A system and method for assessing the health index of a power distribution network
By using a spatiotemporal graph attention network based on topology-based data quality governance and health decay memory mechanism, combined with a variational autoencoder with physical constraints, the problems of insufficient data quality verification and temporal consistency in existing distribution network health assessments are solved, achieving accurate health assessments and remaining lifetime predictions.
Patent Information
- Application Number
- CN202611193229.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-07
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for assessing the health of power distribution networks rely on the accuracy of raw measurement data and lack a mechanism for verifying data quality. They are easily affected by sensor accuracy drift and communication interference, leading to distorted health index assessment results. Furthermore, they lack the memory and modeling of long-term physical degradation trajectories of equipment, resulting in assessment results that lack temporal consistency and physical interpretability.
A data quality governance mechanism based on electrical topology is adopted. Data quality is evaluated by calculating topological constraint residuals and graph attention networks. Health assessment is performed by combining a spatiotemporal graph attention network with a health decay memory mechanism. Remaining lifetime is predicted by using a variational autoencoder with physical constraints, and a closed-loop optimization architecture is constructed.
It significantly improves the accuracy and robustness of health assessment, effectively distinguishes between real fault characteristics and measurement errors, dynamically and accurately assesses the health status of equipment, and achieves physically interpretable remaining life prediction and adaptive optimization.
Smart Images

Figure CN123117767A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system condition monitoring technology, and in particular to a distribution network health index assessment system and method. Background Technology
[0002] As a critical link in the power system directly facing users, the health assessment of the distribution network is crucial for ensuring power supply reliability. Existing methods for assessing the health status of distribution networks primarily rely on the collection and analysis of electrical operating data such as voltage and current, determining the health level of equipment by setting thresholds or constructing statistical models. With the development of smart grids, some solutions are beginning to incorporate deep learning technologies such as graph neural networks, attempting to leverage the topological information of the distribution network to improve the accuracy of the assessment.
[0003] However, existing technologies still have significant shortcomings in practical applications. Current assessment methods often rely excessively on the accuracy of raw measurement data, lacking mechanisms to verify the quality of the data itself. Under actual operating conditions, affected by factors such as sensor accuracy drift, communication interference, or packet loss, the collected data often contains noise or even errors. Directly inputting such "dirty data" into the assessment model can easily lead to distorted health index assessment results and misjudgments. Existing data cleaning or anomaly detection techniques are mostly based on statistical principles, focusing only on the rationality of data distribution while ignoring the strict electrical and physical constraints of the distribution network itself (such as Kirchhoff's laws). This makes it difficult for existing technologies to effectively distinguish between "real power grid fault characteristics" and "false measurement errors," easily misjudging physical anomalies as data noise and smoothing them out, or vice versa. Furthermore, most existing health assessment models are based only on static mapping of data at the current moment or within a short time window, lacking the memory and modeling of the long-term physical degradation trajectory of equipment. This makes it difficult to accurately reflect the dynamic evolution of equipment aging over time, resulting in assessment results lacking temporal consistency and physical interpretability. Summary of the Invention
[0004] To overcome the shortcomings of the prior art, the present invention discloses a power distribution network health index assessment system and method.
[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:
[0006] A method for assessing the health index of a power distribution network includes the following steps:
[0007] S1. Data Acquisition and Preprocessing Steps: Collect multi-source heterogeneous data from each monitoring node in the power distribution network, including electrical operation data, environmental stress data, and equipment ledger data; perform spatiotemporal alignment and normalization processing on the multi-source heterogeneous data.
[0008] S2. Data Quality Governance Steps: Based on the electrical topology of the distribution network, calculate the topology constraint residuals of each monitoring node; input the topology constraint residuals and the original electrical measurement data into a graph attention network to calculate the data quality score of each monitoring node; calculate the feeder-level topology constraint data quality index (TDQI) based on the data quality score and the topology constraint residuals; and execute a tiered data repair strategy based on the numerical range of the TDQI.
[0009] S3. Health Index Assessment Steps: Model the power distribution network as a spatiotemporal graph, with power distribution equipment as nodes and electrical connection relationships as edges; use the data after data quality governance as node state features, input it into the spatiotemporal graph attention network with a health decay memory mechanism, and calculate the health index HI of each device; obtain the feeder-level and regional-level health indices through weighted aggregation of device criticality.
[0010] S4. Remaining lifetime prediction step: Input the historical health index sequence and current state characteristics of each device into the physical constraint variational autoencoder PC-CVAE, use the Weibull degradation model as a prior constraint, and output the probability distribution of the future health index; calculate the mean and confidence interval of the remaining healthy lifetime RUL based on the probability distribution.
[0011] S5. Predictive Maintenance Decision-Making Steps: Calculate the Maintenance Urgency Index (MUI) by combining the current health index, remaining health life prediction results, and equipment criticality; generate tiered maintenance work orders based on the MUI, and feed back the maintenance execution results to the data quality governance step and the health index assessment step to form a closed-loop optimization.
[0012] Preferably, in step S2, the method for calculating the topological constraint residual is as follows:
[0013] For any monitoring node in the distribution network Calculate its topological constraint residual vector ,include:
[0014] Kirchhoff's Current Law Residual ,in For nodes The set of electrically adjacent nodes, For a moment From adjacent nodes Inflow node The measured value of the branch current. For a moment node The measured value of the load current;
[0015] Kirchhoff's voltage law residual ,in For containing nodes A set of independent loops For the first in the loop Measured voltage drop of the branch circuit.
[0016] Preferably, in step S2, the calculation method for the Topology Constraint Data Quality Index (TDQI) is as follows:
[0017]
[0018] in, For feeder The collection of all monitoring nodes included. The total number of monitoring nodes, For nodes computed through a graph attention network Data quality score, For nodes In the time window The average topological constraint residual norm within the range, This is the residual attenuation coefficient;
[0019] The hierarchical data repair strategy includes: directly using the original data when the TDQI is within a first preset range; performing local repair based on graph attention interpolation when the TDQI is within a second preset range; performing iterative correction based on physical constraints when the TDQI is within a third preset range; and marking the data as unreliable and triggering sensor self-test when the TDQI is within a fourth preset range.
[0020] Preferably, in step S3, the health decline memory mechanism includes:
[0021] For each device node Maintain a healthy decaying memory state Its update rules are as follows:
[0022]
[0023] in, The decay gate vector is obtained by performing a nonlinear transformation on the concatenation of the current state features and historical memory using the Sigmoid function, and is used to control the proportion of historical degradation information retained. The candidate memory vector is obtained by performing a nonlinear transformation on the concatenation of the current state features and historical memories using the Tanh function, and is used to encode the new degradation information at the current moment; For Hadamard's element-wise product;
[0024] The physical meaning of the attenuation gating vector is as follows: when the equipment experiences severe overload or harsh environmental stress, the gating value approaches zero, and the historical memory is updated rapidly; when the equipment operates smoothly, the gating value approaches one, and the historical memory is retained to reflect the slow degradation trend.
[0025] Preferably, in step S3, the spatiotemporal graph attention network further includes a multi-scale temporal aggregation step:
[0026] For the spatial convolution output sequence of each device node, multi-scale temporal features are extracted using temporal convolutional networks with different kernel sizes, respectively. The different kernel sizes correspond to short-term fluctuation scale, medium-term trend scale, and long-term degradation scale. The temporal features at each scale are weighted and summed using learnable weights to obtain multi-scale temporal aggregated features. Based on the multi-scale temporal aggregated features, the health index is normalized to a preset range using the Sigmoid function.
[0027] Preferably, in step S4, the physical constraint variational autoencoder (PC-CVAE) includes:
[0028] The encoder, employing a bidirectional LSTM structure, encodes the historical health index sequence and current state features into the mean vector and standard deviation vector of the posterior distribution.
