Fault Prediction Method and System for Power Transformers Driven by State Holographic Perception Data

Abstract

The present invention discloses a method and system for power transformer fault prediction driven by state holographic perception data. A historical fault data set of the transformer is established; the concentration of dissolved gases in the oil during the operation of the transformer, as well as environmental meteorological data and other operation data reflecting the state of the transformer, are collected. An optimal data set is formed through mutation point detection and phase space reconstruction. A SARIMA model for short-term prediction of the concentration of dissolved gases in the transformer oil is built to obtain the prediction results of the change in the concentration of dissolved gases in the transformer oil in the near future; based on the relationship between the concentration of dissolved gases in the oil and other state variables of the transformer, the concentration of dissolved gases in the oil that changes in real time is obtained; a transformer fault diagnosis model is built based on the DBN network; the prediction results of the change in the concentration of dissolved gases in the oil in the near future or the concentration of dissolved gases in the oil that changes in real time are used as the input feature quantities of the transformer fault diagnosis model to achieve transformer fault prediction. It realizes the comprehensive real-time analysis and auxiliary decision-making of the operation risk level of the power grid transformer.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment state prediction and fault diagnosis, and particularly relates to a power transformer fault prediction method and system driven by state holographic perception data. Background Art

[0002] To adapt to the trend of clean energy development and low-carbon transformation, a large number of new energy installations will be incorporated into the power grid. This brings greater challenges to the safety and stability of the power grid and operation optimization, especially for the "dual-high" power system, which poses unprecedented challenges to the conventional prediction and dispatching control of the power grid operation situation. With the increasing requirement for the reliability of the power system in social construction, the multi-dimensional detection, state prediction, and system control capabilities of the power system all urgently need to be improved.

[0003] In terms of data processing, with the construction and development of the power Internet of Things, a large amount of historical data has been accumulated during the construction and operation of power grid equipment. A large amount of equipment data has accumulated and is in a "dormant state", and its value has not been effectively mined. There is a lack of a means to scientifically analyze and make efficient decisions on numerous data such as historical tests and online detections, and it is difficult to accurately guide the current production operations using big data. Among them, power transformers, as precision and complex electrical equipment, involve a large number of related parameters characterizing their operating states, such as online monitoring data, electrical test data, power grid operation data, meteorological environment data, dissolved gas in oil data, and equipment quality records. Due to the strong coupling and close relationship between various characteristic attributes, scientifically evaluating and predicting the current and future operating states of power transformers through data-driven methods helps to reasonably and specifically arrange equipment maintenance and formulate operation and maintenance strategies.

[0004] In terms of cost reduction, although power enterprises currently invest a large amount of manpower, material resources, and financial resources in equipment operation and maintenance every year to ensure the safe operation of equipment, the traditional maintenance mode lacks timeliness and pertinence for power transformers with faults or defects, and is mandatory for power transformers without defects, which is likely to cause new faults and thus result in waste of resources. Therefore, the traditional maintenance mode has great limitations, and the level of cost lean management needs to be improved. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a power transformer fault prediction method and system driven by state holographic perception data for the deficiencies in the above-mentioned prior art, so as to realize the comprehensive real-time analysis and auxiliary decision-making of the operation risk level of grid transformers, and help to improve the reliability assessment, global situation perception, and adaptive coordinated control capabilities of the power system in multiple aspects.

[0006] The present invention adopts the following technical solutions:

[0007] A method for fault prediction of power transformers driven by state holographic perception data, comprising the following steps:

[0008] S1. During the operation of the transformer, collect the concentration of dissolved gases in the transformer oil, environmental meteorological data and operation data reflecting the state of the transformer, and form a historical data set for prediction. Sort out and collect the concentration of dissolved gases in the oil and the corresponding fault types when the transformer failed in the past to form a historical fault data set;

[0009] S2. Clean and complement the historical data set obtained in step S1, and perform mutation point detection. Delete the historical data with the largest difference from the current transformer working conditions to obtain a historical data set that correctly reflects the current transformer working conditions. Select the optimal time series length for the historical data set through chaos theory and phase space reconstruction theory to obtain an optimal data set for real-time prediction. Decompose the optimal data set based on the seasonal trend decomposition method of locally weighted regression to obtain the trend terms of the dissolved gases in each transformer oil, and form an optimal data set for short-term prediction;

[0010] S3. Based on the SARIMA model, use the optimal data set obtained in step S2 to build a short-term prediction SARIMA model for the concentration of dissolved gases in the transformer oil, and obtain the prediction results of the change in the concentration of dissolved gases in the transformer oil in the near future based on the short-term prediction SARIMA model;

[0011] S4. Establish a real-time prediction model for the concentration of dissolved gases in the transformer oil. Use the optimal data set obtained in step S2, based on the multiple linear regression model, determine the relationship between the concentration of dissolved gases in the oil and other state variables of the transformer, obtain other state variables of the transformer monitored in real time, and obtain the concentration of dissolved gases in the oil that changes in real time to complete real-time prediction;

[0012] S5. Use the historical fault data set obtained in step S1 to build a transformer fault diagnosis model based on the DBN network; use the prediction results of the change in the concentration of dissolved gases in the transformer oil in the near future obtained in step S3, or the concentration of dissolved gases in the transformer oil that changes in real time obtained in step S4 as the input feature quantity of the transformer fault diagnosis model to realize transformer fault prediction.

[0013] Specifically, in step S1, the state operation data of the transformer includes active power, reactive power, ambient temperature and oil temperature; the historical fault data set includes the historical data of the concentration of dissolved gases in the transformer oil when it is normal and the concentration of dissolved gases in the transformer oil corresponding to when the transformer fails, where the fault types include partial discharge, spark discharge, arc discharge, low-temperature overheating, medium-temperature overheating and high-temperature overheating.

[0014] Specifically, in step S2, the mutation point detection is specifically:

[0015] S201. Use the Mann-Kendall algorithm to determine whether there are mutation points in the trend term of the historical data set obtained in step S1. If not, execute step S202; otherwise, execute step S203;

[0016] S202. Take the historical data set obtained in step S1 as the optimal data set for real-time prediction;

[0017] S203. Detect the position of the mutation point in the historical data set obtained in step S1 based on the Pettitt mutation point detection theory to obtain the position of the mutation point;

[0018] S204. Delete the pre-temporal data at the position of the mutation point obtained in step S203, and take the formed new sequence as the historical data set reflecting the current working condition of the transformer.

[0019] Specifically, in step S2, the optimal length of the time series selected through the chaos theory and the phase space reconstruction theory is specifically as follows:

[0020] S205. Take the historical data set as the data set to be processed, and use it as the original data set after data cleaning and complementation;

[0021] S206. Judge the Lyapunov exponent of the original data set selected in step S205;

[0022] S207. If the Lyapunov exponent obtained in step S206 is greater than or equal to 0, use the C-C algorithm to calculate the embedding dimension m of the original data set; otherwise, take the original data set obtained in step S301 as the optimal data set;

[0023] S208. Calculate the length of the optimal data set using the embedding dimension m obtained in step S207;

[0024] S209. Judge the length of the optimal data set obtained in step S208;

[0025] S210. If the length of the optimal data set obtained in step S209 is less than the length of the original data set selected in step S205, take the data of the length of the optimal data set from the back to the front of the original data set to form the optimal data set; if the length of the optimal data set obtained in step S209 is greater than the length of the original data set selected in step S205, take the original data set as the optimal data set.

