First-second fusion sensor full life cycle drift prediction method based on deep learning

Abstract

The invention discloses a first-second fusion sensor full life cycle drift prediction method based on deep learning, and the method comprises the steps: carrying out the modeling of a power grid as an attribute heterogeneous information network sequence with a timestamp, and carrying out the deep fusion of multi-modal data, such as a primary equipment standing book, secondary equipment measurement, and the like, through an adaptive attention mechanism, and generating node features; the sequence is input to a space-time diagram neural network model, the model captures collaborative drifting through a space aggregation module of relation perception, learns a drifting rule through a time sequence evolution module of topology perception, and decouples intrinsic drifting and collaborative drifting components through orthogonal constraint; and performing multi-task prediction based on the decoupling characteristics, and outputting a drift value, a grade and residual life. According to the method, the adaptability of the model to the dynamic topology is improved by adopting a meta-learning framework, and drift attribution and operation and maintenance suggestions are provided through a graph interpretation algorithm. According to the method, sensor drift can be accurately predicted, deep cause analysis is realized, and powerful support is provided for predictive maintenance of an intelligent power grid.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and power system monitoring technology, and in particular to a deep learning-based method for predicting the drift of a single- and two-element fusion sensor throughout its entire lifecycle. Background Technology

[0002] With the rapid development of smart grids, massive sensors, acting as the "nerve endings" of the power grid, rely on accurate measurement data as the cornerstone for precise situational awareness, advanced state estimation, and safe and stable operation of the grid. However, during long-term service, sensors inevitably experience slow and continuous degradation in measurement performance due to multiple stresses, including equipment aging, environmental factors, and changes in operating conditions—a phenomenon known as performance drift. To address this issue, existing technologies primarily employ prediction methods based on traditional statistical models (such as Autoregressive Integrated Moving Average (ARIMA) and Kalman filtering) or time-series deep learning models (such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU)). These methods can, to some extent, capture the nonlinear temporal evolution of individual sensor drift, providing some technical support for condition-based maintenance.

[0003] However, existing technologies still have significant shortcomings when applied to complex power grid systems: First, existing methods generally treat sensors as independent entities, ignoring the strong spatial coupling effects inherent in the power grid topology. In physically interconnected power grids, the drift of a single sensor can affect the measurement performance of neighboring nodes through power flow, forming a complex "cooperative drift" phenomenon. Existing methods cannot effectively model this spatiotemporal correlation, easily leading to misjudgment of drift sources. Second, existing technologies have limitations in data utilization, relying heavily on secondary electrical quantity measurements from sensors and failing to effectively integrate multimodal heterogeneous information that characterizes the physical health of equipment, such as primary equipment ledgers and historical maintenance records, resulting in an incomplete characterization of the sensor's entire lifecycle status. Third, existing predictive models typically output a general drift result, failing to deeply decouple the causes of drift and making it difficult to distinguish between drift caused by intrinsic equipment performance degradation and apparent drift caused by network cooperative effects, which is detrimental to the accurate formulation of maintenance strategies. Fourth, the power grid topology is not static. Line switching, maintenance and other operations will cause dynamic changes in the network structure. However, the existing models have insufficient generalization ability. Once the topology changes, the prediction accuracy drops sharply, and they lack the ability to quickly adapt to dynamic topologies. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a deep learning-based method for predicting the drift of a fusion sensor throughout its entire lifecycle, in order to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a deep learning-based method for predicting the entire lifecycle drift of a fusion sensor, comprising: Sensors and associated physical devices in the power grid are defined as different types of nodes, and the physical or electrical connections between these nodes are defined as different types of relationships. Based on real-time operating data and static ledger data, a time-stamped attributed heterogeneous information network sequence is generated. For each node in the network sequence, the multimodal data of the node is fused to generate the feature vector of the node; The attributed heterogeneous information network sequence is input into a preset spatiotemporal graph neural network model. The model captures the collaborative drift effect between nodes at each time step through a spatial information aggregation module, and learns the drift pattern of node features over time through a temporal evolution module. Based on the output of the spatiotemporal graph neural network model, the drift state of the target sensor at future time points is predicted.

[0007] As a preferred embodiment of the deep learning-based sensor lifecycle drift prediction method described in this invention, the multimodal data includes primary equipment static ledger data, secondary equipment dynamic measurement data, and environmental and operating condition data.

[0008] As a preferred embodiment of the deep learning-based sensor lifecycle drift prediction method based on fusion of one- and two-modal data as described in this invention, the method involves fusing the multimodal data of the nodes to generate feature vectors for the nodes, including: Data from different modalities are embedded separately to obtain their respective modal feature vectors; The weight coefficients of each modality feature vector are calculated using an attention mechanism. These weight coefficients characterize the relative importance of each modality data in describing the current node state. The modal feature vectors are weighted and aggregated to generate the feature vectors of the nodes.

[0009] As a preferred embodiment of the deep learning-based one-to-two sensor lifecycle drift prediction method described in this invention, the spatial information aggregation module is a relation-aware graph attention network module. For each relation type existing in the heterogeneous information network sequence, the module learns independent information transformation parameters and attention weights, and differentially aggregates information from neighboring nodes of different relation types.

[0010] As a preferred embodiment of the deep learning-based one-to-two fusion sensor lifecycle drift prediction method described in this invention, the temporal evolution module is a gated cyclic unit network. When the gated unit in the gated cyclic unit network updates the hidden state of the current node, it not only fuses the current information of the node itself, but also fuses the historical hidden state information of its neighboring nodes to model the temporal transmission effect of drift state in the topological network.

