Transformer fault early warning method based on stacked self-encoding network and electronic equipment

Abstract

The invention relates to a transformer fault early warning method based on a stacked self-encoding network and electronic equipment. The method comprises the following steps of S1, acquiring monitoring variable historical values including a transformer variable, an environment variable, a cable incoming line variable and a bus variable; s2, training a stacked self-encoding network based on the historical values of the monitoring variables, reconstructing the measured values of the monitoring variables obtained in real time, and calculating a reconstruction error; s3, performing mean filtering and window filtering processing of a time dimension on the reconstruction error; and S4, judging whether the reconstruction error and the change rate thereof are greater than a threshold value determined in advance through a random forest model, and realizing transformer fault early warning. Compared with the prior art, the method fully considers the influence of the variables of multiple dimensions on the state of the transformer, and has the advantages of wide abnormal type monitoring, accurate threshold determination and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault early warning, and in particular to a transformer fault early warning method and electronic equipment based on a stacked autoencoding network. Background Art

[0002] In recent years, a large number of non-electrical variable monitoring devices have been widely installed in power systems, opening up new possibilities for using multi-dimensional monitoring data for transformer safety assessment. However, existing transformer safety assessments still mainly rely on the threshold judgment of a single monitoring signal, which not only leads to serious delays in fault warnings, but also makes the results less reliable.

[0003] Chinese patent application number CN202211466260 discloses a multi-parameter fusion method for defect identification, state classification and active safety protection of power transformers. This method integrates multi-source data such as the operating environment conditions and operating parameters of the power transformer to form a knowledge graph, conducts a safety assessment of the transformer, and takes corresponding safety regulation and protection measures based on the assessment results. However, this method only considers the impact of the transformer's own factors on its safety, and does not involve the impact of transformer-related equipment and its monitoring variables on the transformer.

[0004] In summary, there is currently a lack of a transformer fault early warning method that fully considers the influence of multiple factors. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a transformer fault warning method and electronic device based on a stacked autoencoder network, which fully considers the impact of variables on the transformer from multiple dimensions and realizes transformer fault warning.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] One aspect of the present invention provides a transformer fault early warning method based on a stacked autoencoder network, comprising the following steps:

[0008] Step S1, obtaining historical values ​​of monitoring variables including transformer variables, environmental variables, cable incoming line variables and busbar variables;

[0009] Step S2, training a stacked autoencoder network based on the historical values ​​of the monitoring variables, reconstructing the measured values ​​of the monitoring variables obtained in real time, and calculating the reconstruction error;

[0010] Step S3, performing mean filtering and window filtering in the time dimension on the reconstruction error;

[0011] Step S4, determining whether the reconstruction error and its rate of change are greater than a threshold value pre-determined by the random forest model, thereby realizing transformer fault early warning.

[0012] As a preferred technical solution, the training process of the stacked autoencoder network includes the following steps:

[0013] Step S201, obtaining multiple groups of historical values ​​of monitoring variables for training and performing data cleaning and normalization processing;

[0014] Step S202, inputting the multiple groups of monitoring variable historical values ​​obtained in step S201 into the encoder in the autoencoding network, and using the output of the encoder as the input of the decoder in the autoencoding network to obtain reconstructed monitoring variable values;

[0015] Step S203, based on the historical values ​​of the monitoring variables used for training and the reconstructed monitoring variable values, a reconstruction error is calculated, and the autoencoder network is trained with the goal of minimizing the reconstruction error.

[0016] As a preferred technical solution, in step S4, the process of determining the threshold value includes the following steps:

[0017] Step S401, obtaining multiple groups of reconstruction errors and corresponding change rates in fault state and normal state, and marking them according to normal or fault;

[0018] Step S402, using the labeled multiple groups of reconstruction errors and the corresponding change rates, using a random forest model to obtain the normal and fault category scores, and obtain the threshold.

[0019] As a preferred technical solution, in step S4, if the reconstruction error is greater than the corresponding threshold or the rate of change of the reconstruction error is greater than the corresponding threshold, a transformer fault warning signal is issued.

