Box-type substation multi-dimensional diagnosis method based on multi-network fusion and intelligent algorithm

Abstract

The invention relates to a box-type substation multi-dimensional diagnosis method based on multi-network fusion and an intelligent algorithm. The method comprises the steps that S1, a provided box-type substation monitoring system comprises a multi-network data acquisition module, a multi-network data fusion module, a composite fault diagnosis and prediction module and a health state quantitative evaluation module; s2, a multi-network data acquisition module obtains data of different dimensions of the box-type substation to obtain multi-network data; s3, the multi-network data fusion module carries out space-time alignment on the multi-network data, extracts cross-network features, calculates the weight of each cross-network feature and carries out fusion to obtain a fusion feature; s4, a composite fault diagnosis and prediction module synchronously diagnoses an electrical fault, a thermal fault and an insulation fault and predicts a composite fault based on the fusion features; and S5, outputting a multi-network health index, an equipment comprehensive health score, a fault positioning result and a maintenance suggestion by a health state quantitative evaluation module. According to the invention, accurate diagnosis and potential fault prediction of the composite fault of the box-type substation can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and particularly relates to a multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms. Background Art

[0002] A box-type substation is a complete set of power distribution and transformation equipment. It has a compact structure, and each box constitutes an independent system. The combination method is flexible and changeable, and can be freely combined according to actual situations to meet the needs of safe operation. It is applicable to places such as urban distribution networks, commercial centers, residential communities, and industrial parks. With the advancement of the construction of smart grids, as a core equipment of the distribution network, the intelligent monitoring demand for box-type substations has increased rapidly.

[0003] However, the monitoring methods of traditional box-type substations usually rely on a single sensor network to monitor specific parameters (such as temperature or current). They can only independently collect multi-dimensional data (conductance parameters, temperature rise characteristics, partial discharge signals, etc.), and lack cross-network association and collaborative analysis, which will form "data islands", making it impossible to grasp the overall operation status of the equipment, difficult to predict potential chain faults, resulting in a lag in operation and maintenance response, unable to predict the health status of the equipment in real time, and the operation and maintenance decision-making relying on manual experience. In addition, due to the insufficient sensitivity of the single-network model to complex faults (such as insulation aging accompanied by local overheating), misjudgment or missed judgment is likely to occur, resulting in low diagnostic accuracy. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms. This multi-dimensional diagnosis method for box-type substations can effectively fuse multi-dimensional data of box-type substations, construct a dynamic diagnosis model, and achieve accurate diagnosis of composite faults and prediction of potential faults of box-type substations. The adopted technical solutions are as follows: A multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms, characterized by including the following steps: S1. Provide a box-type substation monitoring system and set it in association with the box-type substation; wherein, the box-type substation monitoring system includes a multi-network data acquisition module, a multi-network data fusion module, a composite fault diagnosis and prediction module, and a health status quantitative evaluation module; S2. The multi-network data acquisition module obtains data of different dimensions of the box-type substation in real time to obtain multi-network data; S3. The multi-network data fusion module performs spatio-temporal alignment on the multi-network data and extracts cross-network features, and then calculates the weights of each cross-network feature through the GRU dynamic attention fusion algorithm and realizes the fusion of cross-network features to obtain fusion features; S4. Based on the fused features, the composite fault diagnosis and prediction module synchronously diagnoses electrical faults, thermal faults, and insulation faults through a multi-task deep learning model, and predicts composite faults using the cross-attention mechanism; S5. The health status quantitative evaluation module outputs the multi-network health index, the comprehensive equipment health score, the fault location result, and the maintenance suggestion in real time.

