus obtaining the calibration parameter set. A cognitive stability control module is introduced into the cognitive stabilization dual-loop assimilation to monitor the rate of change, convergence amplitude and system stability of the assimilation residual. Based on the monitoring results, the parameter update step size and topology correction intensity are adjusted, and a second calibration parameter set is output. Based on the monitoring dataset, the first twin prediction result, and the second calibration parameter set, a counterfactual triggering and spatiotemporal inversion update mechanism is established. When the deviation between the monitoring data and the first twin prediction result exceeds a set threshold, a virtual working condition dataset containing different control variable conditions is generated, and spatiotemporal inversion calculation is performed on the virtual working condition dataset to obtain the inversion correction parameter set. The inverse correction parameter set is input into the evolution graph neural operator model to update the model parameters and topology, generate the second twin prediction result, and output multi-field monitoring results data of the water conservancy project operation status.
[0008] Optionally, the multi-source monitoring data includes water level data, flow rate data, seepage pressure data, pore water pressure data, strain data, vibration data, displacement data, temperature data, humidity data, rainfall data, wind speed data, wind direction data, soil moisture content data, groundwater level data, motor current data, gate opening data, and equipment operating status data.
[0009] Optionally, the preprocessing of multi-source monitoring data includes time synchronization processing, noise filtering processing, missing data imputation processing, outlier removal processing, unit unification conversion processing, and feature normalization processing of the collected multi-source monitoring data.
[0010] Optionally, obtaining the first twin prediction result includes: Based on the structural information of water conservancy projects, engineering component diagrams are established, and dam bodies, galleries, curtain walls, foundations, seepage channels and monitoring holes are marked as nodes. The actual connection relationships between nodes are marked as edges. The fixed node attributes include material type, geometric dimensions, boundary conditions and location identifiers. The fixed edge attributes include connection type, seepage parameters, boundary flux identifiers and control association identifiers. The monitoring dataset is aligned with the engineering component diagram in terms of time and topology. Water level data, seepage pressure data, strain data, temperature data and displacement data at the same time are written at the node level, and flow rate data, pore water pressure gradient data and gate opening data at the same time are written at the edge level. A graph input data package corresponding to the current time is generated and a one-to-one correspondence with the engineering component diagram is maintained. An evolutionary graph neural operator model is constructed and embedded with a spatiotemporal evolutionary memory kernel. The evolutionary graph neural operator model consists of a three-layer structure: Component mapping layer: Taking the engineering component diagram as input, the input data packet is encoded according to the grouping channels of nodes and edges, and a sub-channel mapping is established according to the component type while keeping the physical quantity dimension unchanged; Physical Consistency Message Layer: In the node-edge-node message transmission process, consistency constraint identifiers of hydraulic field, seepage field and structural field are introduced to pair and transmit multi-physical quantity messages within the same component, and the directionality and conservation of cross-component messages are checked and the verification results are recorded at the edge level. Memory Coupling Aggregation Layer: The spatiotemporal evolution memory kernel updates the memory state of the verification results and graph input data packets at consecutive time steps, performs cross-scale aggregation with component subgraphs as units, and generates the graph state output at the current time step; Multi-field coupling calculations are performed using the output graph. Following the correspondence between node and edge levels, the head component of the hydraulic field, the flux component of the seepage field, and the displacement and stress components of the structural field are calculated sequentially, forming a model for engineering components. Figure 1 The result set of multi-field coupling results; The multi-field coupling result set is recorded as the first twin prediction result and associated with the engineering component diagram and the current graph input data packet.
[0011] Optionally, obtaining the calibration parameter set includes: Records that are at the same time as the first twin prediction result are selected from the monitoring dataset. The same monitoring quantity is paired one by one between the monitoring dataset and the first twin prediction result. A deviation fingerprint set is generated according to a fixed field order, and a deviation fingerprint entry is established for each node and each side. The inner loop parameter assimilation is performed. Based on the deviation fingerprint set, the material parameters, permeability coefficient and boundary condition parameters are adjusted item by item according to the preset update order table. The subdomain with the highest deviation fingerprint amplitude is processed first using a domain hierarchical strategy. A parameter snapshot is generated for each parameter adjustment, and a temporary freeze flag is registered for the subgraph that shows a sudden increase in two consecutive time points. Perform outer loop topology assimilation, establish a reversible topology candidate pool, add edges that continuously exceed limits in the deviation fingerprint set, edges that correspond to historical defect records, and edges that are highly sensitive to changes in control variables to the candidate pool, and perform connectivity reassessment, channel weight redistribution, or edge replacement operations in sequence according to the priority queue, generate a topology snapshot for each topology modification, and perform shadow connectivity tests on the sandbox graph before writing it into the main graph. Cognitive stabilization control is implemented, and parameter snapshots and topology snapshots are checked in three levels in the order of boundary conservation check, cross-field consistency check and time continuity check. Items that fail the check are rolled back according to the most recent snapshot and the rollback index is registered. Items that pass the check are unfrozen and the unfreezing timestamp is registered. The verified parameters and topology corrections are summarized to form a calibration parameter set. The calibration parameter set is labeled with the time index and engineering component map version number corresponding to the monitoring dataset and the first twin prediction result.
