Power transformer temperature monitoring method, device and equipment

Abstract

The invention discloses a power transformer temperature monitoring method, device and equipment, and relates to the technical field of power equipment state monitoring and intelligent modeling. Firstly, according to a load factor, an average winding temperature and a position of a tap switch of the power transformer at the current moment, a physical model of the power transformer is utilized to calculate heating power at the current moment; the heating power at the current moment is substituted into a heat balance equation for solving, and a top oil temperature simulation value at the current moment is obtained; forming an enhanced feature sequence by the physical features of the current moment and a preset number of moments before the current moment; and inputting the enhanced feature sequence into the deviation prediction model to obtain the dynamic deviation of the top oil temperature at the current moment, and calibrating the simulation value of the top oil temperature at the current moment to obtain a monitoring value of the top oil temperature. Priori knowledge is provided through a physical model, the deviation dynamic characteristics are learned in combination with data driving, stability constraint is applied, and high-precision and stable oil temperature prediction is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring and intelligent modeling, in particular to a power transformer temperature monitoring method, device and equipment. BACKGROUND

[0002] With the continuous advancement of urbanization, the demand for electricity in society is increasing year by year, while the number of transformers is relatively limited, which leads to a continuous increase in the load rate. Especially in the summer high-temperature weather and winter central heating period, the power load increases sharply, and the transformer overloads and heavy loads frequently occur. In the long run, it will pose a serious threat to the safe operation of the power grid. At present, most of the transformers in the power grid are oil-immersed transformers. This type of equipment mainly relies on natural oil circulation combined with sheet-shaped radiators to achieve heat dissipation, but this heat dissipation method is low in efficiency. When the temperature rise of the insulating oil exceeds a certain limit value due to sudden load increase and other reasons, problems such as electrification in the oil and reduced insulation capacity will occur, and in extreme cases, the transformer may even explode and catch fire, causing serious safety accidents. As one of the most important and most expensive devices in the power system generation, transformation, transmission and use, the safety and reliability of the power transformer are directly related to the operational integrity of the entire power system, and have a profound impact on the benefits of the power transformation enterprise, the interests of the general public and the daily life of the people. Statistics show that among all kinds of internal faults of the transformer, the overheat fault accounts for more than 73%, which will directly or indirectly shorten the service life of the equipment. Therefore, it is of great significance to discover and monitor overheat faults in a timely manner, accurately analyze the causes and specific positions of the faults, and ensure the safe and stable operation of the transformer.

[0003] The transformer winding hot spot temperature, as the name implies, is the highest point of the winding temperature of the transformer during operation. Usually, the position of the transformer winding hot spot is difficult to accurately determine, and needs to be measured by a sensor, which requires high operation and has high cost and low measurement accuracy, so it is less used in actual power grids. Since the top layer oil temperature can be regarded as a substitute for representing the winding hot spot temperature, how to accurately calculate the top layer oil temperature becomes the key to the transformer thermal state evaluation.

[0004] At present, the methods for calculating the top oil temperature mainly focus on physical model method and data-driven method. The semi-physical model method is based on the simplification of the heat balance relationship, covering the thermal model method and the load guide method. The more popular method is the IEEE Std C57.91 guide and the GB / T1094.7-2008 guide recommended method. The GB / T1094.7-2008 guide considers that the temperatures of the insulating oil and the winding both show a linear growth trend, which can be regarded as two parallel straight lines, and the closer to the top of the transformer, the higher the temperature value. However, the model rarely involves the nonlinear characteristics in the heat transfer process, resulting in a large error in the calculation result. With the continuous emergence of artificial intelligence technology, the prediction of the top oil temperature based on the data-driven algorithm provides a new idea for the acquisition of the thermal state of the transformer. The method needs to establish an evaluation index for the prediction result to reflect the accuracy of the model. Among them, the back propagation (BP) neural network, support vector machine, Kalman filter model and the like are used to predict the top oil temperature and have achieved certain effect. However, although the traditional neural network can establish the mapping relationship between different data to a certain extent, it ignores the correlation between the front and back of the time series data, resulting in relatively limited prediction performance. Then, the recurrent neural network (RNN) emerges as the times require, and its unique network structure makes it have a satisfactory performance in processing time series problems. However, for long-time dependent sequences, RNN is prone to defects such as gradient dissipation or gradient explosion. Therefore, the model based on data-driven has achieved satisfactory results in oil temperature prediction, but due to the use of deep neural network by most models, there are still problems such as a large number of subjective given parameters, long model training time and the quality of the original input data to be improved, which will hinder the improvement of the accuracy and rapidity of the model.