[0005] In the above multi-dimensional diagnosis method for box-type substations, in step S2, the multi-network data acquisition module can obtain data in different dimensions (such as electrical parameters, temperature, partial discharge signals, etc.) through their respective sensor deployments to obtain multi-network data, thereby achieving multi-dimensional monitoring and solving the blind area problem of traditional single-network monitoring. In step S3, the multi-network data fusion module can perform spatio-temporal alignment (such as timestamp synchronization, spatial calibration) on the multi-network data obtained in step S3 and extract cross-network features (such as leakage magnetic field strength gradient, temperature change rate, discharge phase concentration), and then use the GRU dynamic attention fusion algorithm to calculate the weights of each cross-network feature and fuse the features; based on the GRU dynamic attention fusion algorithm, the multi-network data fusion module can adaptively adjust the weights of each cross-network feature, solve the data complementarity problem, and the core purpose of fusing features is to achieve the deep collaboration of multi-network data. Through the dynamic collaboration and weight allocation of multi-dimensional features, it solves the problems of "data islands", "false judgment and missed judgment", and "response lag" in traditional single-network monitoring, realizes accurate diagnosis of composite faults, prediction of potential faults, and quantitative evaluation of health status, and supports intelligent operation and maintenance decision-making. In step S4, the composite fault diagnosis and prediction module can diagnose electrical faults, thermal faults, and insulation faults synchronously based on the fused features through a multi-task deep learning model. By combining dynamic attention fusion with a multi-task model, the accuracy of composite fault diagnosis can be significantly improved (measured ≥ 15%), reducing false judgment and missed judgment. The multi-task deep learning model can also introduce transfer learning technology to adapt to the monitoring requirements of different models of box-type substations; at the same time, it also uses the cross-attention mechanism to predict composite faults, predicts potential chain faults through the cross-attention mechanism, triggers the operation and maintenance response in advance, and reduces the equipment failure risk. In step S5, the multi-network health index is By a quantitative evaluation index generated after preprocessing the original data, feature extraction, and algorithm analysis on the multi-network data obtained in step S2 ； The health status quantitative evaluation module can output the multi-network health index, the comprehensive equipment health score, the fault location result, and the maintenance suggestion in real time. Thus, it can realize the real-time monitoring of the equipment operation status, accurate fault location, and quantitative health evaluation, support intelligent operation and maintenance decision-making, and by predicting faults in advance and optimizing the operation and maintenance plan, it can reduce the equipment downtime loss by 30% - 50%, reduce the maintenance cost by 20% - 30%, and increase the asset utilization rate by more than 15%.

[0006] As a preferred embodiment of the present invention, the multi-network data acquisition module in step S1 includes a conductance magnetization network unit, a temperature rise network unit, and a partial discharge network unit; in step S2, the conductance magnetization network unit collects three-phase current and voltage data in real time; the temperature rise network unit monitors the temperature distribution and thermal gradient in real time; the partial discharge network unit collects partial discharge signals and PRPD pattern features in real time. Specifically, the conductance magnetization network unit installs high-precision current transformers and voltage transformers on the high-voltage and low-voltage sides of the transformer to collect three-phase current and voltage data in real time; the temperature rise network unit deploys fiber optic temperature sensors (or thermocouple sensors), RFID wireless temperature sensors, and infrared thermal imagers at key parts such as windings, cable joints, and busbars to monitor the temperature distribution and thermal gradient in real time; the partial discharge network unit can deploy ultra-high frequency (UHF), transient earth voltage (TEV), and ultrasonic (AA) sensor arrays in the ring main unit and transformer room to collect partial discharge signals and PRPD pattern features, and can also combine acoustic emission sensors to improve the signal capture ability.

[0007] As a further preferred embodiment of the present invention, in step S3, the multi-network data fusion module performs spatio-temporal alignment on the multi-network data of the conductance magnetization network unit, the temperature rise network unit, and the partial discharge network unit based on the spatio-temporal alignment algorithm.

[0008] As a preferred embodiment of the present invention, in step S3, the multi-network data fusion module implements the GRU dynamic attention fusion algorithm based on the PyTorch framework, calculates the cross-network feature weights of each network, and fuses the cross-network features to form a fusion feature. The formula is: F fused =Σ iE(C,T,P) δ(W i ·h t-1 +b i )·F i , where C / T / P represents the multi-network features, δ is the Sigmoid function, h t-1 is the GRU hidden state, W i is the weight matrix, F i is the cross-network feature vector of the i-th network, and b i is the bias term. The multi-network data fusion module adopts the GRU dynamic attention fusion algorithm, which can adaptively adjust the multi-network feature weights and solve the data complementarity problem.

[0009] As a preferred embodiment of the present invention, in step S4, the multi-task deep learning model is constructed based on the TensorFlow framework, using CNN (Convolutional Neural Network) to identify electrical faults, LSTM (Long Short-Term Memory Network) to predict the trend of thermal faults, and GNN (Graph Neural Network) to classify insulation fault types, and sharing multi-network association information through a cross-attention mechanism. By combining the multi-task deep learning model (CNN / LSTM / GNN) with the cross-attention mechanism, the accuracy of composite fault diagnosis can be improved.