[0011] As a preferred embodiment of the deep learning-based one-to-two sensor lifecycle drift prediction method described in this invention, the method further includes a step of decoupling the drift components before the output of the spatiotemporal graph neural network model. This step includes: The output of the time-series evolution module is mapped to the first drift feature subspace and the second drift feature subspace, respectively; By adding an orthogonality penalty term to the total loss function of model training, the feature representations in the first drift feature subspace and the second drift feature subspace are made independent of each other, thereby quantifying the intrinsic drift caused by the aging of the equipment itself and the cooperative drift caused by the influence of the power grid environment.

[0012] As a preferred embodiment of the deep learning-based one-to-two fusion sensor lifecycle drift prediction method described in this invention, wherein: the predicted drift state of the target sensor at a future time point includes: Based on the intrinsic drift and the cooperative drift, multi-task prediction is performed, which includes at least: regression prediction of sensor drift value, classification prediction of sensor drift level, and prediction of sensor remaining service life.

[0013] As a preferred embodiment of the deep learning-based one-to-two sensor lifecycle drift prediction method described in this invention, the model training process of the method employs a meta-learning framework, which includes: Data under different power grid topologies or operating conditions are divided into multiple learning tasks; Through the gradient optimization process of the inner and outer loops, a model initial parameter set is learned that can quickly adapt to new tasks using a small amount of new data. The learning tasks are divided based on at least one of the following: power grid topology change events, power grid operating mode switching events, and extreme weather events.

[0014] As a preferred embodiment of the deep learning-based one-to-two fusion sensor lifecycle drift prediction method described in this invention, the method further includes, after predicting the drift state of the target sensor at a future time point, a step of attributing the causes of drift in relation to other factors. This step includes: Based on the drift prediction results of the target sensor, a pre-defined graph interpretation algorithm is used to identify and extract a network subgraph and key node features that contribute the most to the prediction results, in order to locate the key sources and key influencing factors of cooperative drift.

[0015] As a preferred embodiment of the deep learning-based one-to-two fusion sensor lifecycle drift prediction method described in this invention, wherein: after the correlation attribution step, a step of generating counterfactual explanations is further included, the step comprising: An optimization algorithm is used to find a minimum perturbation to the node features, such that under the minimum perturbation, the drift prediction result of the spatiotemporal graph neural network model for the target sensor changes from an unsatisfactory state to a satisfactory state, and the minimum perturbation is used as an executable operation and maintenance control suggestion.

[0016] Compared with existing technologies, the beneficial effects of the invention are as follows: 1. This invention constructs a spatiotemporally heterogeneous graph network and employs a relation-aware graph neural network to explicitly capture the "cooperative drift" effect among sensors in the power grid topology at the model level. Compared to existing analytical methods that treat sensors as independent entities, this invention can more accurately identify apparent drift caused by network conduction. Through orthogonal decoupling, it quantifies drift into "intrinsic drift" caused by the aging of the equipment itself and "cooperative drift" under the influence of the power grid environment. This provides early warning insights for fault diagnosis and liability determination, solving the pain point of existing technologies that cannot distinguish the root cause of drift.

[0017] 2. To address the challenges posed by frequent changes in power grid topology, this invention introduces a meta-learning framework for model training. By constructing data from different topologies or operating conditions into a series of learning tasks, the model learns "how to learn," enabling it to quickly adapt to the drift evolution patterns in new environments with only a small amount of new data after topology changes. This overcomes the shortcomings of traditional deep learning models, which suffer from a sharp performance decline after topology changes, ensuring the robustness and practicality of the method in real, dynamic power grid environments.

[0018] 3. Furthermore, this invention not only provides accurate prediction results but also transforms the "black box" model into a "white box" decision support tool through integrated graph interpretability algorithms and counterfactual interpretation generation methods. This tool is used to locate the key sources and paths of influence causing collaborative drift and to generate specific, actionable operation and maintenance control suggestions, such as "reducing the load rate of a certain line by 5% can restore the target sensor drift to normal," thus enhancing the application value of the technical solution. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the overall process of a deep learning-based one-to-two sensor lifecycle drift prediction method according to an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0024] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0025] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0026] Example 1 Reference Figure 1 This is the first embodiment of the present invention, which provides a deep learning-based method for predicting the drift of a sensor throughout its entire lifecycle, including: S1. Define the sensors and associated physical devices in the power grid as different types of nodes, define the physical or electrical connections between nodes as different types of relationships, and generate a time-stamped attributed heterogeneous information network sequence based on real-time operation data and static ledger data.

[0027] It should be noted that this step aims to transform raw, unstructured power grid data into a graph data format that can be effectively processed by deep learning models. In other words, sensors are no longer treated as isolated data points, but rather placed within the complete physical and electrical environment of the power grid, explicitly modeling the complex dependencies between devices by constructing a Heterogeneous Information Network (HIN).

[0028] Furthermore, based on the physical structure and business logic of the power grid, the network's meta-structure is abstracted and defined. It should be explained that the meta-structure is a blueprint describing the node types and relationship types in a heterogeneous network. In this embodiment, this meta-structure... Defined as follows: This is represented as a set of node types, which includes all key entity types in the power grid related to sensor performance drift, for example, A = {sensor, transformer, line, bus}. This approach overcomes the limitation of focusing solely on the sensor itself, incorporating primary physical equipment that directly supports the sensor, influences its operating environment, and transmits electrical disturbances into the modeling scope, thus enabling a more comprehensive characterization of the potential influencing factors of drift.