[0005] In summary, the temperature monitoring work of the transformer has achieved certain results, but the traditional physical model method has large prediction deviation, and the data-driven method relies on massive labeled data and has poor generalization ability when the data is sparse or the working condition is suddenly changed. SUMMARY

[0006] The purpose of the present application is to provide a power transformer temperature monitoring method, device and equipment to improve the accuracy and generalization ability of power transformer temperature monitoring.

[0007] To achieve the above purpose, the present application provides the following solutions.

[0008] In a first aspect, the present application provides a power transformer temperature monitoring method, comprising: According to the load factor, the average winding temperature and the position of the tap changer of the power transformer at the current time, the physical model of the power transformer is used to calculate the heat generation power at the current time. The heat generation power at the current moment is brought into the heat balance equation of the power transformer to solve, to obtain the top layer oil temperature simulation value at the current moment; The physical characteristics at the current moment and a preset number of moments before the current moment are combined to form an enhanced feature sequence; the physical characteristics include an ambient temperature, a load factor and the top layer oil temperature simulation value; The enhanced feature sequence is input into a deviation prediction model to obtain a top layer oil temperature dynamic deviation at the current moment; the deviation prediction model is obtained by training a machine learning model that fuses a residual network and a long short-term memory network; The top layer oil temperature simulation value at the current moment is calibrated by using the top layer oil temperature dynamic deviation at the current moment to obtain a calibrated top layer oil temperature simulation value as a top layer oil temperature monitoring value.

[0009] In a second aspect, the present application provides a power transformer temperature monitoring device, which applies the power transformer temperature monitoring method described above, and the power transformer temperature monitoring device comprises: A heat generation power calculation module is configured to calculate the heat generation power at the current moment by using a physical model of the power transformer according to a load factor, an average winding temperature and a position of a tap changer of the power transformer at the current moment; A top layer oil temperature simulation module is configured to bring the heat generation power at the current moment into the heat balance equation of the power transformer to solve, to obtain the top layer oil temperature simulation value at the current moment; An enhanced feature sequence construction module is configured to combine the physical characteristics at the current moment and a preset number of moments before the current moment to form an enhanced feature sequence; the physical characteristics include an ambient temperature, a load factor and the top layer oil temperature simulation value; A top layer oil temperature dynamic deviation prediction module is configured to input the enhanced feature sequence into a deviation prediction model to obtain a top layer oil temperature dynamic deviation at the current moment; the deviation prediction model is obtained by training a machine learning model that fuses a residual network and a long short-term memory network; A calibration module is configured to calibrate the top layer oil temperature simulation value at the current moment by using the top layer oil temperature dynamic deviation at the current moment to obtain a calibrated top layer oil temperature simulation value as a top layer oil temperature monitoring value.

[0010] In a third aspect, the present application provides a computer device, which comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the power transformer temperature monitoring method described above.

[0011] According to the specific embodiments provided in the present application, the present application has the following technical effects.

[0012] The application provides a power transformer temperature monitoring method, device and equipment. The application first calculates the heat generation power at the current moment according to the load factor, average winding temperature and position of a tap changer of the power transformer at the current moment, and the physical model of the power transformer; the heat generation power at the current moment is brought into a heat balance equation of the power transformer to solve, and the top oil temperature simulation value at the current moment is obtained; then physical characteristics at the current moment and a preset number of moments before the current moment are combined to form an enhanced feature sequence; the enhanced feature sequence is input into a deviation prediction model to obtain the top oil temperature dynamic deviation at the current moment; and the top oil temperature dynamic deviation at the current moment is used to calibrate the top oil temperature simulation value at the current moment to obtain the top oil temperature monitoring value. The application provides prior knowledge through the heat balance equation, combines data-driven learning of deviation dynamic characteristics, and applies stability constraints to realize high-precision and stable oil temperature prediction. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0014] Figure 1 A flowchart of a power transformer temperature monitoring method provided by an embodiment of the present application is shown.

[0015] Figure 2 A network structure diagram provided by an embodiment of the present application is shown.

[0016] Figure 3 A structural diagram of a computer device provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some 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.

[0018] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail in combination with the drawings and specific embodiments.