[0010] As a preferred embodiment of the present invention, in step S5, the health status quantitative evaluation module outputs the comprehensive health score of the device from 0 to 100 points by using the fuzzy comprehensive evaluation method. The fuzzy comprehensive evaluation method has the characteristics of clear results and strong systematicness, can better solve fuzzy and difficult-to-quantify problems, and is suitable for solving various non-deterministic problems.

[0011] Compared with the prior art, the present invention has the following advantages: (1) The box-type substation monitoring system in the present invention can deeply integrate conductance magnetization, temperature rise, and partial discharge network units, and through the conductance magnetization-temperature rise-partial discharge multi-network collaborative architecture, construct multi-dimensional monitoring of electrical, thermodynamic, and insulation faults, covering three fault dimensions of electrical faults, thermal faults, and insulation faults, breaking through the limitations of traditional single-network monitoring; (2) The present invention obtains data of different dimensions of the box-type substation in real time through the multi-network data acquisition module, eliminates data differences through the spatio-temporal alignment algorithm, uses the dynamic attention fusion algorithm to achieve adaptive complementarity of multi-network features, improves the data utilization efficiency, combines the multi-task deep learning model (CNN / LSTM / GNN) to synchronously diagnose electrical, thermal, and insulation faults, combines the dynamic attention fusion with the multi-task model, effectively fuses multi-network data such as conductance magnetization, temperature rise, and partial discharge, constructs a dynamic diagnosis model, realizes the accurate diagnosis of composite faults and the prediction of potential faults in the box-type substation, significantly improves the accuracy of composite fault diagnosis (measured ≥ 15%), and reduces misjudgment and missed judgment; (3) The present invention supports intelligent operation and maintenance decision-making by outputting the comprehensive health score of the device through the fuzzy comprehensive evaluation method through multi-source heterogeneous data collaborative analysis and intelligent algorithms, solves problems such as "data island", "fault misjudgment", and "response lag" of traditional single-network monitoring, and realizes real-time monitoring of the device operation state, accurate fault location, and health quantitative evaluation; (4) The present invention shares feature association information through multi-task learning and cross-attention mechanism to solve the problem of composite fault diagnosis, filling the gap in the field of complex fault prediction in the prior art; by predicting faults in advance and optimizing the operation and maintenance plan, the equipment downtime loss can be reduced by 30%-50%, the maintenance cost can be reduced by 20%-30%, and the asset utilization rate can be increased by more than 15%. Brief Description of the Drawings

[0012] Figure 1 is the overall architecture diagram of the box-type substation monitoring system provided by the preferred embodiment of the present invention; Figure 2 is the schematic flowchart of steps S2 - S3 of the multi-dimensional diagnosis method for box-type substations provided by the preferred embodiment of the present invention; Figure 3 is the schematic flowchart of step S4 of the multi-dimensional diagnosis method for box-type substations provided by the preferred embodiment of the present invention; Figure 4 is the schematic diagram of the process of the dynamic attention fusion algorithm in step S3 of the multi-dimensional diagnosis method for box-type substations provided by the preferred embodiment of the present invention. Detailed Description of the Preferred Embodiment

[0013] As Figures 1 - 4 shown, this multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms includes the following steps: S1. Provide a box-type substation monitoring system and set it associated with the box-type substation; the box-type substation monitoring system includes a multi-network data acquisition module 1, a multi-network data fusion module 2, a composite fault diagnosis and prediction module 3, and a health status quantitative evaluation module 4. The multi-network data acquisition module 1 includes a conductance magnetization network unit 11, a temperature rise network unit 12, and a partial discharge network unit 13; S2. The multi-network data acquisition module obtains data of different dimensions of the box-type substation in real time to obtain multi-network data: the conductance magnetization network unit 11 collects three-phase current and voltage data in real time; the temperature rise network unit 12 monitors the temperature distribution and thermal gradient in real time; the partial discharge network unit 13 collects partial discharge signals and PRPD pattern features in real time; S3. The multi-network data fusion module 2 performs spatio-temporal alignment (such as timestamp synchronization and spatial calibration) on the obtained multi-network data and extracts cross-network features 20 (such as leakage magnetic field intensity gradient, temperature change rate, discharge phase concentration), and then uses the GRU dynamic attention fusion algorithm (i.e., the dynamic attention fusion algorithm based on the GRU model) to calculate the weights of each cross-network feature 20 and fuse the cross-network features 20 to obtain the fused feature 21; S4. The composite fault diagnosis and prediction module 3 synchronously diagnoses electrical faults 32, thermal faults 33, and insulation faults 34 based on the fused feature 21 through a multi-task deep learning model 31, and uses the cross-attention mechanism to predict composite faults; S5. The health status quantification and evaluation module 4 outputs in real time the multi-network health index (such as conductance magnetization characteristics, temperature rise characteristics, partial discharge signal characteristics), the comprehensive health score of the equipment, the fault location result, and the optimized maintenance suggestion, and supports the operation and maintenance personnel to view through the PC side or the mobile side, so as to realize the automatic generation of fault early warning and maintenance plan.