[0029] This is represented as a set of relationship types, which defines the physical or electrical connections between different types of nodes. For example, ={Sensor-Measurement at-Bus, Bus-Connection to-Line, Bus-Connection to-Transformer, Line-Connection to-Bus, Transformer-Connection to-Bus}. The advantage of this definition is that each relationship contains a unique physical meaning. For example, "Sensor-Measurement at-Bus" indicates that voltage or current fluctuations on the bus will be directly reflected in the sensor's measurement value. Another example is the "Bus-Connection to-Line" relationship, which represents the transmission path of power flow disturbances between different buses.

[0030] Furthermore, after defining the metastructure, it is necessary to utilize real-time and historical data from the power grid at each discrete timestamp. (For example, the sampling interval can be set to 5 minutes, 15 minutes, or 1 hour, which can be determined based on data availability and the rate of drift change.) Instantiate a specific property graph. Then a timestamp A snapshot of a graph can be formally represented as: in, It is a set of all nodes, whose type follows a set principle. Definition. For example, all the sensors, transformers, lines, and busbars within a substation constitute a set. . In timestamp The set of edges, the existence and type of which follow the set rules. The definition is determined based on the real-time topology of the power grid. For example, if a certain line is in time... If the line is in operation, it means that there is a "connection" type edge between the nodes of the line and the bus nodes they are connected to; if the line is disconnected, the corresponding edge will be... The removal of the graph structure allows the graph to dynamically reflect changes in the power grid topology, thus enabling the model to adapt to drastic topological changes. It is an optional edge feature matrix. This refers to the dimension of the edge features. For example, for an edge of the "line" type, its features can include static information such as the line's impedance parameters and length, as well as dynamic information such as the real-time power flow magnitude on the line. It is the node feature matrix, where, It is the total number of nodes. It is the dimension of the node feature vector, and the th dimension of the node feature matrix. row vector Representative node In time Attributes.

[0031] Furthermore, the construction of this node feature matrix incorporates multimodal data. This multimodal data includes, but is not limited to: Primary equipment static ledger data: For nodes such as transformers and lines, this includes the years of operation, equipment model, insulation class, and historical overhaul records. This ledger data is encoded as numerical values ​​or one-hot vectors, reflecting the inherent health condition and aging degree of the equipment.

[0032] Secondary equipment dynamic measurement data: For sensor nodes, this includes directly measured values ​​of the sensor node (such as voltage, current amplitude / phase), waveform distortion rate, harmonic content, etc.; while for other equipment nodes, it includes dynamic data collected by the SCADA system, such as temperature, active / reactive power, load rate, etc.

[0033] Environmental and operating condition data: This refers to environmental data shared by all nodes, such as ambient temperature and humidity, as well as operating condition data that characterizes the overall operating status of the system, such as total load level.

[0034] It is important to note that, based on the aforementioned set of node types, in this embodiment, the primary components (such as pole-mounted switches and ring main units) in the integrated primary and secondary equipment are defined as "source nodes," and the integrated secondary terminals and sensors are defined as "sensing nodes." When constructing the edge set, strong connections of the "parasitic / integrated" type are established between the source nodes and the sensing nodes to explicitly model the direct thermal conduction or electromagnetic interference effects of the physical state of the primary equipment (such as vibration and heat generation) on the sensor performance.

[0035] Specifically, after cleaning and normalizing these multimodal data from different sources, they are concatenated together to form an information-rich initial node feature vector. .

[0036] Furthermore, through continuous time windows By repeatedly performing the above instantiation, a time-stamped, attribute-based heterogeneous information network sequence can be generated. This sequence provides a complete record of the evolution of the power grid topology and equipment status over time.

[0037] S2. For each node in the network sequence, fuse the multimodal data of the node to generate the feature vector of the node.

[0038] It should be noted that, due to the significant differences in scale, physical units, and semantic level between directly concatenated multimodal data, inputting them directly into the model without processing will hinder effective learning, and less important modalities may be overwhelmed by noise. Therefore, this invention designs an adaptive multimodal fusion module. This module uses deep learning methods to map data from different sources to a unified semantic space and utilizes an attention mechanism to dynamically evaluate the contribution of each modality to describing the current node state, ultimately generating a node feature representation with higher information density and richer semantics.

[0039] Specifically, regarding timestamps Attribute diagram any node in The original splicing feature vector generated in step S1 is Therefore, this feature vector can be decomposed according to the data source into... In this embodiment, one modal component is taken. , respectively corresponding to: : Primary equipment static ledger data components; Components of dynamic measurement data from secondary equipment; Components of environmental and operating condition data; Furthermore, in order to eliminate heterogeneity between different modalities, it is first necessary to project the data of each modality into a shared, low-dimensional latent feature space through a dedicated embedding layer.

[0040] Specifically, for categorical data (such as device model, manufacturer, etc.), a learnable embedding lookup table is used to transform it into a dense vector representation. For numerical data (such as voltage, current, temperature, etc.), a nonlinear transformation network (e.g., a small multilayer perceptron, MLP) consisting of one or more fully connected layers is used for embedding. The above process can be formally represented as follows: in, It is the modal index. It is the first Modality-specific embedding networks. For example, , , These are MLP networks designed for the static ledger data components of primary equipment, the dynamic measurement data components of secondary equipment, and the environmental and operating condition data components. These network structures can differ to adapt to the characteristics of their respective data, but their output dimensions are unified, i.e. ,in, It is the dimension of the shared potential space (e.g., it can be set to 64 or 128). It is a node In time The Embedding vectors for each modality.

