GIS equipment state evaluation method based on cloud station collaboration

Abstract

The invention discloses a GIS equipment state evaluation method based on cloud station collaboration, and belongs to the technical field of high-voltage electrical equipment state evaluation. Comprising the following steps: constructing a cloud sample library according to data collected by different sensors at a station end; constructing a physical field model related to GIS equipment operation at the cloud and performing order reduction calculation; inputting the collected real-time data into a virtual digital twinborn model to predict the complex behavior of the GIS equipment; obtaining a differentiation evaluation threshold value, and performing differentiation evaluation on each GIS device according to a real-time data optimization clustering result; extracting a situation quantity capable of representing a state quantity trend and a state trend at the same time, and predicting the situation quantity of the GIS equipment by referring to a differentiation evaluation result; a sample library and traditional expert experience are combined at a station end, the state of GIS equipment is evaluated, and equipment scheduling is optimized. According to the invention, the intelligent control and efficient operation of the power grid are improved through sharing decision information and cooperative work of the cloud end and the station end.

Description

Technical Field

[0001] The present invention relates to the technical field of high-voltage electrical equipment condition assessment, and particularly to a GIS equipment condition assessment method based on cloud-station collaboration. Background Art

[0002] With the development of China's smart grid, the modular construction of smart substations has become a trend. Coupled with factors such as grid transformation requirements and land shortages, smart modular substations with GIS equipment as the core have become the mainstream direction. Due to the special operating conditions of GIS equipment, when the equipment condition shows a deteriorating trend, it is impossible to carry out power-off maintenance in a timely manner. Therefore, various sensing devices need to be installed during the transportation of electrical equipment to obtain various state variables that reflect the equipment operation in real time, and based on this, corresponding evaluation models are established to timely grasp the operation status of the equipment.

[0003] However, in current practical applications, there are still problems with the poor quality of equipment state variable data. Due to factors such as the poor stability of sensing devices, harsh on-site operating environments, and complex electromagnetic environments, there are a large number of abnormal data and noises in the state variable data that reflect the equipment operation. In addition, existing condition assessment models usually rely on expert experience or general algorithms, ignoring the personalized characteristics of equipment, resulting in weak model expression ability and poor generalization ability. At the same time, traditional prediction methods cannot effectively characterize the time-series characteristics of GIS equipment when predicting the equipment condition, resulting in low prediction accuracy and difficulty in achieving accurate condition assessment.

[0004] Most of the existing technologies conduct digital modeling on the fault diagnosis of a certain type of equipment in a substation, and there are few models that simultaneously model multiple devices in a substation. Moreover, the concept of digital twin models is not introduced, and the application scope is relatively narrow. These models cannot realize real-time perception of the equipment condition in a GIS smart substation, cannot use fault diagnosis technology to judge the equipment condition, and the calculated results have little reference significance and need further conversion when guiding practical work. Now the on-site detection of GIS disconnectors has gradually developed from off-line monitoring to on-line monitoring. Therefore, it is particularly important to evaluate the condition of GIS equipment based on the collaborative work of the cloud and the station end, combined with digital twin technology. Summary of the Invention

[0005] The object of the present invention is to propose a GIS equipment condition assessment method based on cloud-station collaboration, including:

[0006] Constructing a cloud sample library according to the data collected by different sensors at the station end;

[0007] Constructing a physical field model related to the operation of GIS equipment in the cloud and performing reduced-order calculation;

[0008] Build a virtual digital twin model and input the collected real-time data into the virtual digital twin model to predict the complex behaviors of GIS devices;

[0009] Obtain the differential evaluation threshold, compare the real-time data with the complex behaviors predicted by the digital twin model, and conduct differential evaluation on each GIS device according to the optimized clustering results of the real-time data;

[0010] Extract the situation quantities that can simultaneously characterize the trends of state variables and state trends, and predict the situation quantities of GIS devices with reference to the differential evaluation results;

[0011] Evaluate the status of GIS devices and optimize device scheduling at the substation terminal by combining the sample library and traditional expert experience.

[0012] Furthermore, the data collected by different sensors at the substation terminal include the intrinsic data of current, voltage, temperature, and material parameters, as well as the diagnostic data of vibration, partial discharge, infrared images, and ultraviolet images of GIS devices.