[0020] As a preferred technical solution, the transformer variables include high-voltage side current value, core grounding current, clamp grounding current, oil level height, oil level ratio, oil temperature, main transformer temperature, main ventilator temperature and main ventilator humidity.

[0021] As a preferred technical solution, the environmental variables include ambient temperature, ambient humidity and water level.

[0022] As a preferred technical solution, the cable incoming line variables include busbar air chamber temperature, switch air chamber temperature, line transformer air chamber temperature and line cable air chamber temperature.

[0023] As a preferred technical solution, the busbar variables include the busbar gas chamber temperature, the switch gas chamber temperature and the line cable gas chamber temperature.

[0024] Another aspect of the present invention provides an electronic device, comprising: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the aforementioned transformer fault warning method based on a stacked autoencoder network.

[0025] Another aspect of the present invention provides a computer-readable storage medium comprising one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the aforementioned transformer fault warning method based on a stacked autoencoding network.

[0026] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0027] (1) Fully consider the impact of variables in multiple dimensions on the transformer state: The present invention comprehensively considers the impact of variables in the transformer, environment, cable feeder and bus dimensions on the transformer state, links electrical variables with non-electrical variables, and constructs a stacked autoencoding network between multi-dimensional monitoring variables associated with the transformer. Without the need to manually generate fault data, the transformer state assessment and fault diagnosis can be performed through normal data.

[0028] (2) Capable of monitoring multiple anomalies: The present invention combines the reconstruction error and its rate of change to comprehensively determine whether to issue a fault warning, thereby avoiding the situation where some faults cannot be monitored due to judgment based solely on the reconstruction error. This overcomes the limitations of the single monitoring variable and single threshold state assessment method in terms of fault alarm. For example, when the temperature of the transformer or its associated equipment continues to rise but does not reach the threshold set by the regulations, the transformer will not issue an alarm message according to the single threshold method. However, if the transformer temperature continues to rise, the transformer will usually fail.

[0029] (3) The threshold is determined more accurately: The present invention realizes transformer fault warning through the threshold determined by the random forest model, obtains the reconstruction error and its relative change rate based on the historical monitoring data containing fault data after labeling, and uses the random forest training model to separate normal and abnormal data. The appropriate threshold is determined according to the category score output by the model to distinguish between normal and abnormal. The determined threshold can better distinguish between normal and fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of a transformer fault early warning method based on a stacked autoencoder network in an embodiment;

[0031] Figure 2 A schematic diagram of a stacked autoencoder network in an embodiment;

[0032] Figure 3 It is an equivalent diagram of the transformer stacking autoencoder network principle in the embodiment. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0034] Example 1

[0035] In response to the problems existing in the aforementioned prior art, this embodiment provides a transformer fault warning method based on a stacked autoencoder network. By comprehensively analyzing the electrical and non-electrical monitoring variables of the transformer and the monitoring variables of the equipment associated with it, the intrinsic relationship between the multi-dimensional monitoring variables is effectively learned, thereby overcoming the limitations of the traditional single-variable single-threshold state evaluation and fault alarm method.

[0036] See also Figure 1 , transformer fault warning based on stacked autoencoder network is a safety status assessment that integrates electrical variables and non-electrical variables. First, multi-dimensional monitoring data of the transformer and its associated equipment during normal operation is collected, and the multi-dimensional monitoring data is input into the model for training to obtain a nonlinear functional relationship model between the multi-dimensional monitoring data during normal operation of the transformer, that is, to build a stacked autoencoder network model. Then, the nonlinear relationship model is used to output the input multi-dimensional real-time monitoring data, and the reconstruction error between the output result and the input data, as well as the relative change rate of the reconstruction error, is calculated to determine whether the transformer is faulty. The main steps of this method include:

[0037] Step S1, select multi-dimensional monitoring variables of the transformer and its related equipment, and clean and normalize these data. Specifically including:

[0038] Step S101, in the stacked autoencoder (SAE) model, the electrical and non-electrical monitoring variables of the transformer are simultaneously used as inputs, aiming to associate these variables, and predict potential electrical faults in advance by monitoring the slow changes and trends of non-electrical variables, thereby preventing losses caused by sudden changes in electrical variables. In addition, this method can also avoid the problem of a single variable being abnormal and its monitoring device failing.

