Transformer current rising test method based on deep learning

Abstract

The invention discloses a transformer current rising test method based on deep learning, and the method comprises the following steps: collecting historical test data, and physical parameters and current operation condition data of a to-be-tested transformer; the performance of the transformer to be tested under different current and voltage conditions is predicted according to the physical parameters of the transformer to be tested by using a calculation model based on the electromagnetic theory; applying a deep learning algorithm to obtain preliminary test parameter suggestions and performance prediction based on historical data; and obtaining comprehensive test parameter suggestions and performance prediction by using a decision fusion technology. According to the method, a traditional physical model and a deep learning technology are combined, a comprehensive and accurate mode is provided to predict the performance of the transformer under specific conditions, and the method is particularly suitable for a current rising test of a high-voltage and high-power transformer. The performance of the transformer can be accurately evaluated before the transformer is put into use, the operation risk is reduced, and the reliability and safety of a power system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformers, and in particular to a transformer current boosting test method based on deep learning. Background Art

[0002] As the power system becomes increasingly complex and high-capacity, the performance and reliability of transformers, as key power transmission and distribution equipment, are crucial to the stable operation of the entire power grid. Transformer current rise test is an important performance evaluation method used to determine the behavior and response of transformers under different current and voltage conditions. This test is of great significance to ensure the safety and effectiveness of transformers in actual operation.

[0003] Traditional transformer current-up test methods rely primarily on laboratory tests and physical model calculations. While these methods are able to provide basic information about transformer performance, they often fail to fully capture the transformer's behavior under complex operating conditions. Especially when dealing with high-voltage, high-capacity transformers, traditional methods may not accurately predict their losses, thermal effects, and dielectric strength changes under extreme conditions. In addition, because the design and operating conditions of transformers can be extremely diverse, traditional physical models and test methods have limitations in personalized evaluation and rapid adaptation to new situations. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a transformer current boost test method based on deep learning.

[0005] A first aspect of the present invention provides a transformer current boost test method based on deep learning, comprising the following steps:

[0006] S1. Data collection: Collect historical test data of transformers similar to the transformer to be tested, including previous current-increasing test results; collect physical parameters of the transformer to be tested, including electrical characteristics and design parameters of the transformer to be tested; collect current operating condition data of the transformer to be tested, including temperature, humidity and load conditions of the test environment;

[0007] S2. Physical model processing: using a computational model based on electromagnetic theory, according to the input physical parameters of the transformer to be tested, the predicted theoretical performance of the transformer to be tested under different currents and voltages is calculated, wherein the predicted theoretical performance includes predicted losses, thermal effects, and insulation strength changes;

[0008] S3, deep learning processing: applying deep learning algorithms to analyze input historical test data and current operating condition data, and obtaining preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis;

[0009] S4. Decision fusion: Apply decision fusion technology to integrate the predicted theoretical performance, preliminary test parameter recommendations and preliminary performance predictions of the transformer to be tested to obtain comprehensive test parameter recommendations and comprehensive performance predictions; wherein the comprehensive parameter recommendations include the current size, frequency and duration in the secondary winding of the transformer to be tested; the comprehensive performance prediction includes the temperature rise and insulation strength of the transformer to be tested under high current conditions.

[0010] Preferably, the use of a calculation model based on electromagnetic theory to calculate the predicted theoretical performance of the transformer to be tested under different currents and voltages according to the input physical parameters of the transformer to be tested includes:

[0011] Calculate the predicted iron loss P of the transformer to be tested using the following formula 1: loss :

[0012]

[0013] Among them, K h is the loss factor of the transformer to be tested, which depends on the core material; f req is the operating frequency of the transformer to be tested; B max is the maximum magnetic flux density of the core of the transformer to be tested; V core is the core volume of the transformer to be tested;

[0014] Use the following formula 2 to calculate the predicted temperature rise ΔT of the transformer under test:

[0015]

[0016] Among them, S area is the heat dissipation surface area of ​​the transformer to be tested; H conv is the convective heat transfer coefficient of the transformer to be tested;

[0017] Use the following formula 3 to calculate the predicted insulation strength R of the transformer to be tested: insulation :

[0018]

[0019] Among them, D material is the initial dielectric strength of the insulating material; V is the voltage applied to the transformer to be tested; T ref is the reference temperature.

[0020] Preferably, the application of a deep learning algorithm to analyze input historical test data and current operating condition data to obtain preliminary test parameter recommendations and preliminary performance predictions based on the historical data analysis includes:

[0021] Build a deep neural network model to analyze input historical test data or current operating condition data;

[0022] Using historical test data to train the deep neural network model;

[0023] Apply the trained deep neural network model, combined with current operating condition data, to obtain preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis.

[0024] Preferably, the deep neural network model includes:

[0025] An input layer, used to receive input data and perform standardization processing on the input data, wherein the input data includes historical test data or current operating condition data;

[0026] The first hidden layer is implemented by a fully connected layer and is used to receive the standardized data provided by the input layer and obtain the primary feature set of the input data;

[0027] The second hidden layer is implemented by a fully connected layer and is used to receive the primary feature set provided by the first hidden layer to obtain a deep feature set;

[0028] The electromagnetic functional layer is implemented by a fully connected layer customized according to the electromagnetic principle, which is used to receive the deep feature set provided by the second hidden layer to obtain the electromagnetic characteristic analysis results;

[0029] Thermodynamic function layer, implemented by a fully connected layer custom designed in combination with thermal principles, is used to receive the deep feature set provided by the second hidden layer to obtain thermal effect analysis results;

[0030] The output layer is used to receive the electromagnetic characteristic analysis results provided by the electromagnetic functional layer and the thermal effect analysis results provided by the thermodynamic functional layer, and to fuse the received electromagnetic characteristic analysis results and thermal effect analysis results to obtain preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis.

[0031] Preferably, the weight of the electromagnetic functional layer adopts a random initialization method of normal distribution, wherein the mean value of the normal distribution is set to 0.0 and the standard deviation is set to 0.05; the electromagnetic functional layer applies a customized activation function, which keeps the input unchanged when the input value is greater than 0.1, and multiplies the input value by 0.01 when the input value is less than or equal to 0.1.

[0032] Preferably, the weight of the thermodynamic functional layer adopts a random initialization method of normal distribution, wherein the average value of the normal distribution is set to 0.0 and the standard deviation is set to 0.1; and the electromagnetic functional layer uses the Softplus function as an activation function.

[0033] Preferably, the data collection step includes using sensors to monitor the operating condition data of the transformer to be tested in real time, wherein the sensors are used to measure temperature, humidity and load conditions.

[0034] Preferably, at least one batch normalization layer is included between the first hidden layer and the second hidden layer, for normalizing neuron outputs and improving the stability of model training.

[0035] Preferably, the output layer includes at least one dropout layer for preventing the model from overfitting and enhancing the generalization ability of the deep neural network model to new data.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The present invention collects historical test data similar to the transformer to be tested, as well as physical parameters and current operating condition data of the transformer to be tested. The comprehensiveness and diversity of these data help to improve the accuracy and reliability of the test results.

[0038] 2. By using computational models based on electromagnetic theory, the performance of the transformer under test at different currents and voltages is predicted based on the input physical parameters. This helps predict the transformer's losses, thermal effects, and dielectric strength changes, providing a deeper understanding of its performance.

[0039] 3. By applying deep learning algorithms and analyzing historical test data and current operating condition data, preliminary test parameter recommendations and preliminary performance predictions are provided for the transformer to be tested. This approach can learn from large amounts of data and provide more accurate recommendations.

[0040] 4. By combining information from different sources, including predicted theoretical performance, preliminary test parameter recommendations, and preliminary performance predictions, this helps to develop comprehensive test parameter recommendations and comprehensive performance predictions, providing a more comprehensive test plan.