[0013] Furthermore, the specific steps for obtaining the differential evaluation threshold, comparing the real-time data with the complex behaviors predicted by the digital twin model, and conducting differential evaluation on each GIS device according to the optimized clustering results of the real-time data are as follows:

[0014] Analyze the influencing factors of the digital twins of the group of devices, extract the optimal influencing factors that can best reflect the differences between devices, and conduct differential partitioning on the digital twins of the group of devices according to the optimal influencing factors to obtain the differential evaluation threshold for a single device;

[0015] Standardize the real-time data to ensure that the data can be compared on the same scale;

[0016] Use the random search clustering algorithm CLARANS for local search to optimize clustering. During each search, calculate the Mahalanobis distance between each cluster center and the current sample, and select the sample with the minimum distance to join the corresponding cluster;

[0017] Compare the real-time data with the cluster centers of the group of GIS devices, and evaluate the deviation of the device status through the Mahalanobis distance. If the Mahalanobis distance exceeds the differential evaluation threshold, it is considered that the behavior of the device significantly deviates from the group and further monitoring or maintenance is required.

[0018] Furthermore, the calculation formula for the Mahalanobis distance is:

[0019]

[0020] In the formula, x is the sample point to be measured, μ is the mean vector of the cluster center, Σ-1 is the inverse of the covariance matrix of the data.

[0021] Furthermore, extract the situation quantities that can simultaneously characterize the trends of state quantities and state trends, and predict the situation quantities of GIS equipment with reference to the differential evaluation results. The specific steps are as follows:

[0022] Considering the importance of state quantities at different times and the correlation relationships between state quantities in different stages in the input state quantities, establish an attention model. For each output data s i , i = 1, 2, 3, when generating each data, replace the original same intermediate code C with the continuously changing C according to the current input i , i = 1, 2, 3, each output is:

[0023] s1 = G(C1)

[0024] s2 = G(C2, s1)

[0025] s3 = G(C3, s1, s2)

[0026] In the formula, C i is set as the attention distribution probability distribution of different state quantity data, and the calculation is as follows:

[0027] C1 = Q′(ω 1,1 ×F′(x1), ω 1,2 ×F′(x2), ω 1,3 ×F′(x3), ω 1,4 ×F′(x4))

[0028] C2 = Q′(ω 2,1 ×F′(x1), ω 2,2 ×F′(x2), ω 2,3 ×F′(x3), ω 2,1 ×F′(x4))

[0029] C3 = Q′(ω 3,1 ×F′(x1), ω 3,2 ×F′(x2), ω 3,3 ×F′(x3), ω 3,4 ×F′(x4))

[0030] In the formula, the function F'(x) represents a certain transformation function of the Encoder for the input state quantity, and Q'(x) represents the transformation function of the intermediate code. ω ij represents the attention distribution coefficient of the jth state quantity in the input state quantity when obtaining the ith state s i ;

[0031] The input process of the timing relationship C(t) is carried out simultaneously with the association relationship A(t), the fuzzy relationship F(t), the mapping relationship M(t), and the expert experience G(t), and they jointly participate in the generation process of each element in {y1,…,y t-1 ,y t}, where y t represents the output of the t-th output layer, that is, the state of the GIS device at time t;

[0032] The attention encoding C i is dimensionally reduced according to the following formula:

[0033]

[0034] In the formula, α s i,j is the degree of attention to the dependence relationship, and are the final encoded vectors of the hidden layer sequences of the timing relationship, association relationship, fuzzy relationship, and mapping relationship respectively, and e g j is the dimensionality reduction result of the expert experience Embedding;

[0035] A single-layer perceptron is used according to the following formula:

[0036]

[0037] In the formula, e i,j is the original attention degree before normalization, and V e , W e , U s are all weight matrices obtained by training the single-layer perceptron, and S i-1 is the output result of the hidden layer at time i - 1;

[0038] Softmax is used for normalization:

[0039]

[0040] In the formula, t is the total number of inputs of any relationship input layer;

[0041] Finally, according to the prediction structure, different relationships are segmented to obtain the prediction of the situation quantity S(t) at the same time node t.