[0026] Specifically, in step S2, the seasonal trend decomposition method based on locally weighted regression is specifically as follows:

[0027] Use the seasonal trend decomposition method based on locally weighted regression to decompose the optimal data set of the dissolved gas in the transformer oil obtained in step S3 into a trend term, a periodic term, and a remainder term, and take the trend term as the optimal data set for short-term prediction.

[0028] Specifically, in step S3, the short-term prediction is as follows:

[0029] Use the historical dataset of the dissolved gas concentration in transformer oil in the past 1 - 2 years to predict the dissolved gas concentration in the oil within the next 1 month; use the seasonal trend decomposition method based on locally weighted regression to extract the trend term of the prediction data, and use the trend term as the input reference quantity of the prediction model; construct a dissolved gas prediction model in transformer oil based on the SARIMA(p,d,q)(P,D,Q,m) model, where p, d, q, P, D, Q are the orders of the SARIMA model; determine the periodic time interval m according to the characteristics of the most recent data used for short-term prediction; perform difference calculation on the non-stationary time series to obtain the difference orders d, D; calculate the AIC values corresponding to different p, q, P, Q of the SARIMA model, and take the p, q, P, Q corresponding to the minimum AIC as the best model order of the transformer prediction model, and the dissolved gas prediction model in transformer oil is constructed; set the number of days to be predicted for the model to obtain the dissolved gas concentration in transformer oil for the predicted number of days.

[0030] Specifically, in step S4, the real-time prediction is as follows:

[0031] Use the real-time monitored transformer operation data and environmental meteorological data to predict the real-time concentration of dissolved gas in the oil; based on the multiple linear regression model, use the operation data such as transformer active power, transformer reactive power, ambient temperature, and grounded core current as independent variables {β1, β2, …, β n}, and the dissolved gas concentration in the oil as the dependent variable Y gas ; use the historical dataset to fit the relationship between the transformer state variables and the dissolved gas concentration in the oil, input the real-time monitored transformer state variables, and obtain the real-time concentration of dissolved gas in the oil.

[0032] Furthermore, the relationship between the transformer state variables and the dissolved gas concentration in the oil is specifically:

[0033]

[0034] Among them, {Y1, Y2, …, Y n} are the concentrations of various dissolved gases in the oil, {β1, β2,..., β n} are the operation parameters related to the transformer state such as transformer active power, transformer reactive power, grounding current, ambient temperature, and oil temperature, {α 11 , α 12 ,..., α nn} are the coefficients of the real-time prediction multiple linear regression model fitted between the transformer state variables and the dissolved gas concentration in the oil, {ε1, ε2,..., ε n} is the remainder of the real-time prediction multiple linear regression model obtained by fitting the transformer state quantity and the concentration of dissolved gases in oil.

[0035] Specifically, in step S5, the use of the DBN fault diagnosis model for fault prediction is as follows:

[0036] S501. Use the historical fault data set obtained in step S1 as the input of the DBN network without encoding the ratio.

[0037] S502. Normalize the historical fault data set obtained in step S501 to the interval [-1, 1], and divide it into a training set and a test set.

[0038] S503. Set the DBN network parameters: the number of neurons in the input layer and the number of hidden layers of the DBN network and the initial values of the number of neurons in each hidden layer. Among them, the number of neurons in the input layer is equal to the number of fault types.

[0039] S504. Determine the number of neurons in the output layer of the DBN network model.

[0040] S505. Conduct unsupervised layer-by-layer training on the multi-layer RBMs in the DBN network, and use the BP algorithm to perform backpropagation fine-tuning on the connection weights and bias thresholds of the DBN network.

[0041] S506. Use the training set in step S502 as the input of the DBN network, adjust the DBN network parameters, and train the DBN fault diagnosis model under each parameter; use the test set in step S702 as the input of the DBN fault diagnosis model under each parameter to obtain the corresponding fault diagnosis results. Compare the fault diagnosis results given by the DBN fault diagnosis model with the actual fault types, and use the parameters of the DBN network when the diagnostic accuracy rate of the test set is the highest as the parameters of the transformer DBN fault diagnosis model, and save the DBN fault diagnosis model at this time.

[0042] S507. Normalize the data to be diagnosed using the normalization rule of the training set in step S502. The data to be diagnosed is the concentration of dissolved gases in the transformer oil obtained by recent prediction or real-time prediction in step S3 or step S4.

[0043] S508. Input the data to be diagnosed in step S507 into the DBN fault diagnosis model.

[0044] S509. Give the fault diagnosis result to obtain the transformer fault prediction result.

[0045] Another technical solution of the present invention is a power transformer fault prediction system driven by state holographic perception data, including:

[0046] The data acquisition module collects the concentration of dissolved gases in transformer oil during the operation of the transformer, environmental meteorological data and operation data reflecting the state of the transformer, and constitutes a historical data set for prediction. It organizes and collects the concentration of dissolved gases in oil and the corresponding fault types when the transformer failed in the past, and constitutes a historical fault data set;

[0047] The data processing module performs data cleaning and complementation on the historical data set obtained by the data acquisition module, and conducts mutation point detection, deletes the historical data with the largest difference from the current transformer working conditions, and obtains a historical data set that correctly reflects the current transformer working conditions. The historical data set selects the optimal time series length through chaos theory and phase space reconstruction theory to obtain an optimal data set for real-time prediction; the optimal data set is decomposed based on the seasonal trend decomposition method of locally weighted regression to obtain the trend terms of each dissolved gas in the transformer oil, and constitutes an optimal data set for short-term prediction;

[0048] The historical prediction module builds a short-term prediction SARIMA model for the concentration of dissolved gases in transformer oil based on the SARIMA model, and uses the optimal data set obtained by the data processing module to obtain the prediction results of the recent change in the concentration of dissolved gases in transformer oil based on the short-term prediction SARIMA model;

[0049] The real-time prediction module establishes a real-time prediction model for the concentration of dissolved gases in transformer oil. Using the optimal data set obtained by the data processing module, based on the multiple linear regression model, it determines the relationship between the concentration of dissolved gases in oil and other state variables of the transformer, obtains other state variables of the transformer monitored in real time, and obtains the concentration of dissolved gases in oil that changes in real time to complete real-time prediction;

[0050] The prediction output module builds a transformer fault diagnosis model based on the DBN network using the historical fault data set obtained by the data acquisition module; uses the prediction results of the recent change in the concentration of dissolved gases in transformer oil obtained by the historical prediction module, or the concentration of dissolved gases in oil that changes in real time obtained by the real-time prediction module as the input feature quantity of the transformer fault diagnosis model to realize transformer fault prediction.

