Transformer outgoing line part partial discharge diagnosis method and device based on random forest

Abstract

The application provides a transformer outgoing line part partial discharge diagnosis method and device based on a random forest, and the method comprises the following steps: obtaining measurement data of each sensor in a transformer to be measured for several days; measurement data of each sensor in a target time interval is predicted through a sliding window method and a random forest model, and a characteristic quantity is calculated; each characteristic quantity is subjected to standardization processing, and a characteristic parameter is calculated according to a standardization processing result and in combination with a sensor type weight; each characteristic parameter is subjected to clustering, and whether the transformer to be measured exists partial discharge is judged. The application predicts measurement data in a target interval through a sliding window and a random forest prediction model, calculates a characteristic quantity, obtains a characteristic parameter of the transformer in each day in the target interval through standardization processing of the characteristic quantity, realizes data fusion of various types of sensors, and the judgment result is beneficial to horizontal comparison of each transformer.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of transformer discharge detection, and in particular to a transformer outgoing line part partial discharge diagnosis method and device based on random forest. BACKGROUND

[0002] Large power transformers are important components of modern power systems, and their safe and stable operation is of great significance to ensuring the effective transmission of electric energy. The outgoing line part of a large transformer has complex structure, high voltage level and concentrated field strength, and is prone to insulation defects and partial discharge faults during operation. Currently, common transformer partial discharge monitoring methods include ultrasonic, ultra-high frequency and high-frequency current, etc. These technologies monitor the parameter changes of each sensor in real time, and compare the changes with historical data to identify partial discharge. However, in actual situations, different types of sensors monitor different types of signals, and the results measured by these technical solutions are difficult to compare horizontally, making it difficult to extract characteristic parameters and to realize data fusion of multiple sensors. In addition, differences in internal insulation structure make it difficult to compare multiple devices horizontally when partial discharge occurs at the outgoing line part. SUMMARY

[0003] The present application provides a transformer outgoing line part partial discharge diagnosis method and device based on random forest, which realizes data fusion of multiple types of sensors to solve the technical problem of how to diagnose multiple transformers and compare them horizontally.

[0004] To solve the above technical problems, the present application provides a transformer outgoing line part partial discharge diagnosis method based on random forest, comprising:

[0005] Obtaining measurement data of each sensor in the transformer to be tested for several days; wherein the types of sensors include ultrasonic sensors, ultra-high frequency current sensors and high-frequency current sensors;

[0006] Predicting the predicted values of the measurement data of each sensor in the target time interval by the sliding window method and the preset random forest prediction model, and calculating the characteristic quantity corresponding to each day in the target time interval for each sensor;

[0007] Calculating the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position of all characteristic quantities, standardizing each characteristic quantity according to the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position, and then calculating the characteristic parameter corresponding to each day in the target time interval for the transformer to be tested according to the standardization result and the weight corresponding to the type of sensor;

[0008] The feature parameters are clustered, and it is judged whether the outgoing line part of the transformer to be measured has partial discharge according to the clustering result.

[0009] As a preferred solution, the prediction values of the measurement data of each sensor target time interval are respectively predicted by the sliding window method and the preset random forest prediction model, specifically:

[0010] The measurement data of the first preset time interval is predicted by the preset random forest base model, and the prediction value of the target time interval measurement data is obtained based on the sliding window method;

[0011] When the target time interval is greater than or equal to the second preset time interval, the model parameters of the random forest base model are trained by the prediction value of the measurement data of the second preset time interval, and a trained random forest prediction model is obtained;

[0012] The prediction values of the measurement data of each sensor target time interval are respectively predicted by the sliding window method and the trained random forest prediction model.

[0013] As a preferred solution, each feature quantity is standardized, specifically:

[0014] The value of each feature quantity after standardization is calculated according to the following formula

[0015]

[0016] Where n is the sensor number, k is the kth day, and Ay n,m is the median, Ay n,25％ is the quartile point not greater than the 75% position, and Ay n,75％ is the quartile point not greater than the 25% position.