[0012] Optionally, the output second calibration parameter set includes: Read the parameter snapshot and topology snapshot of the current moment from the calibration parameter set, pair the monitoring dataset with the first twin prediction result under the same time index and the same field order, generate residual records and establish a fixed-length sliding time window; A cognitive stability control module is constructed, which consists of three units: The residual convergence sensing unit performs statistics and sorts the change sequences of residual records and parameter snapshots within the sliding time window, and generates residual change level and convergence level. The cross-field conservation consistency verification unit performs boundary conservation verification, cross-field consistency verification, and time continuity verification on the hydraulic field, seepage field, and structural field at the node and edge levels, and generates verification result levels and a list of failures; The adaptive scheduling unit generates parameter step size suggestions, topology correction quota suggestions, rollback lists, and freeze lists based on residual change level, convergence level, and verification result level. The results of the adaptive scheduling unit are invoked to establish a stability classification for the stabilization-enhancing dual-loop assimilation, which is divided into four levels: stable, controllable, alert, and unstable. A corresponding control instruction set is generated for each subgraph and the global graph. The control instruction set includes parameter step size, topology correction quota, rollback list, and freeze list. The stability-enhancing dual-loop assimilation is scheduled based on stability classification and control instruction set: Under the stability level, it is executed according to the normal parameter step size and normal topology correction amount of the control instruction set; Reduce parameter step size and limit the number of topology changes per cycle under controllable conditions; Under the alert level, further reduce the parameter step size, perform topology correction only in the triggered area and roll back the items that failed the verification according to the rollback list, and apply a freeze list to the corresponding subgraph; At the instability level, pause topology correction, roll back to the most recent parameter snapshot according to the rollback list, and extend the observation window; The parameter snapshot after execution is merged with the topology snapshot to generate a second calibration parameter set. The second calibration parameter set is labeled with a time index and the engineering component drawing version number.
[0013] Optionally, obtaining the inversion correction parameter set includes: Based on the monitoring dataset and the first twin prediction results, residual records are generated under the same time index and the same field order. Residual statistical indicators are calculated within a fixed-length sliding time window and an adaptive threshold is generated. When the residual statistical indicators reach the threshold, a trigger flag is set. When the trigger flag is in the enabled state, a candidate set of control variable sequences is established, and external input data corresponding to the current time index is extracted from the monitoring dataset. The external input data includes rainfall, upstream flow, meteorological parameters and boundary water level. The current parameter snapshot and topology snapshot are extracted from the second calibration parameter set and bound to the engineering component drawing version number to generate a virtual working condition dataset. For each record in the virtual working condition dataset, spatiotemporal inversion calculation is performed. The inversion variable set is set to include initial condition variables, boundary condition variables, adjustable parameter variables and local topology adjustment variables. Based on the parameter snapshot and topology snapshot in the second calibration parameter set, the candidate correction parameter set is obtained one by one under the constraint of the engineering component drawing version number. The candidate correction parameter set is subjected to consistency verification and priority sorting. The consistency verification includes boundary conservation verification, cross-field consistency verification and time continuity verification. Among the candidates that pass the consistency verification, the candidates with the smallest residual with the monitoring dataset under the current time index and the limited deviation from the benchmark are selected first. The inverse correction parameter set is obtained by summarizing them. Establish an association index between the inversion correction parameter set and the monitoring dataset, the first twin prediction result, the second calibration parameter set, and the version number of the engineering component drawing.
[0014] Optionally, the multi-site monitoring results data of the output water conservancy project operation status include: Read the inversion correction parameter set, parse the inversion correction parameter set into parameter snapshots and topology snapshots corresponding to the current time index, and establish a one-to-one association with the engineering component drawing version number; Inject parameter snapshots and topology snapshots into the evolution graph neural operator model, replace the material parameters, permeability coefficients, boundary condition parameters and edge connectivity weights in the model evolution graph neural operator model, and perform memory state reset and time index alignment. The graph input data packet is invoked, and the updated evolution graph neural operator model performs multi-field coupling calculations to generate a second twin prediction result. Based on the second twin prediction results, the hydraulic head data, seepage flux data, structural displacement data and structural stress data are summarized according to the correspondence between nodes and edges in the engineering component diagram to form multi-field monitoring results data of the water conservancy project operation status. The second twin prediction results are associated and stored with multiple monitoring results data, and bound to time index, engineering component drawing version number, monitoring dataset, first twin prediction results, second calibration parameter set and inversion correction parameter set.