[0019] ​The embodiment of the application provides a power transformer temperature monitoring method, device and equipment, a data-physical fusion error calibration scheme for transformer temperature monitoring, prior knowledge is provided through a physical model, dynamic bias characteristics are combined with data-driven learning, and stability constraints are applied, high-precision and stable oil temperature prediction is realized, and the scheme is suitable for power transformer online temperature monitoring, state evaluation and fault early warning scenes.

[0020] In one exemplary embodiment, a power transformer temperature monitoring method is provided, as shown in the following steps 101-105. Figure 1

[0021] Step 101, according to the load factor, average winding temperature and tap switch position of the power transformer at the current moment, the heat generation power at the current moment is calculated by using the physical model of the power transformer; Step 102, the heat generation power at the current moment is brought into the heat balance equation of the power transformer to solve, and the top layer oil temperature simulation value at the current moment is obtained; Step 103, the physical characteristics of the current moment and a preset number of moments before the current moment are combined to form an enhanced feature sequence; the physical characteristics include ambient temperature, load factor and top layer oil temperature simulation value; Step 104, the enhanced feature sequence is input into a bias prediction model to obtain the top layer oil temperature dynamic bias at the current moment; the bias prediction model is obtained by training a machine learning model that fuses a residual network and a long short-term memory network; Step 105, the top layer oil temperature simulation value at the current moment is calibrated by using the top layer oil temperature dynamic bias at the current moment, and the calibrated top layer oil temperature simulation value is obtained as the top layer oil temperature monitoring value.

[0022] In another exemplary embodiment, in the above steps 101 and 102, the ambient temperature is first collected by a temperature sensor, the load current is obtained by a load monitoring device, and the load factor is calculated by combining the rated current as the input of the physical model; then a model is constructed based on the principle of thermal dynamics, the heat generation power is calculated by calculating the iron loss, copper loss and additional loss through the load factor, the heat dissipation characteristics are described by using the thermal resistance and the heat capacity, and the dynamic relationship between the top layer oil temperature and the ambient temperature and the load factor is established; finally, the heat balance equation is solved, and the top layer oil temperature simulation value is output.

[0023] In another exemplary embodiment, the heat balance equation is constructed based on the physical model of the power transformer, and the input parameters of the physical model include ambient temperature, transformer load factor, tap switch position, etc., and the heat generation power can be calculated by calculating the iron loss, copper loss and additional loss based on the physical model. The constructed heat balance equation is shown in formula (1).​

[0024] (1) wherein, P is the heating power of the power transformer, K is the load factor; T is the average winding temperature, P is the position of the tap changer, and C and R are the heat capacity and thermal resistance of the equivalent thermal circuit of the power transformer, respectively; DT is the temperature difference between the top oil temperature simulation value and the ambient temperature; T is the top oil temperature simulation value; T is the ambient temperature; DT is the difference between the average winding temperature and the top oil temperature simulation value, k is the ambient temperature conversion coefficient, and t is the time variable.

[0025] Further, in the above formula, represents the iron loss, copper loss and additional loss, the power for heating the oil is composed of two parts: the loss (iron loss, assumed to be a constant value) independent of the transformer load and the loss (copper loss and additional loss) dependent on the transformer load, the latter depends on the average winding temperature and the load factor, and the specific relationship is as follows: wherein , and ( is a correction coefficient related to the material resistivity, is the total Joule loss of the transformer under rated load, is the rated copper loss of the transformer at the reference temperature, is the additional loss of the transformer under rated load). The above formula integrates the coupled effects of the load factor, winding temperature and tap changer position on the loss, rather than a single load function, which is closer to the actual heating mechanism, and with and dynamically depicting the relationship between thermal resistance, heat capacity and temperature, overcoming the defect of traditional constant value, improving the prediction accuracy, in addition to introducing capturing the coupling of the ambient temperature change rate and the winding-oil temperature difference, realizing the linkage of the environment and the internal thermal characteristics, and being more suitable for complex scenarios.

[0026] To solve the above heat balance equation, the existing experimental data needs to be calibrated so that the top oil temperature prediction value output by the model is as close as possible to the measured value, and the core is to determine and . By using two physical models, one is a linear model in which and are assumed to be constants; the other is a nonlinear physical model in which​ and The two models are applicable to different scenarios, and have different effects on the accuracy of the predicted top oil temperature output by the physical model.