[0051] Compared with the prior art, the present invention has at least the following beneficial effects:

[0052] A fault prediction method for power transformers driven by state holographic perception data of the present invention is based on the principle of dissolved gas analysis (DGA) in oil, uses the concentration of dissolved gases in oil to achieve fault prediction of the transformer, and comprehensively grasps the operation and development trend of the transformer through the recent and real-time fault prediction results of the transformer. The recent prediction model is trained using the historical data set to achieve the recent prediction of the concentration of dissolved gases in oil. To prevent sudden changes in the transformer state, the relationships between other key parameters such as environmental meteorological data, transformer operation data, and state monitoring data and the concentration of dissolved gases in oil are explored, and the concentration of dissolved gases in oil is reflected in real time through other key parameters of the transformer monitored in real time. Using the recent and real-time prediction results of the concentration of dissolved gases in oil, the recent and real-time fault prediction of the transformer is achieved through the transformer fault diagnosis model.

[0053] Furthermore, the concentration of dissolved gases in transformer oil changes with its operating environment and operating state. Introducing the environmental meteorological data of the area where the transformer is located and the transformer operation data related to the transformer state into the real-time prediction model of dissolved gases in oil will improve the prediction accuracy. There are complex non-linear relationships between environmental temperature, transformer load, grounding current, oil temperature, etc. and the load performance and insulation performance of the transformer. For example, the oil temperature is related to the adsorption capacity of insulating oil paper for dissolved gases in oil. When the oil temperature is too high, the adsorption capacity in the oil paper decreases, the gas adsorbed in the oil paper is released, and the concentration of dissolved gases in oil increases. When the transformer has an overheating fault, the heat generated at the fault point raises the oil temperature, accelerating the deterioration and decomposition of the oil to generate fault gases. For example, when the transformer has high-temperature overheating, the content of acetylene in the oil increases significantly.

[0054] Furthermore, when the transformer has undergone a disassembly experiment or its operating state has changed suddenly, the premature historical data can no longer correctly reflect the current working conditions of the transformer. The premature historical data that does not conform to the current transformer state is eliminated through mutation point detection. The Mann-Kendall mutation point detection method is a non-parametric statistical test method with little influence of sample characteristics on its detection effect and convenient calculation. The Pettitt mutation point detection method is a non-parametric mutation detection method. By calculating the statistics of each point in the time series, the time point corresponding to the point with the largest absolute value of the statistic is taken as the mutation point. The historical data set is detected by combining the Mann-Kendall mutation point detection method and the Pettitt mutation point detection method. If there is a mutation point, the data before the mutation point is eliminated to form a new historical data set. If there is no mutation point, the original historical data set is retained.

[0055] Furthermore, obtaining the optimal length of the historical dataset is beneficial to improving the accuracy and computational efficiency of the prediction of the dissolved gas concentration in oil. The dissolved gas concentration in the oil of the vast majority of transformers has chaotic characteristics, and the chaotic characteristics can be judged by the Lyapunov exponent. When the Lyapunov exponent is less than 0, the time series changes stably. When the Lyapunov exponent is equal to 0, the time series change is at the stable boundary. When the Lyapunov exponent is greater than 0, the time series change is unstable. The phase space reconstruction method is a commonly used method for studying chaotic structure, which can extract a shorter time series that can reflect the system law from a longer time series.

[0056] Furthermore, the Seasonal-Trend decomposition procedure based on Loess (STL) of locally weighted regression is a widely used and strongly robust time series decomposition method, which can decompose a time series into a trend term, a seasonal term, and a remainder term. The dissolved concentration in the transformer oil under long-term monitoring fluctuates continuously, and its changing characteristics cannot be grasped. The STL is used to extract its trend term from the historical data of the dissolved gas concentration in the oil, and the changing trends of the dissolved gas concentrations in each oil are obtained.

[0057] Furthermore, making a short-term prediction of the dissolved gas concentration in the transformer oil is beneficial to obtaining the short-term changing trend of the gas concentration in advance. The prediction result is used as the input of the DBN fault diagnosis model to obtain the short-term fault prediction result of the transformer. If the prediction result contains a fault, corresponding measures should be taken in advance to avoid the occurrence of the fault as much as possible. The Seasonal Autoregressive Integrated Moving Average model (SARIMA) is a classic time series prediction model with high stability, which consists of three parts: autoregression, differencing, and moving average. The autoregressive model uses the lagged values of the target variable for prediction; differencing means that the difference value of the original target value is used during prediction, which makes the prediction result more stable; the moving average takes the lagged prediction error as the input, making the prediction result more accurate.

[0058] Furthermore, the monitoring frequency of the dissolved gas concentration in oil is relatively low, generally 1 - 2 times a day, while the monitoring frequency of other operation data related to the transformer status is relatively high. For example, the grounded core current is monitored once an hour. When the transformer status suddenly changes, relying solely on the historical data set of the dissolved gas concentration in oil for prediction cannot reflect the transformer status in a timely manner. Therefore, environmental meteorological data with a high monitoring frequency and other operation data related to the transformer status are introduced into the real-time prediction model. Taking the environmental meteorological data and transformer operation data as independent variables, and the historical data of the dissolved gas concentration in oil for real-time prediction as the dependent variable, a real-time prediction model of the transformer is obtained by fitting based on the multiple linear fitting model. When the real-time monitored environmental meteorological data and transformer operation data are input into the model, the real-time predicted dissolved gas concentration in the transformer oil can be obtained.

[0059] Furthermore, the Deep Belief Networks (DBN) is a deep learning method based on a large number of training data sets. This network consists of several layers of Restricted Boltzmann Machines (RBM) and a classification output layer. The model weights are obtained by using the unsupervised greedy layer-by-layer pre-training method, the network is fine-tuned by using the gradient descent method, and the ReLU (Rectified Linear Units) activation function is used to improve the convergence performance of the deep structure of the network. After case testing, the classification accuracy of the deep belief network is higher than that of the feedforward neural network and the support vector machine. The DBN network is used as the basic network for training the transformer fault diagnosis model.

[0060] In summary, the present invention has at least the following advantages:

[0061] (1) Combining the real-time fault prediction and the near-term fault prediction of the transformer, on the one hand, the fault status of the transformer within one day is tracked and predicted, and on the other hand, the near-term fault status of the transformer is predicted, so as to comprehensively grasp the future fault development trend of the transformer.

[0062] (2) Introducing environmental meteorological data and other operation data related to the transformer status into the real-time prediction model of the dissolved gas concentration in the transformer oil improves the accuracy of gas concentration prediction, thus better reflecting the real-time status of the transformer, and has a fast calculation speed, which is convenient for on-site application.