[0039] The reason for using the non-electrical monitoring variables of transformer-related equipment and the monitoring variables of the transformer itself as model inputs at step S102 is that when the electrical signal of the associated equipment detects an abnormality, the fault has often occurred and may affect the transformer. Therefore, by monitoring the change trend of non-electrical variables such as temperature, the transformer can be warned in advance to prevent abnormalities of other equipment from affecting the transformer.

[0040] Step S103, combining steps S101 and S102, select monitoring variables of the transformer and its related equipment as shown in Table 1.

[0041] Table 1 Key monitoring variables for transformers and related equipment

[0042]

[0043] Step S104, considering that some monitoring variable data are missing, the mean before and after the missing value is used for filling. In view of the large difference in the value ranges of different monitoring variables, the historical monitoring data of the transformer under normal operation is selected for normalization processing to serve as the training data of the SAE model.

[0044] Step S2, constructing a transformer stacked autoencoder network (SAE) model. Specifically including:

[0045] a) The principle of autoencoding network;

[0046] The autoencoder learns the feature representation of the input data through two parts: the encoder and the decoder. The encoder is responsible for compressing the input data into a lower-dimensional hidden layer representation to capture the structure and features of the input data. The decoder maps the hidden layer representation back to the input space to reconstruct the original input data. The difference between the input and output is called the reconstruction error. The parameters of the encoder and decoder are continuously adjusted through back propagation to make the reconstruction value As close as possible to the original input value x, so as to achieve effective reconstruction of the data. The encoding and decoding process of the autoencoder can be expressed as

[0047] y=f y (W y x+b y )(1)

[0048]

[0049] In the formula, x, are the input and output values ​​of the autoencoder respectively; y is the output value of the encoder and the input value of the decoder, that is, the hidden layer representation; f y (·), are the activation functions of the encoder and decoder respectively; W y and b y are the weight and bias of the decoder respectively; and are the weight and bias of the encoder respectively.

[0050] The goal of this model is to minimize the reconstruction error by continuously back-propagating the optimization parameters. The required optimization parameters and reconstruction error are: Among them, θ is the autoencoder parameter, L(·) is the loss function for selecting the mean square error, and E is the minimum reconstruction error.

[0051] b) The principle of stacked autoencoder networks;

[0052] Stacked Auto-Encoder (SAE) is an unsupervised deep learning network that replaces the single-layer encoder and decoder in the autoencoder model with multiple neural layers. The encoding and decoding of the SAE network are symmetrical in structure, such as Figure 2 As shown in the figure. SAE uses the Sigmoid activation function in the encoding and decoding process. In the encoding process, the input data is forward propagated through multiple layers of fully connected layers to obtain low-dimensional features; in the decoding process, the low-dimensional features are input into the decoder symmetrical to the encoder, and finally the reconstructed value of the input data is obtained. The mean square error between the input data and the output reconstructed data of the SAE network is used as the loss function, and the parameters of SAE are optimized and updated again through the back-propagation algorithm, and finally the SAE network model that meets the error requirements is obtained.

[0053] c) Construction of the transformer stacked autoencoder network model.

[0054] The internal principle equivalent diagram of the SAE network model constructed using the transformer-related multidimensional monitoring variables, that is, the nonlinear functional relationship model between the monitoring variables, is as follows: Figure 3 As shown. The properties of the multidimensional monitoring data obtained by the model and the relationship between different monitoring variables are reflected in the parameters of the transformer SAE model. Since the data input into the SAE model are all monitoring data when the transformer is in normal operation, the model can characterize the functional relationship and data characteristics between different monitoring variables under normal operating conditions of the transformer. When real-time monitoring data is input into the transformer SAE model, if the transformer fails or is about to fail, due to abnormalities inside the transformer, the internal balance relationship of different monitoring variables during normal operation is destroyed, and the reconstruction error between the reconstructed data output by the model and the input data will continue to increase. In addition, when the transformer fails, the rate of change of the reconstruction error also increases significantly. Therefore, the reconstruction error and its rate of change can be used to conduct real-time safety assessment of the transformer and realize transformer failure warning.