[0041] 5. Through comprehensive parameter recommendations, including current magnitude, frequency, and duration, and comprehensive performance predictions, including temperature rise and insulation strength, this method can help optimize the design of the current-up test and ensure the stability and reliability of the transformer under high current conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a principle flow chart of the present invention;

[0043] Figure 2 It is a schematic diagram of the deep neural network model in the present invention. DETAILED DESCRIPTION

[0044] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0045] like Figure 1 As shown, the first embodiment of the present application provides a transformer current boosting test method based on deep learning, comprising the following steps:

[0046] S1. Data collection steps: Collect historical test data of transformers similar to the transformer to be tested, including previous current increase test results; collect physical parameters of the transformer to be tested, including electrical characteristics and design parameters of the transformer to be tested; collect current operating condition data of the transformer to be tested, including temperature, humidity and load conditions of the test environment.

[0047] In this embodiment, "transformers similar to the transformer to be tested" refer to transformers that are similar to the transformer to be tested in design, specifications, functions or operating conditions. This similarity is determined based on a series of key factors, including but not limited to the type, capacity, design purpose, manufacturer and use environment of the transformer. Transformers similar to the transformer to be tested must meet at least one of the following key factors.

[0048] 1. Type similarity: Transformers can be classified into different types according to their functions or structures, such as distribution transformers, power transformers, etc. Transformers of similar types usually show certain consistency in electrical characteristics and performance.

[0049] 2. Capacity and specification similarity: This involves the transformer's rated power, voltage level, current capacity, etc. Transformers with similar specifications will perform similarly in terms of handling performance under different loads and pressures.

[0050] 3. Design purpose and application similarity: Transformers are designed according to their specific application in the power system, such as some transformers are designed for high load efficiency, while others may focus on stability or durability. Transformers in similar application areas will have commonalities in their operating and performance characteristics.

[0051] 4. Manufacturer and manufacturing standard similarity: Transformers from the same manufacturer or following the same manufacturing standards will often have similarities in material selection, build quality, and performance standards.

[0052] 5. Similarity of operating environment and historical operating conditions: This includes the operating environment of the transformer (such as temperature, humidity, altitude) and historical load and stress conditions. The performance and durability of transformers that have worked in similar operating environments can provide important references for the transformer to be tested.

[0053] By analyzing historical test data and response records of similar transformers, valuable reference information can be provided for predictive performance analysis of the transformer under test. This approach can enhance the understanding of the likely behavior of the transformer under test, especially in the absence of direct historical data, thereby improving the accuracy and reliability of the entire test.

[0054] In the data collection step of the transformer current rise test method, a detailed data collection strategy is adopted to ensure that the data obtained are both comprehensive and accurate. This step critically involves the collection of data in three main areas:

[0055] The first involves collecting historical test data on transformers similar to the one being tested. This data includes the results of historically conducted ramp-up tests, such as the behavior of the transformer at different current and voltage levels, recorded temperature changes, and loss data. This historical information provides an important reference for understanding the typical behavior and performance trends of transformers under similar conditions.

[0056] Next, the physical parameters of the transformer to be tested are collected, including electrical characteristics and design parameters. Specifically, this includes the transformer's inductance, capacitance, resistance, designed maximum load capacity, voltage level, etc. These parameters are the key to analyzing the basic electromagnetic characteristics of the transformer and are an indispensable part of the subsequent analysis process.

[0057] Finally, data about the current operating conditions of the transformer under test is also collected, such as the temperature, humidity and load conditions of the test environment. These real-time data help depict the specific conditions that the transformer may face in the actual operating environment, making the test setup closer to the actual application scenario.

[0058] Through the comprehensive collection of these three types of data, a detailed and comprehensive data set can be constructed, providing a solid foundation for subsequent physical model processing, deep learning analysis, and decision fusion and optimization steps. In-depth analysis of these data is crucial to accurately predict the performance of transformers under various conditions, thereby ensuring the effectiveness and reliability of the test.

[0059] Furthermore, the data collection step includes using sensors to monitor the operating condition data of the transformer to be tested in real time, wherein the sensors are used to measure temperature, humidity and load conditions.

[0060] This step involves using sensors to measure and record key parameters of the transformer in its actual operating environment, including temperature, humidity, and load conditions.

[0061] Implementation of real-time monitoring:

[0062] 1. Choose the appropriate sensor:

[0063] Temperature sensor: Select a sensor that can accurately measure the temperature on the surface or inside the transformer. These sensors can be thermocouples, RTDs, or infrared sensors, depending on the required measurement accuracy and response speed.

[0064] Humidity sensor: Use capacitive or resistive humidity sensors to measure the humidity of the air surrounding the transformer. Changes in humidity may affect the cooling efficiency and insulation performance of the transformer.

[0065] Load monitoring sensors: Current and voltage sensors are installed to monitor the load condition of the transformer in real time. These data help analyze the behavior of the transformer under different loads.

[0066] 2. Installation and configuration of sensors:

[0067] Install sensors at critical locations on the transformer, such as close to the windings, core or cooling system.

[0068] Make sure the sensor is well connected to the data logging device or control system so that the data can be transmitted and recorded in real time.

[0069] The sensor is calibrated to ensure the accuracy of the measurement data.

[0070] 3. Data collection and processing:

[0071] Set up a data acquisition system to receive data from the sensors in real time.

[0072] The collected data are preprocessed, such as filtering, denoising and normalization, to facilitate subsequent analysis.

[0073] Store data for later use while monitoring changes in key parameters in real time during the experiment to quickly respond to any abnormal situations.

[0074] Application scenarios:

[0075] These sensor data can be used to assess the initial state and environmental conditions of the transformer before conducting a current ramp-up test.

[0076] During the test, real-time monitoring data is crucial to ensure test safety, adjust test parameters and detect potential problems in a timely manner.

[0077] After the experiment is completed, these data are very useful for analyzing the experimental results, optimizing the design, and improving future experimental methods.

[0078] S2. Physical model processing step: using a computational model based on electromagnetic theory, according to the input physical parameters of the transformer to be tested, the predicted theoretical performance of the transformer to be tested under different currents and voltages is calculated, wherein the predicted theoretical performance includes predicted losses, thermal effects, and insulation strength changes.

[0079] In the transformer current-up test method, the physical model processing step (S2) is a key link, which involves using a computational model based on electromagnetic theory to predict the performance of the transformer under test under different current and voltage conditions. The main purpose of this step is to use the physical parameters of the transformer under test to estimate its theoretical performance under specific conditions, including predicted losses, thermal effects, and insulation strength changes.

[0080] In this step, the transformer's predicted iron losses are first calculated using a specific formula that takes into account the transformer's operating frequency, maximum magnetic flux density, and core volume. This calculation reflects the loss characteristics of the transformer's core material, helping to predict the transformer's energy loss levels during operation.

[0081] The predicted temperature rise of the transformer is then calculated by another formula, which takes into account losses, heat dissipation surface area, and convection heat transfer coefficient. This calculation helps estimate the temperature change of the transformer during long-term operation or under high load conditions, which is crucial for evaluating the thermal stability and safety of the transformer.

[0082] Finally, it involves the calculation of the predicted insulation strength of the transformer, where the formula used includes the dielectric strength of the insulation material, the applied voltage, and the temperature effect. This calculation is critical to understanding the insulation performance of the transformer under high voltage and high temperature conditions, helping to prevent insulation failures and improve transformer reliability.

[0083] The entire physical model processing step is based on electromagnetic theory, which combines the basic physical characteristics of the transformer with the actual operating conditions, thus providing a comprehensive and scientific method to predict the theoretical performance of the transformer under different conditions. Through this method, the performance of the transformer can be accurately estimated before the actual physical test, providing an important basis for subsequent test design and performance evaluation.