[0063] The technical solution of the present invention will be further described in detail below through the accompanying drawings and embodiments. Description of the Drawings

[0064] Figure 1 It is a schematic diagram of the whole process of the transformer fault prediction driven by the state holographic perception data of the present invention;

[0065] Figure 2Flow chart of data selection based on mutation point detection algorithm of the present invention;

[0066] Figure 3 Flow chart of selecting the optimal length of the data set based on the phase space reconstruction method of the present invention;

[0067] Figure 4 Flow chart of the steps for establishing the short-term prediction SARIMA model of the present invention;

[0068] Figure 5 Flow chart of the training of the fault diagnosis DBN network model of the present invention;

[0069] Figure 6 Flow chart of the use of the fault diagnosis DBN network model of the present invention;

[0070] Figure 7 Original data graph of dissolved hydrogen in the oil of a certain 220 kV transformer;

[0071] Figure 8 Data graph of the processed dissolved hydrogen data in the oil of a certain 220 kV transformer;

[0072] Figure 9 Data graph of the trend item of dissolved hydrogen in the oil of a certain 220 kV transformer;

[0073] Figure 10 Data graph of the predicted dissolved hydrogen in the oil of a certain 220 kV transformer. Detailed implementation manners

[0074] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0075] In the description of the present invention, it should be understood that the terms "include" and "comprise" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0076] It should also be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0077] It should also be further understood that the term "and / or" used in the specification and appended claims of the present invention refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0078] Various structural schematic diagrams according to the disclosed embodiments of the present invention are shown in the drawings. These figures are not drawn to scale, where for the purpose of clear expression, some details are enlarged and some details may be omitted. The shapes of various regions and layers shown in the figures, as well as their relative sizes and positional relationships, are merely exemplary. In practice, there may be deviations due to manufacturing tolerances or technical limitations, and those skilled in the art can design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0079] The present invention provides a method for predicting power transformer faults driven by state holographic perception data. During the operation of the transformer, a series of operations such as selection and processing are performed on a large amount of historical holographic perception data deposited. A recent prediction model, a real-time prediction model, and a fault diagnosis model are respectively established. First, through the recent prediction model or the real-time prediction model, the concentration of dissolved gases in oil predicted recently or in real time for the transformer is obtained. Then, through the fault diagnosis model, recent fault prediction or real-time fault prediction of the transformer is realized. Subsequently, in production practice, only by taking the data monitored in real time as the model input after data processing can the purpose of data-driven fault prediction be achieved. Based on historical operation data and real-time data, the present invention realizes the comprehensive real-time analysis and auxiliary decision-making of the operation risk level of grid transformers, which helps to rapidly improve the reliability assessment, global situation awareness, and adaptive coordinated control capabilities of the digital twin power system in many aspects.

[0080] The high proportion of clean energy grid connection brings greater uncertainties to the safety and stability of the power grid and the operation optimization, posing a huge challenge to the conventional prediction and dispatching control of the power grid operation situation. Through the research on the transformer fault prediction technology driven by state holographic perception data, power grid situation awareness, fault diagnosis, and analysis and deduction are carried out based on real-time data, and precise control of the power system dispatching operation and equipment operation and maintenance is realized by this as the driving force.

[0081] Please refer to Figure 1, the present invention is a method for predicting power transformer faults driven by state holographic perception data. Based on the DGA theory as the basis for transformer fault diagnosis, first, various state data during the operation of the transformer are collected, and after data processing, it provides effective information for real-time prediction, short-term prediction, and fault prediction. Then, based on the SARIMA model, the changing trend of the dissolved gas concentration in the transformer oil in the near future is predicted. Based on the multiple linear regression model, the relationship between the dissolved gas concentration in the transformer oil in the near future and other state parameters that can be monitored in real time is explored, and the real-time dissolved gas concentration in the transformer oil is obtained according to the state variables that can be monitored in real time. Finally, based on the DBN network, a DBN fault diagnosis model is established with the uncoded ratio of the dissolved gas concentration in the transformer oil as the input feature quantity to realize the fault prediction of the transformer. The specific steps are as follows:

[0082] S1. Use the on-line oil chromatograph monitoring device to collect the volume data of the dissolved gas concentration in the transformer oil; use sensors to collect the state parameters of the remaining transformer. During the commissioning and operation of the transformer, a large amount of data is generated, and the types of data that can be retrieved are as follows:

[0083] (1) Transformer ledger information, including the commissioning time, equipment parameters, and usage conditions of the transformer;

[0084] (2) Transformer state detection information, including oil-gas detection, vibration signals, and partial discharges;

[0085] (3) Transformer operating state information, including working voltage and working current;

[0086] (4) Transformer wide-area environment information, including temperature, humidity, air pressure, and sunshine;

[0087] (5) Transformer state change information, including defect records, abnormal operating conditions, and maintenance records;

[0088] (6) Information on the dissolved gas concentration in the transformer oil,

[0089] The historical data sets for fault prediction include the following:

[0090] (1) Dissolved gas concentration in the transformer oil, including H2, CH4, C2H6, C2H4, C2H2, CO, CO2, and total hydrocarbons. The content values of gas components in the transformer oil are obtained by the on-line oil chromatograph monitoring device at a fixed sampling period to form an on-line oil chromatograph monitoring time series;

[0091] (2) Other operating state variables and environmental meteorological data of the transformer: active power, reactive power, oil temperature, grounding current, environmental temperature, etc., measured by various different sensors;

[0092] (3) Collect the dissolved gas concentration in the transformer oil and the corresponding fault types when the transformer has failed in the past to form a historical fault data set.

[0093] S2. Process the historical dataset collected in step S1 to form an optimal dataset;

[0094] The data processing includes data cleaning, data completion, and data selection.

[0095] Among them, data cleaning eliminates duplicate data, null values, outliers in the data, and unifies the time format;

[0096] Data completion fills in the information of the transformer state quantity for the missing dates through numerical interpolation methods;

[0097] Data selection includes mutation point detection and phase space reconstruction to select the optimal length of the dataset;

[0098] Please refer to Figure 2 , perform mutation point detection on the dataset after data cleaning and data completion in step S2, delete the historical dataset with a large difference from the current transformer working condition, and obtain the effective historical data. As shown in Figure 2 (a), the hydrogen concentration fluctuates around 10 (μL / L) between January 2019 and May 2020, and then suddenly changes to fluctuate around 7 (μL / L) after May. Therefore, the historical data before May can no longer correctly reflect the current working condition of the transformer. Thus, mutation point detection is performed, and the data before May is eliminated after detection. As shown in Figure 2 (b).

[0099] Please refer to Figure 2 (c). The mutation point detection is specifically as follows:

[0100] S201. Perform the MK test on the historical dataset after data cleaning and data completion;

[0101] Use the Mann - Kendall algorithm to determine whether there is a mutation point in the data trend term. If not, go to step S202; otherwise, go to step S203;

[0102] S202. The original historical dataset is the optimal dataset;

[0103] S203. Perform the Pettitt mutation point position detection, and then go to step S204;

[0104] S204. Delete the time - series data before the mutation point to form a new historical dataset, and end.