[0015] The beneficial effects of this invention are: This invention introduces the concept of digital twins to map the physical entity of a water conservancy project to a virtual model in real time, establishing an evolutionary graph neural operator model driven by multi-source monitoring data and engineering structural information. This model achieves dynamic interaction and coupled calculation of multiple physical quantities at the node and edge levels, comprehensively reflecting the evolutionary laws of the hydraulic field, seepage field, and structural field. Compared to traditional monitoring methods based on static numerical models, this invention can achieve spatiotemporal alignment and continuous updating of monitoring information with the support of multi-source data, significantly improving the accuracy and real-time performance of water conservancy project operation status simulation.
[0016] This invention constructs a cognitively stabilizing dual-loop assimilation mechanism. The inner loop enables adaptive parameter adjustment, while the outer loop dynamically corrects the engineering topology, thus forming a model self-learning and structural optimization process driven by monitoring data. This mechanism can dynamically monitor assimilation residuals, parameter change rates, and system stability, automatically adjusting parameter step sizes and topology correction intensity to ensure model convergence and stability under complex boundary conditions. Compared to existing models that rely on manual parameter tuning and offline correction, this invention improves the robustness and adaptability of the monitoring system while reducing human intervention and maintenance costs.
[0017] The counterfactual triggering and spatiotemporal inversion update mechanism proposed in this invention enables the model to generate virtual operating condition data and perform inverse correction when the deviation of monitoring data exceeds a threshold, thus possessing self-diagnosis and early warning capabilities for potential anomalies and emergencies. Through inversion analysis of multi-field results and parameter constraint optimization, this invention achieves dynamic perception and predictive control of the operating status of water conservancy projects. This invention possesses advantages such as high-precision prediction, strong stability, good interpretability, and fast intelligent early warning response, providing an efficient, intelligent, and sustainable monitoring method for the safe operation of water conservancy projects. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a water conservancy project operation monitoring method based on digital twin proposed in this invention; Figure 2 This is a schematic diagram of the multi-field coupled computational structure based on the evolution graph neural operator model of the water conservancy project operation monitoring method based on digital twin proposed in this invention. Detailed Implementation
[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0020] refer to Figure 1 and Figure 2 A method for monitoring the operation of water conservancy projects based on digital twins, comprising: Acquire multi-source monitoring data collected by sensors deployed at key locations in water conservancy projects, preprocess the multi-source monitoring data, and form a monitoring dataset; Based on monitoring datasets and hydraulic engineering structural information, an evolutionary graph neural operator model is constructed. A spatiotemporal evolution memory kernel is embedded in the evolutionary graph neural operator model to complete multi-field coupling calculations of hydraulic field, seepage field and structural field, and obtain the first twin prediction result. Using the monitoring dataset and the first twin prediction results, a cognitive stabilization double-loop assimilation is performed. The inner loop updates the material parameters, permeability coefficient, and boundary condition parameters based on the deviation between the monitoring data and the first twin prediction results, while the outer loop performs topological correction on the edge connectivity of the engineering component diagram based on the deviation distribution, thus obtaining the calibration parameter set. A cognitive stability control module is introduced into the cognitive stabilization dual-loop assimilation to monitor the rate of change, convergence amplitude and system stability of the assimilation residual. Based on the monitoring results, the parameter update step size and topology correction intensity are adjusted, and a second calibration parameter set is output. Based on the monitoring dataset, the first twin prediction result, and the second calibration parameter set, a counterfactual triggering and spatiotemporal inversion update mechanism is established. When the deviation between the monitoring data and the first twin prediction result exceeds a set threshold, a virtual working condition dataset containing different control variable conditions is generated, and spatiotemporal inversion calculation is performed on the virtual working condition dataset to obtain the inversion correction parameter set. The inverse correction parameter set is input into the evolution graph neural operator model to update the model parameters and topology, generate the second twin prediction result, and output multi-field monitoring results data of the water conservancy project operation status.
[0021] In this embodiment, the multi-source monitoring data includes water level data, flow rate data, seepage pressure data, pore water pressure data, strain data, vibration data, displacement data, temperature data, humidity data, rainfall data, wind speed data, wind direction data, soil moisture content data, groundwater level data, motor current data, gate opening data, and equipment operating status data.
[0022] In this embodiment, the preprocessing of multi-source monitoring data includes time synchronization processing, noise filtering processing, missing data imputation processing, outlier removal processing, unit conversion processing, and feature normalization processing of the collected multi-source monitoring data.