[0084] Furthermore, the calculation model based on electromagnetic theory is used to calculate the predicted theoretical performance of the transformer under different currents and voltages according to the input physical parameters of the transformer under test, including:

[0085] Calculate the predicted iron loss P of the transformer to be tested using the following formula 1: loss :

[0086]

[0087] Among them, K h is the loss factor of the transformer to be tested, which depends on the core material; f req is the operating frequency of the transformer to be tested; B max is the maximum magnetic flux density of the core of the transformer to be tested; V core is the core volume of the transformer to be tested;

[0088] Use the following formula 2 to calculate the predicted temperature rise ΔT of the transformer under test:

[0089]

[0090] Among them, S area is the heat dissipation surface area of ​​the transformer to be tested; H conv is the convective heat transfer coefficient of the transformer to be tested;

[0091] Use the following formula 3 to calculate the predicted insulation strength R of the transformer to be tested: insulation :

[0092]

[0093] Among them, D material is the initial dielectric strength of the insulating material; V is the voltage applied to the transformer to be tested; T ref is the reference temperature.

[0094] Iron losses in transformers are mainly composed of hysteresis losses and eddy current losses, which depend on the material properties of the transformer core and its behavior in an alternating magnetic field. The predicted iron loss P of the transformer to be tested can be calculated using the following formula 1: loss :

[0095]

[0096] K h It is the loss factor of the transformer to be tested, which is usually obtained from the transformer's technical manual or data provided by the manufacturer. It is determined based on the characteristics of the transformer core material and is usually determined during the design and testing stage of the transformer.

[0097] f req is the operating frequency of the transformer to be tested. The operating frequency is determined when the transformer is designed and is usually reflected in the transformer specification or technical documentation. For transformers used in power grids, this is usually a standard power supply frequency, such as 50Hz or 60Hz.

[0098] B max is the maximum magnetic flux density of the core of the transformer to be tested. This parameter can be obtained through electromagnetic design calculations or from the transformer design specifications. It reflects the magnetic field strength of the transformer core at the maximum load or maximum efficiency operating point.

[0099] V core is the core volume of the transformer to be tested. The core volume is calculated based on the physical size of the transformer and the core construction. This can be obtained by directly measuring the core size or from the design drawings. For standard transformers, this information is usually provided in the technical specification.

[0100] The temperature rise of a transformer is closely related to its heat dissipation capacity. Iron losses and other losses (such as copper losses) are converted into heat energy, increasing the temperature of the transformer. The predicted temperature rise ΔT of the transformer under test can be calculated using the following formula 2:

[0101]

[0102] In formula 2, the iron loss P loss It is calculated according to Formula 1, taking into account the material properties, operating frequency, magnetic flux density and volume of the transformer core.

[0103] S area is the heat dissipation surface area of ​​the transformer to be tested. The heat dissipation surface area of ​​the transformer can be calculated by directly measuring the external dimensions of the transformer, or obtained from the design specifications or technical documents of the transformer. This area is one of the key factors in determining the heat dissipation efficiency of the transformer.

[0104] H conv is the convective heat transfer coefficient of the transformer to be tested. The convective heat transfer coefficient is usually determined based on the cooling method and installation environment of the transformer. This coefficient can be determined experimentally or obtained from relevant engineering manuals or cooling system design documents. For specific types and models of transformers, the manufacturer may provide this data.

[0105] Combining these parameters, the formula Used to estimate the predicted temperature rise of the transformer during operation. Accurately estimating the temperature rise is critical to ensuring that the transformer operates within a safe temperature range, especially under high load or extreme environmental conditions. By understanding the heat dissipation capacity of the transformer and the maximum temperature that may be reached during actual operation, problems such as performance degradation and shortened life caused by overheating can be effectively avoided.

[0106] The insulation strength of a transformer is related to the dielectric strength of its material and the operating conditions (especially temperature). Increased temperature may degrade the performance of the insulation material. Use the following formula 3 to calculate the predicted insulation strength R of the transformer to be tested: insulation :

[0107]

[0108] D material The initial dielectric strength of the insulation material is an important parameter based on the electrical properties of transformer insulation materials, usually given in the material's technical manual or data sheet. It describes the material's ability to resist electrical breakdown under voltage.

[0109] V is the voltage applied to the transformer to be tested. This voltage can be the rated voltage determined when the transformer is designed, and can be obtained directly from the transformer specification or technical documentation. This voltage is the maximum voltage that the transformer may withstand under normal operating conditions. It should be pointed out that in the current increase test, the focus is on evaluating the performance of the transformer under high current operating conditions, and the secondary winding (i.e. the winding with higher current) is the main object of analysis. Therefore, the voltage V applied to the transformer to be tested here refers to the secondary winding voltage of the transformer to be tested.

[0110] The predicted temperature rise ΔT can be obtained using the previous formula 2.

[0111] T ref The reference temperature is usually taken as the ambient temperature of the transformer to be tested or the temperature under standard test conditions. This parameter can be obtained from laboratory test records, technical specifications or engineering manuals.

[0112] Combining these parameters, the formula Used to estimate the insulation strength of a transformer. This calculation is critical because it helps predict the insulation reliability of the transformer under actual operating conditions, especially in high temperature environments. Knowing the insulation strength is very important for evaluating the safety and reliability of transformers, especially under high load or high voltage operating conditions. By accurately predicting the insulation strength, insulation failures can be effectively prevented, thereby improving the operating safety and life of the transformer.

[0113] S3, deep learning processing step: Apply deep learning algorithms to analyze the input historical test data and current operating condition data to obtain preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis.

[0114] Furthermore, the deep learning algorithm is applied to analyze the input historical test data and current operating condition data to obtain preliminary test parameter recommendations and preliminary performance predictions based on the historical data analysis, including:

[0115] Build a deep neural network model to analyze the input historical test data and current operating condition data;

[0116] Using historical test data to train the deep neural network model;

[0117] Apply the trained deep neural network model, combined with current operating condition data, to obtain preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis.

[0118] The preliminary test parameter recommendations include the current magnitude, frequency and duration in the secondary winding of the transformer to be tested; the preliminary performance prediction includes the temperature rise and insulation strength of the transformer to be tested under high current conditions.

[0119] The third step in the transformer current boost test method, deep learning processing step S3, involves applying a deep learning algorithm to analyze the input historical test data and current operating condition data. The goal of this step is to obtain preliminary test parameter recommendations and preliminary performance predictions based on these data analyses in order to provide accurate guidance for subsequent tests. The following is a detailed description of this step:

[0120] The construction and training of deep learning models include:

[0121] 1. Model construction:

[0122] Build a deep neural network model that is specifically designed to process and analyze the historical test data and current operating condition data of the transformer. This model should include multiple layers, each focusing on extracting different levels of data features.

[0123] 2. Data preprocessing:

[0124] Preprocess the collected historical test data and current operating condition data to ensure that the data format is suitable for deep learning models. This may include steps such as data normalization, denoising, and missing value processing.

[0125] 3. Model training:

[0126] The deep neural network model is trained using the preprocessed data. By providing historical test results and corresponding operating conditions, the model can learn the performance patterns of the transformer under different conditions.

[0127] Performance prediction and parameter recommendations include:

[0128] 1. Performance prediction:

[0129] By applying the trained model, combined with current operating condition data (such as temperature, humidity, and load, etc.), the model can predict the performance of the transformer under similar conditions. This may include predicted efficiency, expected temperature rise, insulation strength, etc.

[0130] 2. Parameter recommendations:

[0131] The model can also make preliminary test parameter recommendations based on the analysis results. This includes the ideal current size, frequency, duration, etc., which are critical for the upcoming current increase test.

[0132] Through this deep learning processing step, a deeper understanding of the transformer behavior can be achieved, especially under different operating conditions. This not only improves the accuracy of the test, but also improves the efficiency of the test, because key factors that may affect the test results can be identified and adjusted in advance. In addition, the use of deep learning models also provides a more flexible and intelligent way to adapt the test to a variety of different types and conditions of transformers.