[0029] Physically constrained prior modules, using characteristic lifetime parameters of the Weibull degradation model. and degenerate shape parameters Given the conditional input, generate the mean vector and standard deviation vector of the prior distribution, replacing the standard normal prior;
[0030] The decoder, using a TransformerDecoder structure, generates a health index prediction sequence for future preset time steps from latent variables and current state features;
[0031] During training, a total loss function is used, which includes reconstruction loss, KL divergence loss and physical consistency loss. The physical consistency loss only takes effect under the condition of no maintenance intervention, and penalizes the recovery prediction of the health index under the condition of no maintenance intervention, so as to ensure that the prediction results meet the monotonic physical constraint of equipment degradation.
[0032] Preferably, in step S4, the method for calculating the remaining healthy lifespan (RUL) is as follows:
[0033] The remaining healthy lifespan is defined as the time when the health index first drops to the warning threshold; multiple future health index trajectories are obtained from the decoder output of the PC-CVAE through Monte Carlo sampling; the time step when each trajectory first touches the warning threshold is statistically analyzed to obtain the empirical probability distribution of RUL; the mean of RUL, the confidence interval composed of preset quantiles, and the failure probability within the preset time window are output.
[0034] Preferably, in step S5, the method for calculating the Maintenance Urgency Index (MUI) is as follows:
[0035]
[0036] in, For equipment Current health index For equipment The average remaining lifespan, As a reference time constant, The equipment criticality weight is obtained by normalizing the product of the power loss due to equipment failure and the average recovery time.
[0037] The closed-loop optimization includes: after maintenance is completed, re-entering the updated status data of the equipment into the data quality governance step to verify the data quality improvement effect, re-entering the health index assessment step to verify the health recovery effect, and encoding the maintenance event and inputting it into the health decay memory mechanism to learn the impact of maintenance intervention on the degradation trajectory.
[0038] A power distribution network health index assessment system, comprising:
[0039] The perception layer module is used to collect multi-source heterogeneous data from various monitoring nodes in the power distribution network, and to perform spatiotemporal alignment and normalization processing on the multi-source heterogeneous data.
[0040] The data quality governance module is used to calculate topology constraint residuals based on the electrical topology of the distribution network, calculate data quality scores and topology constraint data quality indices (TDQI) through graph attention networks, and execute hierarchical data repair strategies based on the TDQI.
[0041] The health index assessment module is used to model the distribution network as a spatiotemporal diagram and calculate the device-level, feeder-level, and regional-level health indices through a spatiotemporal diagram attention network with a health decay memory mechanism.
[0042] The remaining life expectancy prediction module is used to output the probability distribution of future health index through a physical constraint variational autoencoder PC-CVAE, and to calculate the mean and confidence interval of the remaining healthy life.
[0043] The predictive maintenance decision module is used to calculate the maintenance urgency index (MUI), generate tiered maintenance work orders, and feed the maintenance execution results back to the data quality governance module and the health index assessment module to form a closed-loop optimization.
[0044] Preferably, the system adopts an edge-cloud collaborative deployment architecture, specifically including:
[0045] The perception layer module is deployed on the field side of the distribution network and includes feeder terminal units (FTU), distribution terminal units (DTU), transformer terminal units (TTU), smart sensors, and meteorological monitoring devices. Each terminal device aggregates data through an edge computing gateway. The edge computing gateway has a built-in lightweight data quality pre-inspection unit, which is used to perform real-time calculation of topology constraint residuals before data is uploaded. When the topology constraint residuals are detected to exceed a preset threshold, data re-sampling or marking of abnormal timestamps is triggered on the edge side.
[0046] The data quality governance module is deployed between the edge computing gateway and the regional server, and adopts a two-level collaborative mechanism: the first level performs lightweight graph attention inference on the edge computing gateway to complete the real-time calculation of data quality scores and the immediate response of hierarchical repair strategies; the second level performs full-topology-wide graph attention network inference on the regional server to complete the global calculation of feeder-level TDQI and cross-feeder data consistency verification.
[0047] The health index assessment module and the remaining lifespan prediction module are deployed on the central cloud platform. The health index assessment module deploys the spatiotemporal graph attention network on the GPU cluster in a model parallel manner according to the feeder partition. The remaining lifespan prediction module adopts an asynchronous batch inference mode and performs batch forward inference on the PC-CVAE model of each device according to a preset cycle.
[0048] The predictive maintenance decision module interacts with the production management system (PMS) and the power outage management system (OMS) through a standardized interface. The closed-loop feedback is achieved through a message queue: after the maintenance work order is completed, the field terminal re-collects the updated status data of the equipment through the perception layer module and pushes it asynchronously to the data quality governance module and the health index assessment module through the message queue, triggering an incremental update of the health decay memory status.
[0049] By employing the technical solution described above, the present invention has the following beneficial effects:
[0050] (1) This invention significantly improves the accuracy and robustness of health index assessment by introducing a data quality governance mechanism based on electrical topology constraints. Existing technologies often directly use raw collected data for assessment, which is easily affected by sensor noise or communication errors. This invention uses Kirchhoff's current law and voltage law to calculate topology constraint residuals and combines them with graph attention networks to calculate data quality scores, which can effectively distinguish between real power grid fault characteristics and false measurement errors from the physical mechanism level. This data cleaning and repair strategy based on physical consistency ensures that the data input to the assessment model is true and reliable, fundamentally avoiding misjudgments of health status caused by "dirty data" and improving the system's anti-interference ability under complex operating conditions.
[0051] (2) This invention employs a spatiotemporal graph attention network with a health decay memory mechanism to achieve dynamic and accurate assessment of equipment health status. Unlike existing technologies that rely solely on static mapping of current-moment data, this invention models the distribution network as a spatiotemporal graph and introduces a memory mechanism to capture the long-term degradation trajectory of equipment. This not only fully utilizes the spatial topological correlation of the distribution network but also effectively characterizes the dynamic evolution of equipment performance over time. Furthermore, by combining the criticality of equipment with weighted aggregation, the health indices at the feeder and regional levels can objectively reflect the actual impact of critical equipment failures on power supply reliability, solving the problem of lack of temporal coherence and physical interpretability in assessment results in traditional methods.
[0052] (3) This invention achieves both physically interpretable remaining lifetime prediction and adaptive optimization through a variational autoencoder with physical constraints and a closed-loop feedback mechanism. This invention embeds the Weibull degradation model as a physical prior into the variational autoencoder, enabling the model to output a remaining lifetime probability distribution that conforms to physical laws even in data-sparse scenarios, overcoming the defect of pure data-driven models that easily produce non-physical predictions when extrapolating distributions. Simultaneously, the constructed "perception-governance-evaluation-prediction-decision" closed-loop architecture can feed back maintenance execution results to model training, allowing the system to autonomously learn the resetting or improvement effects of maintenance interventions on the degradation trajectory, thereby achieving continuous iteration of the evaluation model and continuous improvement of decision-making accuracy. Attached Figure Description
[0053] Figure 1 This is an overall flowchart of the power distribution network health index assessment method provided in the embodiments of the present invention;
[0054] Figure 2 A detailed flowchart of the data quality governance steps and health index assessment steps provided for embodiments of the present invention;
[0055] Figure 3 This is a schematic diagram of the edge-cloud collaborative deployment architecture and closed-loop feedback provided in an embodiment of the present invention. Detailed Implementation
[0056] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0057] Combined with appendix Figures 1-3The present invention discloses a distribution network health index assessment system and method in some embodiments, comprising the following steps: S1, data acquisition and preprocessing step: acquiring multi-source heterogeneous data from each monitoring node in the distribution network, the multi-source heterogeneous data including electrical operation data, environmental stress data, and equipment ledger data; performing spatiotemporal alignment and normalization processing on the multi-source heterogeneous data; S2, data quality governance step: calculating the topology constraint residuals of each monitoring node based on the electrical topology of the distribution network; inputting the topology constraint residuals and the original electrical measurement data into a graph attention network to calculate the data quality score of each monitoring node; calculating the feeder-level topology constraint data quality index (TDQI) based on the data quality score and the topology constraint residuals; executing a hierarchical data repair strategy based on the numerical range of the TDQI; S3, health index assessment step: modeling the distribution network as a spatiotemporal graph, with distribution equipment as nodes and electrical... The connection relationship is represented by edges; the data after data quality governance is used as the node state feature and input into a spatiotemporal graph attention network with a health decay memory mechanism to calculate the health index HI of each device; the feeder-level and regional-level health indices are obtained through weighted aggregation of device criticality; S4, Remaining Life Prediction Step: the historical health index sequence and current state features of each device are input into a physical constraint variational autoencoder PC-CVAE, with a Weibull degradation model as a prior constraint, to output the probability distribution of the future health index; the mean and confidence interval of the remaining healthy life RUL are calculated based on the probability distribution; S5, Predictive Maintenance Decision Step: the maintenance urgency index MUI is calculated by combining the current health index, the remaining healthy life prediction result, and the device criticality; a graded maintenance work order is generated according to the MUI, and the maintenance execution result is fed back to the data quality governance step and the health index evaluation step to form a closed-loop optimization.
