Reservoir health state dynamic prediction method based on digital twinning

Abstract

The present application relates to the technical field of water conservancy twin prediction, and discloses a reservoir health state dynamic prediction method based on digital twinning. The method comprises starting a digital twin body synchronized with a physical reservoir, configuring an independent analysis unit with state perception and memory function for each type of monitoring signal of the digital twin body, and generating a standardized signal vector containing a time sequence context. A collaborative arbitration module is used to integrate all vectors to form a comprehensive state description matrix. A health state simulator calls a pre-trained implicit correlation network to perform state deduction on the matrix, generates a preliminary health trajectory, and compares it with a dynamically updated historical experience knowledge base. If the deviation exceeds the tolerance threshold, an external risk prediction agent is activated, and the current state matrix and difference characteristics are packaged as a task package for transmission. The method improves the accuracy of reservoir state description and early risk identification ability.

Description

Technical Field

[0001] This invention relates to the field of water conservancy digital twin prediction technology, specifically a method for dynamic prediction of reservoir health status based on digital twins. Background Technology

[0002] Current reservoir health prediction largely relies on a centralized data processing architecture within a digital twin framework. Various monitoring signals are aggregated to a unified platform for format conversion and time synchronization, forming a standardized dataset. This approach homogenizes signals with different physical characteristics and dynamic response features, resulting in the stripping away of inherent temporal patterns and historical state information during preprocessing. The data input to the analysis model lacks a deep characterization of the long-term behavioral characteristics of individual parameters. The digital twin's perception of the physical entity's state remains at the current numerical level, making it difficult to establish a refined state description with memory and contextual understanding.

[0003] In the prediction and assessment phases, existing technologies primarily rely on predefined physical models or statistical regression methods. The former struggles to accurately model the coupled effects of complex environments and material aging, while the latter fails to effectively capture the undefined inherent correlations within high-dimensional data. Furthermore, assessments of health status are typically based on fixed thresholds or static design indicators, lacking a dynamically evolving database of normal operating conditions formed during long-term reservoir operation as a benchmark. This results in insufficient system capability to identify potential risk states deviating from classic failure modes and in their early stages of development; early warnings are often based on obvious anomalies after the fact, limiting the predictive foresight and sensitivity. Summary of the Invention

[0004] The purpose of this invention is to provide a method for dynamic prediction of reservoir health status based on digital twins, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a method for dynamic prediction of reservoir health status based on digital twins, the method comprising:

[0006] A digital twin synchronized with the physical reservoir is activated. Within the digital twin, an independent analysis unit with state awareness and memory functions is configured for each type of signal in the monitoring signal stream. The analysis unit preprocesses the input signal to generate a standardized signal vector with a unified timestamp.

[0007] A collaborative arbitration module is preset in the digital twin. The collaborative arbitration module collects the standardized signal vectors generated by all analysis units at a set period, and integrates the vectors according to the preset fusion rules to form a comprehensive state description matrix.

[0008] The comprehensive state description matrix is ​​submitted to the health state simulator embedded in the digital twin. The health state simulator calls the offline pre-trained implicit association network to perform state inference on the comprehensive state description matrix and generate a preliminary health state trajectory.

[0009] The health status simulator compares the generated preliminary health status trajectory with the normal status pattern in the dynamically updated historical experience knowledge base. If the comparison result exceeds the preset deviation tolerance threshold, the health status simulator activates the risk prediction agent located outside the digital twin and encapsulates the current comprehensive status description matrix and comparison difference features into a prediction task package, which is then transmitted to the risk prediction agent.

[0010] Preferably, the digital twin that is activated and synchronized with the physical reservoir includes:

[0011] The digital twin continuously receives and verifies the multimodal monitoring signal stream reported by the sensor array;

[0012] A high-fidelity 3D virtual model of the reservoir is initialized, and the reservoir's structural parameters, historical operational data, and environmental baseline data are imported into the 3D virtual model to complete the baseline construction of the digital twin. A real-time data channel is established in the digital twin, and the real-time data channel is securely connected to the distributed sensor array of the physical reservoir to receive the multimodal monitoring signal stream in a streaming processing manner. The multimodal monitoring signal stream includes water quality parameter sequences, dam stress and strain readings, seepage field data point sets, reservoir bank deformation monitoring, dam appearance quality monitoring, reservoir water level changes, and ecological flow data monitoring.

[0013] Preferably, the preprocessing of the input signal by the analysis unit includes: each analysis unit loading a dedicated signal cleaning rule according to its corresponding signal type, the signal cleaning rule being used to remove outliers in the monitored signal stream and compensate for missing data segments; after signal cleaning, applying a normalization template corresponding to the signal type to convert the original signal values ​​of different dimensions and magnitudes into a preset numerical range to generate normalized signal segments; stamping each normalized signal segment with a high-precision timestamp and arranging them in chronological order to form the standardized signal vector, the standardized signal vector being temporarily stored in a dedicated buffer associated with each analysis unit, awaiting invocation by the collaborative arbitration module.

[0014] Preferably, the collaborative arbitration module simultaneously collects standardized signal vectors generated by all analysis units at a set period and performs vector integration according to a preset fusion rule, including: at the end of each analysis period, the collaborative arbitration module sends a vector retrieval instruction to the dedicated buffer of all analysis units to read each standardized signal vector in parallel in a non-blocking manner; after reading, it verifies the timestamp alignment and data integrity of each standardized signal vector, and for vectors with misaligned timestamps, it uses an interpolation algorithm for time synchronization processing; it concatenates all standardized signal vectors that have completed synchronization processing according to a predefined vector stacking order to form an initial fusion tensor; it applies a configurable weighted fusion kernel to the initial fusion tensor, and the weighted fusion kernel dynamically adjusts the weight coefficients according to the data confidence of each signal type in the current period, and outputs the comprehensive state description matrix after weighted calculation.

[0015] Preferably, the step of the health state simulator calling an offline pre-trained implicit association network to perform state inference on the comprehensive state description matrix includes: after receiving the comprehensive state description matrix, the health state simulator loads the structural parameters of the offline pre-trained implicit association network from its internal storage. The implicit association network is a deep neural network trained on a large amount of historical normal and abnormal state data. The comprehensive state description matrix is ​​used as input and fed into the implicit association network. The network performs nonlinear transformation and feature extraction layer by layer, and finally generates a vector sequence representing the reservoir health status at multiple consecutive time points in the future at the output layer. The vector sequence is the preliminary health state trajectory, and each vector element in the trajectory corresponds to a comprehensive health score at a predicted time point.