[0133] Furthermore, the deep neural network model includes:

[0134] The input layer 201 is used to receive input data and perform standardization processing on the input data, wherein the input data includes historical test data or current operating condition data;

[0135] The first hidden layer 202 is implemented as a fully connected layer and is used to receive the standardized data provided by the input layer and obtain a primary feature set of the input data;

[0136] The second hidden layer 203 is implemented by a fully connected layer and is used to receive the primary feature set provided by the first hidden layer to obtain a deep feature set;

[0137] The electromagnetic functional layer 204 is implemented by a fully connected layer customized according to electromagnetic principles, and is used to receive the deep feature set provided by the second hidden layer to obtain electromagnetic characteristic analysis results;

[0138] Thermodynamic function layer 205, implemented by a fully connected layer custom designed in combination with thermal principles, is used to receive the deep feature set provided by the second hidden layer to obtain thermal effect analysis results;

[0139] The output layer 206 is used to receive the electromagnetic characteristic analysis results provided by the electromagnetic functional layer and the thermal effect analysis results provided by the thermodynamic functional layer, and to fuse the received electromagnetic characteristic analysis results and thermal effect analysis results to obtain preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis.

[0140] In the deep neural network model 200, the main responsibility of the input layer 201 is to receive and preprocess data so that the model can effectively learn and predict. The input data of the model is usually divided into two categories, but their processing methods and purposes are consistent. Whether it is historical test data or current operating condition data, they all go through the same preprocessing steps to meet the needs of the model.

[0141] Historical test data (for model training):

[0142] The historical test data includes the records of the transformer in previous tests, such as parameters such as current, voltage, temperature, humidity and load conditions, usually focusing on the secondary winding. These data are used to train the deep neural network model to enable it to identify the behavior and performance patterns of the transformer under different operating conditions. The input layer 201 performs standardization on these historical data, including normalization, denoising, filling missing values, etc., to ensure that the data format is suitable for model learning.

[0143] Current operating condition data (for model prediction):

[0144] Once the model is trained with historical data, it can then receive current operating condition data, such as current, voltage, temperature, humidity, load, etc. of the test environment, which usually focuses on the secondary winding, to predict the performance of the transformer under these conditions.

[0145] The current operating condition data undergoes the same normalization process as the historical test data in the input layer 201 to ensure data consistency and accurate prediction of the model.

[0146] Data transmission and output:

[0147] The preprocessed data, whether historical test data or current operating condition data, is passed to the next layer of the network (such as the first hidden layer 202) for further feature extraction and analysis.

[0148] In the deep neural network model, the first hidden layer 202 plays a crucial role. It is the first layer in the model for feature extraction and processing. The design and implementation details of this layer are key to ensuring the effectiveness of the model. The following is a detailed description of the first hidden layer 202:

[0149] The first hidden layer 202 is mainly responsible for receiving the standardized data from the input layer 201 and starting a deeper analysis of the data. At this stage, the layer extracts a primary feature set of the input data through a series of neurons.

[0150] This layer usually contains a large number of neurons, and the specific number depends on the complexity of the data to be processed and the diversity of features. When implementing, you can consider using dozens to hundreds of neurons. For example, you can set 128 neurons. This number is obtained through experiments and tuning, aiming to balance computational efficiency and feature extraction capabilities.

[0151] In each neuron, a nonlinear activation function is usually applied to enhance the expressiveness of the model. ReLU (Rectified Linear Unit) is a common choice because it works well when training deep networks and is computationally efficient.

[0152] This layer is implemented as a fully connected (Dense) layer. This means that every neuron in the input layer is connected to every neuron in the first hidden layer. This comprehensive connection ensures that every feature of the input data can be fully utilized.

[0153] The first hidden layer receives the standardized data preprocessed by the input layer, which may include information about the historical test parameters or current operating conditions of the transformer.

[0154] Neurons transform the data through a combination of weighted sums and activation functions to extract meaningful primary features. After preliminary processing and feature extraction, the data is then passed to the next layer of the deep neural network, the second hidden layer 203. In this layer, the data will be further analyzed.

[0155] In summary, the first hidden layer 202 acts as a bridge in the deep neural network model, connecting the input layer with the subsequent deeper network structure. By carefully designing the number of neurons and activation functions, and implementing an effective fully connected structure, this layer ensures that data can be properly converted and utilized, laying a solid foundation for in-depth analysis and accurate prediction.

[0156] The following is a reference implementation code of the first hidden layer 202:

[0157] import tensorflow as tf

[0158] from tensorflow.keras.layers import Dense

[0159] #Assume that input_data is the output of the input layer

[0160] input_data = tf.random.normal([10,20]) # Example input, assuming the input layer output dimension is 20

[0161] #Create the first hidden layer

[0162] first_hidden_layer=Dense(units=128,activation='relu')#Use 128 neurons and ReLU activation function

[0163] # Apply the first hidden layer

[0164] output_first_hidden_layer=first_hidden_layer(input_data)

[0165] In this sample code, the Dense layer of Keras is used to implement the first hidden layer. The Dense layer is a fully connected layer in which each neuron is connected to all neurons in the previous layer. We set 128 neurons here and use relu as the activation function.

[0166] units=128 specifies the number of neurons in the layer. This number is optimized to provide enough capacity to learn and extract features while maintaining computational efficiency.

[0167] activation = 'relu' specifies the use of the ReLU activation function. ReLU (Rectified Linear Unit) is a simple but effective nonlinear activation function that can enhance the expressiveness of the model while avoiding the problem of gradient vanishing, and is particularly suitable for deep learning models.

[0168] In the above code, input_data represents the data passed from the input layer. These data are first normalized and then passed to the first hidden layer for further processing and feature extraction. The processed data (output_first_hidden_layer) will be passed to the next layer of the model, that is, the second hidden layer 203.

[0169] In the deep neural network model 200, the second hidden layer 203 is a key component that is responsible for further processing and refining the primary feature set extracted from the first hidden layer 202. The purpose of this layer is to further deepen the understanding of the data and capture more abstract and complex features, thereby providing a richer information basis for subsequent electromagnetic and thermodynamic analysis. The following is a detailed description of the second hidden layer 203:

[0170] The second hidden layer 203 continues to process the data from the first hidden layer 202 to further refine and deepen the features. This layer is designed to capture more subtle and complex patterns in the data that may not be fully revealed in the first hidden layer.

[0171] This layer usually contains fewer neurons than the first hidden layer in order to focus on more specific data features. For example, 64 neurons can be configured. This number is determined based on experiments and optimization, with the goal of providing in-depth data analysis capabilities while maintaining sufficient computational efficiency.

[0172] Similar to the first hidden layer, the second hidden layer also uses a nonlinear activation function, such as ReLU, to enhance the expressive power of the network.

[0173] The second hidden layer also uses a fully connected structure, that is, each neuron in this layer is connected to all neurons in the first hidden layer. This fully connected structure ensures that all features extracted by the first hidden layer can be further analyzed and processed by the second layer.

[0174] This layer receives the primary feature set extracted by the first hidden layer, which contains important information about the input data. Neurons further process the received features using weighted sums and activation functions to extract deeper data features. For example, some neurons may begin to recognize specific data patterns, such as the relationship between current and voltage, or the temperature change pattern under different operating conditions. The processed data is passed to the next layer of the network.

[0175] The reference implementation code of the second hidden layer 203 is as follows:

[0176] import tensorflow as tf

[0177] from tensorflow.keras.layers import Dense

[0178] #Assume that first_hidden_layer_output is the output of the first hidden layer

[0179] first_hidden_layer_output = tf.random.normal([10,128]) # Example input, assuming the first hidden layer output dimension is 128

[0180] # Create the second hidden layer

[0181] second_hidden_layer=Dense(units=64,activation='relu')#Use 64 neurons and ReLU activation function

[0182] # Apply the second hidden layer

[0183] output_second_hidden_layer=second_hidden_layer(first_hidden_layer_output)

[0184] In this example, the Dense layer is used to implement the second hidden layer. 64 neurons are set and relu is used as the activation function.