[0058] Specifically, the power distribution network health index assessment method provided in this embodiment constructs a five-layer closed-loop architecture of "perception-governance-assessment-prediction-decision," with each step forming a complete closed-loop link through data flow and feedback flow. In step S1, the collection of multi-source heterogeneous data covers all nodes in the power distribution network that are equipped with monitoring terminals. Electrical operation data includes real-time measurements such as voltage amplitude, current amplitude, active power, reactive power, power factor, frequency deviation, current harmonic distortion rate, and transformer winding temperature rise of each node; environmental stress data includes ambient temperature, relative humidity, wind speed, precipitation, number of thunderstorm days, and salt spray concentration in coastal areas; equipment ledger data includes static or semi-static information such as equipment model, commissioning date, rated parameters, cumulative overload count, historical insulation resistance test value, most recent maintenance date, cumulative defect record count, and family defect markers. Spatiotemporal alignment refers to unifying data sources with different sampling frequencies and communication delays to a common time reference, and interpolating to fill in missing timestamps caused by packet loss. Normalization refers to dimensionless normalization of data in each dimension according to their physical dimensions, eliminating the interference of different dimensions on subsequent model training. Specifically, the min-max normalization method is used to linearly map the data in each dimension to... Interval.
[0059] In step S2, the core idea of data quality governance is to use the electrical and physical laws of the distribution network as hard constraints on data quality. Unlike schemes that rely solely on statistical distribution for anomaly detection, this embodiment utilizes Kirchhoff's current law and voltage law to perform physical consistency verification on measured data, distinguishing between "real power grid anomalies" and "false anomalies caused by sensor measurement errors or communication interference." The introduction of graph attention networks allows data quality assessment to move beyond viewing individual nodes in isolation. Instead, it aggregates information from electrically adjacent nodes through an attention mechanism, thereby determining at the topology level whether a node's data contradicts the measurements of its upstream and downstream nodes. The calculation of feeder-level TDQI aggregates the data quality information of a single node into a data reliability index for the entire feeder, providing data quality assurance for subsequent health index assessment. The tiered data repair strategy employs repair measures of varying strengths based on the TDQI value range, avoiding over-processing of high-quality data and the indiscriminate use of low-quality data.
[0060] In step S3, modeling the distribution network as a spatiotemporal graph is the key modeling method in this embodiment. Spatially, the nodes of the graph correspond to distribution equipment (including distribution transformers, pole-mounted switches, ring main units, cable line segments, overhead line segments, etc.), the edges of the graph correspond to the direct electrical connections between equipment, and the adjacency matrix reflects the radial or ring topology of the distribution network. Temporally, each node has a state feature vector at each time step, which is composed of electrical operation features, environmental stress features, aging features, and maintenance history features. The introduction of the health decay memory mechanism enables the spatiotemporal graph attention network to have the ability to remember the long-term degradation trajectory of equipment, rather than relying solely on information within the current moment or short-term time window for judgment. The weighted aggregation of equipment criticality enables the feeder-level and regional-level health indices to reflect the "actual impact of critical equipment failures on power supply reliability," rather than a simple arithmetic average.
[0061] In step S4, the physical constraint variational autoencoder fuses the data-driven deep learning model with the physics-driven Weibull degradation model. The Weibull degradation model provides a physical prior that the device health index monotonically decreases over time. This prior is injected into the model training process through the prior distribution of the conditional variational autoencoder, enabling the model to output predictions that conform to physical laws even in scenarios with sparse data or distribution extrapolation. The output is a probability distribution rather than a single point estimate, allowing maintenance personnel to obtain information about the uncertainty of remaining lifespan, thus considering risk factors in maintenance decisions.
[0062] In step S5, the Maintenance Urgency Index (MUI) integrates the current health status, future degradation trend, and equipment importance into a single decision indicator, making the maintenance priorities of different devices comparable. The closed-loop optimization is implemented as follows: after a maintenance work order is completed, the equipment undergoes repair or replacement, and its status data is re-collected at the perception layer. After data quality governance, the health index is recalculated, and the "maintenance event" information is encoded and input into the health decay memory mechanism. This allows the model to learn the causal relationship that "maintenance intervention can reset or improve the degradation trajectory," thereby reflecting the maintenance effect in subsequent lifespan predictions and avoiding overly pessimistic lifespan predictions for maintained equipment.
[0063] In some embodiments, in step S2, the method for calculating the topology constraint residual is as follows: for any monitoring node in the distribution network Calculate its topological constraint residual vector This includes: Kirchhoff's current law residuals ,in For nodes The set of electrically adjacent nodes, For a moment From adjacent nodes Inflow node The measured value of the branch current. For a moment node Measured load current; Kirchhoff voltage law residual ,in For containing nodes A set of independent loops For the first in the loop Measured voltage drop of the branch circuit.
[0064] Specifically, the physical meaning of the topological constraint residual vector is to measure the extent to which measured data violates fundamental circuit laws. For Kirchhoff's current law residuals, this embodiment adopts the unified notation convention of "positive for inflow nodes." Under this convention, Indicates time From adjacent nodes Inflow node The branch current, taking a positive value indicates that the actual direction is inflow into the node. A negative value indicates that the actual direction is an outflow node. According to Kirchhoff's current law, under ideal, error-free conditions, the current flowing into the node... The sum of all branch currents should equal the load current at that node, i.e. When there is a deviation between the two sides of this equation, the absolute value of the deviation is the KCL residual. The sources of this residual may include: a current transformer in a branch having a turns ratio error or saturation, electromagnetic interference affecting the current sampling channel on the load side, or unmonitored distributed source injection current at the node. For the Kirchhoff voltage law residual, its calculation logic is as follows: along the path containing the node... When any independent loop completes one revolution, the algebraic sum of the voltage drops in each branch should be zero. The voltage drops are taken as positive when the positive direction is the same as the loop's direction, and negative when it is opposite. When this algebraic sum is not zero, its absolute value is the KVL residual. The sources of this residual may include: the angle error of the voltage transformer, inaccurate identification of line parameters, or physical faults such as abnormally increased contact resistance in a certain section of the line.
[0065] It should be noted that in actual distribution networks, due to limited measurement accuracy, topology constraint residuals cannot be strictly zero. Therefore, this embodiment does not use whether the residual is zero as the criterion, but rather inputs the magnitude of the residual as a continuous value into the subsequent graph attention network, allowing the network to learn "what level of residual belongs to normal measurement noise and what level of residual indicates a real data quality problem." Furthermore, for radial distribution networks, the selection of independent loops is determined according to the mesh method; for ring distribution networks containing tie switches, independent loops are determined according to the ring topology when the tie switch is closed, and degenerate into a radial topology when the tie switch is open, adjusting the set of independent loops accordingly. The composition of.