[0042] The beneficial effects of the present invention are as follows:

[0043] When predicting the device state, the present invention can effectively characterize the timing characteristics of the GIS device and achieve accurate state assessment. By sharing decision-making information, the cloud side and the station side work together to improve the intelligent control and efficient operation of the power grid. Description of the Drawings

[0044] Figure 1 Schematic diagram of the GIS equipment status evaluation method based on cloud-station collaboration of the present invention;

[0045] Figure 2 Flowchart of the GIS equipment status evaluation method based on cloud-station collaboration of the present invention. Specific implementation manner

[0046] The present invention proposes a GIS equipment status evaluation method based on cloud-station collaboration. The following further describes the present invention with reference to the accompanying drawings and specific embodiments.

[0047] Figure 1 、 Figure 2 Are respectively the schematic diagram and flowchart of the GIS equipment status evaluation method based on cloud-station collaboration of the present invention, including the following steps:

[0048] Step 1: Construct a sample library of status variables of GIS equipment.

[0049] Make the intrinsic and diagnostic data measured at the station end into a table and fuse them to form a unified format data information library. After the data is unified in format, the prepared data processing program can be used to read and process the data by setting the corresponding retrieval commands according to requirements.

[0050] The station end data includes intrinsic data such as current, voltage, temperature, material parameters, etc., and diagnostic data such as vibration, partial discharge, infrared image, ultraviolet image, etc. of GIS equipment. It is classified differently according to the voltage level of the GIS switch unit gas chamber. At the same time, considering the influence of factors such as on-site sensor failure and noise interference, the curve fitting method of weighted least squares is used to statistically analyze the data and construct a sample library according to the past equipment status.

[0051] The electrical behavior of GIS equipment has complex non-linear characteristics. The selected model form is:

[0052] y i = f(x i ,θ) = θ0 + θ1x i + θ2x i 2 + ε

[0053] In the formula, y i is the observed value of the i-th data point; x i is the corresponding independent variable, f(x i ,θ) is the electrical characteristic of the parameter, θ0, θ1, θ2 are the model parameters to be fitted, and ε is the error term.

[0054] In the non-linear model, the objective function form of weighted least squares is:

[0055]

[0056] where w i is the weight of the i-th data point.

[0057] For data points with large noise, smaller weights are assigned to reduce their influence on the fitting result; for reliable data points, larger weights are assigned.

[0058] Step 2: Perform accurate modeling and reduced-order calculation of the multi-physical field model.

[0059] First, based on the device data collected at the station end, construct a physical field model related to the operation of GIS equipment, generate the corresponding mesh model, define the geometric structure of GIS equipment in the simulation software, set appropriate physical field conditions, and at the same time configure the parameters and probes of the model, and set the input and output of the control matrix.

[0060] Then, use finite element analysis for simulation to obtain the process matrix. Next, through the LiveLink function of MATLAB, export the process matrix data obtained from the simulation, input the exported matrix data into the state space equation calculation toolbox of MATLAB, and use this toolbox for reduced-order calculation.

[0061] In the modeling of material properties, special attention is paid to the influence of environmental factors on electrical properties and thermal behavior. Taking conductivity as an example, the conductivity of the material is not only affected by temperature, but may also be affected by factors such as humidity and pressure. Considering these influences, the change in conductivity can be expressed as:

[0062]

[0063] where σ0 is the conductivity constant, E a is the activation energy, k B is the Boltzmann constant, T is the temperature, P is the pressure, H is the humidity, and α and β are the influence coefficients of pressure and humidity on conductivity respectively.

[0064] On this basis, further consider the electric field distribution and heat conduction characteristics during the operation of the equipment. In terms of thermal characteristics, the temperature field distribution of the equipment in the working state is:

[0065]

[0066] where α T is the thermal diffusivity, Q is the heat source generated inside the equipment, ρ is the density of the material, and c is the specific heat capacity. By coupling with the electric field distribution and current density equation, the interaction simulation between the temperature field and the electric field can be realized.

[0067] Step 3: Use the digital twin model to predict the complex behaviors of GIS devices.

[0068] By simulating the dynamic behaviors of GIS devices such as electrical characteristics, mechanical characteristics, temperature, and pressure, a virtual digital twin model is constructed, and a reference model for group behaviors is established through historical data, design parameters, and the behaviors of typical devices.