[0105] Please refer to Figure 3 , based on the chaos theory, select the optimal length of the time series for the historical dataset after mutation point detection using the phase space reconstruction theory. The specific steps are as follows:

[0106] The historical dataset obtained in S205, step S202, or step S204 is the historical dataset to be processed, and proceed to step S206;

[0107] S206. Determine the Lyapunov exponent of the historical dataset. If the Lyapunov exponent is greater than 0, proceed to step S207; otherwise, proceed to step S211;

[0108] S207. Calculate the embedding dimension m of the original dataset using the C-C algorithm, and then proceed to step S208;

[0109] S208. Find the optimal dataset length, and the optimal dataset length is 2 m (m + 1) m , and then proceed to step S209;

[0110] S209. Determine whether the optimal dataset length is greater than the length of the original data. If so, proceed to step S211; otherwise, proceed to step S210;

[0111] S210. Obtain the optimal dataset. The optimal dataset is composed of the data with the optimal dataset length taken from the original dataset from back to front, and then the algorithm ends;

[0112] S211. The original dataset is the optimal dataset, and then the algorithm ends.

[0113] S3. The short-term prediction is based on the SARIMA(p,d,q)(P,D,Q,m) model. Through the seasonal trend decomposition method STL of locally weighted regression, the trend term of the historical dataset of the dissolved gas concentration in oil is extracted and used as the input feature quantity of the prediction model. The short-term prediction model SARIMA(p,d,q)(,D,Q,m) of each dissolved gas in oil is trained and fitted, the number of days to be predicted is set, and the prediction result is given;

[0114] Please refer to Figure 4 , and the specific steps of the short-term prediction of the dissolved gas concentration in transformer oil are as follows:

[0115] S301. Extract the trend term of the historical dataset of the dissolved gas concentration in oil based on the seasonal trend decomposition method STL of locally weighted regression. The trend terms of each gas obtained are used as the input feature quantities of the prediction model, and then proceed to step S302;

[0116] S302. Determine the periodic time interval m, and then proceed to step S303;

[0117] S303. Perform differential calculation on the input feature quantities to obtain the differential orders d, D, and then proceed to step S304;

[0118] S304. Parameter optimization calculation. Take p, q, P, and Q when the AIC value is the smallest as the model orders, that is, obtain the prediction model SARIMA(p, d, q)(P, D, Q, m) for the recent prediction of the dissolved gas concentration in oil, and then go to step S305;

[0119] Among them, the expression of AIC is shown in Equation (1), and L is the maximum likelihood function

[0120] AIC = -2log(L) + 2(p + q + P + Q) (1)

[0121] S305. Trend prediction. Set the number of prediction days to obtain the recent prediction result of the dissolved gas concentration in oil.

[0122] S4. Establish a multiple linear regression model for real-time prediction of the dissolved gas concentration in transformer oil. Take the transformer active power, transformer reactive power, ambient temperature, grounding current, etc. as independent variables {β1, β2,..., β n}, and the dissolved gas concentration in oil as the dependent variable Y gas , and use the historical data set of independent variables and dependent variables to fit the relationship between the transformer state variables and the dissolved gas concentration in oil, as shown in Equation (2);

[0123]

[0124] Among them, {Y1, Y2,..., Y n} are the dissolved gas concentrations in oil, {β1, β2,..., β n} are the operating parameters related to the transformer state such as the transformer active power, transformer reactive power, grounding current, ambient temperature, and oil temperature, {α 11 , α 12 ,..., α nn} are the coefficients of the multiple linear regression model for real-time prediction fitted between the transformer state variables and the dissolved gas concentration in oil, and {ε1, ε2,..., ε n} are the remainders of the multiple linear regression model for real-time prediction fitted between the transformer state variables and the dissolved gas concentration in oil.

[0125] The specific steps for real-time prediction of the dissolved gas concentration in transformer oil are as follows:

[0126] S401. Take the historical data set of the dissolved gas concentration in oil obtained in step S2 as the dependent variable, and take the historical data set of other operating state parameters related to the transformer state obtained in step S2 as the independent variables, and then go to step S402;

[0127] S402. Use the historical data set of independent and dependent variables in step S401 to fit a real-time prediction multiple linear regression model for the dissolved gas concentration in transformer oil, and then go to step S403;

[0128] S403. Input the real-time monitored operating state parameters of the transformer {βt 1， β t2 , …, β tn}, and the real-time concentration of dissolved gas in the oil can be obtained, and the process ends.

[0129] S5. Use the historical fault data set obtained in step S1 to build a transformer fault diagnosis model based on DBN; use the recent prediction result or real-time prediction result of the dissolved gas concentration in the oil obtained in step S3 or step S4 as the input feature quantity of the DBN fault diagnosis model to realize transformer fault prediction;

[0130] Fault diagnosis includes two parts: the establishment of a fault model and the use of a fault model.

[0131] Among them, the establishment of a fault model based on the DBN network is to determine the number of neurons in the input and output layers according to the historical fault data set, determine the number of hidden layers of the DBN and the number of neurons in each layer, perform unsupervised training on the multi-layer RBM, and use the BP algorithm to repeatedly fine-tune the connection weights and biases of each neuron in the DBN model. During the process of tuning the parameters of the DBN network, obtain the fault diagnosis results of the test set under each DBN network parameter, record the network parameters corresponding to the DBN network with the highest diagnostic accuracy of the test set, and save the DBN fault diagnosis model.

[0132] The use of the fault diagnosis model means using the recent prediction result or real-time prediction result of the dissolved gas concentration in the oil obtained in step S3 or step S4 as the input feature quantity of the DBN fault diagnosis model, and using the DBN fault diagnosis model to diagnose the fault type corresponding to the prediction result of the dissolved gas concentration in the oil to realize transformer fault prediction.

[0133] Please refer to Figure 5 for the specific method of building a fault diagnosis model based on DBN:

[0134] S501. Perform non-coding ratio on the historical fault data set obtained in step S1 as the data to be trained for the DBN fault diagnosis model, and then go to step S502;

[0135] S502. Normalize the data to be trained in step S501 to the interval [-1, 1], and divide the training set and test set in a ratio of 8:2, and then go to step S503;

[0136] S503. Determine the number of neurons in the input layer (the number of neurons in the input layer is equal to the number of fault types), the number of hidden layers of the DBN, the number of neurons in each hidden layer, and the number of neurons in the output layer of the DBN model, and then proceed to step S504;

[0137] S504. Conduct unsupervised layer-by-layer training on the multi-layer RBM, and use the BP algorithm to perform reverse fine-tuning on the connection weights and bias thresholds of the DBN model, and then proceed to step S505;

[0138] S505. Test the trained DBN model with the test set, record the fault diagnosis accuracy rate of the test set under each DBN network model parameter, and then proceed to step S506;

[0139] S506. Repeat steps S503 to S504, take the network parameters when the fault diagnosis accuracy rate of the test set is the highest as the final network parameters of the DBN fault diagnosis model, and save the trained DBN fault diagnosis model, and then the algorithm ends.