[0066] In some embodiments, the method for calculating the Topology Constraint Data Quality Index (TDQI) in step S2 is as follows: ,in, For feeder The collection of all monitoring nodes included. The total number of monitoring nodes, For nodes computed through a graph attention network Data quality score, For nodes In the time window The average topological constraint residual norm within the range, The residual attenuation coefficient is used. The hierarchical data repair strategy includes: directly using the original data when the TDQI is in the first preset range; performing local repair based on graph attention interpolation when the TDQI is in the second preset range; performing iterative correction based on physical constraints when the TDQI is in the third preset range; and marking the data as unreliable and triggering sensor self-test when the TDQI is in the fourth preset range.
[0067] Specifically, data quality scoring The calculation is implemented through a graph attention network. This network uses a distribution network topology as its input graph structure, and the input feature vector of each node is composed of a topological constraint residual vector, the original electrical measurement vector, and an environmental feature vector. The graph attention network assigns different attention coefficients to each neighboring node through an attention mechanism. These coefficients reflect the degree to which neighboring nodes contribute to the current node's data quality judgment. For example, when a node... When the KCL residual is large, the network will assign higher attention weights to its directly electrically adjacent nodes to determine whether the residual originates from the node. The sensor malfunction itself stems from measurement biases propagating through the topology from neighboring nodes. Attention coefficients are calculated using the standard scaled dot product attention method, incorporating non-zero gradients in negative regions via the LeakyReLU activation function to prevent gradient vanishing during training. The output layer maps the aggregated features to [the desired input] using the Sigmoid activation function. The interval is used as a data quality score.
[0068] The training of each weight matrix in the graph attention network adopts the following strategy: all weight matrices (including , , , The Xavier uniform initialization method is used for initialization to ensure that the variance of the output of each layer remains consistent during the early stages of training, avoiding gradient explosion or gradient vanishing. The training optimizer uses the Adam optimizer, with an initial learning rate of [value missing]. The learning rate is decayed using a cosine annealing strategy. An early stopping mechanism is employed during training; training stops when the loss on the validation set no longer decreases for a preset number of consecutive epochs to prevent overfitting. The training data consists of manually labeled data quality data from historical datasets, categorized as "normal," "mildly abnormal," and "severely abnormal."
[0069] In the TDQI formula, the exponentially decaying term Its function is to impose a nonlinear penalty on the degree of topology constraint violation. When the average topology constraint residual of a node... As the index approaches zero, the exponential term approaches one, at which point the TDQI is primarily determined by the data quality score. The decision is made that as the average residual increases, the exponential term decays rapidly. Even if the graph attention network gives a high data quality score, the node's contribution to the overall TDQI of the feed will be significantly suppressed. (Residual decay coefficient) The sensitivity of the penalty is controlled, and its value is determined on a historical labeled dataset through cross-validation. Time window. The selection of a window needs to take into account both response speed and noise resistance. A window that is too short is easily affected by transient interference, while a window that is too long will delay the detection of persistent data quality problems.
[0070] The four-level classification of the tiered data repair strategy is based on: the first preset range corresponds to a TDQI value of... The range indicates that the overall data quality of the feeder is excellent, and the topology constraint residuals of each node are within the normal measurement noise range, requiring no repair operations. The raw data can be directly used for subsequent health index assessment; the second preset range corresponds to a TDQI value within... The interval indicates that some nodes have mild data quality issues. In this case, the topological relationships learned by the graph attention network are used to interpolate and repair missing or abnormal data of the problematic nodes using measurements from their electrically adjacent nodes. This interpolation process considers the weighting of topological distance and attention weights. The third preset range corresponds to a TDQI value in... The range indicates a serious data quality problem, and simple interpolation is insufficient to restore data reliability. In this case, a physical constraint iterative correction based on the Newton-Raphson method is used. This involves using the power flow equations of the distribution network as constraints and the measured values of normal nodes as boundary conditions to iteratively solve for the most likely true value of the problematic nodes. The fourth preset range corresponds to the TDQI value within... A certain interval indicates that the data quality problem is extremely serious and has exceeded the algorithm's repair capabilities. At this point, the data in that time period is marked as unreliable, and a self-test command is sent to the corresponding sensor terminal to check the sensor hardware status and communication link quality.
[0071] In some embodiments, in step S3, the health decay memory mechanism includes: for each device node Maintain a healthy decaying memory state Its update rules are as follows: ,in, The decay gate vector is obtained by performing a nonlinear transformation on the concatenation of the current state features and historical memory using the Sigmoid function, and is used to control the proportion of historical degradation information retained. The candidate memory vector is obtained by performing a nonlinear transformation on the concatenation of the current state features and historical memories using the Tanh function, and is used to encode the new degradation information at the current moment; The element-wise product of Hadamard is used; the physical meaning of the decay gating vector is: when the device experiences severe overload or harsh environmental stress, the gating value approaches zero, and the historical memory is updated quickly; when the device operates smoothly, the gating value approaches one, and the historical memory is preserved to reflect the slow degradation trend.
[0072] Specifically, the design motivation for the health decay memory mechanism stems from two fundamental physical characteristics of power distribution equipment health degradation: First, equipment degradation is a cumulative process; the current health state depends not only on the current operating conditions but also on the cumulative effect of all historical stresses. Second, the degradation rate is not constant but accelerates or decelerates with changes in operating conditions. While traditional recurrent neural networks possess memory capabilities, their gating mechanisms are general sequence modeling tools and are not designed specifically for the physical process of "health degradation." The decay gating vector in this embodiment... The output of the Sigmoid function has a range of 1. Physically, this corresponds to "the proportion of historically degraded information retained at the current moment." When When all components are close to one, it indicates that the equipment is operating stably at the current moment, without any new significant degradation, and historical memory is fully preserved; when When all components are close to zero, it indicates that the equipment has experienced an abnormal stress event (such as severe overload, lightning overvoltage, extreme high temperature, etc.) at the current moment, and the historical memory is selected as the candidate memory vector. Significant replacements are made to reflect abrupt changes in health status.
[0073] Candidate memory vector The range of values output by the Tanh function is: The decay gating vector and the candidate memory vector encode the direction and intensity of newly generated degradation information at the current moment. The symmetry of the Tanh function allows the candidate memory to encode both "health deterioration" (positive direction) and "health improvement" (negative direction, such as after maintenance), which aligns with the fluctuations in health status that may occur throughout the device's lifespan. Both the decay gating vector and the candidate memory vector are inputs to the current state features. Memory state of the previous moment The splicing of these elements allows gating decisions to consider both "what is happening now" and "what has been accumulated before," thereby enabling adaptive adjustment of the degradation rate.
[0074] This mechanism complements the subsequent multi-scale temporal convolution: the health decay memory mechanism is responsible for capturing the cumulative degradation effect over long periods (timescale from months to years), while the multi-scale temporal convolution is responsible for capturing fluctuations and trends over short and medium periods (timescale from hours to months). Together, they constitute a complete temporal modeling capability for health status.
[0075] In some embodiments, in step S3, the spatiotemporal graph attention network further includes a multi-scale temporal aggregation step: for the spatial convolution output sequence of each device node, multi-scale temporal features are extracted using temporal convolutional networks with different kernel sizes, where the different kernel sizes correspond to short-term fluctuation scale, medium-term trend scale, and long-term degradation scale, respectively; the temporal features at each scale are weighted and summed using learnable weights to obtain multi-scale temporal aggregated features; based on the multi-scale temporal aggregated features, the health index is normalized to a preset range using the Sigmoid function.
[0076] Specifically, the spatial convolution step is performed at each time step. Independent execution involves fusing the state features of each device node with those of its electrically adjacent nodes. Spatial convolution employs a graph attention mechanism to calculate the aggregation weights of adjacent nodes, giving higher aggregation weights to nodes that are electrically close to the current device and have a high correlation with its operating state. The output of spatial convolution is a spatially fused feature vector for each node at each time step, which contains both the device's own operating state information and the operating state information of its local topology region.