[0041] It should be noted that through the above projection operation, all the original, heterogeneous data are transformed into feature vectors with the same dimension and comparable semantics.

[0042] Furthermore, after obtaining the embedding vectors for each modality, it is necessary to determine the current time step. For nodes In this context, it depends on which modality's information is more critical. For example, for an older sensor, its static register (service life) information may be more important than the ambient temperature; while during a power grid disturbance, the fluctuation of its dynamic measurements is the key information. Based on this, to enable the model to automatically learn the importance of this dynamic, the present invention introduces an attention mechanism. The calculation process for the attention weights is as follows: S201.1 Calculate a nonlinear transformation for each modality embedding vector to enhance its expressive power: in, and It consists of a learnable weight matrix and a bias vector. It is the dimension after attention transformation. It is the hyperbolic tangent activation function.

[0043] S201.2 Calculate the relationship between each transformed vector and a shared, learnable context vector. The similarity is used to obtain the attention score: Among them, context vector Shared across all nodes and modalities, it functions as a "query" to measure the importance of each modal's information.

[0044] S201.3. Use the Softmax function to normalize the scores of all modalities to obtain the final weight coefficients. : Among them, the weighting coefficient The value is between 0 and 1, and This weighting coefficient intuitively represents the first... Each modal data point describes a node. In time The relative importance of the overall state.

[0045] Furthermore, after obtaining the weight coefficients of each modality, these weight coefficients are used as weights to perform a weighted summation on the corresponding modality embedding vectors, thus obtaining the fused node feature vector. : It should be noted that the feature vector of this node It is a highly refined representation that adaptively integrates all information from primary equipment, secondary measurements, and the external environment. Its superscript... This represents the input features of layer 0 (the first layer) of the spatiotemporal graph neural network, i.e., the initial node representation. After the above processing, the updated node feature matrix can be obtained. Each row of this matrix represents the fused feature vector of the corresponding node. .

[0046] S3. Input the attributed heterogeneous information network sequence into a pre-defined spatiotemporal graph neural network model. The model captures the collaborative drift effect between nodes at each time step through a spatial information aggregation module, and learns the drift pattern of node features over time through a temporal evolution module.

[0047] Furthermore, a deep learning model capable of simultaneously understanding the spatial topology dependence and temporal dynamic evolution of the power grid is constructed. This model receives the spatiotemporal graph sequence generated in steps S1 and S2. As input, where The historical time window length can be set to 24, meaning it uses data from the past 6 hours (assuming a sampling interval of 15 minutes) to predict the future. The overall architecture of the model is a recurrent neural network (RNN) framework, in which, at each time step... Each loop body embeds a Graph Neural Network (GNN) module. The GNN is responsible for propagating spatial information on the graph snapshot at the current moment, while the RNN is responsible for connecting states at different moments and learning the temporal evolution pattern. In this embodiment, the spatiotemporal graph neural network model consists of a spatial information aggregation module, a temporal evolution module, and a drift component decoupling module.

[0048] Furthermore, to accurately capture the effects of cooperative drift under different physical / electrical connection relationships, this invention employs a relation-aware graph attention network (RGAT) as the spatial information aggregation module. Unlike standard graph attention networks (GAT), RGAT can distinguish different types of neighbors and learn a unique information transmission pattern for each relationship.

[0049] Specifically, at any given time step ,for The central node in (assuming it is) The process by which it aggregates information from neighboring nodes to update its representation is as follows (taking the model number as an example). (For example, a layer) First, for nodes Every neighbor If they are related Connection (i.e.) ,in, Represents a node In relationship (the set of neighbors below), first check the neighbor nodes. The upper layer feature representation Perform a relationship Related linear transformations: in, It is the first Layers are exclusive to relation types The learnable weight matrix. The physical meaning of this weight matrix is ​​that the model acknowledges that the physical meaning of the information conveyed by the two relationships, "sensor-measurement-bus" and "bus-connection-line", is completely different, and therefore different parameter matrices are needed to extract and transform the information.

[0050] Then, the model computation nodes To the neighbors (in relationship) Attention coefficient (below) This coefficient is also relation-aware: in, It is the central node The characteristics are represented. It is a relationship Learnable attention weight vectors. This indicates a vector concatenation operation. It is a non-linear activation function.

[0051] It should be noted that the design of this attention coefficient allows the learning of attention weights to take into account the type of relationship. Therefore, the model can learn which neighboring nodes are more worthy of attention under a certain relationship.

[0052] Finally, the Softmax function is used to process the nodes. The attention coefficients of all neighbors under all relations are normalized to obtain the final attention weights. : The features of neighboring nodes after relation-specific transformations are weighted and summed using their corresponding attention weights to obtain an intermediate representation that aggregates spatial neighborhood information. : in, It is another non-linear activation function (such as ELU).

[0053] It should be noted that by stacking multiple layers of RGAT (two layers in this embodiment), the model can capture the cooperative drift effects within higher-order neighborhoods. In this embodiment, the output of the RGAT module is... (For brevity, the number of layers has been omitted) This output condenses the time... node All information about the local network environment.

[0054] Furthermore, in obtaining the spatial embedding at each time step Subsequently, a temporal evolution module is needed to learn its evolutionary patterns. Based on this, the present invention employs a topology-aware gated cyclic unit network as this temporal evolution module.