[0016] Preferably, the comparison between the preliminary health status trajectory generated by the health status simulator and the normal status patterns in the dynamically updated historical experience knowledge base includes: the health status simulator accessing an external database, which serves as the dynamically updated historical experience knowledge base, storing various typical health status patterns extracted from historical normal operation data, each pattern being represented as a feature vector with an occurrence frequency and seasonal label; extracting key features from the preliminary health status trajectory, calculating the multidimensional distance between the feature vectors and all normal status patterns in the dynamically updated historical experience knowledge base; selecting several normal status patterns with the smallest multidimensional distance as a reference set, and calculating a weighted average distance based on the occurrence frequency of the reference set, using the weighted average distance as the input comparison value for the deviation tolerance threshold.

[0017] Preferably, the step of activating the risk prediction agent located outside the digital twin by the health status simulator and encapsulating the current comprehensive state description matrix and contrast difference features into a prediction task package and transmitting it to the risk prediction agent includes: when the weighted average distance exceeds the deviation tolerance threshold, the health status simulator generates a trigger instruction, which is sent to the dormant risk prediction agent through an encrypted message queue to wake up the risk prediction agent; simultaneously, the health status simulator packages the current comprehensive state description matrix, the preliminary health status trajectory, the calculated multidimensional distance, and the weighted average distance together and attaches a digital signature to form the prediction task package; and pushes the prediction task package to the task receiving buffer of the awakened risk prediction agent through a protected asynchronous communication link.

[0018] Preferably, the processing of the risk prediction agent after receiving the prediction task package includes: the risk prediction agent obtaining the prediction task package from the task receiving buffer, verifying its digital signature, unpacking it, and extracting the data; based on the extracted data, the risk prediction agent initiating a simulation process containing multi-stage reasoning, the simulation process constructing multiple possible evolution paths of the current state, calling a path evaluation model based on a mixture of physical mechanisms and statistical laws, and quantitatively scoring the risk level of each possible evolution path; integrating the quantitative scores of all possible evolution paths to generate a dynamic risk spectrum, the dynamic risk spectrum describing the probability distribution and severity of various risk events within different future time windows.

[0019] Preferably, after generating the dynamic risk spectrum, the method further includes: the risk prediction agent encodes the dynamic risk spectrum into a machine-readable warning instruction sequence, and transmits the warning instruction sequence and the corresponding risk interpretation summary back to the digital twin through a security gateway; the instruction parsing module built into the digital twin receives and parses the warning instruction sequence, and drives the reservoir 3D virtual model to perform visual highlight rendering based on the parsing results, and overlays and displays the risk level and trend on the corresponding dam location, reservoir area or hydrological node.

[0020] Preferably, the step of extracting key features from the preliminary health status trajectory and calculating the multidimensional distance between the feature vectors of all normal state patterns in the dynamically updated historical experience knowledge base includes:

[0021] The health status simulator performs dual feature analysis in the time and frequency domains on the preliminary health status trajectory, and identifies inflection points in the trajectory where the rate of change exceeds a preset threshold.

[0022] Centered on each inflection point, extract adjacent trajectory segments before and after it, and calculate the mean, variance, skewness, and kurtosis of each trajectory segment to form a local feature vector;

[0023] The initial health status trajectory is divided into multiple equal-length sub-intervals, and the linear fitting slope and periodic intensity of each sub-interval are calculated to form a global feature vector.

[0024] The key features are obtained by combining the local feature vectors with the global feature vectors.

[0025] Calculate the absolute difference between the key feature and each normal state pattern feature vector in the dynamically updated historical experience knowledge base in each dimension;

[0026] The absolute differences in each dimension are multiplied by preset dimension weight coefficients and then summed to obtain the multidimensional distance between the feature vector and each normal state pattern.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] Each type of monitoring signal is equipped with an independent analysis unit with state awareness and memory capabilities, replacing the traditional centralized preprocessing data pool. Each unit continuously learns and memorizes the long-term trends, periodic characteristics, and event-related contexts of the signals it is responsible for, so that the output standardized signal vector embeds the "state history" and "behavioral characteristics" of the signal. This processing preserves the individual deep characteristics of the signals, providing the upper layer with information primitives rich in temporal context, enabling the mapping of the digital twin to the physical reservoir to move from the data synchronization level to the state cognition level, enhancing the system's sensitivity to and accuracy in describing subtle state evolution.

[0029] The health status simulator utilizes an offline pre-trained implicit correlation network to perform state inference. This network, through deep learning, encapsulates the complex and nonlinear implicit coupling mechanism between multiple reservoir parameters, enabling it to deduce a state evolution trajectory that more closely aligns with real physical laws based on the current comprehensive state. The simulator compares this preliminary trajectory with a dynamically updated historical experience knowledge base, which continuously absorbs and refines the reservoir's historical normal operation patterns. This mechanism performs real-time verification between the prediction results based on deep data correlation and the dynamic health benchmark based on historical experience. When the inferred trajectory deviates from the normal pattern in the experience base, the system not only triggers an early warning but also automatically extracts specific discrepancies, achieving early, accurate, and interpretable identification of potential risks and shifting the early warning process from threshold judgment to pattern deviation recognition. Attached Figure Description

[0030] Figure 1This is a schematic diagram illustrating the working principle of the dynamic prediction method for reservoir health status based on digital twins as described in this invention.

[0031] Figure 2 This is a flowchart illustrating the preprocessing of the input signal by the analysis unit.

[0032] Figure 3 A flowchart for integrating vectors for the collaborative arbitration module;

[0033] Figure 4 Heat map of the comprehensive state description matrix of the reservoir;

[0034] Figure 5 This is a comparison chart of risk thresholds for multimodal monitoring signals in a reservoir digital twin system. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see Figure 1 This invention provides a method for dynamic prediction of reservoir health status based on digital twins. The method includes: configuring independent analysis units with state perception and memory functions for each type of signal in the monitoring signal stream within the digital twin; these analysis units preprocess the input signals to generate standardized signal vectors with uniform timestamps; a pre-set collaborative arbitration module within the digital twin collects all standardized signal vectors at fixed intervals and integrates them according to preset rules to form a comprehensive state description matrix; this matrix is ​​submitted to an embedded health status simulator, which calls an offline pre-trained implicit association network to perform state inference on the matrix and generate a preliminary health status trajectory; the health status simulator compares this trajectory with normal state patterns in a dynamically updated historical experience knowledge base; if the comparison result exceeds a preset deviation tolerance threshold, a risk prediction agent located outside the digital twin is activated, and the current comprehensive state description matrix and comparison difference features are encapsulated into a prediction task package and transmitted to the agent.