[0027] In another exemplary embodiment, the heat capacity and thermal resistance of the equivalent thermal circuit of the power transformer are core parameters of the heat balance equation; the core parameters are obtained in the following manner: According to the measured value of the top oil temperature obtained by the experiment, the core parameters in the heat balance equation are adjusted, so that the absolute value of the difference between the simulated value of the top oil temperature obtained by solving the heat balance equation and the measured value of the top oil temperature is less than a preset threshold.

[0028] In the process of obtaining the core parameters, the core parameters are assumed to be constant or the core parameters are assumed to be dynamic changes related to the top oil temperature.

[0029] In another exemplary embodiment, in steps 103 and 104 described above, the machine learning framework model includes: (1) an LSTM layer: used to process the time sequence dependence of the environmental temperature, the load factor and the predicted value of the physical model, output a feature vector containing historical information, and the historical information at least contains data of the previous 4 time steps; (2) a ResNet layer: based on the output features of the LSTM, learning the incremental change of the bias, and realizing the dynamic prediction of the bias through residual connection.

[0030] In the process of training the machine learning model fused with the residual network and the long short-term memory network to obtain the bias prediction model, the measured value of the top oil temperature is collected, and the bias between the measured value of the top oil temperature and the simulated value of the top oil temperature is calculated as the learning target of the bias prediction dynamic model; a machine learning architecture fused with the residual network (Residual Network, ResNet) and the long short-term memory network (Long Short-Term Memory, LSTM) is adopted, and the simulated value of the top oil temperature, the environmental temperature, the load factor, and the historical time sequence data of the above three types of core features (at least containing records of the previous 4 time steps) are taken as inputs to learn the dynamic characteristics of the bias evolution over time, and obtain the bias prediction model, which specifically includes: The measured value of the top oil temperature is collected by the temperature sensor installed on the top of the transformer and the simulated value of the top oil temperature obtained by solving the heat balance equation The bias is calculated, as shown in equation (2).

[0031] (2) Further, the architecture fused with the long short-term memory network (LSTM) and the residual network (ResNet) is adopted to learn the dynamic bias of the top oil temperature of the transformer The dynamic change rule over time can be obtained by two different methods, namely linear residual network, which is suitable for scenarios with slow deviation change and stable working conditions, and its dynamic evolution equation is as follows: (3) Wherein, represents the time rate of change of the top oil temperature dynamic deviation, reflecting the evolution trend of the deviation over time; is a linear constraint term function only dependent on the enhanced feature sequence , which is used to apply a fixed strength linear stability constraint to the deviation to avoid divergence of the deviation; is the top oil temperature dynamic deviation; is a deviation dynamic driving term function only dependent on the enhanced feature sequence , which is used to fit the active driving effect of the enhanced feature set on the deviation change.

[0032] In addition, there is a nonlinear residual network, which is suitable for scenarios with severe deviation fluctuations and complex working conditions, and its dynamic evolution equation is as follows: (4) Wherein, is a nonlinear adaptive constraint term function dependent on the enhanced feature sequence and , which can dynamically adjust the constraint strength according to the deviation size (the larger the deviation, the stronger the constraint strength), and adapt to the nonlinear deviation evolution characteristics under sudden working condition changes; the definitions of the remaining parameters , are consistent with the linear residual network equation.

[0033] In the above two formulas, represents the enhanced feature sequence, which is based on the physical features X (including the ambient temperature , the transformer load factor ) and superimposes the top oil temperature simulation value obtained by solving the heat balance equation, and covers the time series data of the current time and the previous 4 time steps, which is used to capture the temporal dependence of the LSTM layer and provide comprehensive support for the deviation dynamic rule.

[0034] For the linear residual network, two LSTM architectures designed for the core functions and are used, both of which consider the extended features involved in the extended feature set at the current time and the previous four time steps. The specific LSTM architectures are described in Table 1 and Table 2, both of which use the enhanced feature sequence For the base, not only the extended features at the current time step are considered, but also the extended features of the previous four time steps, where the augmented feature sequence contains physical features such as ambient temperature, load factor, etc., and the simulated top oil temperature value obtained by solving the heat balance equation, to ensure that the LSTM can fully capture the correlation logic of feature changes and bias dynamics evolution in the time dimension, and to support the linear residual network to accurately learn the dynamic characteristics of the bias over time.