[0055] Specifically, when updating the hidden state of the current node, not only is the node's own historical information considered, but the historical state information of its neighboring nodes is also explicitly incorporated, thereby directly modeling the temporal propagation effect of drift states in the power grid topology. For nodes In time Hidden state The update process is as follows: Aggregation Node All the neighbors at the last moment The hidden state. In this embodiment, for ease of calculation, the process uses a simple average aggregation, which is expressed as: in, Represents a node The overall state of the surrounding environment at the previous moment.

[0056] It should be explained that the standard GRU has two gates, namely the reset gate. and Update Gate In the temporal evolution module of this invention, the calculation of both gates additionally considers the historical state of the neighborhood. : in, It is the spatial embedding obtained from the temporal evolution module at the current moment. It is a node The hidden state of oneself in the previous moment. It is the learnable parameter matrix and bias of the time-series evolution module. It is used to learn how to adjust the gate control signal based on the historical status of the neighbors. It is the Sigmoid activation function.

[0057] Meanwhile, candidate hidden state The calculation and final hidden state The updates are similar to those of the standard GRU: in, It is the tanh activation function. It is an element-wise product.

[0058] Furthermore, by inputting the spatiotemporal graph sequence into this module and iterating... By taking a time step, we can obtain the time of each node. The final hidden state of time The final hidden state encodes the node and its neighborhood from time 1 to... All spatiotemporal dynamic information.

[0059] Furthermore, to achieve in-depth analysis of the causes of drift, this invention also designs a decoupling module before the output of the temporal evolution module. This decoupling module mainly forces the module to learn two independent drift representations by constraining the orthogonality of the two feature subspaces.

[0060] Specifically, the final hidden state output by the temporal evolution module Through two independent linear projection layers (or a small MLP) and These are mapped to the first drift feature subspace (intrinsic drift) and the second drift feature subspace (cooperative drift), respectively: in, , It is the dimension of the decoupled feature vector. These are the decoupled feature vectors, represented as the first drift feature subspace and the second drift feature subspace, respectively.

[0061] Specifically, in order to make the physical meanings represented by these two subspaces independent of each other, the total loss function during model training... Add an orthogonal constraint loss to The loss term aims to minimize the covariance of the two feature vector sets. Therefore, a simple and efficient implementation is to minimize the covariance of all samples within a batch. and Cosine similarity squared of vectors: in, This is the loss from downstream forecasting tasks. It is a hyperparameter that controls the strength of orthogonal constraints. It can be selected in [0.01, 0.1, 1.0] through cross-validation. In this embodiment, a typical effective value is 0.1. Indicates the batch number. This represents the total number of samples in the batch.

[0062] It should be noted that by minimizing This motivates the model to seek a mapping that makes the feature representations of intrinsic drift and co-drift tend to be orthogonal in the vector space, thereby achieving decoupling at the feature level.

[0063] Furthermore, in order to train the aforementioned spatiotemporal graph neural network model, a historical dataset containing real labels must be constructed. In this embodiment, the training dataset is constructed from at least one of the following sources: (i) Temperature and humidity cycling and electrical stress aging tests were conducted on the same type of fusion sensor under controlled environment, and the measurement error was recorded in real time as a drift tag by a high-precision standard source.

[0064] (ii) Error data at discrete time points obtained by the field calibration instrument of the power grid metering center are used to fit a continuous drift trend curve as a label through spline interpolation.

[0065] (iii) Establish a physical circuit simulation model of the sensor, artificially inject different degrees of component parameter degradation (such as capacitor drying, coil short circuit) and external interference, and generate synthetic data with accurate drift values ​​for model training.

[0066] It is important to note that during the prediction phase, the model only takes real-time running data and ledger data as input and outputs the predicted drift status.

[0067] S4. Based on the output of the spatiotemporal graph neural network model, predict the drift state of the target sensor at future time points.

[0068] It should be noted that this step aims to utilize the intrinsic drift characteristics obtained from the decoupling in step S3. ) and cooperative drift characteristics ( Using a multi-task learning (MTL) framework as an information source, it simultaneously predicts multiple key indicators of the sensor's future state.

[0069] Specifically, for any node Following the main structure of the spatiotemporal graph neural network model, an independent prediction head is constructed for each prediction task. Each prediction head is a simple neural network consisting of one or more fully connected layers, and receives decoupled feature vectors as input.

[0070] Furthermore, in order to allow each task to utilize information from both drift causes, the intrinsic drift features are first... and cooperative drift characteristics The features are concatenated to form a comprehensive feature vector. : in, This indicates a vector concatenation operation.

[0071] It should be noted that this comprehensive feature vector This will serve as a common input for all prediction tasks, ensuring that each task can learn from complete information about the causes of drift.

[0072] Furthermore, based on the aforementioned comprehensive feature vectors, a regression prediction head is constructed. The prediction head is a multilayer perceptron (MLP) used to predict sensor performance in the future. Each time step (e.g., The drift value of the regression prediction head can be expressed as: in, It is a predicted future A sequence of drift values ​​at each time step. For example, It is the drift prediction value for the next moment. It typically contains 1 to 2 hidden layers and uses a linear activation function as the output layer.

[0073] Furthermore, based on the aforementioned comprehensive feature vectors, a classification prediction head is constructed. This is used to predict the drift level of a sensor at a critical future point in time (e.g., the end of the next maintenance cycle). The drift level can be predefined according to the operation and maintenance procedures, such as three categories: "Normal," "Concern," and "Abnormal." The output of this classification prediction head can be expressed as: in, It is a probability distribution vector. It is the number of grade categories (in this embodiment, Each element of each probability distribution vector represents the probability that the sensor belongs to the corresponding level. The output layer uses the Softmax activation function.