[0037] In one embodiment of the present invention, see [reference] Figure 2The digital twin continuously receives and verifies the multimodal monitoring signal streams reported by the sensor array. Simultaneously, it initializes a high-fidelity 3D virtual model of the reservoir, importing reservoir structural parameters, historical operational data, and environmental baseline data into this model to complete the baseline construction of the digital twin. A real-time data channel is established within the digital twin, securely connected to the distributed sensor array of the physical reservoir. This channel receives the multimodal monitoring signal streams in a streaming manner. These signal streams include water quality parameter sequences, dam stress and strain readings, seepage field data points, reservoir bank deformation monitoring, dam appearance quality monitoring, reservoir water level changes, and ecological flow data monitoring. During the preprocessing of the input signals by the analysis units, each unit loads a specific signal cleaning rule based on its corresponding signal type. This rule is used to remove outliers from the monitoring signal stream and compensate for missing data segments. After signal cleaning, a normalization template corresponding to the signal type is applied to convert the original signal values ​​of different dimensions and magnitudes to a preset numerical range, generating normalized signal segments. Each normalized signal segment is tagged with a high-precision timestamp and arranged in chronological order to form a standardized signal vector, which is temporarily stored in a dedicated buffer associated with each analysis unit.

[0038] In practical implementation, the digital twin continuously receives and verifies multimodal monitoring signal streams reported by sensor arrays deployed in the physical reservoir. These multimodal monitoring signal streams are a continuous data sequence containing measurements from sensors at different physical locations and of different types. In an example scenario, for the construction of the digital twin of the Danjiangkou Reservoir, the sensor array includes multi-parameter water quality meters deployed at different depths within the reservoir area, strain gauge arrays installed on key sections of the dam, a network of piezometers deployed on the dam foundation and bank slopes, and GNSS monitoring stations located at potential landslide sites on the reservoir bank. During the startup phase, the digital twin initializes a high-fidelity 3D virtual model of the reservoir. The construction of this model imports reservoir structural parameters, including the dam's geometric dimensions, material zoning properties, and the outline of the spillway structures. Historical operational data covers the water level change process line and spillway flow records for the past ten years. Environmental baseline data includes reservoir topographic data, geological structure maps, and multi-year meteorological data. These data collectively complete the baseline construction of the digital twin, providing a static background for subsequent real-time data fusion.

[0039] A real-time data channel is established within the digital twin. This real-time data channel establishes a secure, two-way authentication connection with the distributed sensor array of the physical reservoir based on digital certificates, receiving multimodal monitoring signal streams in a streaming manner. The specific data content of the multimodal monitoring signal stream continuously flows in over time. For example, at a certain moment, the water quality parameter sequence includes numerical sequences of temperature, pH value, dissolved oxygen, turbidity, and ammonia nitrogen concentration sampled by second or minute; the dam body stress and strain readings are a set of digital signals reflecting the internal mechanical response of the dam body under different loads; the seepage field data point set is the pressure head measurement value from dozens of piezometers; the reservoir bank deformation monitoring quantity is the three-dimensional coordinate change output by the GNSS monitoring station; the dam appearance quality monitoring is the dam surface image data and characteristic parameter sequence collected at preset time intervals by a visual monitoring array composed of high-definition industrial cameras, infrared thermal imagers, and crack monitors deployed in key areas of the dam surface; the reservoir water level change is the water level value sequence collected in real time by radar water level gauges and pressure water level gauges installed at key sections such as the reservoir intake, spillway, dam front, and reservoir tail; and the ecological flow data monitoring is the flow monitoring data sequence collected by electromagnetic flow meters and ultrasonic flow meters deployed in the reservoir discharge channel, ecological water release pipe, and downstream key hydrological control sections. The real-time data channel performs format parsing and preliminary verification on the received raw data stream, removes data packets that obviously do not conform to the communication protocol, and distributes the verified data to the corresponding downstream analysis unit.

[0040] During the preprocessing stage of the input signals in the analysis unit, each analysis unit loads its own signal cleaning rules based on its corresponding signal type. These rules are stored locally in the form of configuration files. For example, the dedicated signal cleaning rules loaded by the water quality parameter analysis unit define outlier detection logic for pH sensors. When the pH value changes by more than 3.0 pH units within two consecutive sampling periods, the latter sampled value is marked as an outlier and discarded. For missing data segments, the dedicated signal cleaning rules stipulate the use of time series prediction methods for compensation. Specifically, this involves training a linear prediction model using valid data from the previous time window of the sensor, and then using the model to generate estimated values ​​for the missing time points to fill in the gaps. The dedicated signal cleaning rules for the dam stress analysis unit set a reasonable range for stress values ​​based on the principles of materials mechanics; readings exceeding this range are considered outliers.

[0041] After signal cleaning, a normalization template corresponding to the signal type is applied. This template defines a linear or nonlinear transformation function that maps the original signal to a uniform numerical range. Taking dam stress-strain readings as an example, the corresponding normalization template first calculates the maximum and minimum stress values ​​at the monitoring point from historical data. Then, it uses a formula to transform the real-time stress readings to the [0,1] interval. A specific transformation formula for this application is as follows:

[0042]

[0043] in: This represents the normalized stress value. This represents the original stress reading. This represents the minimum stress value at that monitoring point extracted from historical data. This represents the maximum stress value at the monitoring point extracted from historical data. Through normalization template processing, the original signal values ​​of different dimensions and magnitudes are transformed into a preset numerical range, generating normalized signal segments.

[0044] Each normalized signal segment is assigned a high-precision timestamp, derived from the unified network time protocol clock within the digital twin, ensuring a consistent time base for the data output by all analysis units. Multiple consecutive normalized signal segments are combined into a standardized signal vector, arranged in chronological order. This standardized signal vector is temporarily stored in a dedicated buffer associated with each analysis unit. The dedicated buffer employs a first-in, first-out (FIFO) queue data structure, with the queue length pre-set based on the signal sampling frequency and the call cycle of the collaborative arbitration module. When the collaborative arbitration module issues a call command, the standardized signal vector is read from the head of the queue in the dedicated buffer. After reading, the storage location is released for new normalized signal segments to be written, thus forming a circular buffer mechanism.