[0069] Create the computational domain:

[0070] mesh = UnitSquareMesh(32, 32)

[0071] Define the finite element space:

[0072] V = FunctionSpace(mesh, 'P', 1) / / Electric potential field space

[0073] Q = FunctionSpace(mesh, 'P', 1) / / Temperature field space

[0074] Define the boundary conditions:

[0075] u_D = Expression('sin(pi*x[0])*sin(pi*x[1])', degree = 2)

[0076] bc = DirichletBC(V, u_D, 'on_boundary')

[0077] Define the trial function and test function:

[0078] u = TrialFunction(V)

[0079] v = TestFunction(V)

[0080] Define the electrical equation (Poisson equation):

[0081] epsilon = Constant(1.0) / / Assume the relative permittivity is 1

[0082] rho = Constant(1.0) / / Assume the charge density is 1

[0083] a_electric = epsilon * dot(grad(u), grad(v)) * dx

[0084] L_electric = -rho * v * dx

[0085] Solve the electrical equation:

[0086] A_electric,b_electric = assemble_system(a_electric,L_electric)

[0087] u_solution = Function(V)

[0088] solve(A_electric,u_solution.vector(),b_electric)

[0089] Define the temperature field:

[0090] T = TrialFunction(Q)

[0091] v_temp = TestFunction(Q)

[0092] Assume the thermal conductivity and heat source:

[0093] k = Constant(1.0) / / Thermal conductivity

[0094] Q_source = Constant(1.0) / / Heat source term (can be calculated based on current loss)

[0095] Heat conduction equation:

[0096] a_thermal = k * dot(grad(T), grad(v_temp)) * dx

[0097] L_thermal = Q_source * v_temp * dx

[0098] Solve the heat equation:

[0099] A_thermal,b_thermal = assemble_system(a_thermal,L_thermal)

[0100] T_solution = Function(Q)

[0101] solve(A_thermal,T_solution.vector(),b_thermal)

[0102] Visualize the results:

[0103] import matplotlib.pyplot as plt

[0104] plot(u_solution)

[0105] plt.title("Electric Potential")

[0106] plt.show()

[0107] plot(T_solution)

[0108] plt.title("Temperature Distribution")

[0109] plt.show()

[0110] Step 4: Compare the actual data with the behavior predicted by the digital twin model, evaluate the difference between the current device state and the expected state, and obtain the differential evaluation result of the GIS device.

[0111] For all GIS devices within a specific range, establish a digital twin model of all devices, and select all possible influencing factors to divide the digital twin. By analyzing the influencing factors of the digital twin of the group of devices, extract the optimal influencing factors that can best reflect the differences between devices, and based on these optimal influencing factors, divide the digital twin of the group of devices differently, so as to obtain the differential evaluation threshold for a single device.

[0112] Subsequently, compare the monitoring data obtained by the device in real time with the corresponding differential threshold, and interact with the digital twin of the device in real time, so as to obtain the differential evaluation result. The CLARANS method of the random search clustering algorithm is adopted, and the Mahalanobis Distance is used as the distance metric to overcome the differences in aspects such as the change scales of various gases.

[0113] For each data sample, the calculation formula of the Mahalanobis distance is:

[0114]

[0115] In the formula, x is the sample point to be measured, μ is the mean vector of the cluster center, and Σ -1 is the inverse of the covariance matrix of the data.

[0116] Standardize the monitoring data of each influencing factor to ensure that the data is compared on the same scale. Each time a search is made, calculate the Mahalanobis distance between each cluster center and the current sample, and select the sample with the minimum distance to join the corresponding cluster.

[0117] Based on the clustering results, differential evaluation can be performed on each GIS device. The monitoring data of the GIS device obtained in real time is compared with the clustering center of the group of GIS devices, and the deviation of the device status is evaluated by calculating the Mahalanobis distance.

[0118] Step 5: According to the GIS device status variables predicted based on the digital twin model and their differential results, further predict the change of the situation variables.

[0119] Prediction process: Prediction of the time series relationship C(t), correlation relationship A(t), fuzzy relationship F(t), mapping relationship M(t), and verification of the expert experience G(t), and the prediction process is constrained based on the verification results.