[0017] As a preferred solution, the corresponding feature parameters of the transformer to be measured on each day in the target time interval are calculated according to the standardization result combined with the weight corresponding to the sensor type, specifically:

[0018] The feature parameters Y n are calculated according to the following formula:

[0019]

[0020] Where K is the number of days corresponding to the target time interval, and w i is the weight corresponding to the sensor type i.

[0021] As a preferred solution, the corresponding feature quantities of each sensor on each day in the target time interval are calculated, specifically:

[0022] Subtract the corresponding actual value from the predicted value of each sensor in each day of the target time interval respectively, and take the difference as the characteristic quantity.

[0023] Correspondingly, the embodiment of the present application also provides a transformer outgoing line part partial discharge diagnosis device based on random forest, comprising a data acquisition module, a prediction module, a characteristic parameter calculation module and a diagnosis module; wherein,

[0024] The data acquisition module is used to acquire measurement data of each sensor in the transformer to be measured for several days; wherein, the types of sensors include ultrasonic sensors, ultrahigh frequency current sensors and high frequency current sensors;

[0025] The prediction module is used to predict the predicted values of the measurement data of each sensor in the target time interval by the sliding window method and the preset random forest prediction model, and calculate the corresponding characteristic quantity of each sensor in each day of the target time interval;

[0026] The characteristic parameter calculation module is used to calculate the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position of all characteristic quantities, to perform standardization processing on each characteristic quantity according to the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position, and to calculate the corresponding characteristic parameter of the transformer to be measured in each day of the target time interval according to the standardization processing result and the weight corresponding to the type of sensor;

[0027] The diagnosis module is used to perform clustering processing on each characteristic parameter, and to judge whether there is partial discharge in the outgoing line part of the transformer to be measured according to the clustering processing result.

[0028] As a preferred scheme, the prediction module predicts the predicted values of the measurement data of each sensor in the target time interval by the sliding window method and the preset random forest prediction model, specifically:

[0029] The prediction module predicts the measurement data of the first preset time interval by the preset random forest base model, and obtains the predicted values of the measurement data of the target time interval based on the sliding window method;

[0030] When the target time interval is greater than or equal to the second preset time interval, the model parameters of the random forest base model are trained by the predicted values of the measurement data of the second preset time interval, and a trained random forest prediction model is obtained;

[0031] The predicted values of the measurement data of each sensor in the target time interval are predicted by the sliding window method and the trained random forest prediction model.

[0032] As a preferred solution, the feature parameter calculation module normalizes each feature quantity, specifically:

[0033] The feature parameter calculation module calculates the value of each feature quantity after normalization according to the following formula

[0034] Wherein, n is the sensor number, k is the kth day, Δy n,m is the median, Δy n,25％ is the quartile point at the position not greater than 75%, Δy n,75％ is the quartile point at the position not greater than 25%.

[0035] As a preferred solution, the feature parameter calculation module calculates the feature parameter corresponding to each day in the target time interval of the transformer to be tested according to the normalized processing result combined with the weight corresponding to the sensor type, specifically:

[0036] The feature parameter calculation module calculates the feature parameter Y n according to the following formula:

[0037]

[0038] Wherein, K is the number of days corresponding to the target time interval, w i is the weight corresponding to the sensor type i.

[0039] As a preferred solution, the prediction module calculates the feature quantity corresponding to each day of each sensor in the target time interval, specifically:

[0040] The prediction module subtracts the corresponding actual value from the predicted value of each sensor in each day of the target time interval, and takes the difference as the feature quantity.