[0023] In this embodiment, obtaining the first twin prediction result includes: Based on the structural information of water conservancy projects, engineering component diagrams are established, and dam bodies, galleries, curtain walls, foundations, seepage channels and monitoring holes are marked as nodes. The actual connection relationships between nodes are marked as edges. The fixed node attributes include material type, geometric dimensions, boundary conditions and location identifiers. The fixed edge attributes include connection type, seepage parameters, boundary flux identifiers and control association identifiers. The monitoring dataset is aligned with the engineering component diagram in terms of time and topology. Water level data, seepage pressure data, strain data, temperature data and displacement data at the same time are written at the node level, and flow rate data, pore water pressure gradient data and gate opening data at the same time are written at the edge level. A graph input data package corresponding to the current time is generated and a one-to-one correspondence with the engineering component diagram is maintained. An evolutionary graph neural operator model is constructed and embedded with a spatiotemporal evolutionary memory kernel. The evolutionary graph neural operator model consists of a three-layer structure: Component mapping layer: Taking the engineering component diagram as input, the input data packet is encoded according to the grouping channels of nodes and edges, and a sub-channel mapping is established according to the component type while keeping the physical quantity dimension unchanged; Physical Consistency Message Layer: In the node-edge-node message transmission process, consistency constraint identifiers of hydraulic field, seepage field and structural field are introduced to pair and transmit multi-physical quantity messages within the same component, and the directionality and conservation of cross-component messages are checked and the verification results are recorded at the edge level. Memory Coupling Aggregation Layer: The spatiotemporal evolution memory kernel updates the memory state of the verification results and graph input data packets at consecutive time steps, performs cross-scale aggregation with component subgraphs as units, and generates the graph state output at the current time step; Multi-field coupling calculations are performed using the output graph. Following the correspondence between node and edge levels, the head component of the hydraulic field, the flux component of the seepage field, and the displacement and stress components of the structural field are calculated sequentially, forming a model for engineering components. Figure 1 The result set of multi-field coupling results; The multi-field coupling result set is recorded as the first twin prediction result and associated with the engineering component diagram and the current graph input data packet.
[0024] In this embodiment, obtaining the calibration parameter set includes: Records that are at the same time as the first twin prediction result are selected from the monitoring dataset. The same monitoring quantity is paired one by one between the monitoring dataset and the first twin prediction result. A deviation fingerprint set is generated according to a fixed field order, and a deviation fingerprint entry is established for each node and each side. The inner loop parameter assimilation is performed. Based on the deviation fingerprint set, the material parameters, permeability coefficient and boundary condition parameters are adjusted item by item according to the preset update order table. The subdomain with the highest deviation fingerprint amplitude is processed first using a domain hierarchical strategy. A parameter snapshot is generated for each parameter adjustment, and a temporary freeze flag is registered for the subgraph that shows a sudden increase in two consecutive time points. Perform outer loop topology assimilation, establish a reversible topology candidate pool, add edges that continuously exceed limits in the deviation fingerprint set, edges that correspond to historical defect records, and edges that are highly sensitive to changes in control variables to the candidate pool, and perform connectivity reassessment, channel weight redistribution, or edge replacement operations in sequence according to the priority queue, generate a topology snapshot for each topology modification, and perform shadow connectivity tests on the sandbox graph before writing it into the main graph. Cognitive stabilization control is implemented, and parameter snapshots and topology snapshots are checked in three levels in the order of boundary conservation check, cross-field consistency check and time continuity check. Items that fail the check are rolled back according to the most recent snapshot and the rollback index is registered. Items that pass the check are unfrozen and the unfreezing timestamp is registered. The verified parameters and topology corrections are summarized to form a calibration parameter set. The calibration parameter set is labeled with the time index and engineering component map version number corresponding to the monitoring dataset and the first twin prediction result.
[0025] In this embodiment, the output of the second calibration parameter set includes: Read the parameter snapshot and topology snapshot of the current moment from the calibration parameter set, pair the monitoring dataset with the first twin prediction result under the same time index and the same field order, generate residual records and establish a fixed-length sliding time window; A cognitive stability control module is constructed, which consists of three units: The residual convergence sensing unit performs statistics and sorts the change sequences of residual records and parameter snapshots within the sliding time window, and generates residual change level and convergence level. The cross-field conservation consistency verification unit performs boundary conservation verification, cross-field consistency verification, and time continuity verification on the hydraulic field, seepage field, and structural field at the node and edge levels, and generates verification result levels and a list of failures; The adaptive scheduling unit generates parameter step size suggestions, topology correction quota suggestions, rollback lists, and freeze lists based on residual change level, convergence level, and verification result level. The results of the adaptive scheduling unit are invoked to establish a stability classification for the stabilization-enhancing dual-loop assimilation, which is divided into four levels: stable, controllable, alert, and unstable. A corresponding control instruction set is generated for each subgraph and the global graph. The control instruction set includes parameter step size, topology correction quota, rollback list, and freeze list. The stability-enhancing dual-loop assimilation is scheduled based on stability classification and control instruction set: Under the stability level, it is executed according to the normal parameter step size and normal topology correction amount of the control instruction set; Reduce parameter step size and limit the number of topology changes per cycle under controllable conditions; Under the alert level, further reduce the parameter step size, perform topology correction only in the triggered area and roll back the items that failed the verification according to the rollback list, and apply a freeze list to the corresponding subgraph; At the instability level, pause topology correction, roll back to the most recent parameter snapshot according to the rollback list, and extend the observation window; The parameter snapshot after execution is merged with the topology snapshot to generate a second calibration parameter set. The second calibration parameter set is labeled with a time index and the engineering component drawing version number.