[0055] For the monitoring variables of transformer-related equipment, when transformer-related equipment such as the 110 kV bus shows a fault trend, the temperature of the busbar gas chamber increases, causing the functional relationship between the temperatures of multiple equipment to be destroyed. Therefore, if the reconstruction error of the SAE model increases, the rate of change of the reconstruction error also increases. By monitoring these changes, the transformer can be warned in advance and it can be checked whether the failure of the associated equipment will affect the normal operation of the transformer.

[0056] Step S3, based on the SAE model, calculate the reconstruction error, and then determine the reconstruction error threshold of the transformer fault alarm through random forest algorithm training. Specifically including:

[0057] Step S301, calculating the reconstruction error.

[0058] When the reconstruction error and its change rate obtained by the model exceed the preset threshold, it is determined that the transformer has a tendency to fail and an alarm signal is issued. When the transformer alarm occurs, the error between the input monitoring variable and its output reconstruction value is determined one by one. k , in order to determine the specific monitoring variables that cause the fault and its cause, so that corresponding measures can be taken more quickly to avoid major power outages. Reconstruction error E and single variable reconstruction error ξ k As shown in the following formula:

[0059]

[0060]

[0061] Step S302, processing the reconstruction error;

[0062] The obtained reconstruction error is filtered in turn. Since the reconstruction error contains noise and has a certain volatility, this paper first uses the mean filtering method to filter the reconstruction error to obtain E f . Specifically, a filter window is set, the data in the window is averaged, and the average is used as the data value of the window center point. Then the filter window is moved back one unit, and the process is repeated. To ensure the integrity of the data, there is overlap between each window, and the data in half of the window at both ends within the monitoring time retains the original value. When the transformer is abnormal, the reconstruction error will increase significantly, but considering that the reconstruction error is volatile, only considering the reconstruction error may lead to misjudgment. This embodiment evaluates the transformer state and predicts faults by judging the reconstruction error after filtering and its rate of change.

[0063] Step S303: determining a reconstruction error threshold for transformer status assessment and fault alarm.

[0064] In order to obtain the standard for transformer fault judgment, the reconstruction error E of the historical monitoring data containing fault data is first obtained. f and its relative rate of change D i , and mark it with 0 if the device is normal or 1 if it is abnormal f and D i, a random forest training model is used, that is, each decision tree in the random forest randomly samples from the historical monitoring data set and its corresponding 0 / 1 mark for training, and obtains the random forest classification model of the reconstruction error and its change rate. When the reconstruction error and its relative change rate of the measured value of the monitoring variable are respectively input into the corresponding training model, each decision tree will output 0 / 1 to indicate whether the transformer is operating abnormally. The mode of the output results of multiple decision trees is the final output result of the random forest. If the final result is 1, the reconstruction error or its relative change rate is judged to be abnormal. And as long as one of the random forest models outputs 1 for the reconstruction error and the rate of change of the reconstruction error, an alarm will be issued to the transformer.

[0065] Step S4, realizing transformer status assessment and fault warning functions.

[0066] Only a single threshold is used to determine whether the equipment is abnormal. When the alarm information appears, the fault condition has lasted for a period of time, which often leads to a delay in the alarm. It may be too late for maintenance personnel to know the abnormality and take measures. When the temperature of the transformer or its associated equipment continues to rise but does not reach the threshold set by the regulations, the transformer will not have an alarm information according to the single threshold method. However, if the transformer temperature continues to rise, the transformer will usually fail.

[0067] In view of the above problems, this embodiment uses a fault warning based on a stacked autoencoder network to detect fault trends early and issue an alarm. When the temperature of the transformer or its associated equipment shows an increasing trend, its reconstruction error and its relative rate of change gradually increase, and the relative rate of change of the reconstruction error will suddenly increase. When the reconstruction error or its rate of change exceeds the threshold value obtained by training in this paper, an alarm signal will be issued to the transformer. Therefore, an early alarm can be issued to improve the safety of the system.

[0068] In summary, this method has the following characteristics:

[0069] (1) The transformer model based on the stacked autoencoder network uses the electrical and non-electrical multidimensional monitoring data of normally operating transformers and their related equipment to construct a nonlinear functional relationship model between the multidimensional monitoring variables associated with the transformer, avoiding the artificial generation of fault data. The transformer status assessment and fault diagnosis can be performed only with normal data.