[0045] In one embodiment of the present invention, see [reference] Figure 3The collaborative arbitration module simultaneously collects standardized signal vectors generated by all analysis units at a set period and integrates them according to preset fusion rules. At the end of each analysis period, the collaborative arbitration module sends vector retrieval instructions to the dedicated buffers of all analysis units to read each standardized signal vector in parallel in a non-blocking manner. After reading, it verifies the timestamp alignment and data integrity of each standardized signal vector. For vectors with misaligned timestamps, an interpolation algorithm is used for time synchronization. All standardized signal vectors that have completed synchronization are concatenated according to a predefined vector stacking order to form an initial fusion tensor. A configurable weighted fusion kernel is applied to the initial fusion tensor. This weighted fusion kernel dynamically adjusts the weight coefficients according to the data confidence of each signal type in the current period. After weighted calculation, a comprehensive state description matrix is ​​output. After receiving the comprehensive state description matrix, the health state simulator loads the structural parameters of an offline pre-trained implicit association network from its internal storage. This implicit association network is a deep neural network trained on a large amount of historical normal and abnormal state data. The comprehensive state description matrix is ​​fed into the implicit correlation network as input. The network performs nonlinear transformation and feature extraction layer by layer. Finally, a vector sequence representing the health status of the reservoir at multiple consecutive time points in the future is generated at the output layer. This vector sequence is the preliminary health status trajectory. Each vector element in the trajectory corresponds to a comprehensive health score at a predicted time point.

[0046] In practical implementation, the collaborative arbitration module simultaneously collects standardized signal vectors generated by all analysis units according to a set period and integrates them according to preset fusion rules. The health state simulator calls an offline pre-trained implicit correlation network to perform state deduction on the comprehensive state description matrix. At the end of each analysis period, the collaborative arbitration module sends a vector retrieval instruction to the dedicated buffers of all analysis units. This instruction is a standardized message containing a period identifier and a request time window, allowing for parallel reading of each standardized signal vector in a non-blocking manner. This non-blocking approach ensures that the collaborative arbitration module can continue processing other requests without waiting for a single buffer response after issuing a read request, thereby improving data collection efficiency. After reading, the timestamp alignment and data integrity of each standardized signal vector are verified. The timestamp alignment check is achieved by comparing whether the start and end timestamps of different vector sequences are in the same physical time interval. For vectors with misaligned timestamps, an interpolation algorithm is used for time synchronization. The interpolation algorithm, such as cubic spline interpolation, constructs a smooth curve based on existing data points and estimates the signal value of the missing time points. All standardized signal vectors that have completed synchronization are spliced ​​together according to a predefined vector stacking order. The predefined vector stacking order is set according to the priority of signal type when the reservoir digital twin is initialized. For example, the order is water quality parameter vector, dam stress and strain vector, seepage field data vector, and reservoir bank deformation vector. The splicing forms an initial fusion tensor, which is a multidimensional array. Its first dimension represents the time step, and the second dimension represents the total feature dimension after splicing.

[0047] A configurable weighted fusion kernel is applied to the initial fusion tensor. This kernel dynamically adjusts the weight coefficients based on the data confidence level of each signal type in the current period. The data confidence level calculation depends on the noise level and data loss rate of the corresponding signal within the recent historical window. Signal types with low noise levels and low data loss rates are assigned higher weight coefficients. The weighted calculation outputs a comprehensive state description matrix. The weighting process is implemented through element-wise multiplication, as expressed by the formula:

[0048]

[0049] in: This represents the output comprehensive state description matrix. Represents the initial fusion tensor. Represents the weighted fusion kernel vector. This represents the element-wise multiplication operator. The transpose of the weighted fusion kernel vector is represented by an element-wise multiplication with the initial fusion tensor via a broadcast mechanism to ensure that the output matrix has the same dimension as the input tensor. In some embodiments, the dynamic adjustment of the weighted fusion kernel is based on real-time monitoring of data quality indicators, such as signal-to-noise ratio or transmission stability, thereby adaptively optimizing the fusion result.

[0050] After receiving the comprehensive state description matrix, the health state simulator loads the structural parameters of an offline pre-trained implicit association network from its internal storage. This internal storage is a non-volatile memory on the digital twin host machine. The offline pre-trained implicit association network is a deep neural network trained on a large amount of historical normal and abnormal state data. The historical data comes from complete monitoring records and event logs of the reservoir over the past few years. The comprehensive state description matrix is ​​fed into the implicit association network as input. Before the feeding operation, the matrix dimensions need to be adjusted to meet the requirements of the network input layer. The implicit association network performs nonlinear transformations and feature extraction layer by layer. The nonlinear transformations are implemented through activation functions, such as using a modified linear unit function to transform the input data. Through element-wise mapping, the network learns the complex spatiotemporal correlation patterns between monitoring signals in the hidden layers, and finally generates a vector sequence representing the reservoir's health status at multiple consecutive time points in the output layer. This vector sequence is the preliminary health status trajectory, and each vector element in the trajectory corresponds to a comprehensive health score at a predicted time point. The comprehensive health score is a value between 0 and 1, with higher values ​​indicating better reservoir health. Optionally, the implicit correlation network can adopt a long short-term memory network or a temporal convolutional network architecture to capture long-term dependencies in the time series. It can be understood that the state inference process uses a deep learning model to encapsulate the implicit coupling mechanism between multiple source parameters of the reservoir, thereby improving the accuracy of the predicted trajectory.