[0140] Please refer to Figure 6 , for specific fault diagnosis:

[0141] S507. Input the data to be diagnosed, that is, the recent prediction or real-time prediction result of the dissolved gas concentration in transformer oil, and then proceed to step S508;

[0142] S508. Perform non-encoding processing on the data to be diagnosed, and then proceed to step S509;

[0143] S509. Normalize the data to be diagnosed using the normalization rule of the training set, and then proceed to step S510;

[0144] S510. Input the data to be diagnosed into the DBN fault diagnosis model, and then proceed to step S511;

[0145] S511. Calculate the probability distribution of each fault category that the data to be diagnosed may occur according to the DBN fault diagnosis model, and then proceed to step S512;

[0146] S512. Obtain the index corresponding to the maximum probability in the probability distribution obtained in step S512, and then proceed to step S513;

[0147] S513. Convert the index obtained in step S512 into the corresponding fault category to obtain the transformer fault prediction result, and then the algorithm ends.

[0148] In another embodiment of the present invention, a power transformer fault prediction system driven by state holographic perception data is provided. This system can be used to implement the above-mentioned power transformer fault prediction method driven by state holographic perception data. Specifically, the power transformer fault prediction system driven by state holographic perception data includes a data acquisition module, a data processing module, a historical prediction module, a real-time prediction module, and a prediction output module.

[0149] Among them, the data acquisition module collects the concentration of dissolved gases in transformer oil, environmental meteorological data and operation data reflecting the state of the transformer during the operation of the transformer, constitutes a historical data set for prediction, and collates and collects the concentration of dissolved gases in the oil and the corresponding fault types when the transformer failed in the past, constituting a historical fault data set;

[0150] The data processing module performs data cleaning and complementation on the historical data set obtained by the data acquisition module, and performs mutation point detection, deletes the historical data with the largest difference from the current transformer working conditions, and obtains a historical data set that correctly reflects the current transformer working conditions. The historical data set is selected by the chaotic theory and the phase space reconstruction theory to obtain the optimal data set for real-time prediction; the optimal data set is decomposed based on the seasonal trend decomposition method of locally weighted regression to obtain the trend items of the dissolved gases in each transformer oil, constituting the optimal data set for short-term prediction;

[0151] The historical prediction module builds a short-term prediction SARIMA model for the concentration of dissolved gases in transformer oil based on the SARIMA model, and obtains the prediction result of the change in the concentration of dissolved gases in the transformer oil in the short term based on the short-term prediction SARIMA model;

[0152] The real-time prediction module establishes a real-time prediction model for the concentration of dissolved gases in transformer oil, uses the optimal data set obtained by the data processing module, and based on the multiple linear regression model, determines the relationship between the concentration of dissolved gases in the oil and other state variables of the transformer, obtains other state variables of the transformer monitored in real time, and obtains the concentration of dissolved gases in the oil that changes in real time to complete real-time prediction;

[0153] The prediction output module builds a transformer fault diagnosis model based on the DBN network using the historical fault data set obtained by the data acquisition module; uses the prediction result of the change in the concentration of dissolved gases in the transformer oil in the short term obtained by the historical prediction module, or the concentration of dissolved gases in the oil that changes in real time obtained by the real-time prediction module as the input feature quantity of the transformer fault diagnosis model to realize transformer fault prediction.

[0154] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Generally, the components described and shown in the accompanying drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0155] Taking a certain actually put into operation 220 kV transformer as an example, part of the dissolved gas in oil monitoring data from March 2019 to March 2021 is shown in Table 1.

[0156] Table 1 Dissolved gas in oil monitoring data of a certain 220 kV transformer

[0157]

[0158] Taking hydrogen as an example, data processing is performed on the input time series through steps such as filtering duplicate values, filling in missing values, and deleting outliers. The original data sequence has a total of 639 groups of data. As Figure 7 shown, the missing data is in the dashed box and is filled in to 733 groups of data after data processing. As Figure 8 shown, the missing data in the dashed box has been filled in.

[0159] After mutation point detection, ethane has a mutation point at the 68th acquisition time. Therefore, the historical data of the concentration of each dissolved gas in oil retains the last 665 groups of data.

[0160] According to the chaos theory and the phase space reconstruction method, the embedding dimension of this case is 3. Therefore, based on the previous step, the last 512 pieces of historical data are taken as the optimal data set.

[0161] Extract the trend term of the optimal data set as the input data of the short-term prediction model. Taking hydrogen as an example, the trend term is as Figure 9 shown; according to the data characteristics of the historical data set of the dissolved gas in oil concentration, the periodic time interval is set to 4, and the short-term prediction model of the concentration of each dissolved gas in oil is determined, as shown in Table 2. Based on this, the change trend of the concentration of each gas in the next 30 days is predicted. Taking the hydrogen concentration as an example, the change situation is given, as Figure 10 shown, and the predicted results of the concentration of the dissolved gas in oil in the next 30 days are shown in Table 3.

[0162] Table 2 Short-term prediction model of the concentration of each dissolved gas in oil

[0163]

[0164]

[0165] Table 3 Recent predicted results of dissolved gas concentrations in each oil

[0166] Prediction time Methane Ethylene Acetylene Ethane Hydrogen Total hydrocarbons Carbon dioxide Carbon monoxide 20210311 1.551 0 0.428 0 7.463 1.769 154.413 77.242 20210312 1.552 0 0.427 0 7.49 1.758 153.309 77.395 20210313 1.553 0 0.426 0 7.518 1.747 152.202 77.547 20210314 1.555 0 0.425 0 7.545 1.736 151.093 77.699 20210315 1.556 0 0.424 0 7.572 1.725 149.981 77.85 20210316 1.557 0 0.423 0 7.6 1.714 148.867 78.001 20210317 1.559 0 0.422 0 7.627 1.703 147.751 78.151 20210318 1.56 0 0.421 0 7.655 1.693 146.633 78.3 20210319 1.561 0 0.42 0 7.682 1.682 145.512 78.448 20210320 1.563 0 0.419 0 7.709 1.672 144.39 78.596 20210321 1.564 0 0.417 0 7.737 1.662 143.265 78.743 20210322 1.565 0 0.416 0 7.764 1.652 142.137 78.889 20210323 1.567 0 0.415 0 7.792 1.643 141.008 79.035 20210324 1.568 0 0.414 0 7.82 1.634 139.876 79.18 20210325 1.569 0 0.413 0 7.847 1.624 138.742 79.325 20210326 1.571 0 0.412 0 7.875 1.616 137.606 79.468 20210327 1.572 0 0.411 0 7.902 1.607 136.467 79.611 20210328 1.574 0 0.41 0 7.93 1.598 135.326 79.754 20210329 1.575 0 0.409 0 7.957 1.59 134.183 79.895 20210330 1.576 0 0.408 0 7.985 1.582 133.038 80.036 20210331 1.578 0 0.406 0 8.013 1.574 131.891 80.177 20210401 1.579 0 0.405 0 8.04 1.567 130.741 80.316 20210402 1.581 0 0.404 0 8.068 1.56 129.589 80.455 20210403 1.582 0 0.403 0 8.096 1.552 128.435 80.594 20210404 1.583 0 0.402 0 8.123 1.545 127.278 80.731 20210405 1.585 0 0.401 0 8.151 1.539 126.12 80.868 20210406 1.586 0 0.4 0 8.179 1.532 124.959 81.005 20210407 1.588 0 0.399 0 8.206 1.526 123.795 81.14 20210408 1.589 0 0.398 0 8.234 1.52 122.63 81.275 20210409 1.591 0 0.397 0 8.262 1.514 121.462 81.409

[0167] Using the non - coded ratio method to process the recent predicted results of dissolved gas concentrations in oil, and using the trained DBN fault diagnosis model to judge, the diagnosis results within the next 30 days are all "spark discharge";

[0168] The actual situation is that the acetylene content shows an increasing trend, and other characteristic gases are normal, without involving defects in the internal windings, insulation, and iron core of the transformer. Other characteristic gases are normal. It is suspected that there is abnormal contact at the connection part of the 220kV high - voltage lead - out line, resulting in "spark discharge". The fault diagnosis result is consistent with the actual situation, and the fault prediction method provided by the present invention is reasonable.