[0077] The multi-scale temporal aggregation step further processes the time series output by the spatial convolution. This embodiment sets three time scales: the first scale uses a temporal convolutional network with a kernel size of 3, corresponding to the short-term fluctuation scale, to capture the instantaneous impact of hourly load fluctuations and short-term overvoltage events on equipment health; the second scale uses a temporal convolutional network with a kernel size of 7, corresponding to the medium-term trend scale, to capture the medium-term cumulative impact of daily to weekly seasonal load changes and continuous high-temperature weather on equipment health; the third scale uses a temporal convolutional network with a kernel size of 30, corresponding to the long-term degradation scale, to capture the slow degradation processes such as monthly insulation aging and mechanical wear. The temporal features of the three scales are weighted and summed using learnable weights, which are automatically learned during model training, allowing the model to adaptively adjust the contribution ratio of each time scale according to the specific equipment type and operating environment. For example, for oil-immersed transformers with long service lives, the weight of the long-term degradation scale may be higher than that of the short-term fluctuation scale; while for frequently operated pole-mounted switches, the weight of the short-term fluctuation scale may be relatively higher.
[0078] The final output of the health index is normalized to the value using the Sigmoid function. The health index is divided into ranges, where 100 indicates that the equipment is in a brand-new operational state, and 0 indicates that the equipment has completely failed. This normalization method ensures that the health indices of different equipment types and voltage levels are comparable, facilitating cross-equipment and cross-feeder comparisons by maintenance personnel. The feeder-level health index is obtained through weighted aggregation of equipment criticality. The equipment criticality weight is determined by normalizing the product of the load loss power caused by equipment failure and the mean recovery time, giving greater weight to equipment that has a greater impact on power supply reliability in the feeder health index. The regional health index is further weighted and aggregated based on the feeder level according to the power supply area or number of users supplied by each feeder, providing a quantitative basis for the overall health management of the regional power grid.
[0079] In some embodiments, in step S4, the physical constraint variational autoencoder (PC-CVAE) includes: an encoder, employing a bidirectional LSTM structure, encoding the historical health index sequence and current state features into a mean vector and a standard deviation vector of a posterior distribution; and a physical constraint prior module, using the characteristic lifetime parameters of the Weibull degradation model. and degenerate shape parameters As conditional inputs, a mean vector and standard deviation vector of the prior distribution are generated to replace the standard normal prior. The decoder adopts a TransformerDecoder structure to generate a health index prediction sequence for future preset time steps from latent variables and current state features. During training, a total loss function including reconstruction loss, KL divergence loss and physical consistency loss is used. The physical consistency loss only takes effect under the condition of no maintenance intervention, penalizing the prediction of health index recovery, so as to ensure that the prediction results meet the monotonic physical constraint of equipment degradation.
[0080] Specifically, the encoder's bidirectional LSTM structure comprises two LSTM layers: a forward layer and a backward layer. The forward layer processes the historical health index sequence in forward chronological order, capturing the degradation trend from the past to the present. The backward layer processes the same sequence in reverse chronological order, capturing the degradation pattern from the present to the past. The final hidden states from both directions are concatenated and then mapped to the mean vector of the posterior distribution through a fully connected layer. and standard deviation vector Current state characteristics As an additional conditional input, it is fed into the encoder along with the historical sequence, so that the posterior distribution not only reflects the historical degradation trajectory, but also reflects the impact of the current operating conditions on the degradation trend.
[0081] The physical constraint prior module is the core innovation that distinguishes this embodiment from the standard conditional variational autoencoder. The prior distribution of a standard CVAE is usually taken as a standard normal distribution. This prior knowledge does not include any physical knowledge about device degradation. This embodiment uses two parameters of the Weibull degradation model—the characteristic lifetime parameter. and degenerate shape parameters —As a conditional input to the prior distribution. The reliability function of the Weibull degradation model is: This indicates that the device is at a certain time. The probability that the device remains in a healthy state. Correspondingly, the physical degradation trend of the device health index is modeled as follows: Characteristic lifetime parameters The health index of the equipment has dropped to its initial value. The time elapsed when the score drops to approximately 36.8 points reflects the overall lifespan of the equipment; degradation shape parameters Characterizing the trend of degradation rate change, when Greater than a temporary acceleration of degradation (wear-out failure), when Equal to a temporary uniform degradation (random fault), when Less than one moment of degradation and deceleration (early failure type).
[0082] Parameters of the Weibull degeneration model and The following method is used to obtain the data: For equipment types with sufficient historical fault data, the maximum likelihood estimation method is used to fit the historical fault time data. Specifically, a log-likelihood function of the Weibull distribution is constructed. ,in This represents the number of historical fault samples. For the first The lifetime value of each fault sample is determined by solving the log-likelihood function using Newton's iteration method. and For new equipment lacking historical fault data, typical parameter values from the reliability manual provided by the equipment manufacturer are used as initial estimates, and gradually corrected through online Bayesian updates after the equipment is put into operation.
[0083] The physical constraint prior module uses logarithmic transformation to... and The mapping to the mean and standard deviation in the latent space allows the shape of the prior distribution to correlate with the physical degradation characteristics of the equipment. Specifically, the prior mean vector is... ,in For the physical prior projection matrix, The physical prior bias vector is used. The logarithmic transformation maps strictly positive Weibull parameters to the entire real number domain, which is convenient for neural network processing.
[0084] The decoder employs a TransformerDecoder architecture, whose self-attention mechanism can model the dependencies between future time steps. For example, if the model predicts the... The health index of the first step will drop sharply, then the second step Step and the first The predictions for each step are adjusted accordingly, rather than being given independently. The decoder's input includes latent variables sampled from the posterior distribution and current state features, and its output is the future... Health index prediction values at each time step.
[0085] In the total loss function, the reconstruction loss measures the mean squared error between the predicted sequence and the true sequence; the KL divergence loss measures the difference between the posterior distribution and the physical prior distribution, its role is to constrain the posterior distribution from deviating too far from the physical prior, preventing the model from overfitting within the training data distribution and failing when extrapolating the distribution; the physical consistency loss, by penalizing the positive increment of the health index in adjacent time steps (i.e., the health index recovering), forces the model to output a monotonically non-increasing predicted sequence under the assumption of no maintenance intervention. It is important to note that the physical consistency loss only takes effect when the training samples do not contain maintenance events. Specifically, in the training data, each sample is labeled with whether a maintenance event occurred within its time window; for samples without maintenance events, the physical consistency loss is calculated normally; for samples containing maintenance events, the weight of the physical consistency loss is reset to zero, allowing the model to learn the pattern of health index recovery after maintenance. This design resolves the logical contradiction between the physical monotonicity constraint and health recovery after maintenance, enabling the model to both ensure that the degradation prediction under no-intervention conditions conforms to physical laws and correctly reflect the improvement effect of maintenance intervention on the health status. The three losses are balanced by their respective weighting coefficients, with the weights of the KL divergence loss and the physical consistency loss being determined as hyperparameters through validation set performance.
[0086] In some embodiments, in step S4, the remaining health life (RUL) is calculated as follows: the remaining health life is defined as the time when the health index first drops to the warning threshold; multiple future health index trajectories are obtained from the decoder output of the PC-CVAE through Monte Carlo sampling; the time steps when each trajectory first touches the warning threshold are statistically analyzed to obtain the empirical probability distribution of RUL; the mean of RUL, the confidence interval composed of preset quantiles, and the failure probability within a preset time window are output.