[0035] Therefore, the linear correction dynamic model can be expressed as: (5) where, is the dynamic bias of the top oil temperature at time n, is the dynamic bias of the top oil temperature at time n-1, is the time interval, is the linear constraint term function only dependent on the augmented feature sequence is the bias dynamic driving term function only dependent on the augmented feature sequence and are both obtained based on the long short-term memory network, , , , , , , , , , , , , , , , , are the ambient temperatures at time n-4, n-3, n-2, n-1, n, respectively, are the load factors at time n-4, n-3, n-2, n-1, n, respectively,

[0036] are the simulated top oil temperature values at time n-4, n-3, n-2, n-1, n, respectively. The above formula (5) is used to describe the neural network architecture of the functions and Figure 2 are both based on the combination of long short-term memory (LSTM) layers and deep dense neural network layers, as described in Table 1 and Table 2 for the linear residual network (ResNet). These architectures are constructed using the Tensorflow Keras library. The inputs involved in formula (5) are shown in

[0037] AsFigure 2 The key coefficient function of the long short-term memory network (LSTM) architecture modeling the bias differential equation of the present application is shown , realizing the time sequence learning of the dynamic evolution characteristics of the top oil temperature bias; each LSTM unit inputs the environmental temperature, load factor, and top oil temperature simulation value of the previous 5 time steps (n-4 to n), combines the initial cell state (storing long-term thermal characteristic dependence) and hidden state (transmitting short-term bias fluctuations), encodes the features through a gating mechanism, and outputs the intermediate variable encoding, i.e., the dynamic correlation of the "environment-load-oil temperature" at this time step with the bias; after the multi-time step cooperative fusion, the learned is obtained , which depicts the change law of the bias with time sequence characteristics, supports subsequent stability constraints, breaks the steady-state assumption, learns multi-scale correlations, and has scalability (modeling can be reused), provides key support for the numerical and physical fusion calibration of oil temperature monitoring, and takes into account mechanism constraints and error compensation.

[0038] Table 1 Construction module of LSTM-based surrogate model

[0039] Table 2 Construction module of LSTM-based surrogate model

[0040] The training of the above model uses the first 80% of the available measurement data, and the test uses the remaining 20%.

[0041] After the above linear residual network bias correction model is constructed, it needs to be emphasized that the and learned through training in this model will be used together to perform integral operations on the bias time evolution dynamics of the linear residual network (i.e., substitute into the linear residual network bias evolution core formula (5)); Specifically, this integral operation is essentially a time sequence recursion process: the top oil temperature dynamic bias at time n is calculated according to the bias at time n-1, which is then used to calculate the bias at the next time step , and so on.

[0042] In order to distinguish between known bias and bias based on learned dynamic characteristics and The deviation calculated by the integrator (i.e., the deviation predicted by the deviation prediction model) is denoted by a hat symbol, i.e., (6) wherein, , are the deviations predicted by the deviation prediction model.

[0043] The deviation calculated by the linear residual network is the correction amount of the correction, which can effectively optimize the model and exhibit excellent accuracy and stability.

[0044] The linearized correction dynamic model exhibits good performance, which proves that most of the nonlinear characteristics of the problem have been captured by the first-order simplified model. For the nonlinear part of the correction work, the dynamic form of the model is similar to equation (5), and at this time, the deviation prediction model is: (7) wherein, is the top oil temperature dynamic deviation at time n, is the top oil temperature dynamic deviation at time n-1, is the time interval, is a nonlinear adaptive constraint term function that depends on both the enhanced feature sequence and the top oil temperature dynamic deviation at time n-1 , is a deviation dynamic driving term function that depends only on the enhanced feature sequence , and are both obtained based on a long short-term memory network, , , , , are the ambient temperatures at times n-4, n-3, n-2, n-1, and n, respectively, , , , , are the load factors at times n-4, n-3, n-2, n-1, and n, respectively, , , , , are the top oil temperature simulation values at times n-4, n-3, n-2, n-1, and n, respectively.

[0045] The integral form of the nonlinear deviation prediction model corresponding to the above equation (6) is: (8) wherein, , are the deviations predicted by the deviation prediction model.

[0046] In another exemplary embodiment, the process of solving the heat balance equation to obtain the top layer oil temperature simulation value, predicting the dynamic deviation of the top layer oil temperature using the oil temperature prediction model, and further calibrating the top layer oil temperature simulation value is referred to as a hybrid model. To illustrate the performance of the hybrid model, a reference model (pure data-driven model) is constructed using the same learning principle as the hybrid model (such as time series recursive integration, the core architecture of "constraint function + driving function"), which is completely identical to the hybrid model. The only difference between the two is the learning object: the hybrid model learns the deviation between the top layer oil temperature simulation value and the measured value, while the reference model directly learns the evolution rule of the measured top layer oil temperature (experimental data). The specific model formula and input feature set definition are as follows: (9) wherein, and are parameterized functions related to the measured oil temperature, both of which only depend on the input features X.