[0185] units=64 specifies the number of neurons in the layer. This number is optimized to provide enough capacity to deeply analyze and refine data features while maintaining computational efficiency.

[0186] activation = 'relu' continues to use the ReLU activation function. This helps to further enhance the expressiveness of the network, especially when dealing with complex data features.

[0187] In the above code, first_hidden_layer_output represents the data passed from the first hidden layer. This data is further processed and refined by the second hidden layer to capture more complex and subtle data patterns. The processed data (output_second_hidden_layer) is then passed to the next layer of the model, namely the electromagnetic functional layer 204 and the thermodynamic functional layer 205.

[0188] In summary, the second hidden layer 203 plays a vital role in the deep neural network model, which not only deepens the features extracted from the input data, but also lays the foundation for more professional and specific data analysis in subsequent layers. By optimizing the number of neurons and activation functions, the second hidden layer ensures that the complex features in the input data can be effectively extracted, providing an accurate prediction basis for the transformer current increase test.

[0189] The electromagnetic functional layer 204 plays a role in the deep neural network model to specifically analyze the electromagnetic characteristics of the transformer. This layer is custom designed based on electromagnetic principles and is intended to further refine and analyze information related to electromagnetic properties from the features extracted in the previous layers. The following is a detailed description of the electromagnetic functional layer 204:

[0190] Furthermore, the weight of the electromagnetic functional layer adopts a random initialization method of normal distribution, in which the mean value is set to 0.0 and the standard deviation is set to 0.05; the electromagnetic functional layer applies a customized activation function, which keeps the input unchanged when the input value is greater than 0.1, and multiplies the input value by 0.01 when the input value is less than or equal to 0.1.

[0191] The main goal of the electromagnetic functional layer 204 is to further refine the electromagnetic characteristics of the transformer from the deep feature set received from the second hidden layer 203. This includes but is not limited to identifying patterns and relationships related to magnetic flux density, inductance, capacitance, etc.

[0192] This layer contains specially designed neurons, which may be small in number, for example, 32 neurons are used. These neurons are specially trained and tuned to focus on electromagnetic related data analysis.

[0193] Each neuron in the electromagnetic functional layer is connected to a neuron in the second hidden layer. The weights of these connections are specially trained to ensure that electromagnetic related features can be effectively extracted.

[0194] When designing the electromagnetic functional layer, the basic principles of electromagnetism are taken into account to ensure that the neurons can interpret the data according to the laws of physics. This may involve incorporating electromagnetic concepts into the activation function of the neuron or the weight adjustment strategy.

[0195] In the electromagnetic functional layer, weight initialization is one of the key steps. Usually, this can be achieved by choosing an appropriate initialization strategy, for example, using a normal distribution with a small standard deviation, ensuring that the initial weights are not too biased towards specific features. Doing so can help the model learn and adjust the electromagnetic properties better.

[0196] The design of custom activation functions should take into account the characteristics of electromagnetic properties. For example, if the electromagnetic properties are associated with a certain range of data, an activation function can be designed to emphasize the signal in this range. For example, a variant of the ReLU function can be designed that enhances the output under certain conditions.

[0197] The electromagnetic function layer 204 will use the special weight initialization and custom activation function defined above. The calculation process within the layer basically follows the process of the standard fully connected layer, but custom elements will be integrated into it.

[0198] The following is a reference implementation code of the electromagnetic functional layer 204:

[0199]

[0200]

[0201] In this example, the ElectromagneticLayer class is a custom layer that uses a special weight initialization strategy (a normal distribution with a small standard deviation) and a custom activation function (a variant of the ReLU function that enhances the output under certain conditions). This design makes this layer particularly suitable for processing and extracting data related to electromagnetic properties.

[0202] In the sample code, the core idea of ​​the activation function custom_activation is to conditionally process the input signal to emphasize features within a certain range, where the signal strength (x value) exceeds 0.1. The function works as follows:

[0203] Condition Check: The function first checks if the input value (x) to each neuron is greater than 0.1.

[0204] Branch processing:

[0205] If the input values ​​are greater than 0.1, the function will leave those values ​​unchanged. This means that for larger input values, the function behaves like a standard linear function, allowing these higher signal values ​​to pass through, thereby emphasizing them.

[0206] If the input values ​​are less than or equal to 0.1, the function multiplies these values ​​by 0.01, greatly reducing their influence. This step effectively suppresses the smaller input values, making the network pay more attention to the larger values.

[0207] In transformer current-up tests, specific electromagnetic characteristics may be more significant under certain conditions, such as high current or high voltage. Therefore, through the above-mentioned custom activation function, the deep neural network model can pay more attention to signals that may indicate key electromagnetic behaviors, such as current or voltage response under high load or specific temperature conditions. Such processing makes the model more accurate in analyzing and predicting the behavior of the transformer, especially in terms of key electromagnetic performance, thereby improving the accuracy and reliability of the current-up test.

[0208] In this code, the electromagnetic layer ElectromagneticLayer uses a special weight initialization strategy, with the following specific features:

[0209] Initialization method: The RandomNormal initialization method is used, which means that the weights are randomly selected from a normal distribution centered at 0 (mean = 0.0). This initialization method helps ensure that the model weights are neither too large nor too small at the beginning of training.

[0210] Standard deviation setting: The standard deviation is set to 0.05 (stddev = 0.05), which is a relatively small value. This small standard deviation means that the initialized weights will be tightly clustered around 0. Doing so can help prevent excessive activation values ​​from occurring in the early stages of training, making the model training process more stable.

[0211] In transformer current-up tests, accurately identifying and analyzing electromagnetic characteristics is critical to predicting the behavior of the transformer. With the weight initialization strategy above, the electromagnetic functional layer is able to start learning without overemphasizing any specific feature. This means that the model can start training in a more neutral state and then gradually adjust and optimize its ability to identify electromagnetic characteristics based on the data provided. This approach helps improve the accuracy of the model in analyzing the electromagnetic characteristics of the transformer, especially when dealing with complex or uncommon current-up test situations.

[0212] Thermodynamic functional layer 205 plays a key role in the deep neural network model, focusing on analyzing complex data related to the thermal characteristics of the transformer. This layer is custom designed by combining thermodynamic principles. Specifically, it receives the deep feature set from the previous layer (the second hidden layer 203 or the electromagnetic functional layer 204) and further processes and analyzes it to extract information related to the thermal effects of the transformer.

[0213] The data received by the thermodynamic functional layer 205 usually contains features extracted from historical tests or current operating conditions of the transformer, such as temperature change, heat loss, cooling efficiency, etc. These features are processed by the second hidden layer 203 and converted into a deeper feature set, providing rich input data for the thermodynamic functional layer.

[0214] In this layer, the data is further analyzed to identify patterns related to thermal effects. This may include identifying trends in temperature variation, evaluating thermal stability under different operating conditions, and predicting thermal response under specific load or environmental conditions.

[0215] Similar to the electromagnetic functional layer 204, the thermodynamic functional layer 205 can also be implemented by a custom designed fully connected layer. This includes defining a specific weight matrix and possibly a custom activation function in order to focus on the analysis of thermal related properties.

[0216] After analysis, the thermodynamic functional layer 205 outputs analysis results related to thermal effects. These outputs may include predicted temperature rise, heat loss assessment, cooling efficiency estimation, etc., which are key indicators for evaluating the performance of the transformer under actual operating conditions.