[0026] In this embodiment, obtaining the inversion correction parameter set includes: Based on the monitoring dataset and the first twin prediction results, residual records are generated under the same time index and the same field order. Residual statistical indicators are calculated within a fixed-length sliding time window and an adaptive threshold is generated. When the residual statistical indicators reach the threshold, a trigger flag is set. When the trigger flag is in the enabled state, a candidate set of control variable sequences is established, and external input data corresponding to the current time index is extracted from the monitoring dataset. The external input data includes rainfall, upstream flow, meteorological parameters and boundary water level. The current parameter snapshot and topology snapshot are extracted from the second calibration parameter set and bound to the engineering component drawing version number to generate a virtual working condition dataset. For each record in the virtual working condition dataset, spatiotemporal inversion calculation is performed. The inversion variable set is set to include initial condition variables, boundary condition variables, adjustable parameter variables and local topology adjustment variables. Based on the parameter snapshot and topology snapshot in the second calibration parameter set, the candidate correction parameter set is obtained one by one under the constraint of the engineering component drawing version number. The candidate correction parameter set is subjected to consistency verification and priority sorting. The consistency verification includes boundary conservation verification, cross-field consistency verification and time continuity verification. Among the candidates that pass the consistency verification, the candidates with the smallest residual with the monitoring dataset under the current time index and the limited deviation from the benchmark are selected first. The inverse correction parameter set is obtained by summarizing them. Establish an association index between the inversion correction parameter set and the monitoring dataset, the first twin prediction result, the second calibration parameter set, and the version number of the engineering component drawing.
[0027] In this embodiment, the multi-site monitoring results data of the output water conservancy project operation status include: Read the inversion correction parameter set, parse the inversion correction parameter set into parameter snapshots and topology snapshots corresponding to the current time index, and establish a one-to-one association with the engineering component drawing version number; Inject parameter snapshots and topology snapshots into the evolution graph neural operator model, replace the material parameters, permeability coefficients, boundary condition parameters and edge connectivity weights in the model evolution graph neural operator model, and perform memory state reset and time index alignment. The graph input data packet is invoked, and the updated evolution graph neural operator model performs multi-field coupling calculations to generate a second twin prediction result. Based on the second twin prediction results, the hydraulic head data, seepage flux data, structural displacement data and structural stress data are summarized according to the correspondence between nodes and edges in the engineering component diagram to form multi-field monitoring results data of the water conservancy project operation status. The second twin prediction results are associated and stored with multiple monitoring results data, and bound to time index, engineering component drawing version number, monitoring dataset, first twin prediction results, second calibration parameter set and inversion correction parameter set.
[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a reservoir, a Class II medium-sized reservoir with a maximum dam height of 56 meters and a total storage capacity of 52.5 million cubic meters, primarily responsible for urban water supply and flood control. This reservoir has been in operation for over 20 years, and in recent years, it has exhibited abnormal fluctuations in seepage pressure and increased local strain in the dam body during the flood season. Traditional static analysis methods are unable to identify potential structural risks in a timely manner, exhibiting problems of response lag and significant assessment errors.
[0029] In this embodiment, a total of 102 multi-source monitoring points were deployed in the dam body, gallery, curtain wall, and foundation areas. These included 14 water level sensors, 28 piezometers, 12 pore water pressure gauges, 20 strain gauges, 8 displacement gauges, 10 thermometers, 4 rain gauges, and 6 anemometers. Data from all monitoring points was sampled at 5-minute intervals, and the monitoring data was aggregated to the monitoring center via a wireless gateway. The system first performed time synchronization, noise filtering, and missing data imputation on the collected multi-source monitoring data according to the method steps described in the specification, forming a high-quality monitoring dataset.
[0030] Subsequently, an engineering component diagram was established based on the dam structure information, and the dam body, curtain wall, gallery, foundation, and seepage channels were modeled as nodes. Node attributes included material type, geometric dimensions, boundary conditions, and location identifiers, while edge attributes included connectivity type, permeability coefficient, and boundary flux identifiers. An evolutionary graph neural operator model was used, embedding a spatiotemporal evolution memory kernel to achieve multi-field coupled calculations of the hydraulic field, seepage field, and structural field. During system operation, a cognitive stabilization dual-loop assimilation mechanism was employed to update material parameters and boundary conditions in real time based on the deviation between monitoring data and twin prediction results, and to perform topological corrections on edges with significant connectivity changes, thereby dynamically optimizing the model structure.