[0051] In some embodiments, the analysis cycle of the collaborative arbitration module can be flexibly configured according to the reservoir's operational characteristics. For example, it can be set to 3 minutes during the flood season to improve response speed and 10 minutes during the non-flood season to reduce computational load. Data comparison shows that when the timestamp misalignment rate is less than 5%, the vector consistency error after synchronization of the interpolation algorithm can be controlled within the allowable range, ensuring fusion quality. The implicit association network output trajectory of the health status simulator covers the next 12 hours, generating health scores for 48 time points at 15-minute intervals. This fine-grained prediction provides sufficient early warning time for risk management. Optionally, the network training process uses a cross-validation strategy to optimize hyperparameters and avoid overfitting. In the offline training process of the implicit association network, the specific implementation of the cross-validation strategy to optimize hyperparameters includes dividing the historical dataset into multiple complementary subsets, taking turns using each subset as a validation set to evaluate model performance, and using the remaining subset as a training set for model fitting. Hyperparameters such as the number of network layers, the number of neurons, and the learning rate are adjusted through multiple iterations of training. Finally, the hyperparameter combination that performs best on the validation set is selected, thereby enhancing the model's generalization ability and avoiding overfitting.

[0052] In one embodiment of the invention, the health status simulator compares the generated preliminary health status trajectory with normal status patterns in a dynamically updated historical experience knowledge base. The health status simulator accesses an external database, which, as a dynamically updated historical experience knowledge base, stores various typical health status patterns extracted from historical normal operation data. Each pattern is represented as a feature vector with an occurrence frequency and seasonality label. Key features are extracted from the preliminary health status trajectory, and the multidimensional distance between the feature vectors and all normal status patterns in the dynamically updated historical experience knowledge base is calculated. Several normal status patterns with the smallest multidimensional distance are selected as a reference set, and a weighted average distance is calculated based on the occurrence frequency of the reference set. This weighted average distance is used as the input comparison value for deviation tolerance threshold. The process of extracting key features from the preliminary health status trajectory involves performing dual feature analysis in the time and frequency domains on the preliminary health status trajectory to identify inflection points in the trajectory where the rate of change exceeds a preset threshold. Adjacent trajectory segments are extracted centered on each inflection point, and the mean, variance, skewness, and kurtosis of each trajectory segment are calculated to form a local feature vector. The initial health status trajectory is divided into multiple equal-length sub-intervals. The linear fitting slope and periodicity of each sub-interval are calculated to form a global feature vector. The local feature vector is combined with the global feature vector to obtain key features. The absolute difference between the key features and the feature vector of each normal state pattern in the dynamically updated historical experience knowledge base is calculated in each dimension. The absolute differences in each dimension are multiplied by preset dimension weight coefficients and then summed to obtain the multidimensional distance between the key features and the feature vector of each normal state pattern.

[0053] In practical implementation, the health status simulator accesses an external database, which serves as a dynamically updated historical experience knowledge base. This knowledge base stores various typical health status patterns extracted from historical normal operation data. Each typical health status pattern is represented as a feature vector with an occurrence frequency and a seasonal label. The seasonal label identifies the season or hydrological period in which the pattern mainly occurs. In an example scenario, for the digital twin of Fengman Reservoir, the dynamically updated historical experience knowledge base includes "stable operation mode during spring snowmelt season," "high water level flood storage mode during summer main flood season," "slow water level decline mode during autumn regulation period," and "low temperature and low load mode during winter ice cover period." The feature vector of each pattern is obtained by extracting the centroids after performing cluster analysis on a large number of comprehensive state description matrices within the corresponding historical period.

[0054] Key features are extracted from the initial health status trajectory. The health status simulator performs dual feature analysis in both the time and frequency domains on the initial health status trajectory. First, a sliding window difference calculation is applied to the initial health status trajectory sequence to identify inflection points where the rate of change exceeds a preset threshold. The preset threshold is set based on the normal fluctuation range of the reservoir health status score. Adjacent trajectory segments are extracted centered on each inflection point, with the length of each segment defined as ten sampling points before and after the inflection point. The mean, variance, skewness, and kurtosis of each trajectory segment are calculated to form a local feature vector. The initial health status trajectory is then divided into multiple equal-length sub-intervals, the number of which is dynamically determined based on the total trajectory length. The linear fitting slope and periodicity intensity of each sub-interval are calculated. The periodicity intensity is obtained by calculating the amplitude of the main frequency components after the Fourier transform of the sub-interval sequence, forming a global feature vector. The local and global feature vectors are combined to obtain the key feature, which is a vector with a dimension much higher than the original initial health status trajectory.

[0055] Calculate the multidimensional distance between the key feature and all normal state pattern feature vectors in the dynamically updated historical experience knowledge base. Calculate the absolute difference in each dimension between the key feature and each normal state pattern feature vector in the dynamically updated historical experience knowledge base. Multiply the absolute differences in each dimension by preset dimension weight coefficients and sum them to obtain the multidimensional distance between the key feature and each normal state pattern feature vector. The dimension weight coefficients are preset based on the discriminative power of different feature dimensions in distinguishing between normal and abnormal states; feature dimensions with higher discriminative power are assigned higher weights. The calculation of the multidimensional distance can be expressed as:

[0056]

[0057] in: This represents the key features and dynamically updated historical experience knowledge base. Multidimensional distance between feature vectors of a normal state pattern Indicates the first The preset dimension weight coefficients for each feature dimension. This indicates that the key features extracted from the initial health status trajectory are in the first... Values ​​of each dimension This indicates the first [item] in the dynamically updated historical experience knowledge base. The feature vector of the normal state pattern at the th in the ... Values ​​of each dimension This represents the total number of dimensions of the key features. This indicates the calculation of the absolute difference. This represents the summation of the weighted absolute differences across all dimensions.

[0058] In some embodiments, a set of normal state patterns with the smallest multidimensional distance is selected as a reference set, and the size of the reference set is fixed at three. A weighted average distance is calculated based on the frequency of occurrence of the reference patterns, and this weighted average distance serves as the input comparison value for deviations from the tolerance threshold. The weighted average distance is calculated by multiplying the multidimensional distance of each normal state pattern selected for the reference set by its normalized weight based on its frequency of occurrence recorded in the historical experience knowledge base, and then summing the results.

[0059] It is understandable that the dynamically updated historical experience knowledge base has an update mechanism. When a reservoir experiences new typical operating conditions and is verified to be in a healthy state, its corresponding data patterns, after feature extraction, will be added to the dynamically updated historical experience knowledge base. Optionally, the seasonal labels include not only natural seasons but also operating condition labels such as "flood discharge scheduling period" and "peak power generation period." In some embodiments, the preset dimension weight coefficients are not completely fixed but are designed with a fine-tuning mechanism. Based on feedback from recent comparison results, the fine-tuning mechanism appropriately reduces the weight coefficients of dimensions with decreased discriminative power.