[0074] Furthermore, based on the aforementioned comprehensive feature vectors, a RUL prediction head is constructed. This is used to predict the remaining time from the current moment until the sensor's performance drifts beyond an acceptable threshold. The output of this RUL prediction head can be expressed as: in, It is the predicted remaining useful life value (e.g., in days or months).

[0075] It should be noted that, since the RUL prediction head is essentially a regression task, but since its target is positive, the output layer can use activation functions such as ReLU (without restriction).

[0076] Furthermore, in order to jointly optimize the loss of all downstream prediction tasks, a comprehensive loss function needs to be designed. This function is a weighted sum of the losses for each individual task: Furthermore, regarding regression loss Using the mean squared error (MSE), we obtain: in, These are the actual drift value labels.

[0077] Furthermore, regarding classification loss Cross-entropy loss is used: in, It is the one-hot encoding of the real grade label. It is the one-hot encoding of the predicted grade label. An index representing a single category.

[0078] Furthermore, regarding RUL losses Mean square error using logarithmic transformation: in, It is a genuine RUL label.

[0079] It should be noted that by using the mean squared error of logarithmic transformation, the model becomes less sensitive to the large error values ​​commonly found in RUL predictions, making the model more stable during training.

[0080] Furthermore, These are the weight hyperparameters for each of the aforementioned individual tasks, used to balance the importance or magnitude of different tasks. These weights can be set according to the priority and magnitude of the tasks; for example, they can be set to... , , Alternatively, an uncertainty-weighted algorithm can be used for adaptive learning during training.

[0081] Furthermore, the final total loss function for model training is the orthogonal constraint loss defined in step S3: Furthermore, end-to-end training minimization is achieved by performing backpropagation and gradient descent optimizers (such as Adam) on the entire model (including the fusion module in step S2, the spatiotemporal graph network model in step S3, and the task prediction head in step S4). This allows the output of a set of structured prediction results for each target sensor.

[0082] For example, this set of structured prediction results is as follows: A series of drift values ​​for the next 15, 30, or 1 hour (e.g., drift percentages for the next 15, 30, or 1 hour).

[0083] A probability distribution of drift level at a future key time point (e.g., "Normal": 90%, "Attention": 8%, "Abnormal": 2%).

[0084] An estimate of the remaining useful life (e.g., 180 days).

[0085] It should be noted that the prediction results constitute a comprehensive health profile of the sensor throughout its entire lifecycle, providing power grid operation and maintenance personnel with decision support ranging from short-term early warning to long-term planning. For example, regression values ​​can be used for online correction of real-time measurement data; classification results can directly trigger early warning work orders; and RUL prediction provides a key basis for formulating equipment replacement and asset management plans. In this way, the present invention closely integrates the output of the deep learning model with the actual business needs of power grid operation and maintenance, thereby enhancing the practical value of the model's prediction results.

[0086] Furthermore, when standard deep learning models are trained on a specific data distribution, their parameters tend to overfit that distribution. When the power grid topology or operating conditions change significantly (i.e., data distribution drift), the model's performance degrades severely. To overcome this deficiency, this invention employs meta-learning, particularly the model-agnostic meta-learning (MAML) framework. Its core idea is not to learn a set of parameters that perform optimally on any single task, but rather to learn a set of initial parameters that possess rapid learning capabilities. Using this set of parameters as a superior starting point, when encountering new, unseen power grid topologies, it can be quickly fine-tuned with only a small amount of new sample data and minimal gradient update steps, thus adapting to new tasks.

[0087] Furthermore, since a prerequisite for meta-learning is constructing the model training data into a series of independent "learning tasks," in the context of this invention, a "task" represents the power grid topology or operating condition within a specific period. The division of tasks (i.e., the events that trigger task switching) includes, but is not limited to: Power grid topology change events: These events, such as the switching on / off of critical lines, and the maintenance shutdown of transformers or buses, are identified through dispatch automation systems (OMS) or power grid topology analysis tools. Data from each stable topology period is divided into an independent task.

[0088] Power grid operation mode switching events: By analyzing global operating parameters of the power grid (such as total load, power flow direction of major interconnections, and renewable energy penetration rate) using clustering algorithms, typical operating modes (such as "summer high load mode" and "nighttime low load mode") are identified. Data belonging to the same operating mode are grouped into one task.

[0089] Extreme weather events: Based on meteorological data, data during periods of extreme weather such as typhoons, icing, and high temperatures are classified as separate tasks.

[0090] Furthermore, for each independent task, its corresponding dataset is randomly split into two parts: One part is the support set: containing a small number of labeled samples, used to quickly simulate the adaptation process to new tasks in the "inner loop" of meta-learning.

[0091] The other part is the query set: it contains another set of samples that do not intersect with the support set, used to evaluate the model's generalization performance after adapting to new tasks in the "outer loop" and to guide the optimization direction of the initial parameters.

[0092] Furthermore, let the learnable parameters of the entire spatiotemporal graph neural network model (including all parameters from steps S2 to S4) be . The goal of meta-learning is to find an optimal set of initial parameters. .

[0093] Furthermore, the inner loop of meta-learning is used to simulate task adaptation. For tasks distributed... A batch of tasks in the middle sampling For each of these tasks Perform the following operations individually: The computational model in the support set of the current task Losses This loss function is the total loss function defined in step S4. .