[0014] In this embodiment, in step S2, the conductance magnetization network unit 11 collects the three-phase current and voltage data of the box-type substation in real time by installing high-precision current transformers and voltage transformers on the high-voltage and low-voltage sides of the transformer; the temperature rise network unit 12 deploys fiber optic temperature sensors (or thermocouple sensors), RFID wireless temperature sensors and infrared thermal imagers at key parts such as windings, cable joints, and busbars to monitor the temperature distribution and thermal gradient in real time; the partial discharge network unit 13 deploys an ultra-high frequency (UHF), transient earth voltage (TEV), and ultrasonic (AA) sensor array in the ring main unit and the transformer room to collect partial discharge signals and PRPD pattern characteristics, and can also combine acoustic emission sensors to improve the signal capture ability.

[0015] In this embodiment, in step S3, the multi-network data fusion module 2 performs spatio-temporal alignment on the multi-network data of the conductance magnetization network unit 11, the temperature rise network unit 12, and the partial discharge network unit 13 based on the spatio-temporal alignment algorithm.

[0016] The formula of the spatio-temporal alignment algorithm is as follows: (1) Layered DTW:

[0017] D(i,j,k): Three-dimensional cumulative distance matrix, storing the minimum alignment cumulative distance of the first i points of the conductance sequence C, the first j points of the temperature rise sequence T, and the first k points of the partial discharge sequence P; i,j,k: Sequence indices, corresponding to the conductance sequence C, the temperature rise sequence T, and the partial discharge sequence P respectively; C i ,T j ,P k : The i / j / k-th data points of each sequence, C is conductance, T is temperature rise, and P is partial discharge; d(C i ,T j ,P k ): Distance metric of three data points, in the form of weighted Manhattan distance.

[0018]

[0019] Among them, w C , w T , wP is the weight coefficient, which is used to adapt the weight of the fault characteristics of power equipment. Since the partial discharge signal is more sensitive to insulation anomalies, usually w is given P a higher weight; C ref , T ref , P ref are reference values, which can be determined according to the normal operation state or historical data of the equipment.

[0020] (2) Spatial calibration (ICP algorithm) P calib = R·P raw + t P raw : represents the original sensor coordinates, covering any position points of the temperature rise, partial discharge, and conductivity sensors; P calib represents the calibrated unified coordinates; R is a 3×3 orthogonal matrix, that is, a rotation matrix, and satisfies det(R)=1, which is used to describe the rotation relationship of the coordinate system; t is a 3D column vector, that is, a translation vector, which is used to describe the offset of the origin of the coordinate system.

[0021] In this embodiment, the multi-network data fusion module 2 implements the GRU dynamic attention fusion algorithm based on the PyTorch framework, calculates the weights of each cross-network feature 20 and fuses the feature 21, and its formula is: F fused = Σ iE(C,T,P) δ(W i · h t-1 + b i ) · F i , where C / T / P represents the multi-network feature, δ is the Sigmoid function, h t-1 is the GRU hidden state, W i is the weight matrix (used to map the hidden state of the previous moment of GRU (□ t-1 ) to the attention weight space of the current network feature, which is a parameter learned during model training), F i is the cross-network feature vector of the i-th network (representing the features extracted from the conductivity magnetization network (C), temperature rise network (T), and partial discharge network (P) (such as the leakage magnetic field strength gradient, temperature change rate, discharge phase concentration, etc.)), b i is the bias term (used to adjust the offset of the weight calculation and improve the model fitting ability, which is also a parameter learned during model training).