[0031] During a period of continuous heavy rainfall in June 2025, the daily rainfall upstream of the reservoir reached 128 mm, and the peak inflow was 147 cubic meters per second. Traditional monitoring systems predicted a 15.2% increase in pore water pressure at the dam foundation, while the system of this invention predicted a 13.7% increase, an error of only 0.2% compared to the measured value of 13.9%. During the same period, the measured offset at the horizontal displacement monitoring point (ZQ-06) in the middle of the dam was 0.62 mm, while the system of this invention predicted 0.60 mm, an error of only 3.2%, compared to the 12.5% error of the traditional finite element static model. The system triggered a counterfactual inversion update mechanism approximately 5 minutes after the occurrence of abnormal data. By generating virtual operating condition data and performing spatiotemporal inversion calculations, it achieved dual correction of parameters and topology, improving the stability score of the reconstructed model from 0.78 to 0.91.
[0032] Table 1. Comparison of Typical Reservoir Monitoring Data and Digital Twin Prediction Results
[0033] As shown in Table 1, from June 8th to June 14th, 2025, the upstream water level of Longxi Reservoir gradually rose from 253.2 meters to 256.0 meters, with an overall water level change of approximately 2.8 meters. Correspondingly, both the seepage pressure and structural displacement within the dam body showed a slow increasing trend with rising water level, reflecting typical hydraulic-structural coupled response characteristics. A comparison of the measured and model prediction results shows that the digital twin model proposed in this invention has high fitting accuracy for both seepage pressure and displacement monitoring quantities. The deviation between the model prediction and measured values is extremely small; the seepage pressure error is stable within 0.2–0.5 kPa, and the displacement prediction error remains on the order of 0.01–0.02 mm. This indicates that the model can effectively capture the nonlinear relationship between dam seepage and structural response.
[0034] Further comparison reveals that the traditional static model's predicted seepage pressure is generally overestimated by approximately 3% to 4%, especially during the period from June 11th to June 13th, when the water level change rate is relatively rapid, the deviation is more pronounced. In contrast, the twin model of this invention adaptively corrects parameters and topology through a cognitive stabilization dual-loop assimilation mechanism during the same stage, ensuring that the predicted results always change synchronously with the measured values. This demonstrates that the model has stronger adaptability and robustness under dynamic boundary conditions. The stability score gradually increased from 0.88 to 0.93, indicating that after continuous assimilation and counterfactual inversion correction, the parameter updates within the model tend to converge, and the system operates in a highly stable state.
[0035] From the changes in model residuals, the residual percentage decreased from 3.1% to 2.2%, showing a continuous convergence trend. This indicates that with the continuous input of monitoring data and the accumulation of spatiotemporal memory in the model, the digital twin's ability to understand the operational status of water conservancy projects gradually improves. Especially under high water level conditions, the model can still maintain a low error level, demonstrating good generalization ability and engineering adaptability. Overall, Table 1 verifies the effectiveness of the method of this invention in achieving multi-field coupled calculation of hydraulic field, seepage field, and structural field driven by multi-source data. It can achieve high-precision prediction and real-time dynamic correction, providing more accurate, stable, and forward-looking technical support for the monitoring of water conservancy project operation than traditional methods.
[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for monitoring the operation of water conservancy projects based on digital twins, characterized in that, include: Acquire multi-source monitoring data collected by sensors deployed at key locations in water conservancy projects, preprocess the multi-source monitoring data, and form a monitoring dataset; Based on monitoring datasets and hydraulic engineering structural information, an evolutionary graph neural operator model is constructed. A spatiotemporal evolution memory kernel is embedded in the evolutionary graph neural operator model to complete multi-field coupling calculations of hydraulic field, seepage field and structural field, and obtain the first twin prediction result. Using the monitoring dataset and the first twin prediction results, a cognitive stabilization double-loop assimilation is performed. The inner loop updates the material parameters, permeability coefficient, and boundary condition parameters based on the deviation between the monitoring data and the first twin prediction results, while the outer loop performs topological correction on the edge connectivity of the engineering component diagram based on the deviation distribution, thus obtaining the calibration parameter set. A cognitive stability control module is introduced into the cognitive stabilization dual-loop assimilation to monitor the rate of change, convergence amplitude and system stability of the assimilation residual. Based on the monitoring results, the parameter update step size and topology correction intensity are adjusted, and a second calibration parameter set is output. Based on the monitoring dataset, the first twin prediction result, and the second calibration parameter set, a counterfactual triggering and spatiotemporal inversion update mechanism is established. When the deviation between the monitoring data and the first twin prediction result exceeds a set threshold, a virtual working condition dataset containing different control variable conditions is generated, and spatiotemporal inversion calculation is performed on the virtual working condition dataset to obtain the inversion correction parameter set. The inverse correction parameter set is input into the evolution graph neural operator model to update the model parameters and topology, generate the second twin prediction result, and output multi-field monitoring results data of the water conservancy project operation status.