[0047] Similarly, the same model shown in Tables 1 and 2 is used for training, and the same proportion (80% and 20%) of training data set and test data set is used.

[0048] By introducing an enhanced feature set containing ambient temperature, load factor, and top layer oil temperature simulation value through linear residual network and nonlinear residual network, deep integration of data and physical information is achieved, breaking through the limitation of pure data-driven model lacking mechanism support; with the help of the extended features of the first four time steps to depict the time series dependence, the deviation prediction model is used to accurately distinguish between known deviations and integral calculation deviations, and fine modeling of the evolution of deviation over time is achieved; at the same time, linear and nonlinear forms are provided to flexibly adapt to different deviation characteristics scenarios, making the model more universal than traditional single models; by comparing with the reference model, it can be shown that the hybrid model of the present application has significant advantages in terms of accuracy and stability.

[0049] In another exemplary embodiment, a stability constraint is imposed during the training process of the deviation prediction model, which is a linear constraint term function or a nonlinear adaptive constraint term function in the deviation prediction model less than or equal to 0, i.e. imposing a stability constraint during the training process of the deviation prediction model, for linear deviation model, the eigenvalue of its parameter matrix is constrained to be ≤0; for nonlinear deviation model, the key coefficient of its nonlinear term is constrained to be ≤0, to ensure the convergence of the deviation model in the process of time integration.

[0050] In this embodiment, to prevent the bias model from diverging in long-term predictions, all bias models ultimately applied to real-world scenarios must be subject to stability constraints in this step. To demonstrate the impact of stability, a residual network (ResNet) is trained. The formula used during training without stability constraints is as follows: (10) No conditions were applied during the calculation of G and F.

[0051] Its discrete form is: (11) While the above formula can provide excellent prediction results, it is important to note that here... Solution of time It is based on Precise deviation of time It was calculated.

[0052] However, in complete integrals, It is based on the previously calculated To calculate, that is: (12) This will produce extremely poor prediction results, which is a direct consequence of the lack of stability constraints.

[0053] The following section introduces the application of stability constraints to linear dynamic systems, starting with a parameterized dynamic system: (13) in, state of time Depends on the state at the previous moment and loads that drive state changes Assume that the loading depends not only on the state but also on a set of parameters (input features), which are combined here into a vector X. Therefore, the learning process aims to compute the regression function. .

[0054] However, to obtain a stable integrator, certain constraints must be satisfied. As mentioned earlier, these constraints are relatively simple in the linear case, so we linearize the forcing terms as follows: (14) Now, as long as the conditions are met Stability can then be guaranteed. This constraint can be achieved by relating it to... This can be easily achieved by introducing a penalty term into the loss function of the relevant neural network.

[0055] Linearized residual network (ResNet) is similar to dynamic mode decomposition (DMD), but it can be more convenient in introducing the parameter dimension.

[0056] Inspired by the above principles, a feasible method that can guarantee stability without compromising nonlinear behavior is to express it as: (15) At the same time, non-positive constraints are imposed when constructing the regression function , that is .

[0057] In another exemplary embodiment, the above-mentioned step 105 feeds the above-mentioned learned dynamic bias to the top oil temperature simulation value to achieve the final prediction, and the corrected top oil temperature is: (16) wherein, is the dynamic bias of the top oil temperature at the current time t output by the bias prediction model.

[0058] The technical solutions provided by the various method embodiments have the following advantages: 1) Through the data-physical fusion strategy, both the physical model is constructed relying on the principle of transformer thermodynamics to provide mechanism support, and the measured data and the bias model are dynamically corrected, solving the problem of large prediction bias of pure physical model and weak generalization ability of pure data-driven method, and significantly improving the accuracy of top oil temperature monitoring.

[0059] 2) In the bias model training, stability constraints are imposed, the parameter matrix eigenvalue of the linear model is limited to ≤0, and the key coefficient of the nonlinear model is constrained to ≤0, ensuring that the model converges in the long-term time integration process, avoiding prediction divergence caused by sudden changes in working conditions or sparse data, and being applicable to complex and variable operating scenarios.