[0169] For real - time prediction, taking an actually put - into - operation 220kV transformer as an example, a historical data set of dissolved gas concentrations in oil and other operating state parameters of the transformer is collected, as shown in Table 4. In the table, the monitoring of dissolved gas concentrations in oil lists hydrogen as an example. The relationship between oil temperature, active power, reactive power, grounded core current, and ambient temperature and the dissolved gas concentration in oil is mined, and the real - time prediction model is as shown in Equation (3):

[0170] Table 4 Monitoring data of operating state parameters of a 220kV transformer's oil

[0171]

[0172]

[0173] According to the prediction model of Equation (3), the real - time prediction results are shown in Table 5:

[0174] Table 5 Real - time prediction results of dissolved gases in transformer oil

[0175] Prediction time Hydrogen Methane Ethylene Acetylene Ethane Carbon monoxide Carbon dioxide Total hydrocarbons 20211031 12:39 3.81 6.78 1.36 0 0.53 243.79 892.21 8.67

[0176] Using the non - coded ratio method to process the real - time prediction results of dissolved gas concentrations in oil, and using the trained DBN fault diagnosis model to judge, the diagnosis result is "normal";

[0177] The actual situation is that the dissolved gases in each oil are normal and do not affect the continued operation of the equipment. The fault diagnosis result is consistent with the actual situation, and the fault prediction method provided by the present invention is reasonable.

[0178] In summary, for the method and system for predicting faults of a power transformer driven by state holographic perception data according to the present invention, a plurality of operation data of the transformer are retrieved, and an effective historical data set is obtained through gradual processing and selection of the historical data. First, short-term prediction and real-time prediction of the concentration of dissolved gases in transformer oil are performed. A short-term prediction model is built using the concentration of dissolved gases in the transformer oil in the historical data set; a real-time prediction model of the transformer is built using the concentration of dissolved gases in the transformer oil and other state data, such as ambient temperature, oil temperature, active power, reactive power, etc. The short-term or real-time prediction results of the concentration of dissolved gases in the transformer oil are diagnosed through the built fault diagnosis model, so as to achieve the purpose of accurately and comprehensively predicting the fault state of the transformer.

[0179] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present 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.

[0180] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0181] These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0182] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide means for implementing the functions in the processFigure 1 one process or multiple processes and / or boxes Figure 1 steps of functions specified in one box or multiple boxes.

[0183] The above content is only to illustrate the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention falls within the protection scope of the claims of the present invention.

Claims

1. A fault prediction method for power transformers driven by state holographic perception data, characterized in that It includes the following steps: S1. During the operation of the transformer, collect the concentration of dissolved gases in the transformer oil, the active power, reactive power, oil temperature, ground current, and ambient temperature that reflect the transformer state, and form a historical data set for prediction. Organize and collect the concentration of dissolved gases in the oil and the corresponding fault types when the transformer failed in the past, and form a historical fault data set. The historical fault data set includes the historical data of the concentration of dissolved gases in the oil when the transformer is normal and the concentration of dissolved gases in the oil corresponding to when the transformer fails. The fault types include partial discharge, spark discharge, arc discharge, low-temperature overheating, medium-temperature overheating, and high-temperature overheating; S2. Clean and complete the historical data set obtained in step S1, and perform mutation point detection. Delete the historical data with the largest difference from the current transformer operating conditions to obtain a historical data set that correctly reflects the current transformer operating conditions. Select the optimal length of the time series from the historical data set through chaos theory and phase space reconstruction theory to obtain an optimal data set for real-time prediction. Decompose the optimal data set based on the seasonal trend decomposition method of locally weighted regression to obtain the trend terms of the dissolved gases in each transformer oil, and form an optimal data set for short-term prediction; S3. Based on the SARIMA model, use the optimal data set obtained in step S2 to build a short-term prediction SARIMA model for the concentration of dissolved gases in the transformer oil, and obtain the prediction results of the short-term change of the concentration of dissolved gases in the transformer oil based on the short-term prediction SARIMA model; S4. Establish a real-time prediction model for the concentration of dissolved gases in the transformer oil. Use the optimal data set obtained in step S2, based on the multiple linear regression model, determine the relationship between the concentration of dissolved gases in the oil and other state variables of the transformer, obtain other state variables of the transformer monitored in real time, and obtain the concentration of dissolved gases in the oil that changes in real time to complete real-time prediction; S5. Use the historical fault data set obtained in step S1 to build a transformer fault diagnosis model based on the DBN network; use the prediction results of the short-term change of the concentration of dissolved gases in the transformer oil obtained in step S3, or the concentration of dissolved gases in the oil that changes in real time obtained in step S4 as the input feature quantity of the transformer fault diagnosis model to realize transformer fault prediction.

2. The state holographic perception data-driven power transformer fault prediction method according to claim 1, wherein In step S2, the mutation point detection is specifically as follows: S201. Use the Mann-Kendall algorithm to judge whether there are mutation points in the trend terms of the historical data set obtained in step S1. If not, execute step S202; otherwise, execute step S203; S202. Use the historical data set obtained in step S1 as the optimal data set for real-time prediction; S203. Detect the position of the mutation point in the historical data set obtained in step S1 based on the Pettitt mutation point detection theory to obtain the position of the mutation point; S204. Delete the pre-temporal data at the position of the mutation point obtained in step S203, and use the formed new sequence as the historical data set that reflects the current transformer operating conditions.

3. The state holographic perception data-driven power transformer fault prediction method according to claim 1, characterized in that In step S2, the selection of the optimal length of the time series through chaos theory and phase space reconstruction theory is specifically as follows: S205. Use the historical data set as the data set to be processed. After data cleaning and completion, it is used as the original data set. S206. Determine the Lyapunov exponent of the original data set selected in step S205. S207. If the Lyapunov exponent obtained in step S206 is greater than or equal to 0, calculate the embedding dimension of the original data set using the C-C algorithm m , otherwise, the original data set obtained in step S301 is used as the optimal data set; S208. Use the embedding dimension obtained in step S207 m to calculate the optimal data set length; S209. Determine the optimal data set length obtained in step S208. S210. If the optimal data set length obtained in step S209 is less than the length of the original data set selected in step S205, the optimal data set is composed of the data of the optimal data set length taken from the back to the front of the original data set; if the optimal data set length obtained in step S209 is greater than the length of the original data set selected in step S205, the original data set is used as the optimal data set.