[0041] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0042] The embodiment of the present application provides a transformer outgoing line part partial discharge diagnosis method and device based on a random forest, the method comprises the following steps: obtaining measurement data of each sensor in a transformer to be measured for several days; wherein the types of the sensors comprise ultrasonic sensors, ultrahigh frequency current sensors and high frequency current sensors; the prediction values of the measurement data of each sensor in a target time interval are predicted through a sliding window method and a preset random forest prediction model respectively, and the corresponding characteristic quantity of each sensor in each day in the target time interval is calculated; the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position of all characteristic quantities are calculated, the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position are used for standardizing each characteristic quantity, and then the corresponding characteristic parameter of the transformer to be measured in each day in the target time interval is calculated according to the standardization processing result and the weight corresponding to the sensor type; the characteristic parameters are subjected to clustering processing, and whether the outgoing line part of the transformer to be measured exists partial discharge is judged according to the clustering processing result. The measurement data of the target interval is predicted through the sliding window and the random forest prediction model, the characteristic quantity is calculated, the characteristic parameter of the transformer in each day in the target interval is obtained through the standardization processing of the characteristic quantity, the data fusion of various types of sensors is realized, the judgment result is beneficial to the horizontal comparison of various transformers; and the types of the sensors in the embodiment of the present application comprise ultrasonic sensors, ultrahigh frequency current sensors and high frequency current sensors, after the characteristic quantity is obtained and subjected to standardization processing, the characteristic parameter is calculated in combination with the corresponding weight, whether the transformer exists partial discharge is diagnosed through the clustering method, the diagnosis method considers various types of sensors, compared with the technical scheme that the single sensor monitoring and the signal are compared with the historical data in the prior art, the present application has higher accuracy; considering that there may be a too large difference in the amplitude of the measurement signal in the actual measurement process, the present application adopts the median and the quartile point as the calculation reference, can avoid the interference of the maximum value or the minimum value, and realizes effective data fusion. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 : the flowchart of one embodiment of the transformer outgoing line part partial discharge diagnosis method provided by the present application based on a random forest.

[0044] Figure 2 : the principle diagram of one embodiment of the sliding window method provided by the present application.

[0045] Figure 3 : the characteristic parameter diagram of the transformer with partial discharge provided by the present application.

[0046] Figure 4 : the characteristic parameter diagram of the transformer without partial discharge provided by the present application.

[0047] Figure 5 Figure 1 is a structural schematic diagram of an embodiment of a transformer outgoing line part partial discharge diagnosis device provided by the present application based on a random forest. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0049] Embodiment one:

[0050] Please refer to Figure 1 , Figure 1 A transformer outgoing line part partial discharge diagnosis method provided by the present application based on a random forest includes steps S1 to S4.

[0051] Step S1, obtaining measurement data of each sensor in a to-be-measured transformer for several days; wherein the types of the sensors include ultrasonic sensors, ultrahigh frequency current sensors and high frequency current sensors.

[0052] The transformer outgoing line part partial discharge diagnosis method provided by the present application based on a random forest can be applied to a server and the like, the to-be-measured transformer includes but is not limited to a large power transformer, and can be one or more, the sensors are mainly sensors for detecting transformer partial discharge, and there is at least one sensor in number and type on the transformer. The server and the like can call the measurement data obtained by the sensors, which is used for prediction of target time interval measurement data.

[0053] Step S2, predicting the predicted values of the measurement data of each sensor in a target time interval by a sliding window method and a preset random forest prediction model, and calculating the corresponding feature quantity of each sensor in each day in the target time interval.

[0054] In the present embodiment, the random forest is a decision-making and prediction model including multiple decision trees. In the prediction process, the model can take N-day data of a sensor as input and N+1-day data as output, and can predict the results of the sensor for several days in the future by loop stepping (preferably a sliding window method, which can effectively improve the diagnosis efficiency in the present embodiment). It has the advantages of high efficiency, no need to reduce dimension for processing high-dimensional features, and can effectively reduce errors for class imbalance data.