[0060] 3) The present application is applicable to different types and capacities of power transformers, and through multi-source data acquisition (temperature, load, tap switch position, etc.) and closed-loop optimization mechanism (bias feedback updates physical model parameters), it can adapt to various working conditions, reduce the temperature prediction error by more than 40% compared with traditional methods, and provide a reliable basis for transformer state evaluation and fault warning.

[0061] Based on the same inventive concept, the embodiments of the present application also provide a power transformer temperature monitoring device for implementing the power transformer temperature monitoring method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more power transformer temperature monitoring device embodiments provided below can refer to the limitations of the power transformer temperature monitoring method in the above text, and will not be repeated here.

[0062] In one exemplary embodiment, a power transformer temperature monitoring device is provided, comprising: a heat generation power calculation module configured to calculate a heat generation power at a current time according to a load factor, an average winding temperature and a position of a tap changer of the power transformer at the current time, and using a physical model of the power transformer; a top layer oil temperature simulation module configured to bring the heat generation power at the current time into a heat balance equation of the power transformer to solve the heat balance equation and obtain a top layer oil temperature simulation value at the current time; an enhanced feature sequence construction module configured to construct an enhanced feature sequence from physical features at the current time and a preset number of times before the current time; the physical features include an ambient temperature, the load factor and the top layer oil temperature simulation value; a top layer oil temperature dynamic deviation prediction module configured to input the enhanced feature sequence into a deviation prediction model to obtain a top layer oil temperature dynamic deviation at the current time; the deviation prediction model is obtained by training a machine learning model that fuses a residual network and a long short-term memory network; a calibration module configured to calibrate the top layer oil temperature simulation value at the current time using the top layer oil temperature dynamic deviation at the current time to obtain a calibrated top layer oil temperature simulation value as a top layer oil temperature monitoring value.

[0063] In one exemplary embodiment, a computer device is provided, which can be a server or a terminal. An internal structure diagram of the computer device can be as shown in Figure 3 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with external terminals through a network connection. The computer program is executed by the processor to implement a power transformer temperature monitoring method.

[0064] Those skilled in the art can understand that, Figure 3The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in the above method embodiments.

[0065] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties.

[0066] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetic variable memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0067] The database involved in each of the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, and the like, without being limited thereto. The processor involved in each of the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, and the like, without being limited thereto.

[0068] The technical features of the above embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but it should be considered that any combination of the technical features is within the scope of the present disclosure, as long as there is no contradiction.

[0069] The principles and implementation manners of the present application are described by using specific examples herein, and the above embodiments are only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the present application. In summary, the content of the present description should not be understood as a limitation of the present application.

Claims

1. A method of monitoring the temperature of a power transformer, characterized by, The method comprises the following steps: According to the load factor, average winding temperature and tap position of the power transformer at the current time, the heat generation power at the current time is calculated by using the physical model of the power transformer; The heat generation power at the current time is brought into the heat balance equation of the power transformer to solve, and the top oil temperature simulation value at the current time is obtained; The physical characteristics at the current time and a preset number of times before the current time are combined to form an enhanced feature sequence; the physical characteristics include ambient temperature, load factor and top oil temperature simulation value; The enhanced feature sequence is input into a deviation prediction model to obtain the top oil temperature dynamic deviation at the current time; the deviation prediction model is obtained by training a machine learning model that fuses a residual network and a long short-term memory network; The top oil temperature simulation value at the current time is calibrated by using the top oil temperature dynamic deviation at the current time, and the calibrated top oil temperature simulation value is obtained as the top oil temperature monitoring value.

2. The power transformer temperature monitoring method of claim 1, wherein, The heat balance equation is: ； wherein, is the heating power of the power transformer, is the load factor; is the average winding temperature, is the position of the tap changer, and are the thermal capacitance and the thermal resistance of the equivalent thermal circuit of the power transformer, respectively; is the temperature difference between the top oil temperature simulation value and the ambient temperature; is the top oil temperature simulation value; is the ambient temperature; is the difference between the average winding temperature and the top oil temperature simulation value, k is the ambient temperature conversion coefficient, and t is the time variable.

3. The power transformer temperature monitoring method of claim 2, wherein, The heat capacity and thermal resistance of the equivalent thermal circuit of the power transformer are core parameters of the heat balance equation; the core parameters are obtained in the following manner: According to the measured top oil temperature value obtained by experiment, the core parameters in the heat balance equation are adjusted so that the absolute value of the difference between the top oil temperature simulation value obtained by solving the heat balance equation and the measured top oil temperature value is less than a preset threshold.