[0087] Specifically, warning threshold The warning threshold is determined based on equipment type and maintenance standards. For example, for oil-immersed distribution transformers, the warning threshold can be set when the health index drops to 60 points, indicating that the equipment has entered a state of alert and requires enhanced monitoring; for pole-mounted switches, the warning threshold can be set when the health index drops to 50 points, indicating that the equipment needs to be scheduled for maintenance. The Monte Carlo sampling process is as follows: from the posterior distribution... Independent sampling Next (in this embodiment) (Taking 1000), each sampling yields a latent variable vector, which is then fed into the decoder to generate a future health index trajectory. Due to the randomness of the latent variables, The trajectories form a family of probability distributions for the future health index. For each trajectory, the time step at which its health index first drops below the warning threshold is recorded, which is the RUL realization value corresponding to that trajectory. The realized values of RUL constitute an empirical probability distribution of RUL, from which the mean can be extracted. As a central estimate of remaining lifespan, the 5th percentile is extracted. and the 95th percentile This constitutes a 90% confidence interval, reflecting the range of uncertainty in the forecast. Furthermore, statistics... In the future of the trajectory The proportion of cases that trigger the warning threshold within a time window is the failure probability within that time window. This probability can be directly used for risk assessment and actuarial science.
[0088] Probabilistic outputs offer significant decision-making advantages over traditional point estimations: a narrow confidence interval for the RUL indicates low prediction uncertainty, allowing maintenance personnel to formulate deterministic maintenance plans; a wide confidence interval indicates high prediction uncertainty, requiring maintenance personnel to allocate a larger safety margin or increase monitoring frequency. This explicit expression of uncertainty information is the core advantage of physically constrained variational autoencoders compared to deterministic prediction models.
[0089] In some embodiments, the method for calculating the Maintenance Urgency Index (MUI) in step S5 is as follows: ,in, For equipment Current health index For equipment The average remaining lifespan, As a reference time constant, The equipment criticality weight is obtained by normalizing the product of the power loss due to equipment failure and the average recovery time. The closed-loop optimization includes: after maintenance is completed, re-inputting the updated equipment status data into the data quality governance step to verify the data quality improvement effect, re-inputting the health index assessment step to verify the health recovery effect, and encoding the maintenance event and inputting it into the health decay memory mechanism to learn the impact of maintenance intervention on the degradation trajectory.
[0090] Specifically, the Maintenance Urgency Index (MUI) is the product of three factors, each corresponding to a dimension of maintenance decision-making. The first factor... The current level of health deterioration, with a value range of [value range missing]. When the equipment health index is 100 (brand new), this factor is zero, indicating no maintenance required; when the health index drops to zero (complete failure), this factor is one, indicating the maximum maintenance requirement. The second factor... The remaining life urgency level has a range of values. When the remaining lifetime is significantly greater than the reference time constant, this factor approaches zero, indicating that time is not urgent; when the remaining lifetime is significantly less than the reference time constant, this factor approaches one, indicating that maintenance needs to be arranged immediately. Reference time constant The selection of [specific timeframe] is related to the time granularity of the operations and maintenance plan. If the operations and maintenance plan is prepared on a monthly basis, then... Use 30 days; if the cycle is quarterly, use 90 days. The third factor. As the criticality weight of the equipment, the power loss due to equipment failure. With average recovery time The product of the three factors, normalized, reflects the degree of impact of equipment failure on power supply reliability. The product form of the three factors means that the MUI will only reach a high value when the equipment is in poor current health, has a short remaining lifespan, and the equipment itself has a significant impact on power supply reliability, thereby triggering a high-priority maintenance work order.
[0091] The specific implementation process of closed-loop optimization is as follows: After the maintenance work order is issued, maintenance personnel go to the site to perform repair or replacement operations. After the operation is completed, the field terminal (FTU / DTU / TTU) re-collects the updated electrical operation data of the equipment. This data is uploaded by the perception layer module and first undergoes a data quality governance step to calculate the TDQI value after maintenance. This value is then compared with the TDQI value before maintenance to verify whether the maintenance operation has improved data quality (for example, after replacing a faulty sensor, the topology constraint residual should be significantly reduced). Subsequently, the updated data enters the health index assessment step to recalculate the equipment's health index and verify whether the maintenance operation has restored the health index to the expected level. Finally, the "maintenance event" (including maintenance type, maintenance time, replacement part information, etc.) is encoded into an event vector and input into the health decay memory mechanism. During the memory update process, the maintenance event, as part of the current state characteristics, participates in the decay gating calculation, causing the gating value to approach zero. This significantly replaces the historical degradation memory with the new state after maintenance, enabling the model to correctly reflect the reset effect of the maintenance on the degradation trajectory in subsequent lifetime predictions. This closed-loop mechanism enables the entire evaluation system to continuously learn, and as maintenance data accumulates, the model's modeling of maintenance effectiveness will become increasingly accurate.
[0092] In some embodiments, a distribution network health index assessment system includes: a perception layer module, used to collect multi-source heterogeneous data from various monitoring nodes in the distribution network, and perform spatiotemporal alignment and normalization processing on the multi-source heterogeneous data; a data quality governance module, used to calculate topology constraint residuals based on the electrical topology of the distribution network, calculate data quality scores and topology constraint data quality indices (TDQI) through a graph attention network, and execute hierarchical data repair strategies based on the TDQI; a health index assessment module, used to model the distribution network as a spatiotemporal graph, and calculate device-level, feeder-level, and regional-level health indices through a spatiotemporal graph attention network with a health decay memory mechanism; a remaining lifetime prediction module, used to output the probability distribution of future health indices through a physical constraint variational autoencoder (PC-CVAE), and calculate the mean and confidence interval of the remaining health lifetime; and a predictive maintenance decision module, used to calculate the maintenance urgency index (MUI), generate hierarchical maintenance work orders, and feed back the maintenance execution results to the data quality governance module and the health index assessment module to form a closed-loop optimization.
[0093] Specifically, the distribution network health index assessment system provided in this embodiment consists of five functional modules at the hardware level. These modules interact via a data bus or network interface, forming a complete processing link from data acquisition to maintenance decision-making. The perception layer module, located at the front end of the system, is responsible for data interaction with field equipment in the distribution network. Its input is the raw measurement signals from various sensors and monitoring terminals, and its output is structured data that has undergone spatiotemporal alignment and normalization. The data quality governance module receives the output data from the perception layer module, performs physical consistency verification and quality assessment, and outputs high-quality data that has been repaired or labeled. The health index assessment module receives the output data from the data quality governance module, performs forward inference using a spatiotemporal graph attention network, and outputs health indices at each level (equipment level, feeder level, and area level). The remaining lifetime prediction module receives the historical health index sequence output by the health index assessment module, performs inference using a PC-CVAE model, and outputs the probability distribution of remaining lifetime. The predictive maintenance decision module receives the health index and remaining lifetime information, calculates the MUI (Mean Income Required) and generates maintenance work orders, while simultaneously feeding back the maintenance execution results to the preceding modules.
[0094] The data flow between the five modules is a superposition of unidirectional forward flow and reverse feedback flow. The forward flow starts from the perception layer module, passes sequentially through the data quality governance module, the health index assessment module, and the remaining lifespan prediction module, and finally reaches the predictive maintenance decision module. The reverse feedback flow starts from the predictive maintenance decision module, sending the maintenance execution results back to the data quality governance module (for verifying data quality improvement) and the health index assessment module (for verifying health recovery effects and updating memory status). This closed-loop structure makes the system no longer an open-loop "monitoring-alarm" mode, but rather possesses a complete closed-loop capability of "monitoring-assessment-prediction-decision-verification-learning".