[0022] In this embodiment, in step S4, the multi-task deep learning model 31 is constructed based on the TensorFlow framework. The CNN (Convolutional Neural Network) is used to identify electrical faults 32, the LSTM (Long Short-Term Memory Network) is used to predict the trend of thermal faults 33, and the GNN (Graph Neural Network) is used to classify the types of insulation faults 34. The cross-attention mechanism is used to share the correlation information of multiple networks. By combining the multi-task deep learning model 31 (CNN / LSTM / GNN) with the cross-attention mechanism, the accuracy of composite fault diagnosis can be improved. The multi-task model can also introduce transfer learning technology to adapt to the monitoring requirements of different types of box-type substations.

[0023] In this embodiment, in step S5, the multi-network health index is a quantitative evaluation index generated by preprocessing the original data, extracting features, and analyzing algorithms for the multi-network data obtained in step S2.

[0024] In this embodiment, in step S5, the health status quantitative evaluation module 4 uses the fuzzy comprehensive evaluation method to output the comprehensive health score of the device from 0 to 100 points. The fuzzy comprehensive evaluation method has the characteristics of clear results and strong systematicness. It can better solve fuzzy and difficult-to-quantify problems and is suitable for solving various non-deterministic problems.

[0025] In addition, it should be noted that for the specific embodiments described in this specification, the names of their respective parts and the like can be different. Any equivalent or simple changes made according to the structure, features, and principles of the inventive concept of this invention patent are included in the protection scope of this invention patent. Those skilled in the art of this invention can make various modifications, supplements, or use similar methods to replace the specific embodiments described as long as they do not deviate from the structure of this invention or exceed the scope defined by this claim book, and they should all belong to the protection scope of this invention.

Claims

1. A multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms, characterized in that It includes the following steps: S1. Provide a box-type substation monitoring system and set it associated with the box-type substation. Among them, the box-type substation monitoring system includes a multi-network data acquisition module, a multi-network data fusion module, a composite fault diagnosis and prediction module, and a health status quantitative evaluation module; S2. The multi-network data acquisition module obtains data of different dimensions of the box-type substation in real time to obtain multi-network data; S3. The multi-network data fusion module performs spatio-temporal alignment on the multi-network data and extracts cross-network features, then calculates the weights of each cross-network feature through the GRU dynamic attention fusion algorithm and realizes the fusion of cross-network features to obtain fusion features; S4. The composite fault diagnosis and prediction module synchronously diagnoses electrical faults, thermal faults, and insulation faults based on the fusion features through a multi-task deep learning model, and predicts composite faults by using a cross-attention mechanism; S5. The health status quantitative evaluation module outputs a multi-network health index, an equipment comprehensive health score, a fault location result, and a maintenance suggestion in real time.

2. The multi-dimensional diagnosis method of the box-type substation based on multi-network fusion and intelligent algorithm according to claim 1, wherein: The multi-network data acquisition module in step S1 includes a conductance magnetization network unit, a temperature rise network unit, and a partial discharge network unit; in step S2, the conductance magnetization network unit collects three-phase current and voltage data in real time; the temperature rise network unit monitors the temperature distribution and thermal gradient in real time; the partial discharge network unit collects partial discharge signals and PRPD pattern features in real time.

3. A multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms according to claim 2, characterized in that: In step S3, the multi-network data fusion module performs spatio-temporal alignment on the multi-network data of the conductance magnetization network unit, the temperature rise network unit, and the partial discharge network unit based on the spatio-temporal alignment algorithm.

4. A multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms according to any one of claims 1-3, characterized in that: In step S3, the multi-network data fusion module implements the GRU dynamic attention fusion algorithm based on the PyTorch framework, calculates the cross-network feature weights, and fuses the cross-network features to form fused features. The formula is: F fused =Σ iE(C,T,P) δ(W i ·h t-1 +b i )·F i , where C / T / P represents the multi-network features, δ is the Sigmoid function, h t-1 is the GRU hidden state, W i is the weight matrix, F i is the cross-network feature vector of the i-th network, and b i is the bias term.

5. A multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms according to any one of claims 1-3, characterized in that: In step S4, the multi-task deep learning model is constructed based on the TensorFlow framework, uses CNN to identify electrical faults, LSTM to predict the trend of thermal faults, uses GNN to classify the types of insulation faults, and shares multi-network association information through a cross-attention mechanism.

6. A multi-dimensional diagnosis method for box-type substations based on multi-network fusion and intelligent algorithms according to any one of claims 1-3, characterized in that: In step S5, the health status quantitative evaluation module uses the fuzzy comprehensive evaluation method to output an equipment comprehensive health score of 0-100 points.

Images