[0060] In one embodiment of the invention, when the weighted average distance exceeds the deviation tolerance threshold, the health state simulator generates a trigger command. This trigger command is sent to the dormant risk prediction agent via an encrypted message queue, waking the risk prediction agent. Simultaneously, the health state simulator packages the current comprehensive state description matrix, the preliminary health state trajectory, the calculated multidimensional distance, and the weighted average distance together and attaches a digital signature to form a prediction task package. The prediction task package is pushed to the task receiving buffer of the awakened risk prediction agent via a protected asynchronous communication link. In a specific implementation, when the weighted average distance exceeds the deviation tolerance threshold, the health state simulator generates a trigger command. The weighted average distance is a single quantitative value calculated by the health state simulator to measure the difference between the current reservoir state and the historical normal pattern. The trigger command is a structured digital command containing an event type identifier "state deviation warning," a trigger timestamp, the associated health state simulator instance number, and the specific calculated weighted average distance value. The trigger command is sent to the dormant risk prediction agent via an encrypted message queue. The encrypted message queue uses a TLS-based secure communication protocol to ensure that the command is not tampered with or eavesdropped on during transmission.

[0061] It is understandable that the risk prediction agent, deployed as an independent computing process on a secure computing node outside the digital twin, remains in a low-power sleep state to conserve system resources when no trigger command is received. The operation to wake up the risk prediction agent is performed by its daemon process, which continuously listens to the encrypted message queue. Upon receiving a trigger command from the health state simulator, it immediately starts the main business logic process of the risk prediction agent. Simultaneously, the health state simulator packages the current comprehensive state description matrix, the preliminary health state trajectory, the calculated multi-dimensional distance, and the weighted average distance together and attaches a digital signature to form a prediction task package. The digital signature is generated by encrypting the hash value of the core metadata of the prediction task package using the health state simulator's private key, and is used by the recipient to verify the legitimacy and integrity of the data source.

[0062] The prediction task package uses a binary encapsulation format, which includes a fixed-length header and multiple data segments. The header records the version number, total length, creation time, and number of data segments contained in the prediction task package. The data segments sequentially store the serialized byte stream of the comprehensive state description matrix, the preliminary health state trajectory array, the multi-dimensional distance vector, the weighted average distance scalar value, and the digital signature string for verification. The prediction task package is pushed to the task receive buffer of the awakened risk prediction agent via a protected asynchronous communication link. This protected asynchronous communication link uses a WebSocket long-lived connection with authentication. The asynchronous nature ensures that the health state simulator can continue executing other computational tasks after sending the prediction task package without waiting for confirmation from the risk prediction agent.

[0063] In some embodiments, the comparison between the weighted average distance and the deviation tolerance threshold is a continuous decision point. A data comparison for a specific scenario is shown in the table below. This table displays the comparison between the weighted average distance calculated by the health status simulator at five consecutive time points within a certain operating cycle of the Three Gorges Reservoir digital twin, and the preset deviation tolerance threshold, as well as the judgment result on whether to generate a trigger command. See Table 1.

[0064] Table 1: Comparison of Weighted Average Distance and Deviation Tolerance Threshold, and Trigger Decision Table

[0065] Point in time (UTC) Weighted average distance Deviation from tolerance threshold Does it exceed the threshold? Generate trigger command? 2023-10-2608:00:00 0.152 0.300 no no 2023-10-2608:05:00 0.287 0.300 no no 2023-10-2608:10:00 0.415 0.300 yes yes 2023-10-2608:15:00 0.502 0.300 yes yes 2023-10-2608:20:00 0.338 0.300 yes yes

[0066] As can be seen from the data comparison in Table 1, at the time points "2023-10-26 08:10:00" and the two time points thereafter, the weighted average distance consistently exceeded the deviation tolerance threshold of 0.300. Therefore, the health state simulator would execute the operations of generating trigger instructions and encapsulating prediction task packages at each time point exceeding the limit. The encapsulated content of the prediction task package can be formally represented by a set:

[0067]

[0068] in: This represents the encapsulated prediction task package. This represents the current comprehensive state description matrix. This indicates a preliminary health status trajectory. This represents the calculated multidimensional distance vector. This represents the weighted average distance scalar value. Optionally, when encapsulating prediction task packages, the health status simulator also adds a brief context description. This context description includes the current reservoir operating water level, seasonal information, and the most recent maintenance record. This information is not used as core inference data but can provide auxiliary background information for the risk prediction agent. The risk prediction agent's task receiving buffer is a circular queue with priority management. Newly received prediction task packages are assigned a corresponding priority based on their associated weighted average distance. Prediction task packages with larger weighted average distances have higher priority, thus ensuring that more urgent risks are analyzed first.

[0069] See Figure 4 This is a heatmap of the comprehensive state description matrix of a reservoir, used to visualize the dynamic changes in the multi-dimensional health status of a reservoir within a digital twin system. As can be seen from the heatmap, the color gradation of the "Overall Health Score" consistently matches the average color gradation of "Structural Stability, Hydrological Safety, and Environmental Adaptability," verifying its "comprehensive quantitative" attribute. It transforms the abstract "comprehensive state description matrix" into an intuitive risk evolution trajectory, demonstrating both the "high-fidelity monitoring capability" of the digital twin system and providing "precise, layered quantitative basis" for operation and maintenance decisions. It can help operation and maintenance personnel quickly locate the time points and core indicators of state deterioration, providing an intuitive basis for the triggering and scheduling of subsequent risk prediction agents.

[0070] In one embodiment of the present invention, after receiving a prediction task package, the risk prediction agent retrieves the package from the task receiving buffer, verifies its digital signature, unpacks it, and extracts the data. Based on the extracted data, the risk prediction agent initiates a simulation process involving multi-stage reasoning. This simulation process constructs multiple evolution paths of the current state and calls a path evaluation model based on a mixture of physical mechanisms and statistical laws to quantify the risk level of each evolution path. The quantified scores of all evolution paths are integrated to generate a dynamic risk spectrum, which describes the probability distribution and severity of various risk events within different future time windows. After generating the dynamic risk spectrum, the risk prediction agent encodes it into a machine-readable warning instruction sequence and transmits the warning instruction sequence and corresponding risk interpretation summary back to the digital twin via a security gateway. The instruction parsing module built into the digital twin receives and parses the warning instruction sequence, and drives the 3D virtual model of the reservoir to perform visual highlight rendering based on the parsing results, overlaying and displaying the risk level and trend on the corresponding dam location, reservoir area, or hydrological node.