[0055] The prediction values of the measurement data of each sensor in the target time interval are respectively predicted by the sliding window method and the preset random forest prediction model, specifically as follows:

[0056] The measurement data in the first preset time interval is predicted by the preset random forest base model, and the prediction value of the measurement data in the target time interval is obtained based on the sliding window method;

[0057] When the target time interval is greater than or equal to the second preset time interval, the model parameters of the random forest base model are trained by the prediction value of the measurement data in the second preset time interval, and a trained random forest prediction model is obtained;

[0058] The prediction values of the measurement data of each sensor in the target time interval are respectively predicted by the sliding window method (refer to Figure 2 ) and the trained random forest prediction model.

[0059] Preferably, the first preset time interval can be 5 days, and the second preset time interval can be 15 days. The total number of sensors is N (the value range is 1 to n), the monitoring time is in days, and the number of days of continuous monitoring is K (the value range is 1 to k), and K is the target time interval. The specific prediction process of this embodiment is as follows:

[0060] Each sensor's first preset time interval monitoring (measured) data y n,1 ~ y n,5 is extracted and input into the random forest base model, and the prediction data y n,6 of the sixth day is obtained by predicting the random forest base model.

[0061] Taking time as the sliding basis, a 1-day selection window is selected and slid forward, the 2nd-5th measured data y n,2 ~ y n,5 of the sensor are traversed, the prediction data y n,6 of the sixth day is obtained as the input of the model, and the prediction data y n,7 of the seventh day is obtained. Continue to slide the selection window to obtain the prediction data of the first to Kth days of the sensor.

[0062] When the target time interval K is greater than or equal to the second preset time interval (15 days), the measurement data of the previous 15 days can be used as input to the random forest base model to update the model parameters, obtain the trained random forest prediction model, and then update the time index (i.e., the predicted day k). The trained random forest prediction model is then used to complete the prediction of all monitoring times of all sensors. When K is less than 15 days, no update is needed. If K is much greater than 15 days, the model parameters of the random forest model can be updated every 15 days to avoid the problem of increased prediction bias due to unupdated model parameters.

[0063] Furthermore, the calculation of the feature quantity corresponding to each sensor for each day in the target time interval specifically involves:

[0064] The predicted value of each sensor for each day within the target time interval is subtracted from the corresponding actual value, and the difference is used as the feature quantity. Specifically, the feature quantity Δy n,k =|y n,k (Prediction)-y n,k (Actual)|.

[0065] Step S3: Calculate the median, quartiles at the 75th percentile, and quartiles at the 25th percentile for all features. Based on the median, quartiles at the 75th percentile, and quartiles at the 25th percentile, standardize each feature. Then, based on the standardization results and the weights corresponding to the sensor type, calculate the feature parameters of the transformer under test for each day in the target time interval.

[0066] In this embodiment, step S3 uses a robust normalization processing method. This is because, considering the actual measurement process, there may be situations where the amplitude of the measurement signal differs too much (this situation may be caused by different types of sensors or other factors). In order to avoid the problem of data fusion not being possible due to large amplitude differences, this embodiment uses the median and two quartiles as the calculation benchmark, which can effectively avoid interference from maximum or minimum values.

[0067] Preferably, the standardization process for each feature quantity specifically involves:

[0068] Calculate the standardized values ​​of each characteristic quantity according to the following formula.

[0069]

[0070] Where n is the sensor number, k is the kth day, and Δy n,m Let Δy be the median.n,25％ is a quartile point not greater than 75% of the position, and Ay n,75％ is a quartile point not greater than 25% of the position.

[0071] The feature parameter corresponding to each day in the target time interval of the transformer under test is calculated according to the standardized processing result combined with the weight corresponding to the sensor type, specifically:

[0072] The feature parameter Y is calculated according to the following formula: n

[0073]

[0074] wherein K is the number of days corresponding to the target time interval, w i is the weight corresponding to the sensor type i. In this embodiment, the weight of the ultrasonic sensor and the high-frequency sensor is preferably 0.4. The weight of the very high frequency sensor is preferably 0.2 due to its greater susceptibility to spatial electromagnetic wave interference. By robustly standardizing the data of different types of sensors according to the embodiments of the present application, the data fusion of different types of sensors can be easily realized. Considering the influence of external interference on the sensor data, different weights are assigned to different types of sensors to improve the calculation accuracy of the feature parameter.