2. The method for monitoring the operation of water conservancy projects based on digital twins according to claim 1, characterized in that, The multi-source monitoring data includes water level data, flow rate data, seepage pressure data, pore water pressure data, strain data, vibration data, displacement data, temperature data, humidity data, rainfall data, wind speed data, wind direction data, soil moisture content data, groundwater level data, motor current data, gate opening data, and equipment operating status data.
3. The method for monitoring the operation of water conservancy projects based on digital twins according to claim 1, characterized in that, The preprocessing of multi-source monitoring data includes time synchronization processing, noise filtering processing, missing data imputation processing, outlier removal processing, unit conversion processing, and feature normalization processing.
4. The method for monitoring the operation of water conservancy projects based on digital twins according to claim 1, characterized in that, The process of obtaining the first twin prediction result includes: Based on the structural information of water conservancy projects, engineering component diagrams are established, and dam bodies, galleries, curtain walls, foundations, seepage channels and monitoring holes are marked as nodes. The actual connection relationships between nodes are marked as edges. The fixed node attributes include material type, geometric dimensions, boundary conditions and location identifiers. The fixed edge attributes include connection type, seepage parameters, boundary flux identifiers and control association identifiers. The monitoring dataset is aligned with the engineering component diagram in terms of time and topology. Water level data, seepage pressure data, strain data, temperature data and displacement data at the same time are written at the node level, and flow rate data, pore water pressure gradient data and gate opening data at the same time are written at the edge level. A graph input data package corresponding to the current time is generated and a one-to-one correspondence with the engineering component diagram is maintained. An evolutionary graph neural operator model is constructed and embedded with a spatiotemporal evolutionary memory kernel. The evolutionary graph neural operator model consists of a three-layer structure: Component mapping layer: Taking the engineering component diagram as input, the input data packet is encoded according to the grouping channels of nodes and edges, and a sub-channel mapping is established according to the component type while keeping the physical quantity dimension unchanged; Physical Consistency Message Layer: In the node-edge-node message transmission process, consistency constraint identifiers of hydraulic field, seepage field and structural field are introduced to pair and transmit multi-physical quantity messages within the same component, and the directionality and conservation of cross-component messages are checked and the verification results are recorded at the edge level. Memory Coupling Aggregation Layer: The spatiotemporal evolution memory kernel updates the memory state of the verification results and graph input data packets at consecutive time steps, performs cross-scale aggregation with component subgraphs as units, and generates the graph state output at the current time step; Multi-field coupling calculations are performed using the output of the graph. The head component of the hydraulic field, the flux component of the seepage field, and the displacement and stress components of the structural field are calculated sequentially according to the correspondence between the node level and the edge level, forming a multi-field coupling result set consistent with the engineering component graph. The multi-field coupling result set is recorded as the first twin prediction result and associated with the engineering component diagram and the current graph input data packet.
5. The method for monitoring the operation of water conservancy projects based on digital twins according to claim 1, characterized in that, The obtained calibration parameter set includes: Records that are at the same time as the first twin prediction result are selected from the monitoring dataset. The same monitoring quantity is paired one by one between the monitoring dataset and the first twin prediction result. A deviation fingerprint set is generated according to a fixed field order, and a deviation fingerprint entry is established for each node and each side. The inner loop parameter assimilation is performed. Based on the deviation fingerprint set, the material parameters, permeability coefficient and boundary condition parameters are adjusted item by item according to the preset update order table. The subdomain with the highest deviation fingerprint amplitude is processed first using a domain hierarchical strategy. A parameter snapshot is generated for each parameter adjustment, and a temporary freeze flag is registered for the subgraph that shows a sudden increase in two consecutive time points. Perform outer loop topology assimilation, establish a reversible topology candidate pool, add edges that continuously exceed limits in the deviation fingerprint set, edges that correspond to historical defect records, and edges that are highly sensitive to changes in control variables to the candidate pool, and perform connectivity reassessment, channel weight redistribution, or edge replacement operations in sequence according to the priority queue, generate a topology snapshot for each topology modification, and perform shadow connectivity tests on the sandbox graph before writing it into the main graph. Cognitive stabilization control is implemented, and parameter snapshots and topology snapshots are checked in three levels in the order of boundary conservation check, cross-field consistency check and time continuity check. Items that fail the check are rolled back according to the most recent snapshot and the rollback index is registered. Items that pass the check are unfrozen and the unfreezing timestamp is registered. The verified parameters and topology corrections are summarized to form a calibration parameter set. The calibration parameter set is labeled with the time index and engineering component map version number corresponding to the monitoring dataset and the first twin prediction result.