[0094] Based on this loss, the model parameters Perform once or several times ( (Number of) gradient descent updates yield a set of values ​​adapted to the task. Temporary parameters For a single update: in, It is the learning rate of the inner loop, a hyperparameter that controls the speed at which the task adapts. Indicates the model parameters Find the gradient.

[0095] It should be noted that the above single update process simulates how to quickly fine-tune the model using a small amount of data (support set) when it encounters a new task.

[0096] Furthermore, the outer loop of meta-learning is used to optimize the initial parameters. Since the goal of the inner loop is rapid adaptation, the goal of the outer loop is to improve the effectiveness of this "adaptation" itself. Specifically, the outer loop optimizes the initial parameters... This makes from The temporary parameters obtained after fine-tuning the inner loop are as follows: In the query set of the corresponding task It performed best.

[0097] Specifically, the query set losses of all tasks are summed to form the meta-objective function. : Then, the meta-objective function with respect to the initial parameters is calculated. The gradient (meta-gradient) is obtained and updated. : in, It is the learning rate of the outer loop (meta-learning rate). The calculation of the gradient requires backpropagation through the update step of the inner loop, which is a process of calculating the second derivative.

[0098] It should be noted that the parameters obtained are obtained by repeatedly executing the inner and outer loops. It possesses cross-task generalization capabilities. When a real topology change event occurs, only a small amount of data under the new topology needs to be obtained as the support set, and a fine-tuning process is performed in the inner loop to obtain a prediction model with excellent performance under the new topology, thus achieving fast and low-cost adaptation to dynamic topologies.

[0099] Furthermore, to "open" the "black box" of the spatiotemporal graph neural network and explain why the model makes a specific drift prediction, when step S4 predicts that the drift of a target sensor node will exceed the limit, maintenance personnel urgently need to know which devices and factors have jointly caused this result. Therefore, this invention introduces a graph interpretability algorithm to automatically identify the key substructures and features most relevant to the prediction result from the complex power grid topology, achieving accurate tracing of the path of collaborative drift influence. In this embodiment, the preset graph interpretability algorithm selected is GNNExplainer. The core idea of ​​this algorithm is to find the smallest subgraph and the most critical feature subset of the original graph through optimization methods, so that the model's prediction results on this subgraph and sub-features are as similar as possible to the prediction results on the complete graph.

[0100] Specifically, for the target sensor node At any moment drift prediction results The goal of GNNExplainer is to maximize model predictions. (This is) ) and an explanation Mutual Information (MI): in, It is a random variable representing the prediction result of the model. It is the key computational subgraph to be found, which consists of the original graph. A subset of edges is defined. It is a subset of node features, expressed through a feature mask. Choose one. This represents information entropy.

[0101] It should be noted that maximizing mutual information is equivalent to minimizing conditional entropy. The objective is to minimize the uncertainty of the original prediction given an interpretation (subgraph and sub-features). To make the interpretation concise and easy to understand, a penalty term (regularization term) is added to the optimization objective for the subgraph size and the number of features. ): in, It is the number of edges in the subgraph. It is the number of features selected. It is a hyperparameter that controls the simplicity of the interpretation.

[0102] Furthermore, by applying a learnable mask to the adjacency matrix and feature matrix, and optimizing the above using gradient descent... This allows us to obtain the optimal edge mask and feature mask.

[0103] Specifically, based on the learned edge mask, a high-contribution subgraph is extracted from the original power grid topology. The nodes and edges in this subgraph constitute the key impact paths of the cooperative drift. For example, the subgraph might contain an overloaded transformer, a heavily loaded line connected to it, and another sensor upstream of this line with abnormal measurement data. This intuitively tells maintenance personnel that the drift of the target sensor is likely the result of equipment malfunctions or operating conditions along this path. Simultaneously, based on the learned feature mask, it's possible to identify which specific features (such as "load rate," "equipment age," and "harmonic content") contribute most to the prediction results among the key nodes of the subgraph. This provides direct clues for in-depth analysis of the physical causes of the drift.

[0104] Furthermore, this invention also generates counterfactual explanations to provide actionable operational and maintenance (O&M) recommendations. This involves finding a minimal, realistic change to the current state that transforms the model's predicted results from "unsatisfactory" to "satisfactory." This "minimum change" directly corresponds to the lowest-cost and easiest-to-implement O&M recommendation.

[0105] It is important to note that if the drift identified by the model is mainly dominated by (intrinsic drift), that is, the equipment itself has undergone irreversible physical aging, then the counterfactual interpretation will no longer generate operation and maintenance adjustment suggestions, but will directly generate maintenance suggestions of "equipment replacement or repair" to avoid logical errors in the model.

[0106] Furthermore, this step constructs and solves a constrained optimization problem, assuming a certain node in the identified key influence subgraph... The original feature vector is (It is part of the fused feature vector generated in step S2). Our goal is to find a perturbation vector. , applied to Above, new feature vectors are formed. Meanwhile, this disturbance Represented as: Based on this disturbance The following conditions also need to be met: Shift in forecast results: in, It is a pre-trained spatiotemporal graph neural network model. It is a target sensor The set of acceptable states. For example, if the task is regression prediction, this condition is... ,in, It is the output threshold for regression prediction; if it is classification prediction, it is the predicted category as "normal".