4. The method for predicting power transformer faults driven by state holographic perception data according to claim 1, characterized in that, In step S2, the seasonal trend decomposition method based on locally weighted regression is specifically as follows: Use the seasonal trend decomposition method based on locally weighted regression to decompose the optimal data set of the dissolved gases in transformer oil obtained in step S3 into a trend term, a periodic term, and a remainder term, and take the trend term as the optimal data set for short-term prediction.

5. The method for predicting faults of a power transformer driven by state holographic perception data according to claim 1, wherein In step S3, the short-term prediction is specifically as follows: Predict the dissolved gas concentration in oil within the next month using the historical dataset of dissolved gas concentrations in transformer oil over the past 1 to 2 years; use the seasonal trend decomposition method based on locally weighted regression to extract the trend term of the prediction data, and use the trend term as the input reference quantity for the prediction model; based on Build a prediction model for dissolved gases in transformer oil, where is the order of the SARIMA model; determine the periodic time interval according to the characteristics of the most recent data used for short-term prediction ; perform differencing calculations on the non-stationary time series to obtain the differencing order ; Calculate the AIC values corresponding to different of the SARIMA model, and take the corresponding to the minimum AIC as the optimal model order of the transformer prediction model, and the construction of the dissolved gas prediction model in transformer oil is completed; set the number of days to be predicted for the model to obtain the concentration of dissolved gas in transformer oil for the predicted number of days.

6. The method for predicting the faults of a power transformer driven by state holographic perception data according to claim 1, wherein In step S4, the real-time prediction is specifically as follows: Using the real-time monitored operation data of the transformer and environmental meteorological data, predict the real-time concentration of dissolved gases in oil; based on the multiple linear regression model, take the operation data including the active power of the transformer, the reactive power of the transformer, the ambient temperature, and the grounded core current as independent variables , and take the concentration of dissolved gases in oil as the dependent variable , use the historical data set to fit the relationship between the transformer status quantity and the concentration of dissolved gases in oil, input the real-time monitored transformer status quantity, and obtain the real-time concentration of dissolved gases in oil.

7. The method for predicting faults of a power transformer driven by state holographic perception data according to claim 6, characterized in that The relationship between the transformer state quantity and the concentration of dissolved gases in oil is specifically as follows: wherein, are the concentrations of dissolved gases in each oil, are operating parameters related to the active power of the transformer, the reactive power of the transformer, the ground current, the ambient temperature, the oil temperature and the transformer status, are the coefficients of the real-time prediction multiple linear regression model fitted by the transformer status and the concentrations of dissolved gases in the oil, is the remainder of the real-time prediction multiple linear regression model fitted by the transformer status and the concentrations of dissolved gases in the oil.

8. The method for predicting faults of a power transformer driven by state holographic perception data according to claim 1, wherein In step S5, the use of the DBN fault diagnosis model for fault prediction is specifically as follows: S501. Perform non-coded ratio on the historical fault data set obtained in step S1 as the input of the DBN network. S502. Normalize the historical fault data set obtained in step S501 to the interval [-1, 1], and divide it into a training set and a test set. S503. Set the DBN network parameters: the number of neurons in the input layer and the number of hidden layers of the DBN network and the initial values of the number of neurons in each hidden layer. Among them, the number of neurons in the input layer is equal to the number of fault types. S504. Determine the number of neurons in the output layer of the DBN network model. S505. Perform unsupervised layer-by-layer training on the multi-layer RBM in the DBN network, and use the BP algorithm to perform backpropagation fine-tuning on the connection weights and bias thresholds of the DBN network. S506. Use the training set in step S502 as the input of the DBN network, adjust the DBN network parameters, and train the DBN fault diagnosis model under each parameter; use the test set in step S702 as the input of the DBN fault diagnosis model under each parameter, obtain the corresponding fault diagnosis results, compare the fault diagnosis results given by the DBN fault diagnosis model with the actual fault types, and use the parameters taken by the DBN network when the test set diagnosis accuracy rate is the highest as the parameters of the transformer DBN fault diagnosis model, and save the DBN fault diagnosis model at this time. S507. Normalize the data to be diagnosed using the normalization rule of the training set in step S502. The data to be diagnosed is the concentration of dissolved gases in transformer oil obtained from the short-term prediction or real-time prediction in step S3 or step S4. S508. Input the data to be diagnosed in step S507 into the DBN fault diagnosis model. S509. Give the fault diagnosis result to obtain the transformer fault prediction result.

9. A power transformer fault prediction system driven by state holographic perception data, characterized in that Including: The data acquisition module collects the concentrations of dissolved gases in transformer oil during the operation of the transformer, the active power, reactive power, oil temperature, ground current, and ambient temperature that reflect the transformer status, and constitutes a historical dataset for prediction. It also sorts out and collects the concentrations of dissolved gases in oil and the corresponding fault types when the transformer failed in the past, and constitutes a historical fault dataset. The historical fault dataset includes the historical data of the concentrations of dissolved gases in oil when the transformer is normal and the concentrations of dissolved gases in oil corresponding to when the transformer fails. The fault types include partial discharge, spark discharge, arc discharge, low-temperature overheating, medium-temperature overheating, and high-temperature overheating; The data processing module performs data cleaning and completion on the historical dataset obtained by the data acquisition module, and conducts mutation point detection, deleting the historical data with the largest difference from the current transformer operating conditions, to obtain a historical dataset that correctly reflects the current transformer operating conditions. The historical dataset is selected for the optimal time series length through the chaos theory and phase space reconstruction theory to obtain an optimal dataset for real-time prediction; the optimal dataset is decomposed based on the seasonal trend decomposition method of locally weighted regression to obtain the trend terms of the dissolved gases in each transformer oil, constituting an optimal dataset for short-term prediction; The historical prediction module, based on the SARIMA model, uses the optimal dataset obtained by the data processing module to build a short-term prediction SARIMA model for the concentration of dissolved gases in transformer oil, and obtains the prediction results of the recent change in the concentration of dissolved gases in transformer oil based on the short-term prediction SARIMA model; The real-time prediction module establishes a real-time prediction model for the concentration of dissolved gases in transformer oil. Using the optimal dataset obtained by the data processing module, based on the multiple linear regression model, it determines the relationship between the concentration of dissolved gases in oil and other status variables of the transformer, obtains other status variables of the transformer monitored in real time, and gets the concentration of dissolved gases in oil that changes in real time to complete real-time prediction; The prediction output module uses the historical fault dataset obtained by the data acquisition module to build a transformer fault diagnosis model based on the DBN network; uses the prediction results of the recent change in the concentration of dissolved gases in transformer oil obtained by the historical prediction module, or the concentration of dissolved gases in oil that changes in real time obtained by the real-time prediction module as the input feature quantity of the transformer fault diagnosis model to achieve transformer fault prediction.

Images