[0070] (2) The stacked autoencoder model comprehensively considers the relationship between the transformer's own electrical monitoring variables and non-electrical monitoring variables, and links the electrical variables with the non-electrical variables. This model links the transformer's non-electrical variables with the non-electrical monitoring variables of its associated equipment, and fully considers the relationship between the transformer and other equipment's non-electrical monitoring variables.

[0071] (3) The transformer, an important equipment in the power system, is modeled through a deep learning stacked autoencoder network, which overcomes the limitations of the single monitoring variable and single threshold state assessment method in fault alarm.

[0072] Example 2

[0073] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the transformer fault warning method based on a stacked autoencoding network as described in Example 1.

[0074] Example 3

[0075] This embodiment provides a computer-readable storage medium, including one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the transformer fault warning method based on a stacked autoencoding network as described in Example 1.

[0076] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A transformer fault early warning method based on a stacked autoencoder network, characterized in that: The steps include: Step S1, obtaining historical values ​​of monitoring variables including transformer variables, environmental variables, cable incoming line variables and busbar variables; Step S2, training a stacked autoencoder network based on the historical values ​​of the monitoring variables, reconstructing the measured values ​​of the monitoring variables obtained in real time, and calculating the reconstruction error; Step S3, performing mean filtering and window filtering in the time dimension on the reconstruction error; Step S4, determining whether the reconstruction error and its rate of change are greater than a threshold value pre-determined by the random forest model, thereby realizing transformer fault early warning.

2. A transformer fault early warning method based on a stacked autoencoder network according to claim 1, characterized in that: The training process of the stacked autoencoder network includes the following steps: Step S201, obtaining multiple groups of historical values ​​of monitoring variables for training and performing data cleaning and normalization processing; Step S202, inputting the multiple groups of monitoring variable historical values ​​obtained in step S201 into the encoder in the autoencoding network, and using the output of the encoder as the input of the decoder in the autoencoding network to obtain reconstructed monitoring variable values; Step S203, based on the historical values ​​of the monitoring variables used for training and the reconstructed monitoring variable values, a reconstruction error is calculated, and the autoencoder network is trained with the goal of minimizing the reconstruction error.

3. A transformer fault early warning method based on a stacked autoencoder network according to claim 1, characterized in that: In step S4, the threshold determination process includes the following steps: Step S401, obtaining multiple groups of reconstruction errors and corresponding change rates in fault state and normal state, and marking them according to normal or fault; Step S402, using the labeled multiple groups of reconstruction errors and the corresponding change rates, using a random forest model to obtain the normal and fault category scores, and obtain the threshold.

4. A transformer fault early warning method based on a stacked autoencoder network according to claim 1, characterized in that: In the step S4, if the reconstruction error is greater than the corresponding threshold or the rate of change of the reconstruction error is greater than the corresponding threshold, a transformer fault warning signal is issued.

5. A transformer fault early warning method based on a stacked autoencoder network according to claim 1, characterized in that: The transformer variables include high-voltage side current value, core grounding current, clamp grounding current, oil level height, oil level ratio, oil temperature, main transformer temperature, main ventilator temperature and main ventilator humidity.

6. A transformer fault early warning method based on a stacked autoencoder network according to claim 1, characterized in that: The environmental variables include ambient temperature, ambient humidity and water level.

7. A transformer fault early warning method based on a stacked autoencoder network according to claim 1, characterized in that: The cable inlet variables include busbar air chamber temperature, switch air chamber temperature, line transformer air chamber temperature and line cable air chamber temperature.

8. A transformer fault early warning method based on a stacked autoencoder network according to claim 1, characterized in that: The busbar variables include busbar gas chamber temperature, switch gas chamber temperature and line cable gas chamber temperature.

9. An electronic device, characterized in that: include: One or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the transformer fault early warning method based on a stacked autoencoder network as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: It comprises one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs comprise instructions for executing a transformer fault early warning method based on a stacked autoencoder network as claimed in any one of claims 1 to 8.

Images