[0120] Considering the importance of the status variables at different times and the correlation relationship between the status variables at different stages in the input status variables, an "attention model" is established. In the attention model, for each output data s i , i = 1, 2, 3, their corresponding intermediate encodings are all different, that is, when generating each data, the original same intermediate encoding C is replaced by C that changes continuously according to the current input i , i = 1, 2, 3. Each output is:

[0121] s1 = G(C1)

[0122] s2 = G(C2, s1)

[0123] s3 = G(C3, s1, s2)

[0124] In the formula, C i can be set as the attention distribution probability distribution of different status variable data, and the calculation method is:

[0125] C1 = Q′(ω 1,1 ×F′(x1), ω 1,2 ×F′(x2), ω 1,3 ×F′(x3), ω 1,4 ×F′(x4))

[0126] C2 = Q′(ω 2,1 ×F′(x1), ω 2,2 ×F′(x2), ω 2,3 ×F′(x3), ω 2,1 ×F′(x4))

[0127] C3 = Q′(ω 3,1 ×F′(x1), ω 3,2 ×F′(x2), ω 3,3 ×F′(x3), ω 3,4 ×F′(x4))

[0128] In the formula, the function F'(x) represents a certain transformation function of the Encoder for the input state quantity, and Q'(x) represents the transformation function of the intermediate encoding, ω ij represents the attention distribution coefficient of the j-th state quantity in the input state quantity when obtaining the i-th state s i .

[0129] The input process of the state quantity C(t) is simultaneous with A(t), M(t), F(t), and G(t). Therefore, they jointly participate in the generation process of each element in {y1,…,y t-1 ,y t}, where y t represents the output of the t-th output layer, that is, the state of the GIS device at time t. The encoding C i of attention is different from the previous encoding process. Among them, the time sequence relationship C(t), the association relationship A(t), the fuzzy relationship F(t), and the mapping relationship M(t) need to be encoded, while the expert experience G(t) is a time-invariant quantity. Therefore, only dimensionality reduction processing is required, and its calculation formula is:

[0130]

[0131] In the formula, α s i,j is the degree of attention to the dependence relationship, that is, the attention degree, and are the final encoded vectors of the hidden layer sequences of the time sequence relationship, the association relationship, the fuzzy relationship, and the mapping relationship, respectively, obtained through matrix splicing. e g j is the dimensionality reduction result of the expert experience Embedding.

[0132] If only a single-layer perceptron is used, the calculation formula is:

[0133]

[0134] In the formula, e i,j is the original attention degree before normalization, V e , W e , U s are all weight matrices obtained by training the single-layer perceptron. S i-1 is the output result of the hidden layer at the previous moment (the (i - 1)-th moment).

[0135] After normalization using softmax, it is:

[0136]

[0137] Wherein, t is the total number of inputs to any relationship input layer. According to different relationships predicted by the prediction structure, the prediction process of the situation quantity S(t) at the same time node t is obtained.

[0138] Step 6: According to the results of the prediction of the situation quantity of the GIS device, combined with the sample library and traditional expert experience, evaluate the possible operating states of the GIS device and make corresponding processing strategies.

[0139] In summary, when the present invention performs equipment state prediction, it can effectively characterize the time series characteristics of the GIS device and achieve accurate state evaluation. By sharing decision-making information, the cloud side and the station side work together to improve the intelligent control and efficient operation of the power grid.

Claims

1. A GIS device status evaluation method based on cloud-station collaboration, characterized in that, Including: Construct a cloud sample library based on data collected by different sensors at the station end; Construct a physical field model related to the operation of GIS devices in the cloud and perform reduced-order calculations; Construct a virtual digital twin model and input the collected real-time data into the virtual digital twin model to predict the complex behaviors of GIS devices; Obtain a differential evaluation threshold, compare the real-time data with the complex behaviors predicted by the digital twin model, and perform differential evaluation on each GIS device according to the results of optimizing clustering based on the real-time data; Extract situation quantities that can simultaneously characterize the trends of state quantities and state trends, and predict the situation quantities of GIS devices with reference to the differential evaluation results; Combine the sample library and traditional expert experience at the station end to evaluate the state of GIS devices and optimize device scheduling.