[0075] Step S4, clustering processing is performed on each feature parameter, and it is determined whether there is partial discharge at the outgoing line part of the transformer under test according to the clustering processing result.

[0076] In this embodiment, existing clustering methods (or methods of existing classification models) can be used to process and analyze each feature parameter, and each feature parameter is divided into two categories: "partial discharge exists" and "no partial discharge", thereby realizing the comparison of different transformers.

[0077] The transformer outgoing line part partial discharge diagnosis method based on random forest of the present application will be described below taking 10 transformers of different locations and different phases as examples. Specifically, in this embodiment, the number of sensors is 12 ultrasonic sensors, 4 very high frequency sensors and 2 high frequency current sensors (n is 18), and the number of continuous monitoring days is 120 (k is 120).

[0078] The transformer outgoing line part partial discharge diagnosis method based on random forest provided by this embodiment can calculate the feature parameter Y n in 120 days, as shown in Figure 3 (partial discharge exists) and Figure 4 (no partial discharge).

[0079] Correspondingly, referring to Figure 5 ​The embodiment of the present application also provides a transformer outgoing line part partial discharge diagnosis device based on a random forest, comprising a data acquisition module 101, a prediction module 102, a feature parameter calculation module 103 and a diagnosis module 104; wherein,

[0080] The data acquisition module 101 is used for acquiring measurement data of each sensor in a to-be-tested transformer for several days; wherein the types of the sensors include ultrasonic sensors, ultrahigh frequency current sensors and high frequency current sensors.

[0081] The prediction module 102 is used for predicting the predicted values of the measurement data of each sensor in a target time interval by a sliding window method and a preset random forest prediction model, and calculating the corresponding feature quantity of each sensor in each day in the target time interval.

[0082] The feature parameter calculation module 103 is used for calculating the median, the quartile point not greater than the 75% position and the quartile point not greater than the 25% position of all feature quantities, performing standardization processing on each feature quantity according to the median, the quartile point not greater than the 75% position and the quartile point not greater than the 25% position, and then calculating the corresponding feature parameter of the to-be-tested transformer in each day in the target time interval according to the standardization processing result and the weight corresponding to the sensor type.

[0083] The diagnosis module 104 is used for performing clustering processing on each feature parameter, and judging whether the outgoing line part of the to-be-tested transformer has partial discharge according to the clustering processing result.

[0084] As a preferred implementation manner, the prediction module 102 predicts the predicted values of the measurement data of each sensor in a target time interval by a sliding window method and a preset random forest prediction model, specifically as follows:

[0085] The prediction module 102 predicts the measurement data of a first preset time interval by a preset random forest base model, and obtains the predicted values of the measurement data of the target time interval based on a sliding window method;

[0086] When the target time interval is greater than or equal to a second preset time interval, the model parameters of the random forest base model are trained by the predicted values of the measurement data of the second preset time interval, and a trained random forest prediction model is obtained;

[0087] The predicted values of the measurement data of each sensor in a target time interval are predicted by a sliding window method and the trained random forest prediction model.

[0088] As a preferred implementation manner, the feature parameter calculation module 103 performs standardization processing on each feature quantity, specifically as follows:

[0089] The feature parameter calculation module 103 calculates the normalized value of each feature according to the following formula

[0090]

[0091] wherein n is the sensor number, k is the kth day, Δy n,m is the median, Δy n,25％ is the quartile point at the position not greater than 75%, Δy n,75％ is the quartile point at the position not greater than 25%.

[0092] As a preferred embodiment, the feature parameter calculation module 103 calculates the feature parameter of the transformer to be measured corresponding to each day in the target time interval according to the normalized processing result combined with the weight corresponding to the sensor type, specifically:

[0093] The feature parameter calculation module 103 calculates the feature parameter Y n according to the following formula:

[0094]

[0095] wherein K is the number of days corresponding to the target time interval, w i is the weight corresponding to the sensor type i.