[0095] In some embodiments, the system adopts an edge-cloud collaborative deployment architecture, specifically including: the perception layer module is deployed on the distribution network field side, including feeder terminals (FTU), distribution terminals (DTU), transformer terminals (TTU), intelligent sensors, and meteorological monitoring devices. Each terminal device aggregates data through an edge computing gateway. The edge computing gateway has a built-in lightweight data quality pre-inspection unit, which is used to perform real-time calculation of topology constraint residuals before data upload. When the topology constraint residuals are detected to exceed a preset threshold, data re-sampling or marking of abnormal timestamps is triggered on the edge side; the data quality governance module is deployed between the edge computing gateway and the regional server, adopting a two-level collaborative mechanism: the first level performs lightweight graph attention inference on the edge computing gateway to complete the real-time calculation of data quality scores and the immediate response of graded repair strategies; the second level performs full-topology-wide graph attention on the regional server. The system employs a force network inference mechanism to complete global calculations of feeder-level TDQI and cross-feeder data consistency verification. The health index assessment module and remaining lifetime prediction module are deployed on a central cloud platform. The health index assessment module uses a model-parallel approach, deploying the spatiotemporal graph attention network across feeder partitions on a GPU cluster. The remaining lifetime prediction module uses an asynchronous batch inference mode, performing batch forward inference on the PC-CVAE models of each device according to a preset cycle. The predictive maintenance decision module interacts with the Production Management System (PMS) and the Outage Management System (OMS) through standardized interfaces. The closed-loop feedback is achieved through a message queue: after a maintenance work order is completed, the field terminal re-collects the updated equipment status data via the perception layer module and asynchronously pushes it to the data quality governance module and the health index assessment module through the message queue, triggering incremental updates to the health decay memory state.
[0096] Specifically, the edge-cloud collaborative deployment architecture is designed following the principle of "deploying tasks with high real-time requirements to the edge and uploading tasks with high computational complexity to the cloud." The perception layer module is deployed on the field side of the distribution network. Its hardware components include: Feeder Terminal Units (FTUs) installed at 10kV feeder switches, responsible for collecting electrical quantities such as voltage, current, and power of the feeder, as well as the switch position status; Distribution Terminal Units (DTUs) installed in ring main units or substations, responsible for collecting electrical quantities and switch status of multiple feeders; Transformer Terminal Units (TTUs) installed on the low-voltage side of the distribution transformer, responsible for collecting parameters such as transformer voltage, current, temperature, and oil level; intelligent sensors, including temperature sensors, partial discharge sensors, and vibration sensors installed on the equipment surface, used to collect equipment status quantities; and meteorological monitoring devices installed on substations or line towers, responsible for collecting meteorological data such as ambient temperature, humidity, wind speed, and precipitation. Each terminal device uploads data to the edge computing gateway via fiber optic Ethernet or 4G / 5G wireless communication.
[0097] The lightweight data quality pre-inspection unit built into the edge computing gateway uses knowledge distillation technology to distill a lightweight student network from the complete graph attention network (i.e., the teacher network). During distillation, the output probability distribution of the teacher network on training samples is used as the soft label, and the KL divergence between the student network's output probability distribution and the soft label is used as the distillation loss. Simultaneously, the cross-entropy loss between the student network's output and the true label is combined as the hard label loss; the total distillation loss is the weighted sum of these two. Through this distillation process, the student network retains as much of the teacher network's data quality assessment capabilities as possible while significantly reducing the number of parameters and computational load, enabling it to achieve millisecond-level topology constraint residual calculation and data quality scoring inference on the gateway's limited computing resources. When the KCL or KVL residual of a node exceeds a preset threshold, the pre-inspection unit immediately triggers one of two responses: if it is determined to be transient communication interference (the residual is a single spike), the timestamp is marked as abnormal and a resampling request is sent to the terminal; if it is determined to be a persistent deviation (the residual exceeds the limit for multiple consecutive sampling periods), the node is marked as a suspected sensor fault and reported to the second level of the data quality governance module for further diagnosis.
[0098] In the two-tiered collaborative mechanism of the data quality governance module, the first tier is deployed at the edge computing gateway, performing lightweight graph attention inference. Its function is to conduct real-time assessment and immediate repair of data quality within a single feeder, with a response time controlled within seconds, meeting the timeliness requirements of real-time monitoring of the distribution network. The second tier is deployed at the regional server, performing full-parameter graph attention network inference. Its function is to conduct global data quality assessment across the entire topology of the power supply area, including cross-feeder data consistency verification (e.g., for two feeders connected by a tie switch, the data at their common node should be consistent when the tie switch is closed), and global calculation of feeder-level TDQI. The collaboration between the two tiers is as follows: the assessment results of the first tier serve as the input prior for the second tier, and the global assessment results of the second tier are fed back to the first tier to correct its local assessment biases.
[0099] The health index assessment module is deployed on a GPU cluster on the central cloud platform. Since the computational complexity of the spatiotemporal graph attention network is superlinearly related to the number of nodes and edges in the graph, a single GPU cannot complete full graph inference within an acceptable time for a regional power distribution network containing a large number of device nodes. Therefore, this embodiment adopts a model parallel approach, dividing the spatiotemporal graph attention network according to feeder partitions. Each feeder's corresponding subgraph is assigned to a GPU for independent inference, and boundary nodes between feeders exchange information through message passing. The remaining lifetime prediction module adopts an asynchronous batch inference mode. Since PC-CVAE inference involves Monte Carlo sampling (1000 samples per device), the computational load is large and unsuitable for real-time triggering. Therefore, this module performs batch forward inference on the PC-CVAE models of all devices according to a preset period (e.g., every early morning), storing the inference results in a database for querying by the predictive maintenance decision module.
[0100] The predictive maintenance decision-making module interacts with the Production Management System (PMS) and the Outage Management System (OMS) via standardized interfaces (such as the CIM model interface of the IEC 61968 / 61970 standard). It pushes maintenance work orders (including equipment identification, maintenance type, urgency level, and suggested maintenance time window) to the PMS and receives maintenance execution receipts (including actual maintenance time, replaced parts, and maintenance personnel) from the PMS. It also pushes outage requests for maintenance to the OMS and receives outage plan approval results from the OMS. Closed-loop feedback is achieved asynchronously through a message queue: after a maintenance work order is executed, the field terminal re-collects and uploads the updated equipment status data via the perception layer module. This data is asynchronously pushed to the data quality governance module and the health index assessment module via the message queue, without waiting for synchronous responses from the preceding modules. Upon receiving the data, the data quality governance module recalculates the TDQI, and the health index assessment module recalculates the health index and triggers incremental updates to the health decay memory state. The asynchronous mechanism of message queues ensures that closed-loop feedback will not block the real-time processing of positive data flow, while message persistence and retry mechanisms ensure that feedback data is not lost.
[0101] It should be noted that the components or steps in the above embodiments can be interleaved, substituted, added, or deleted. Therefore, the combinations formed by these reasonable permutations and transformations should also fall within the protection scope of the present invention, and the protection scope of the present invention should not be limited to the above embodiments.
[0102] The parts of this invention not described in detail are prior art. It will be apparent to those skilled in the art that this invention is not limited to the details of the above exemplary embodiments, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and are intended to encompass all changes falling within the meaning and scope of equivalents within this invention.
Claims
1. A method for assessing the health index of a power distribution network, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing Steps: Collect multi-source heterogeneous data from each monitoring node in the power distribution network, including electrical operation data, environmental stress data, and equipment ledger data; perform spatiotemporal alignment and normalization processing on the multi-source heterogeneous data. S2. Data quality governance steps: Based on the electrical topology of the distribution network, calculate the topology constraint residuals of each monitoring node; The topology constraint residuals and the original electrical measurement data are input into a graph attention network to calculate the data quality score for each monitoring node; the feeder-level topology constraint data quality index (TDQI) is calculated based on the data quality score and the topology constraint residuals; and a hierarchical data repair strategy is executed based on the numerical range of the TDQI. S3. Health Index Assessment Steps: Model the power distribution network as a spatiotemporal graph, with power distribution equipment as nodes and electrical connection relationships as edges; use the data after data quality governance as node state features, input it into the spatiotemporal graph attention network with a health decay memory mechanism, and calculate the health index HI of each device; obtain the feeder-level and regional-level health indices through weighted aggregation of device criticality. S4. Remaining lifetime prediction step: Input the historical health index sequence and current state characteristics of each device into the physical constraint variational autoencoder PC-CVAE, use the Weibull degradation model as a prior constraint, and output the probability distribution of the future health index; calculate the mean and confidence interval of the remaining healthy lifetime RUL based on the probability distribution. S5. Predictive Maintenance Decision Steps: Calculate the Maintenance Urgency Index (MUI) by combining the current health index, remaining health life prediction results, and equipment criticality; generate tiered maintenance work orders based on the MUI, and feed back the maintenance execution results to the data quality governance step and the health index assessment step to form a closed-loop optimization.