[0217] The output of the thermodynamic functional layer is then passed to the output layer 206 for comprehensive analysis and fusion with the electromagnetic characteristic analysis results provided by the electromagnetic functional layer. This step ensures that the model can fully consider the electromagnetic and thermodynamic characteristics of the transformer, thereby providing more accurate test parameter recommendations and performance predictions.

[0218] Through the above design, the thermodynamic functional layer 205 plays a vital role in the deep neural network model, especially when processing complex data in the transformer current increase test, and can provide an in-depth understanding and accurate prediction of the thermal characteristics of the transformer.

[0219] In order to implement the thermodynamic functional layer 205 and ensure that it has specific functions in processing data related to the thermal characteristics of the transformer, a customized neural network layer can be designed. In this layer, using a specific weight matrix and a customized activation function, these designs will focus on extracting and analyzing features related to thermal effects. The following is an example code to implement this functional layer:

[0220]

[0221]

[0222] In this example, the ThermalDynamicsLayer class is a layer specifically tailored for thermal dynamics analysis. Its key features are as follows:

[0223] Weight initialization: Use the `RandomNormal` initializer, with a relatively large standard deviation (e.g. `stddev=0.1`) chosen here to ensure that the initial distribution of weights is wider, thus helping to capture a wider range of thermodynamic properties.

[0224] Custom activation function: Use the `thermal_activation` function. The Softplus function (`tf.math.log(1+tf.math.exp(x))`) is used here. This is a smooth nonlinear function that is particularly suitable for processing positive data such as temperature or thermal energy.

[0225] Through such a design, `ThermalDynamicsLayer` can focus more on processing complex data related to transformer thermal characteristics and improve the prediction accuracy and efficiency in transformer current increase tests.

[0226] The output layer 206 plays a crucial role in the deep neural network model. It is responsible for integrating and outputting the final prediction results. This layer receives the analysis results from the electromagnetic functional layer 204 and the thermodynamic functional layer 205, and combines these results to form a comprehensive prediction of the transformer current increase test. The following is a detailed description of the output layer 206 and a specific implementation example:

[0227] The output layer 206 receives two key inputs: one is the electromagnetic characteristics analysis results provided by the electromagnetic functional layer 204, and the other is the thermal effect analysis results provided by the thermodynamic functional layer 205. The core function of this layer is to fuse these two types of analysis results. Both electromagnetic data and thermal data are very important in the performance analysis of the transformer. They complement each other and jointly determine the overall behavior of the transformer. Based on these comprehensive data, the output layer provides prediction results. This may include recommendations for test parameters (such as current size, frequency and duration) and expected performance (such as expected temperature rise and insulation strength changes).

[0228] An implementation example of the output layer 206 is as follows:

[0229]

[0230]

[0231] In this example, the `OutputLayer` class combines the outputs of the electromagnetic function layer and the thermodynamic function layer through the `Concatenate` layer, and then uses a fully connected layer (`Dense` layer) to further process the combined data. Here, the ReLU activation function is selected, but other activation functions can also be selected according to the specific situation. This design ensures that the model can integrate the key information of electromagnetic and thermal properties, thereby providing more comprehensive and accurate test parameter recommendations and performance predictions.

[0232] Training the deep neural network model 200 involves several key steps, including data preparation, model building, compilation, training, and evaluation. The following is a detailed description of the training process for the deep neural network model 200:

[0233] Step S201, data preparation, including:

[0234] 1. Collect data: First, you need to collect enough historical test data, which should include parameters such as current, voltage, temperature, humidity and load of the transformer under different conditions. The data should be of high quality and cover different operating conditions.

[0235] 2. Data preprocessing: The collected data is preprocessed, which may include cleaning (removing outliers, missing value processing), standardization or normalization (making the data of the same magnitude), and dividing the data into training and test sets.

[0236] Step S202: Model construction, including:

[0237] 1. Define the model structure: construct a deep neural network model 200 including an input layer, multiple hidden layers (including an electromagnetic functional layer and a thermodynamic functional layer), and an output layer.

[0238] 2. Layer implementation: Implement the specific functions of each layer. For example, the first hidden layer is used to extract primary features, the second hidden layer is used to further refine features, the electromagnetic functional layer focuses on the analysis of electromagnetic characteristics, the thermodynamic functional layer focuses on the analysis of thermal effects, and finally the output layer combines these characteristics for prediction.

[0239] Step S203, compiling the model, including:

[0240] 1. Choose an optimizer: Choose a suitable optimizer, such as Adam or SGD, to tune the model parameters.

[0241] 2. Set the loss function: Define the loss function, such as mean squared error (MSE) or cross entropy, which is used to calculate the difference between the predicted value and the actual value.

[0242] 3. Choose the evaluation metric: Set the metric to evaluate the model performance, such as accuracy or MAE (mean absolute error).

[0243] Step S204: training the model, including:

[0244] 1. Provide training data: Input the preprocessed training data into the model for training.

[0245] 2. Set training parameters: determine the number of training epochs and the number of samples per training batch (batch size).

[0246] 3. Model training: Use the fit method to train the model. The model will learn how to extract features from the input data and make effective predictions.

[0247] Step S205, evaluation and tuning, includes:

[0248] 1. Test the model: Evaluate the performance of the model on the test set to ensure that the model performs well on unseen data.

[0249] 2. Model tuning: Adjust model parameters or structure based on test results, such as adjusting the number of neural network layers, the number of neurons, or the learning rate.

[0250] 3. Iterative training: Depending on the evaluation results and tuning strategy, multiple iterative training processes may be required to achieve optimal performance.

[0251] Step S206, implementation and application, includes:

[0252] After the model training is completed and optimized, it can be used for actual transformer current boost test prediction to help determine the optimal test parameters and predict transformer performance.

[0253] Furthermore, at least one batch normalization layer is included between the first hidden layer and the second hidden layer, which is used to normalize neuron outputs and improve the stability of model training.

[0254] At least one batch normalization layer is introduced between the first hidden layer and the second hidden layer. This step is aimed at the optimization of the deep neural network model. The purpose is to normalize the output of neurons and improve the stability of the entire model training process.

[0255] The main function of the batch normalization layer is to adjust and normalize the output of the neurons in the previous layer. In this way, the internal covariate shift (i.e. the change in the distribution of activations of the network layer) can be reduced, thereby speeding up the model training process.

[0256] Introducing batch normalization layers can make the model training process more stable and less sensitive to initial weights and learning rates. This is especially important when training deep networks, which are susceptible to vanishing or exploding gradients.

[0257] Batch Normalization layers also help the model converge faster, allowing higher learning rates to be used, thus reducing training time.

[0258] The implementation steps of batch normalization layer are:

[0259] 1. Layer insertion position: Insert the batch normalization layer between the first hidden layer and the second hidden layer. This means that the output of the first hidden layer is first processed by batch normalization before being fed into the second hidden layer.

[0260] 2. Processing flow: The batch normalization layer will normalize the data of each batch, that is, first calculate the mean and variance of the data in the batch, and then use these statistical information to normalize the data.

[0261] 3. Parameter setting: In the process of batch normalization, two trainable parameters are usually introduced, namely the scale factor and the offset. These parameters are optimized during the training process to ensure that the normalization process does not destroy the original expression ability of the network.

[0262] 4. Code implementation: When implementing, you can use the batch normalization function or layer provided by the deep learning framework. For example, in TensorFlow, you can use `tf.keras.layers.BatchNormalization` to implement it.