6. The method for monitoring the operation of water conservancy projects based on digital twins according to claim 1, characterized in that, The output second calibration parameter set includes: Read the parameter snapshot and topology snapshot of the current moment from the calibration parameter set, pair the monitoring dataset with the first twin prediction result under the same time index and the same field order, generate residual records and establish a fixed-length sliding time window; A cognitive stability control module is constructed, which consists of three units: The residual convergence sensing unit performs statistics and sorts the change sequences of residual records and parameter snapshots within the sliding time window, and generates residual change level and convergence level. The cross-field conservation consistency verification unit performs boundary conservation verification, cross-field consistency verification, and time continuity verification on the hydraulic field, seepage field, and structural field at the node and edge levels, and generates verification result levels and a list of failures; The adaptive scheduling unit generates parameter step size suggestions, topology correction quota suggestions, rollback lists, and freeze lists based on residual change level, convergence level, and verification result level. The results of the adaptive scheduling unit are invoked to establish a stability classification for the stabilization-enhancing dual-loop assimilation, which is divided into four levels: stable, controllable, alert, and unstable. A corresponding control instruction set is generated for each subgraph and the global graph. The control instruction set includes parameter step size, topology correction quota, rollback list, and freeze list. The stability-enhancing dual-loop assimilation is scheduled based on stability classification and control instruction set: Under the stability level, it is executed according to the normal parameter step size and normal topology correction amount of the control instruction set; Reduce parameter step size and limit the number of topology changes per cycle under controllable conditions; Under the alert level, further reduce the parameter step size, perform topology correction only in the triggered area and roll back the items that failed the verification according to the rollback list, and apply a freeze list to the corresponding subgraph; At the instability level, pause topology correction, roll back to the most recent parameter snapshot according to the rollback list, and extend the observation window; The parameter snapshot after execution is merged with the topology snapshot to generate a second calibration parameter set. The second calibration parameter set is labeled with a time index and the engineering component drawing version number.
7. The method for monitoring the operation of water conservancy projects based on digital twins according to claim 1, characterized in that, The obtained inversion correction parameter set includes: Based on the monitoring dataset and the first twin prediction results, residual records are generated under the same time index and the same field order. Residual statistical indicators are calculated within a fixed-length sliding time window and an adaptive threshold is generated. When the residual statistical indicators reach the threshold, a trigger flag is set. When the trigger flag is in the enabled state, a candidate set of control variable sequences is established, and external input data corresponding to the current time index is extracted from the monitoring dataset. The external input data includes rainfall, upstream flow, meteorological parameters and boundary water level. The current parameter snapshot and topology snapshot are extracted from the second calibration parameter set and bound to the engineering component drawing version number to generate a virtual working condition dataset. For each record in the virtual working condition dataset, spatiotemporal inversion calculation is performed. The inversion variable set is set to include initial condition variables, boundary condition variables, adjustable parameter variables and local topology adjustment variables. Based on the parameter snapshot and topology snapshot in the second calibration parameter set, the candidate correction parameter set is obtained one by one under the constraint of the engineering component drawing version number. The candidate correction parameter set is subjected to consistency verification and priority sorting. The consistency verification includes boundary conservation verification, cross-field consistency verification and time continuity verification. Among the candidates that pass the consistency verification, the candidates with the smallest residual with the monitoring dataset under the current time index and the limited deviation from the benchmark are selected first. The inverse correction parameter set is obtained by summarizing them. Establish an association index between the inversion correction parameter set and the monitoring dataset, the first twin prediction result, the second calibration parameter set, and the version number of the engineering component drawing.
8. The method for monitoring the operation of water conservancy projects based on digital twins according to claim 1, characterized in that, The multi-site monitoring results data of the output water conservancy project operation status include: Read the inversion correction parameter set, parse the inversion correction parameter set into parameter snapshots and topology snapshots corresponding to the current time index, and establish a one-to-one association with the engineering component drawing version number; Inject parameter snapshots and topology snapshots into the evolution graph neural operator model, replace the material parameters, permeability coefficients, boundary condition parameters and edge connectivity weights in the model evolution graph neural operator model, and perform memory state reset and time index alignment. The graph input data packet is invoked, and the updated evolution graph neural operator model performs multi-field coupling calculations to generate a second twin prediction result. Based on the second twin prediction results, the hydraulic head data, seepage flux data, structural displacement data and structural stress data are summarized according to the correspondence between nodes and edges in the engineering component diagram to form multi-field monitoring results data of the water conservancy project operation status. The second twin prediction results are associated and stored with multiple monitoring results data, and bound to time index, engineering component drawing version number, monitoring dataset, first twin prediction results, second calibration parameter set and inversion correction parameter set.