[0096] As a preferred embodiment, the prediction module 102 calculates the feature corresponding to each sensor in each day in the target time interval, specifically:

[0097] The prediction module 102 subtracts the actual value corresponding to each sensor in each day in the target time interval from the predicted value, and takes the difference as the feature.

[0098] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0099] The embodiment of the present application provides a transformer outgoing line part partial discharge diagnosis method and device based on a random forest, the method comprises the following steps: obtaining measurement data of each sensor in a transformer to be measured for several days; wherein the types of the sensors comprise ultrasonic sensors, ultrahigh frequency current sensors and high frequency current sensors; the predicted values of the measurement data of each sensor in a target time interval are predicted through a sliding window method and a preset random forest prediction model respectively, and the corresponding characteristic quantity of each sensor in each day in the target time interval is calculated; the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position of all characteristic quantities are calculated, the standardization processing is performed on each characteristic quantity according to the median, the quartile point not greater than 75% position and the quartile point not greater than 25% position, and then the corresponding characteristic parameter of the transformer to be measured in each day in the target time interval is calculated according to the standardization processing result and the weight corresponding to the sensor type; the characteristic parameters are subjected to clustering processing, and whether the outgoing line part of the transformer to be measured has partial discharge is judged according to the clustering processing result. The present application predicts the measurement data in the target interval through the sliding window and the random forest prediction model, calculates the characteristic quantity, obtains the characteristic parameter of the transformer in each day in the target interval through the standardization processing of the characteristic quantity, realizes the data fusion of various types of sensors, and the judgment result is beneficial to the horizontal comparison of various transformers; and the types of the sensors in the embodiment of the present application comprise ultrasonic sensors, ultrahigh frequency current sensors and high frequency current sensors, after the characteristic quantity is obtained and subjected to standardization processing, the characteristic parameter is calculated in combination with the corresponding weight, whether the transformer has partial discharge is diagnosed through the clustering method, the diagnosis method considers various types of sensors, compared with the single sensor monitoring and the technical scheme of comparing the signal with historical data in the prior art, has higher accuracy; considering that there may be a case that the measurement signal amplitude difference is too large in the actual measurement process, the present application adopts the median and the quartile point as the calculation reference, can avoid the interference of the maximum value or the minimum value, and realizes effective data fusion.

[0100] The above-described specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application, and it should be understood that the above-described specific embodiments are only specific embodiments of the present application and are not used to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for diagnosing partial discharge at transformer outgoing lines based on random forest, characterized in that, include: Acquire measurement data for several days from each sensor in the transformer under test; the types of sensors include ultrasonic sensors, ultra-high frequency current sensors, and high frequency current sensors. The predicted values ​​of the measurement data of each sensor in the target time interval are predicted by the sliding window method and the preset random forest prediction model, and the feature quantity corresponding to each sensor for each day in the target time interval is calculated. Calculate the median, quartiles at the 75th percentile, and quartiles at the 25th percentile for all features. Based on the median, quartiles at the 75th percentile, and quartiles at the 25th percentile, standardize each feature. Then, based on the standardization results and the weights corresponding to the sensor type, calculate the feature parameters of the transformer under test for each day in the target time interval. Clustering is performed on each of the aforementioned characteristic parameters, and the presence of partial discharge at the output of the transformer under test is determined based on the clustering results. The method of predicting the target time interval measurement data of each sensor using the sliding window method and a preset random forest prediction model is as follows: The measurement data in the first preset time interval is predicted using a preset random forest basic model, and the predicted value of the measurement data in the target time interval is obtained based on the sliding window method. When the target time interval is greater than or equal to the second preset time interval, the model parameters of the random forest base model are trained using the predicted values ​​of the measurement data in the second preset time interval to obtain a trained random forest prediction model. Using the sliding window method and the trained random forest prediction model, the predicted values ​​of the measurement data of each sensor target time interval are predicted respectively. The calculation of the feature quantity corresponding to each sensor for each day in the target time interval is specifically as follows: The predicted value of each sensor for each day in the target time interval is subtracted from the corresponding actual value, and the difference is used as the feature quantity.