[0263] The sample code is as follows:

[0264] import tensorflow as tf

[0265] from tensorflow.keras.layers import Dense,BatchNormalization

[0266] #Assume that input_data is the output of the input layer

[0267] input_data = tf.random.normal([10,20]) # Example input, assuming the input layer output dimension is 20

[0268] #Create the first hidden layer

[0269] first_hidden_layer=Dense(units=128,activation='relu')

[0270] #Apply the first hidden layer and add a batch normalization layer

[0271] output_first_hidden_layer=first_hidden_layer(input_data)

[0272] batch_norm_layer=BatchNormalization()

[0273] output_batch_norm=batch_norm_layer(output_first_hidden_layer)

[0274] #Create and apply the second hidden layer

[0275] second_hidden_layer=Dense(units=64,activation='relu')

[0276] output_second_hidden_layer=second_hidden_layer(output_batch_norm)

[0277] In this way, the batch normalization layer can effectively improve the stability and efficiency of model training, which is crucial for the application of deep learning models in transformer current boosting test methods.

[0278] Furthermore, the output layer includes at least one discard layer for preventing the model from overfitting and enhancing the generalization ability of the deep neural network model to new data.

[0279] The output layer contains at least one dropout layer to prevent the model from overfitting and enhance the model's ability to generalize to new data.

[0280] In deep learning models, especially when there are a large number of parameters, the model may over-learn the details and noise in the training data, resulting in overfitting. An overfitted model performs well on the training data, but performs poorly on new, unseen data. Dropout layers can effectively alleviate this problem by randomly "dropping" (i.e. temporarily removing) a portion of neurons during training.

[0281] By discarding some neurons, the dropout layer forces different parts of the network to learn more robust feature representations, thereby improving the model's ability to generalize to new data.

[0282] Steps to implement the dropout layer:

[0283] 1. Choose the location: insert the dropout layer before the output layer. This is usually after the last fully connected layer or after a specific functional layer (such as electromagnetic functional layer or thermodynamic functional layer).

[0284] 2. Set the dropout rate: The key parameter of the dropout layer is the dropout rate, which is the ratio of randomly dropped neurons. Common dropout rates are set between 0.2 and 0.5. Choosing an appropriate dropout rate depends on the specific circumstances of the experiment and model.

[0285] 3. Training and testing phases: Apply the dropout layer during the training phase, and turn off the dropout function during the testing or actual application phase to ensure that all neurons are involved in the decision-making process.

[0286] The sample code is as follows:

[0287] import tensorflow as tf

[0288] from tensorflow.keras.layers import Dense,Dropout

[0289] #Assume that previous_layer_output is the output of the previous layer

[0290] previous_layer_output = tf.random.normal([10,64]) # Example input, assuming the output dimension of the previous layer is 64

[0291] #Create a dropout layer before the output layer

[0292] dropout_layer = Dropout(0.3) #Set the dropout rate to 30%

[0293] output_dropout=dropout_layer(previous_layer_output,training=True)

[0294] #Create output layer

[0295] output_layer=Dense(units=32,activation='relu')#The output layer is set according to specific needs

[0296] final_output=output_layer(output_dropout)

[0297] In this sample code, the dropout layer is implemented using the TensorFlow `Dropout` class, where the dropout rate is set to 30%. This means that in each training iteration, 30% of the neurons will be randomly dropped to prevent the model from overfitting. It should be noted that the dropout layer only works during the training phase, and all neurons will be retained when the model is evaluated or predicted.

[0298] By introducing the discard layer, the generalization ability of the model for unknown data can be effectively improved, which is crucial for the application of deep learning models in transformer current boosting test methods.

[0299] S4, decision fusion step: applying decision fusion technology, integrating the predicted theoretical performance, preliminary test parameter recommendations and preliminary performance prediction of the transformer to be tested, and obtaining comprehensive test parameter recommendations and comprehensive performance prediction; wherein the comprehensive parameter recommendations include the current magnitude, frequency and duration in the secondary winding of the transformer to be tested; and the comprehensive performance prediction includes the temperature rise and insulation strength of the transformer to be tested under high current conditions. The high current here refers to the current that is higher than a predetermined threshold.

[0300] In step S4 of the transformer current boost test method, decision fusion and optimization is a crucial link. This step involves integrating information from previous steps and proposing final test parameter recommendations and performance predictions. The following is the detailed implementation process of step S4:

[0301] First, this step receives the physical model processing results from step S2, which include the predicted theoretical performance of the transformer calculated based on electromagnetic theory, such as loss, thermal effect and change in insulation strength. These theoretical performances are obtained by analyzing the physical parameters of the transformer, such as current, voltage, core material, operating frequency, maximum magnetic flux density and other factors.

[0302] At the same time, step S4 also includes the results of the deep learning processing in step S3, that is, preliminary test parameter recommendations and performance predictions based on the analysis of historical test data and current operating condition data. This part usually includes the behavior of the transformer under different operating conditions predicted by the deep neural network model, such as performance under specific current and temperature conditions.

[0303] After collecting this data, decision fusion is performed. This process involves combining theoretical calculation results with predictions from deep learning models to obtain a comprehensive test parameter recommendation and performance prediction. This may involve comparing and weighing data from different sources, as well as considering their feasibility and accuracy in practical applications.

[0304] In order to achieve decision fusion, a variety of technical approaches can be adopted, such as using data fusion algorithms or multi-criteria decision making methods. These techniques can help integrate different data sources and understand the relationship between them and their importance in various situations.

[0305] In the decision fusion process, the focus is to obtain comprehensive test parameter recommendations and comprehensive performance predictions. This includes determining the most appropriate current size, frequency, and duration in the transformer secondary winding, as well as predicting temperature rise and insulation strength changes under high current conditions. By comprehensively considering the results of theoretical calculations and deep learning predictions, more accurate and practical test parameters and performance predictions can be obtained.

[0306] To implement decision fusion in step S4, the following detailed methods and steps can be taken to ensure that the model can achieve its goals:

[0307] (I) Application of data fusion algorithm:

[0308] Integrate multi-source data: First, it is necessary to integrate data from the physical model (step S2) and the deep learning model (step S3). This includes the theoretical calculation output of the transformer (such as loss, thermal effect, and insulation strength change) and the predicted output based on historical data (such as current size, frequency, duration, etc.).

[0309] Determine weight assignment: For different data sources, assign appropriate weights, which can be determined based on the accuracy of historical data or the reliability of each model. For example, if a deep learning model has shown higher accuracy in past experiments, it can be given a higher weight.

[0310] Fusion algorithm selection: Use methods such as weighted average, voting system, or more advanced fusion techniques such as Bayesian model fusion to integrate the results of different data sources. The selection of an appropriate fusion algorithm depends on the nature of the data and the expected output.

[0311] (II) Decision-making process:

[0312] Comprehensive test parameter recommendations: Based on the fused data, generate comprehensive test parameter recommendations for the current magnitude, frequency, and duration in the transformer secondary winding. This step may involve comparing and weighing the results from different data sources to find the most reasonable recommendations.

[0313] Comprehensive performance prediction: Utilize the fused data to make comprehensive performance predictions, including predictions of temperature rise and insulation strength changes under high current conditions.

[0314] Decision support: Provide sufficient supporting information or visualization to help understand why specific test parameters and performance predictions were chosen, ensuring transparency and traceability of decisions.

[0315] Ultimately, through this step, it can be ensured that the transformer current increase test method not only takes into account the accuracy of theoretical calculations, but also utilizes the powerful predictive ability of deep learning models to obtain the most optimized test design and performance prediction. This comprehensive approach makes the test more efficient and accurate, and helps to improve the overall quality of transformer design and testing.

[0316] In the transformer current increase test, the comprehensive parameter recommendations and performance predictions obtained are the key guiding factors for the implementation of the test. The following is a detailed description of how to apply these parameters and predictions to the transformer current increase test:

[0317] Application comprehensive parameter recommendations:

[0318] 1. Current size:

[0319] During the test, the input current of the transformer should be set according to the recommended current size.

[0320] This current value is a best estimate based on safety and efficiency considerations to ensure that the transformer can achieve the required test conditions without exceeding its tolerance limits.

[0321] 2. Frequency

[0322] Adjust the frequency of the power supply to match the recommended frequency.