[0071] In practice, the risk prediction agent obtains prediction task packets from the task receiving buffer, a shared storage area located in the risk prediction agent process's memory space. After verifying the digital signature of the prediction task packet, it unpacks and extracts the data. The verification operation uses a public key paired with the health state simulator to decrypt the digital signature and compares it with the hash value recalculated from the unpacked core data. If the comparison matches, the data is confirmed to be complete and of a trustworthy origin. Based on the extracted data, the risk prediction agent initiates a simulation process involving multi-stage inference. The multi-stage inference simulation process uses the current comprehensive state description matrix and the preliminary health state trajectory as initial state inputs.

[0072] The simulation process, incorporating multi-stage reasoning, constructs multiple evolution paths for the current state. This construction is achieved by perturbing key parameters in the initial state input and applying predefined physical evolution rules. For an example using a digital twin of the Jinping I Hydropower Station, key parameters include upstream inflow, reservoir ground motion parameters, and the attenuation coefficient of the dam's concrete elastic modulus. A path assessment model based on a hybrid of physical mechanisms and statistical laws is invoked to quantify the risk level of each evolution path. The physical mechanism part simulates the reservoir system's response under given parameters based on hydraulic and structural mechanics equations, while the statistical law part modifies the probability of specific consequences based on prior probabilities from a database of similar historical events. The quantified score is a scalar value that integrates the probability of an event occurring with the severity of its potential consequences.

[0073] A dynamic risk spectrum is generated by integrating the quantitative scores of all evolution paths. The integration operation categorizes and statistically analyzes the quantitative scores of all evolution paths according to risk type and time window. The dynamic risk spectrum describes the probability distribution and severity of various risk events within different future time windows. The dynamic risk spectrum can be mathematically expressed in the form of a three-dimensional surface or contour plot, with one dimension being time and the other being risk type, and the numerical axis representing the risk level. The generation of the dynamic risk spectrum can be formally represented as follows:

[0074]

[0075] in: Indicates within the time window and risk type Dynamic risk spectrum value on This represents the total number of evolution paths generated in the simulation. It is an indicator function, when the risk type With the Risk types dominated by the evolution path The value is 1 if the match is consistent, and 0 otherwise. Indicates the first Evolutionary path within the time window Quantitative scoring on the platform.

[0076] In some embodiments, after generating a dynamic risk spectrum, the risk prediction agent encodes the dynamic risk spectrum into a machine-readable sequence of warning instructions. The encoding process transforms areas in the dynamic risk spectrum that exceed a preset risk threshold into a series of structured instructions with time stamps, spatial location identifiers, risk level codes, and recommended action codes. The warning instruction sequence and its corresponding risk interpretation summary are transmitted back to the digital twin via a security gateway. The security gateway performs protocol conversion and access control, allowing only specific format data packets from authorized risk prediction agents to enter the digital twin's internal network.

[0077] The built-in instruction parsing module of the digital twin receives and parses the early warning instruction sequence, breaking it down into its constituent fields according to predefined syntax rules. Based on the parsing results, the module drives visual highlighting rendering of the reservoir's 3D virtual model. This highlighting rendering involves changing the color, transparency, or adding dynamic flashing effects to specific areas or components within the model. Risk levels and trends are overlaid on the corresponding dam locations, reservoir areas, or hydrological nodes, with the overlaid information appearing as semi-transparent text boxes or legends floating in the 3D scene. It's understood that the risk prediction agent, as an independent process, does not consume resources from the main digital twin process due to its complex calculations, thus ensuring the real-time performance of the digital twin's basic synchronization and simulation functions. Optionally, the early warning instruction sequence is also simultaneously sent to the reservoir management personnel's monitoring terminal, triggering audible and visual alarms. In some embodiments, visual highlighting rendering uses color gradient mapping to indicate risk levels, for example, from green and yellow to red to represent risk from low to high.

[0078] See Figure 5 This is a risk threshold comparison chart of multimodal monitoring signals in a reservoir digital twin system. Its core function is to quantify whether the current state of a single signal dimension exceeds the limit. This chart corresponds to the core output of the "analysis unit preprocessing" in the digital twin system. It converts raw signals of different dimensions such as "water quality, stress, seepage, and deformation" into standardized values ​​of 0-100 through a "normalization template," achieving "comparable quantification across signal dimensions." Subsequently, the "collaborative arbitration module" will integrate these standardized values ​​into a "comprehensive state description matrix." This chart is a key intermediate link in the system's transition from "single signal monitoring" to "global state assessment."

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic prediction of reservoir health status based on digital twins, characterized in that, The method includes: A digital twin synchronized with the physical reservoir is activated. Within the digital twin, an independent analysis unit with state awareness and memory functions is configured for each type of signal in the monitoring signal stream. The analysis unit preprocesses the input signal to generate a standardized signal vector with a unified timestamp. A collaborative arbitration module is preset in the digital twin. The collaborative arbitration module collects the standardized signal vectors generated by all analysis units at a set period, and integrates the vectors according to the preset fusion rules to form a comprehensive state description matrix. The comprehensive state description matrix is ​​submitted to the health state simulator embedded in the digital twin. The health state simulator calls the offline pre-trained implicit association network to perform state inference on the comprehensive state description matrix and generate a preliminary health state trajectory. The health status simulator compares the generated preliminary health status trajectory with the normal status pattern in the dynamically updated historical experience knowledge base. If the comparison result exceeds the preset deviation tolerance threshold, the health status simulator activates the risk prediction agent located outside the digital twin and encapsulates the current comprehensive status description matrix and comparison difference features into a prediction task package, which is then transmitted to the risk prediction agent.

2. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 1, characterized in that, The digital twin that is activated and synchronized with the physical reservoir includes: The digital twin continuously receives and verifies the multimodal monitoring signal stream reported by the sensor array; A high-fidelity 3D virtual model of the reservoir is initialized, and the reservoir structural parameters, historical operation data, and environmental basic data are imported into the 3D virtual model to complete the baseline construction of the digital twin. A real-time data channel is established in the digital twin, and the real-time data channel is securely connected to the distributed sensor array of the physical reservoir to receive the multimodal monitoring signal stream in a streaming processing mode. The multimodal monitoring signal stream includes water quality parameter sequences, dam stress and strain readings, seepage field data point sets, reservoir bank deformation monitoring, dam appearance quality monitoring, reservoir water level changes, and ecological flow data monitoring.

3. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 2, characterized in that, The analysis unit preprocesses the input signal by: each analysis unit loading a dedicated signal cleaning rule according to its corresponding signal type, the signal cleaning rule being used to remove outliers in the monitored signal stream and compensate for missing data segments; after signal cleaning, applying a normalization template corresponding to the signal type to convert the original signal values ​​of different dimensions and magnitudes into a preset numerical range to generate normalized signal segments; stamping each normalized signal segment with a high-precision timestamp and arranging them in chronological order to form the standardized signal vector, the standardized signal vector being temporarily stored in a dedicated buffer associated with each analysis unit, awaiting invocation by the collaborative arbitration module.

4. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 3, characterized in that, The collaborative arbitration module simultaneously collects standardized signal vectors generated by all analysis units at a set period and integrates the vectors according to a preset fusion rule. This includes: at the end of each analysis period, the collaborative arbitration module sends a vector retrieval instruction to the dedicated buffer of all analysis units to read each standardized signal vector in parallel in a non-blocking manner; after reading, it verifies the timestamp alignment and data integrity of each standardized signal vector. For vectors with misaligned timestamps, it uses an interpolation algorithm for time synchronization; it concatenates all standardized signal vectors that have completed synchronization processing according to a predefined vector stacking order to form an initial fusion tensor; it applies a configurable weighted fusion kernel to the initial fusion tensor, and the weighted fusion kernel dynamically adjusts the weight coefficients according to the data confidence of each signal type in the current period, and outputs the comprehensive state description matrix after weighted calculation.

5. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 1, characterized in that, The step of the health state simulator calling an offline pre-trained implicit association network to perform state inference on the comprehensive state description matrix includes: after receiving the comprehensive state description matrix, the health state simulator loads the structural parameters of the offline pre-trained implicit association network from its internal storage. The implicit association network is a deep neural network trained on a large amount of historical normal and abnormal state data. The comprehensive state description matrix is ​​used as input and fed into the implicit association network. The network performs nonlinear transformation and feature extraction layer by layer, and finally generates a vector sequence representing the reservoir health status at multiple consecutive time points in the future at the output layer. The vector sequence is the preliminary health state trajectory, and each vector element in the trajectory corresponds to a comprehensive health score at a predicted time point.

6. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 5, characterized in that, The health status simulator compares the generated preliminary health status trajectory with normal status patterns in a dynamically updated historical experience knowledge base. This comparison includes: the health status simulator accessing an external database, which serves as the dynamically updated historical experience knowledge base and stores various typical health status patterns extracted from historical normal operation data. Each pattern is represented as a feature vector with an occurrence frequency and seasonality label. Key features are extracted from the preliminary health status trajectory, and the multidimensional distance between the trajectory and the feature vectors of all normal status patterns in the dynamically updated historical experience knowledge base is calculated. Several normal status patterns with the smallest multidimensional distance are selected as a reference set, and a weighted average distance is calculated based on the occurrence frequency of the reference set. This weighted average distance is used as the input comparison value for the deviation tolerance threshold.

7. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 6, characterized in that, The process of activating a risk prediction agent located outside the digital twin by the health status simulator and encapsulating the current comprehensive state description matrix and contrast difference features into a prediction task package and transmitting it to the risk prediction agent includes: when the weighted average distance exceeds the deviation tolerance threshold, the health status simulator generates a trigger instruction, which is sent to the dormant risk prediction agent through an encrypted message queue to wake up the risk prediction agent; simultaneously, the health status simulator packages the current comprehensive state description matrix, the preliminary health status trajectory, the calculated multidimensional distance, and the weighted average distance together and attaches a digital signature to form the prediction task package; and pushes the prediction task package to the task receiving buffer of the awakened risk prediction agent through a protected asynchronous communication link.

8. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 7, characterized in that, The processing of the risk prediction agent after receiving the prediction task package includes: the risk prediction agent obtains the prediction task package from the task receiving buffer, verifies its digital signature, unpacks it, and extracts the data; based on the extracted data, the risk prediction agent initiates a simulation process containing multi-stage reasoning, the simulation process constructs multiple possible evolution paths of the current state, calls a path evaluation model based on a mixture of physical mechanisms and statistical laws, and quantifies the risk level of each possible evolution path; integrates the quantified scores of all possible evolution paths to generate a dynamic risk spectrum, the dynamic risk spectrum describing the probability distribution and severity of various risk events within different future time windows.

9. The method for dynamic prediction of reservoir health status based on digital twins as described in claim 8, characterized in that, After generating the dynamic risk spectrum, the process further includes: the risk prediction agent encoding the dynamic risk spectrum into a machine-readable warning instruction sequence, and transmitting the warning instruction sequence and the corresponding risk interpretation summary back to the digital twin through a security gateway; the instruction parsing module built into the digital twin receives and parses the warning instruction sequence, and drives the reservoir 3D virtual model to perform visual highlight rendering based on the parsing results, overlaying and displaying the risk level and trend on the corresponding dam location, reservoir area or hydrological node.

10. The method for dynamic prediction of reservoir health status based on digital twins according to claim 6, characterized in that, The step of extracting key features from the preliminary health status trajectory and calculating the multidimensional distance between the feature vectors of all normal state patterns in the dynamically updated historical experience knowledge base includes: The health status simulator performs dual feature analysis in the time and frequency domains on the preliminary health status trajectory, and identifies inflection points in the trajectory where the rate of change exceeds a preset threshold. Centered on each inflection point, extract adjacent trajectory segments before and after it, and calculate the mean, variance, skewness, and kurtosis of each trajectory segment to form a local feature vector; The initial health status trajectory is divided into multiple equal-length sub-intervals, and the linear fitting slope and periodic intensity of each sub-interval are calculated to form a global feature vector. The key features are obtained by combining the local feature vectors with the global feature vectors. Calculate the absolute difference between the key feature and each normal state pattern feature vector in the dynamically updated historical experience knowledge base in each dimension; The absolute differences in each dimension are multiplied by preset dimension weight coefficients and then summed to obtain the multidimensional distance between the feature vector and each normal state pattern.

Images