2. The method for assessing the health index of a distribution network as described in claim 1, characterized in that, In step S2, the method for calculating the topological constraint residual is as follows: For any monitoring node in the distribution network Calculate its topological constraint residual vector ,include: Kirchhoff's Current Law Residual ,in For nodes The set of electrically adjacent nodes, For a moment From adjacent nodes Inflow node The measured value of the branch current. For a moment node The measured value of the load current; Kirchhoff's voltage law residual ,in For containing nodes A set of independent loops For the first in the loop Measured voltage drop of the branch circuit.
3. The method for assessing the health index of a distribution network as described in claim 2, characterized in that, In step S2, the calculation method for the Topology Constraint Data Quality Index (TDQI) is as follows: in, For feeder The collection of all monitoring nodes included. The total number of monitoring nodes, For nodes computed through a graph attention network Data quality score, For nodes In the time window The average topological constraint residual norm within the range, This is the residual attenuation coefficient; The hierarchical data repair strategy includes: directly using the original data when the TDQI is within a first preset range; performing local repair based on graph attention interpolation when the TDQI is within a second preset range; performing iterative correction based on physical constraints when the TDQI is within a third preset range; and marking the data as unreliable and triggering sensor self-test when the TDQI is within a fourth preset range.
4. The method for assessing the health index of a distribution network as described in claim 1, characterized in that, In step S3, the health decay memory mechanism includes: For each device node Maintain a healthy decaying memory state Its update rules are as follows: in, The decay gate vector is obtained by performing a nonlinear transformation on the concatenation of the current state features and historical memory using the Sigmoid function, and is used to control the proportion of historical degradation information retained. The candidate memory vector is obtained by performing a nonlinear transformation on the concatenation of the current state features and historical memories using the Tanh function, and is used to encode the new degradation information at the current moment; For Hadamard's element-wise product; The physical meaning of the attenuation gating vector is as follows: when the equipment experiences severe overload or harsh environmental stress, the gating value approaches zero, and the historical memory is updated rapidly; when the equipment operates smoothly, the gating value approaches one, and the historical memory is retained to reflect the slow degradation trend.
5. The method for assessing the health index of a distribution network as described in claim 4, characterized in that, In step S3, the spatiotemporal graph attention network further includes a multi-scale temporal aggregation step: For the spatial convolutional output sequence of each device node, multi-scale temporal features are extracted using temporal convolutional networks with different kernel sizes. The different kernel sizes correspond to short-term fluctuation scale, medium-term trend scale, and long-term degradation scale, respectively. The temporal features of each scale are weighted and summed using learnable weights to obtain multi-scale temporal aggregated features. Based on the multi-scale time aggregation features, the health index is normalized to a preset range using the Sigmoid function.
6. The method for assessing the health index of a distribution network as described in claim 1, characterized in that, In step S4, the physical constraint variational autoencoder (PC-CVAE) includes: The encoder, employing a bidirectional LSTM structure, encodes the historical health index sequence and current state features into the mean vector and standard deviation vector of the posterior distribution. Physically constrained prior modules, using characteristic lifetime parameters of the Weibull degradation model. and degenerate shape parameters Given the conditional input, generate the mean vector and standard deviation vector of the prior distribution, replacing the standard normal prior; The decoder, using a TransformerDecoder structure, generates a health index prediction sequence for future preset time steps from latent variables and current state features; During training, a total loss function is used, which includes reconstruction loss, KL divergence loss and physical consistency loss. The physical consistency loss only takes effect under the condition of no maintenance intervention, and penalizes the recovery prediction of the health index under the condition of no maintenance intervention, so as to ensure that the prediction results meet the monotonic physical constraint of equipment degradation.
7. The method for assessing the health index of a distribution network as described in claim 6, characterized in that, In step S4, the method for calculating the remaining healthy lifespan (RUL) is as follows: The remaining healthy lifespan is defined as the time when the health index first drops to the warning threshold; multiple future health index trajectories are obtained from the decoder output of the PC-CVAE through Monte Carlo sampling; the time step when each trajectory first touches the warning threshold is statistically analyzed to obtain the empirical probability distribution of RUL; the mean of RUL, the confidence interval composed of preset quantiles, and the failure probability within the preset time window are output.
8. The method for assessing the health index of a distribution network as described in claim 1, characterized in that, In step S5, the maintenance urgency index (MUI) is calculated as follows: in, For equipment Current health index For equipment The average remaining lifespan, As a reference time constant, The equipment criticality weight is obtained by normalizing the product of the power loss due to equipment failure and the average recovery time. The closed-loop optimization includes: after maintenance is completed, re-entering the updated status data of the equipment into the data quality governance step to verify the data quality improvement effect, re-entering the health index assessment step to verify the health recovery effect, and encoding the maintenance event and inputting it into the health decay memory mechanism to learn the impact of maintenance intervention on the degradation trajectory.
9. A power distribution network health index assessment system, characterized in that, include: The perception layer module is used to collect multi-source heterogeneous data from various monitoring nodes in the power distribution network, and to perform spatiotemporal alignment and normalization processing on the multi-source heterogeneous data. The data quality governance module is used to calculate topology constraint residuals based on the electrical topology of the distribution network, calculate data quality scores and topology constraint data quality indices (TDQI) through graph attention networks, and execute hierarchical data repair strategies based on the TDQI. The health index assessment module is used to model the distribution network as a spatiotemporal diagram and calculate the device-level, feeder-level, and regional-level health indices through a spatiotemporal diagram attention network with a health decay memory mechanism. The remaining life expectancy prediction module is used to output the probability distribution of future health index through a physical constraint variational autoencoder PC-CVAE, and to calculate the mean and confidence interval of the remaining healthy life. The predictive maintenance decision module is used to calculate the maintenance urgency index (MUI), generate tiered maintenance work orders, and feed the maintenance execution results back to the data quality governance module and the health index assessment module to form a closed-loop optimization.
10. The power distribution network health index assessment system as described in claim 9, characterized in that, The system adopts an edge-cloud collaborative deployment architecture, specifically including: The perception layer module is deployed on the field side of the distribution network and includes feeder terminal units (FTU), distribution terminal units (DTU), transformer terminal units (TTU), smart sensors, and meteorological monitoring devices. Each terminal device aggregates data through an edge computing gateway. The edge computing gateway has a built-in lightweight data quality pre-inspection unit, which is used to perform real-time calculation of topology constraint residuals before data is uploaded. When the topology constraint residuals are detected to exceed a preset threshold, data re-sampling or marking of abnormal timestamps is triggered on the edge side. The data quality governance module is deployed between the edge computing gateway and the regional server, and adopts a two-level collaborative mechanism: the first level performs lightweight graph attention inference on the edge computing gateway to complete the real-time calculation of data quality scores and the immediate response of hierarchical repair strategies; the second level performs full-topology-wide graph attention network inference on the regional server to complete the global calculation of feeder-level TDQI and cross-feeder data consistency verification. The health index assessment module and the remaining lifespan prediction module are deployed on the central cloud platform. The health index assessment module deploys the spatiotemporal graph attention network on the GPU cluster in a model parallel manner according to the feeder partition. The remaining lifespan prediction module adopts an asynchronous batch inference mode and performs batch forward inference on the PC-CVAE model of each device according to a preset cycle. The predictive maintenance decision module interacts with the production management system (PMS) and the power outage management system (OMS) through a standardized interface. The closed-loop feedback is achieved through a message queue: after the maintenance work order is completed, the field terminal re-collects the updated status data of the equipment through the perception layer module and pushes it asynchronously to the data quality governance module and the health index assessment module through the message queue, triggering an incremental update of the health decay memory status.