[0323] The proper frequency is important because it affects the electromagnetic behavior and loss characteristics of the transformer.

[0324] 3. Duration:

[0325] Set the duration of the experiment to match the recommended length of time.

[0326] This helps evaluate the performance and reliability of the transformer under long-term load or overload conditions.

[0327] Using comprehensive performance prediction:

[0328] 1. Temperature rise prediction:

[0329] The temperature of the transformer is monitored during the test to ensure that it complies with the predicted temperature rise range.

[0330] If the temperature rise is too rapid or beyond expectations, it may be necessary to adjust the test parameters or take cooling measures.

[0331] 2. Insulation strength prediction:

[0332] Evaluate the stability of the transformer insulation under test based on the predicted change in dielectric strength.

[0333] Pay special attention to whether the insulation can withstand the expected voltage levels under high current conditions.

[0334] Overall implementation process:

[0335] 1. Preparation before the test:

[0336] Adjust the test equipment settings based on the comprehensive parameter recommendations.

[0337] Make sure all monitoring equipment (such as temperature sensors, current and voltage meters) are ready.

[0338] 2. Conduct the test:

[0339] Start the experiment while monitoring key parameters (such as current, temperature) in real time.

[0340] If the monitored data is beyond the predicted range, adjust or interrupt the test in time.

[0341] 3. Data recording and analysis:

[0342] Record all key data during the test.

[0343] After the test, the data is analyzed and compared with the predicted values ​​to evaluate the success of the test and the performance of the transformer.

[0344] This method can ensure that the transformer current increase test is carried out under safe and controllable conditions, while maximizing the use of predicted data to improve the accuracy and effectiveness of the test.

[0345] The following is the reference code for step S4:

[0346]

[0347]

[0348] In this example, we first define sample outputs from the physical model and the deep learning model. Then, we create a decision_fusion function that takes the outputs of these two models as input and fuses them together to generate the final trial parameter recommendations and performance predictions. The fusion logic here is simplified, and more complex and sophisticated processing may be required in actual applications.

[0349] The code simulates the decision fusion process, proposing final test parameters and performance predictions by comprehensively considering the theoretical predictions of the physical model and the predictions of the deep learning model based on historical data. This helps ensure that the design and prediction of the experiment are based on multiple data sources and analysis methods, thereby improving the accuracy and reliability of the experiment.

[0350] A second embodiment of the present application provides an electronic device, the electronic device comprising:

[0351] processor;

[0352] The memory is used to store a program, and when the program is read and executed by the processor, the transformer current boost test method provided in the first embodiment of the present application is executed.

[0353] The third embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the transformer current boost test method provided in the first embodiment of the present application is executed.

[0354] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A transformer current boost test method based on deep learning, characterized in that: The following steps are involved: S1. Data collection: Collect historical test data of transformers similar to the transformer to be tested, including previous current rise test results; Collect the physical parameters of the transformer to be tested, including the electrical characteristics and design parameters of the transformer to be tested; collect the current operating condition data of the transformer to be tested, including the temperature, humidity and load conditions of the test environment; S2. Physical model processing: using a computational model based on electromagnetic theory, according to the input physical parameters of the transformer to be tested, the predicted theoretical performance of the transformer to be tested under different currents and voltages is calculated, wherein the predicted theoretical performance includes predicted losses, thermal effects, and insulation strength changes; S3, deep learning processing: applying deep learning algorithms to analyze input historical test data and current operating condition data, and obtaining preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis; S4. Decision fusion: Apply decision fusion technology to integrate the predicted theoretical performance, preliminary test parameter recommendations and preliminary performance prediction of the transformer to be tested to obtain comprehensive test parameter recommendations and comprehensive performance prediction; wherein the comprehensive parameter recommendations include the current magnitude, frequency and duration in the secondary winding of the transformer to be tested; The comprehensive performance prediction includes the temperature rise and insulation strength of the transformer to be tested under high current conditions.

2. A transformer current boosting test method based on deep learning according to claim 1, characterized in that: The method uses a calculation model based on electromagnetic theory to calculate the predicted theoretical performance of the transformer under different currents and voltages according to the input physical parameters of the transformer under test, including: Calculate the predicted iron loss P of the transformer to be tested using the following formula 1: loss : Among them, K h is the loss factor of the transformer to be tested, which depends on the core material; f req is the operating frequency of the transformer to be tested; B max is the maximum magnetic flux density of the core of the transformer to be tested; V core is the core volume of the transformer to be tested; Use the following formula 2 to calculate the predicted temperature rise ΔT of the transformer under test: Among them, S area H is the heat dissipation surface area of ​​the transformer to be tested; conv is the convective heat transfer coefficient of the transformer to be tested; Use the following formula 3 to calculate the predicted insulation strength R of the transformer to be tested: insulation : Among them, D material is the initial dielectric strength of the insulating material; V is the voltage applied to the transformer to be tested; T ref is the reference temperature.

3. A transformer current boosting test method based on deep learning according to claim 2, characterized in that: The deep learning algorithm is applied to analyze the input historical test data and current operating condition data to obtain preliminary test parameter recommendations and preliminary performance predictions based on the historical data analysis, including: Build a deep neural network model to analyze input historical test data or current operating condition data; Using historical test data to train the deep neural network model; Apply the trained deep neural network model, combined with current operating condition data, to obtain preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis.

4. A transformer current boosting test method based on deep learning according to claim 3, characterized in that: The deep neural network model includes: An input layer, used to receive input data and perform standardization processing on the input data, wherein the input data includes historical test data or current operating condition data; The first hidden layer is implemented by a fully connected layer and is used to receive the standardized data provided by the input layer and obtain the primary feature set of the input data; The second hidden layer is implemented by a fully connected layer and is used to receive the primary feature set provided by the first hidden layer to obtain a deep feature set; The electromagnetic functional layer is implemented by a fully connected layer customized according to the electromagnetic principle, which is used to receive the deep feature set provided by the second hidden layer to obtain the electromagnetic characteristic analysis results; Thermodynamic function layer, implemented by a fully connected layer custom designed in combination with thermal principles, is used to receive the deep feature set provided by the second hidden layer to obtain thermal effect analysis results; The output layer is used to receive the electromagnetic characteristic analysis results provided by the electromagnetic functional layer and the thermal effect analysis results provided by the thermodynamic functional layer, and to fuse the received electromagnetic characteristic analysis results and thermal effect analysis results to obtain preliminary test parameter recommendations and preliminary performance predictions based on historical data analysis.

5. A transformer current boosting test method based on deep learning according to claim 4, characterized in that: The weight of the electromagnetic functional layer adopts a random initialization method of normal distribution, wherein the mean value of the normal distribution is set to 0.0 and the standard deviation is set to 0.05; the electromagnetic functional layer applies a customized activation function, which keeps the input unchanged when the input value is greater than 0.1, and multiplies the input value by 0.01 when the input value is less than or equal to 0.

1.

6. A transformer current boosting test method based on deep learning according to claim 5, characterized in that: The weight of the thermodynamic functional layer adopts a random initialization method of normal distribution, wherein the average value of the normal distribution is set to 0.0 and the standard deviation is set to 0.1; the electromagnetic functional layer uses the Softplus function as an activation function.

7. A transformer current boosting test method based on deep learning according to claim 6, characterized in that: The data collection step includes using sensors to monitor the operating condition data of the transformer to be tested in real time, wherein the sensors are used to measure temperature, humidity and load conditions.

8. A transformer current boosting test method based on deep learning according to claim 7, characterized in that: At least one batch normalization layer is included between the first hidden layer and the second hidden layer, which is used to normalize neuron outputs and improve the stability of model training.

9. A transformer current boosting test method based on deep learning according to claim 8, characterized in that: The output layer includes at least one discard layer for preventing the model from overfitting and enhancing the generalization ability of the deep neural network model for new data.

Images