CN-121980341-A - Deep learning-based full-life-cycle drift prediction method for two-integration sensor

CN121980341ACN 121980341 ACN121980341 ACN 121980341ACN-121980341-A

Abstract

The invention discloses a full life cycle drift prediction method of a two-integration sensor based on deep learning, which comprises the steps of modeling a power grid into an attribute heterogeneous information network sequence with a timestamp, deeply integrating multi-mode data such as primary equipment standing accounts, secondary equipment measurement and the like through a self-adaptive attention mechanism to generate node characteristics, inputting the sequence into a space-time diagram neural network model, capturing cooperative drift through a space aggregation module of relational perception, learning drift rules through a time sequence evolution module of topological perception, decoupling intrinsic drift and cooperative drift components through orthogonal constraint, executing multi-task prediction based on the decoupling characteristics, and outputting drift values, grades and residual life. The invention also adopts a meta learning framework to promote the adaptability of the model to the dynamic topology, and provides drift attribution and operation and maintenance suggestions through a graph explanatory algorithm. The method can accurately predict the sensor drift, realize cause deep analysis and provide powerful support for predictive maintenance of the intelligent power grid.

Inventors

KONG DONGBO
GUO KAI
WANG FEIFEI
WEN JIANQIANG
TANG GUOPING
TAN MINGYI

Assignees

江苏丹通电气有限公司

Dates

Publication Date: 20260505
Application Date: 20260109

Claims (10)

1. The full life cycle drift prediction method of the two-integration sensor based on deep learning is characterized by comprising the following steps of: Defining sensors and associated physical devices in a power grid as nodes of different types, defining physical or electrical connection among the nodes as relations of different types, and generating an attribute heterogeneous information network sequence with a time stamp according to real-time operation data and static ledger data; Aiming at each node in the network sequence, fusing the multi-mode data of the node to generate a feature vector of the node; inputting the attribution heterogeneous information network sequence into a preset space-time diagram neural network model, capturing cooperative drift influence among nodes in each time step by the model through a space information aggregation module, and learning drift rules of node characteristics along with time change through a time sequence evolution module; based on the output of the space-time diagram neural network model, a drift state of the target sensor at a future point in time is predicted.
2. The deep learning-based full life cycle drift prediction method of a two-component sensor of claim 1, wherein the multi-modal data comprises primary device static ledger data, secondary device dynamic measurement data, and environmental and operating condition data.
3. The deep learning-based two-fusion sensor full life cycle drift prediction method of claim 1, wherein fusing the multi-modal data of the node to generate the feature vector of the node comprises: Respectively embedding data of different modes to obtain respective mode feature vectors; Calculating weight coefficients of the modal feature vectors through an attention mechanism, wherein the weight coefficients represent the relative importance of the modal data for describing the current node state; And carrying out weighted aggregation on the modal feature vectors to generate feature vectors of the nodes.
4. The deep learning-based two-fusion sensor full life cycle drift prediction method of claim 1, wherein the spatial information aggregation module is a relational-aware graph-annotation-force network module that learns independent information transformation parameters and attention weights for each relationship type existing in the heterogeneous information network sequence, differentially aggregates information from neighboring nodes of different relationship types.
5. The deep learning-based full life cycle drift prediction method of a two-fusion sensor according to claim 1, wherein the time sequence evolution module is a gating circulation unit network, and when a gating unit in the gating circulation unit network updates the hidden state of a current node, not only the current information of the node itself but also the historical hidden state information of neighboring nodes thereof are fused, and the time sequence transfer effect of the drift state in the topology network is modeled.
6. The deep learning based two-fused sensor full life cycle drift prediction method of claim 1, further comprising a step of decoupling the drift component prior to the output of the space-time neural network model, said step comprising: mapping the output of the time sequence evolution module to a first drift characteristic subspace and a second drift characteristic 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 independent of each other, and intrinsic drift caused by equipment aging and cooperative drift caused by the influence of the power grid environment are respectively quantified.
7. The deep learning-based two-fused sensor full lifecycle drift prediction method of claim 6, wherein predicting the drift state of the target sensor at a future point in time comprises: Based on the intrinsic drift and the collaborative drift, a multi-tasking prediction is performed, the multi-tasking prediction comprising at least a regression prediction of sensor drift values, a classification prediction of sensor drift levels, and a prediction of remaining useful life of the sensor.
8. The deep learning-based two-fused sensor full life cycle drift prediction method of claim 1, wherein a model training process of the method employs a meta-learning framework, the framework comprising: Dividing data under different power grid topological structures or operation conditions into a plurality of learning tasks; Learning a model initial parameter set which is quickly adapted to a new task by using a small amount of new data through the gradient optimization process of the inner loop and the outer loop; the division basis of the learning tasks comprises at least one of a power grid topology change event, a power grid operation condition mode switching event and an extreme weather event.
9. The deep learning based two-fused sensor full life cycle drift prediction method of claim 1, further comprising the step of associatively attributing a drift cause after the predicted target sensor's drift state at a future point in time, said step comprising: aiming at a drift prediction result of a target sensor, identifying and extracting a network subgraph and key node characteristics which have the greatest contribution to the prediction result through a preset graph interpretation algorithm, and positioning key influence sources and key influence factors of cooperative drift.
10. The deep learning based two-fused sensor full lifecycle drift prediction method of claim 9, further comprising the step of generating a negative fact interpretation after the relevance attributing step, the step comprising: And searching a minimum disturbance to the node characteristics through an optimization algorithm, so that under the minimum disturbance, a drift prediction result of the space-time diagram neural network model to the target sensor is converted from an unqualified state to a qualified state, and the minimum disturbance is used as an executable operation and maintenance regulation suggestion.