4. The power transformer temperature monitoring method of claim 3, wherein, In the process of obtaining the core parameters, the core parameters are assumed to be constants or the core parameters are assumed to be dynamic changes related to the top oil temperature.

5. The power transformer temperature monitoring method of claim 1, wherein, The deviation prediction model is: ； wherein, is the top layer oil temperature dynamic deviation at time n output by the deviation prediction model, is the top layer oil temperature dynamic deviation at time n-1 output by the deviation prediction model, is the time interval, is a linear constraint term function only dependent on the enhanced feature sequence , is a deviation dynamic driving term function only dependent on the enhanced feature sequence , and are both obtained based on a long short-term memory network, , , , , are respectively the ambient temperatures at time n-4, time n-3, time n-2, time n-1 and time n, , , , , are respectively the load factors at time n-4, time n-3, time n-2, time n-1 and time n, , , , , are respectively the top layer oil temperature simulation values at time n-4, time n-3, time n-2, time n-1 and time n.

6. The power transformer temperature monitoring method of claim 1, wherein, The deviation prediction model is: ； wherein, is the top layer oil temperature dynamic deviation at time n output by the deviation prediction model, is the top layer oil temperature dynamic deviation at time n-1 output by the deviation prediction model, is a time interval, is a nonlinear adaptive constraint term function of the top layer oil temperature dynamic deviation at time n-1, is a deviation dynamic driving term function of the enhanced feature sequence, are both obtained based on a long short-term memory network, are respectively an ambient temperature at time n-4, an ambient temperature at time n-3, an ambient temperature at time n-2, an ambient temperature at time n-1, and an ambient temperature at time n, are respectively a load factor at time n-4, a load factor at time n-3, a load factor at time n-2, a load factor at time n-1, and a load factor at time n, are respectively a top layer oil temperature simulation value at time n-4, a top layer oil temperature simulation value at time n-3, a top layer oil temperature simulation value at time n-2, a top layer oil temperature simulation value at time n-1, and a top layer oil temperature simulation value at time n.​​​​​​​​​​​​​​​​ 7. The power transformer temperature monitoring method of claim 1, wherein, In the training process of the deviation prediction model, a stability constraint is applied, which is that a linear constraint term function or a nonlinear adaptive constraint term function in the deviation prediction model is less than or equal to 0.

8. The power transformer temperature monitoring method of claim 1, wherein, The formula for calibrating the top oil temperature simulation value at the current time by using the top oil temperature dynamic deviation at the current time to obtain the calibrated top oil temperature simulation value as the top oil temperature monitoring value is: ； wherein, is a top layer oil temperature monitoring value at a current time t, is a top layer oil temperature simulation value at a current time t, is a top layer oil temperature dynamic deviation at a current time t output by the deviation prediction model.

9. A power transformer temperature monitoring device, characterized by, The power transformer temperature monitoring device applies the power transformer temperature monitoring method of any one of claims 1-8, and the power transformer temperature monitoring device comprises: A heat generation power calculation module is configured to calculate the heat generation power at the current time by using the physical model of the power transformer according to the load factor, average winding temperature and tap position of the power transformer at the current time; A top oil temperature simulation module is configured to bring the heat generation power at the current time into the heat balance equation of the power transformer to solve, and obtain the top oil temperature simulation value at the current time; An enhanced feature sequence construction module is configured to combine the physical characteristics at the current time and a preset number of times before the current time to form an enhanced feature sequence; the physical characteristics include ambient temperature, load factor and top oil temperature simulation value; A top oil temperature dynamic deviation prediction module is configured to input the enhanced feature sequence into a deviation prediction model to obtain the top oil temperature dynamic deviation at the current time; the deviation prediction model is obtained by training a machine learning model that fuses a residual network and a long short-term memory network; A top oil temperature dynamic deviation prediction module is configured to input the enhanced feature sequence into a deviation prediction model to obtain the top oil temperature dynamic deviation at the current time; the deviation prediction model is obtained by training a machine learning model that fuses a residual network and a long short-term memory network; The calibration module is configured to calibrate the top oil temperature simulation value at the current time by using the dynamic deviation of the top oil temperature at the current time, to obtain a calibrated top oil temperature simulation value as a top oil temperature monitoring value.

10. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the power transformer temperature monitoring method of any one of claims 1-8.

Images