[0107] Disturbance minimization: due to It is a cost function designed to measure the "magnitude" of the disturbance. Therefore, the L1 or L2 norm is usually chosen. In this embodiment, the L1 norm is chosen, and weights can be applied based on the adjustable difficulty of different features: in, It is an L1 norm, which encourages sparse modifications (i.e., changing only a few features). It is a diagonal weight matrix that assigns great weight to features that are difficult to change (such as equipment age) and smaller weight to features that can be adjusted (such as load factor).

[0108] Practical feasibility: The disturbance must conform to physical and operational constraints. For example, for the characteristic component representing the "load rate". The modified value must be within a reasonable range: in, These are the minimum and maximum allowable values ​​for this feature (e.g., load factor between 0 and 1.2).

[0109] Furthermore, due to the model Since it is differentiable, the above conditional optimization problem can be solved using gradient-based methods, such as iterative adjustment using the Lagrange multiplier method or projective gradient descent. This continues until a solution that satisfies all constraints and minimizes cost is found. .

[0110] Furthermore, the optimized solution yields... It is a numerical vector that needs to be translated into natural language instructions that operations and maintenance personnel can understand and execute.

[0111] For example, if If the value of the "active power" characteristic component corresponding to node "transformer T1" is -0.1 (normalized value), then after querying the normalized parameter of this characteristic, the system can generate the following suggestion: "Recommendation: Reduce the active power output of transformer T1 by 10MW, which is expected to restore the drift level of sensor A1 to the 'normal' state." For the "power flow" characteristic corresponding to "Line L2", it is recommended to adjust the power flow allocation of this line.

[0112] It should be noted that, through the above methods, the present invention can provide proactive and operable decision support for power grid operation and maintenance.

[0113] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0118] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A deep learning-based method for predicting the drift of a sensor throughout its entire lifecycle, characterized in that: include: Sensors and associated physical devices in the power grid are defined as different types of nodes, and the physical or electrical connections between these nodes are defined as different types of relationships. Based on real-time operating data and static ledger data, a time-stamped attributed heterogeneous information network sequence is generated. For each node in the network sequence, the multimodal data of the node is fused to generate the feature vector of the node; The attributed heterogeneous information network sequence is input into a preset spatiotemporal graph neural network model. The model captures the collaborative drift effect between nodes at each time step through a spatial information aggregation module, and learns the drift pattern of node features over time through a temporal evolution module. Based on the output of the spatiotemporal graph neural network model, the drift state of the target sensor at future time points is predicted.

2. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 1, characterized in that, The multimodal data includes static ledger data of primary equipment, dynamic measurement data of secondary equipment, and environmental and operating condition data.

3. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 1, characterized in that, By fusing the multimodal data of the nodes, a feature vector for each node is generated, including: Data from different modalities are embedded separately to obtain their respective modal feature vectors; The weight coefficients of each modality feature vector are calculated using an attention mechanism. These weight coefficients characterize the relative importance of each modality data in describing the current node state. The modal feature vectors are weighted and aggregated to generate the feature vectors of the nodes.

4. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 1, characterized in that, The spatial information aggregation module is a relation-aware graph attention network module. For each relation type existing in the heterogeneous information network sequence, the module learns independent information transformation parameters and attention weights, and aggregates information from neighboring nodes of different relation types in a differentiated manner.

5. The deep learning-based method for predicting the drift of a fusion sensor throughout its entire lifecycle as described in claim 1, characterized in that, The temporal evolution module is a gated cyclic unit network. When the gated unit in the gated cyclic unit network updates the hidden state of the current node, it not only integrates the current information of the node itself, but also integrates the historical hidden state information of its neighboring nodes to model the temporal transmission effect of the drift state in the topological network.

6. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 1, characterized in that, Before the output of the spatiotemporal graph neural network model, a step of decoupling the drift component is included, the step comprising: The output of the time-series evolution module is mapped to the first drift feature subspace and the second drift feature subspace, respectively; By adding an orthogonality penalty term to the total loss function of model training, the feature representations in the first drift feature subspace and the second drift feature subspace are made independent of each other, thereby quantifying the intrinsic drift caused by the aging of the equipment itself and the cooperative drift caused by the influence of the power grid environment.

7. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 6, characterized in that, The predicted drift state of the target sensor at future points in time includes: Based on the intrinsic drift and the cooperative drift, multi-task prediction is performed, which includes at least: regression prediction of sensor drift value, classification prediction of sensor drift level, and prediction of sensor remaining service life.

8. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 1, characterized in that, The model training process of the method employs a meta-learning framework, which includes: Data under different power grid topologies or operating conditions are divided into multiple learning tasks; Through the gradient optimization process of the inner and outer loops, a model initial parameter set is learned that can quickly adapt to new tasks using a small amount of new data. The learning tasks are divided based on at least one of the following: power grid topology change events, power grid operating mode switching events, and extreme weather events.

9. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 1, characterized in that, Following the prediction of the drift state of the target sensor at future time points, the method further includes a step of attributing the causes of the drift, the step of which includes: Based on the drift prediction results of the target sensor, a pre-defined graph interpretation algorithm is used to identify and extract a network subgraph and key node features that contribute the most to the prediction results, in order to locate the key sources and key influencing factors of cooperative drift.

10. The deep learning-based one-to-two sensor lifecycle drift prediction method as described in claim 9, characterized in that, Following the attribution step, a step of generating a counterfactual explanation is also included, the step comprising: An optimization algorithm is used to find a minimum perturbation to the node features, such that under the minimum perturbation, the drift prediction result of the spatiotemporal graph neural network model for the target sensor changes from an unsatisfactory state to a satisfactory state, and the minimum perturbation is used as an executable operation and maintenance control suggestion.

Images