Description

Deep learning-based full-life-cycle drift prediction method for two-integration sensor Technical Field The invention relates to the technical field of computer data processing and power system monitoring, in particular to a full life cycle drift prediction method of a two-component fusion sensor based on deep learning. Background With the rapid development of smart power grids, the mass sensor is used as a 'nerve ending' of the power grid, and the accuracy of measurement data of the mass sensor is a foundation stone for realizing accurate situation awareness, advanced state estimation and safe and stable operation of the power grid. However, in the long-term service process, the sensor is affected by multiple stresses such as equipment aging, environmental factors, running condition changes and the like, and the measurement performance of the sensor is inevitably degraded slowly and continuously, namely, the performance drifts. Aiming at the problem, the prior art mainly adopts a method based on a traditional statistical model (such as an autoregressive integral moving average model ARIMA, kalman filtering and the like) or a time sequence deep learning model (such as a long and short term memory network LSTM, a gating circulating unit GRU and the like) to predict. The method can capture the nonlinear time sequence evolution rule of the single sensor drift to a certain extent, and provides a certain technical support for state maintenance. However, the prior art has significant drawbacks when applied to complex grid systems, in that firstly, the prior art generally regards the sensor as an individual for independent analysis, ignoring the strong spatial coupling effect involved in the grid topology. In a power grid with tight physical connection, the drift of a single sensor can influence the measurement performance of adjacent nodes through tide conduction, a complex 'cooperative drift' phenomenon is formed, the space-time correlation cannot be effectively modeled by the existing method, and misjudgment of a drift source is easily caused. Secondly, the prior art has limitation on data utilization, and depends on secondary electric measurement values of sensors, so that multi-mode heterogeneous information such as primary equipment ledgers, historical operation and maintenance records and the like capable of representing the physical health state of equipment cannot be effectively fused, and the full life cycle state of the sensors is not fully described. Thirdly, the existing prediction model generally outputs a general drift result, the drift causes cannot be deeply decoupled, and whether the drift caused by the intrinsic performance degradation of equipment or apparent drift caused by the network synergistic effect is difficult to distinguish, so that the accurate establishment of an operation and maintenance strategy is not facilitated. Fourth, the topology of the power grid is not static, the operation of line switching, maintenance and the like can cause dynamic change of the network structure, the generalization capability of the existing model is insufficient, the prediction precision is suddenly reduced once the topology is changed, and the rapid self-adaption capability to the dynamic topology is lacking. Disclosure of Invention This section is intended to outline some aspects of embodiments of the application and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section as well as in the description of the application and in the title of the application, which may not be used to limit the scope of the application. The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides a full life cycle drift prediction method of a two-integration sensor based on deep learning, which is used for solving the problems in the background technology. In order to solve the technical problems, the invention provides the following technical scheme that the full life cycle drift prediction method of the two-integration sensor based on deep learning comprises the following steps: Defining sensors and associated physical devices in a power grid as nodes of different types, defining physical or electrical connection among the nodes as relations of different types, and generating an attribute heterogeneous information network sequence with a time stamp according to real-time operation data and static ledger data; Aiming at each node in the network sequence, fusing the multi-mode data of the node to generate a feature vector of the node; inputting the attribution heterogeneous information network sequence into a preset space-time diagram neural network model, capturing cooperative drift influence among nodes in each time step by the model through a space information aggregation module, and learning drift rules of node characteristics along with time change through a time sequence evolution modul