2. The GIS device status evaluation method based on cloud-station collaboration according to claim 1, characterized in that Data collected by different sensors at the station end include intrinsic data such as current, voltage, temperature, and material parameters, as well as diagnostic data such as vibration, partial discharge, infrared images, and ultraviolet images of GIS devices.

3. The GIS device status evaluation method based on cloud-station collaboration according to claim 1, wherein The specific steps for obtaining a differential evaluation threshold, comparing the real-time data with the complex behaviors predicted by the digital twin model, and performing differential evaluation on each GIS device according to the results of optimizing clustering based on the real-time data are as follows: Analyze the influencing factors of the digital twins of group devices, extract the optimal influencing factors that can best reflect the differences between devices, and perform differential partitioning on the digital twins of group devices according to the optimal influencing factors to obtain a differential evaluation threshold for a single device; Perform standardization processing on the real-time data to ensure that the data is compared on the same scale; Use the random search clustering algorithm CLARANS for local search to optimize clustering. During each search, calculate the Mahalanobis distance between each cluster center and the current sample, and select the sample with the minimum distance to join the corresponding cluster; Compare the real-time data with the cluster centers of group GIS devices, and evaluate the deviation of the device state through the Mahalanobis distance. If the Mahalanobis distance exceeds the differential evaluation threshold, it is considered that the behavior of the device significantly deviates from the group and further monitoring or maintenance is required.

4. The GIS device status evaluation method based on cloud-station collaboration according to claim 3, characterized in that, The calculation formula of the Mahalanobis distance is: where x is the sample point to be measured, μ is the mean vector of the cluster centers, and Σ -1 is the inverse of the covariance matrix of the data.

5. The GIS device status evaluation method based on cloud-station collaboration according to claim 1, characterized in that The specific steps for extracting situation quantities that can simultaneously characterize the trends of state quantities and state trends, and predicting the situation quantities of GIS devices with reference to the differential evaluation results are as follows: Considering the importance of state variables at different times and the correlation between state variables in different stages in the input state variables, an attention model is established. For each output data s i , where i = 1, 2, 3, when generating each data, the original same intermediate encoding C is replaced with an attention encoding C that changes continuously according to the current input i , where i = 1, 2, 3, each output is: s1 = G(C1) s2 = G(C2, s1) s3 = G(C3, s1, s2) where C i is set as the attention allocation probability distribution of different state quantity data and is calculated as follows: C1 = Q′(ω 1,1 ×F′(x1), ω 1,2 ×F′(x2), ω 1,3 ×F′(x3), ω 1,4 ×F′(x4)) C2 = Q′(ω 2,1 ×F′(x1), ω 2,2 ×F′(x2), ω 2,3 ×F′(x3), ω 2,1 ×F′(x4)) C3 = Q′(ω 3,1 × F′(x1), ω 3,2 × F′(x2), ω 3,3 × F′(x3), ω 3,4 × F′(x4)) wherein, the function F'(x) represents a certain transformation function of the Encoder for the input state quantity, and Q'(x) represents the transformation function of the intermediate encoding, ω ij represents the attention allocation coefficient of the j-th state quantity in the input state quantity when obtaining the i-th state s i ; The input process of the timing relationship C(t) is carried out simultaneously with the association relationship A(t), the fuzzy relationship F(t), the mapping relationship M(t), and the expert experience G(t), and they jointly participate in the generation process of each element in {y1,…,y t-1 ,y t}, where y t represents the output of the t-th output layer, that is, the state of the GIS device at time t; Perform dimensionality reduction on the attention encoding C i according to the following formula: where α s i,j is the degree of attention to the dependency relationship, h c j 、h a j 、h m j and h f j are the final encoded vectors of the hidden layer sequences of the temporal relationship, association relationship, fuzzy relationship, and mapping relationship respectively, and e g j is the dimensionality reduction result of the expert experience Embedding; Use a single-layer perceptron according to the following formula: where, e i,j is the original attention before normalization, V e , W e , U s are all weight matrices, obtained by training a single-layer perceptron, and S i-1 is the output result of the hidden layer at the (i - 1)th moment; Use softmax for normalization: In the formula, t is the total number of inputs of any relationship input layer; Finally, perform segmentation according to different relationships of the prediction structure to obtain the prediction of the situation quantity S(t) at the same time node t.

Images