2. The method for diagnosing partial discharge at transformer outgoing lines based on random forest as described in claim 1, characterized in that, The standardization process for each feature is specifically as follows: Calculate the standardized values ​​of each characteristic quantity according to the following formula. : ； Where n is the sensor number, and k is the kth day. The median is... Interquartiles are those located at no more than 75% of their positions. The interquartiles are those located at positions not exceeding 25%.

3. The method for diagnosing partial discharge at transformer outgoing lines based on random forest as described in claim 2, characterized in that, The characteristic parameters of the transformer under test for each day in the target time interval are calculated based on the standardization results and the weights corresponding to the sensor types. Specifically: The characteristic parameter Y is calculated according to the following formula. n : ； Where K is the number of days corresponding to the target time interval. The weights are for sensor type i.

4. A partial discharge diagnostic device for transformer outgoing lines based on random forest, characterized in that, It includes a data acquisition module, a prediction module, a feature parameter calculation module, and a diagnosis module; among which, The data acquisition module is used to acquire measurement data from various sensors in the transformer under test over several days; the types of sensors include ultrasonic sensors, ultra-high frequency current sensors, and high frequency current sensors. The prediction module is used to predict the predicted values ​​of the measurement data of each sensor in the target time interval using the sliding window method and the preset random forest prediction model, and to calculate the feature quantity corresponding to each sensor for each day in the target time interval. The feature parameter calculation module is used to calculate the median, the quartiles at no more than 75% of the positions, and the quartiles at no more than 25% of the positions for all feature quantities. Based on the median, the quartiles at no more than 75% of the positions, and the quartiles at no more than 25% of the positions, each feature quantity is standardized. Then, based on the standardization results and the weights corresponding to the sensor type, the feature parameters of the transformer under test for each day in the target time interval are calculated. The diagnostic module is used to perform clustering processing on each of the characteristic parameters, and to determine whether there is partial discharge at the output part of the transformer under test based on the clustering processing results. The prediction module predicts the target time interval measurement data of each sensor using the sliding window method and a preset random forest prediction model, respectively. Specifically: The prediction module predicts the measurement data of the first preset time interval using a preset random forest basic model, and obtains the predicted value of the measurement data of the target time interval based on the sliding window method. When the target time interval is greater than or equal to the second preset time interval, the model parameters of the random forest base model are trained using the predicted values ​​of the measurement data in the second preset time interval to obtain a trained random forest prediction model. Using the sliding window method and the trained random forest prediction model, the predicted values ​​of the measurement data of each sensor target time interval are predicted respectively. The prediction module calculates the feature quantity corresponding to each sensor for each day in the target time interval, specifically: The prediction module subtracts the corresponding actual value from the predicted value of each sensor for each day in the target time interval, and uses the difference as the feature quantity.

5. The partial discharge diagnostic device for transformer outgoing lines based on random forest as described in claim 4, characterized in that, The feature parameter calculation module performs standardization processing on each feature quantity, specifically as follows: The feature parameter calculation module calculates the standardized value of each feature quantity according to the following formula. : ； Where n is the sensor number, and k is the kth day. The median is... Interquartiles are those located at no more than 75% of their positions. The interquartiles are those located at positions not exceeding 25%.

6. The partial discharge diagnostic device for transformer outgoing lines based on random forest as described in claim 5, characterized in that, The feature parameter calculation module calculates the feature parameters of the transformer under test for each day in the target time interval based on the standardization results and the weights corresponding to the sensor type. Specifically: The feature parameter calculation module calculates the feature parameter Y according to the following formula. n : ； Where K is the number of days corresponding to the target time interval. The weights are for sensor type i.

Images