Transformer operation state monitoring method and system

Abstract

The invention relates to the technical field of transformers, and particularly discloses a transformer operation state monitoring method and system, and the method comprises the steps: carrying out the feature engineering and time sequence joint analysis of the load current, cooling mode, cooling medium flow velocity and environment temperature of a transformer through employing an artificial intelligence algorithm based on deep learning; the method comprises the following steps: extracting the thermal behavior characteristics of a transformer, and on this basis, further performing transformer temperature prediction by using a deep learning model obtained by training a large amount of healthy transformer operation data to obtain the expected temperature of each key monitoring point of the transformer in a normal state; and performing overheating risk early warning based on the difference between the expected temperature and the actually measured temperature of the transformer in the normal state. According to the method, real-time accurate early warning of the overheating risk of the transformer can be realized, the limitation of a traditional fixed temperature threshold monitoring method is overcome, and the sensitivity and robustness of overheating risk identification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformers, and more particularly, to a transformer operating state monitoring method and system. BACKGROUND

[0002] As the core equipment of power systems, the operating state of transformers is directly related to the safety and stability of power grids. With the continuous growth of power load and the large-scale grid connection of new energy, transformers are long-term under complex and changeable working conditions. The insulation aging problems of key parts such as windings and cores caused by overheating have become the main inducement of equipment failure. Therefore, real-time and accurate monitoring of the operating state of transformers, timely early warning of potential risks, has irreplaceable value for improving the level of equipment health management, prolonging the service life of equipment, and ensuring the safety of power systems.

[0003] Currently, the state monitoring methods for transformers mainly include periodic inspection, dissolved gas analysis (DGA) in oil, and threshold alarm systems based on temperature and other single or few parameters. However, these traditional methods have many limitations: periodic inspection and offline analysis (such as DGA) often have a lag, making it difficult to capture transient abnormalities and early signs of failure during the operation of the transformer; and the alarm system based on fixed temperature thresholds cannot fully consider the state differences of the transformer under different loads and environmental conditions, and is easily disturbed by noise, resulting in false alarms or missed alarms, making it difficult to accurately assess the true health status of the transformer.

[0004] Therefore, an optimized transformer operating state monitoring method and system are expected. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. The embodiments of the present application provide a transformer operating state monitoring method and system, which uses an artificial intelligence algorithm based on deep learning to perform feature engineering and time series joint analysis on the load current, cooling method, cooling medium flow rate, and environmental temperature of the transformer, to mine the thermal behavior characteristics of the transformer. On this basis, a deep learning model trained on a large amount of healthy transformer operating data is further used for transformer temperature prediction to obtain the expected temperature of each key monitoring point of the transformer under normal conditions. Then, based on the difference between the expected temperature and the actual measured temperature of the transformer under normal conditions, an overheating risk early warning is performed. This method can realize real-time and accurate early warning of the overheating risk of the transformer, overcome the limitations of traditional fixed temperature threshold monitoring methods, and improve the sensitivity and robustness of overheating risk identification.

[0006] According to one aspect of the present application, a transformer operating state monitoring method is provided, which comprises:

[0007] obtain transformer operation state data in a predetermined time window in real time, the transformer operation state data including a time series of load current, a time series of ambient temperature, a time series of transformer cooling mode and cooling medium flow rate;

[0008] perform operation state feature engineering on the transformer operation state data to obtain a time series of transformer operation state feature vectors;

[0009] input the time series of transformer operation state feature vectors into a trained transformer normal thermal state prediction model to obtain a temperature reference value of a transformer key monitoring point at a current time point;

[0010] obtain a temperature measurement value of the transformer key monitoring point at the current time point in real time;

[0011] calculate a deviation value between the temperature measurement value of the transformer key monitoring point at the current time point and the temperature reference value, and determine whether the transformer has an overheating risk based on a comparison between the deviation value and a preset threshold.

[0012] According to another aspect of the present application, a transformer operation state monitoring system is provided, which comprises:

[0013] an operation state data acquisition module configured to obtain transformer operation state data in a predetermined time window in real time, the transformer operation state data including a time series of load current, a time series of ambient temperature, a time series of transformer cooling mode and cooling medium flow rate;

[0014] an operation state feature engineering module configured to perform operation state feature engineering on the transformer operation state data to obtain a time series of transformer operation state feature vectors;

[0015] a normal thermal state prediction module configured to input the time series of transformer operation state feature vectors into a trained transformer normal thermal state prediction model to obtain a temperature reference value of a transformer key monitoring point at a current time point;

[0016] a real-time temperature measurement module configured to obtain a temperature measurement value of the transformer key monitoring point at the current time point in real time;

[0017] an overheating risk assessment module configured to calculate a deviation value between the temperature measurement value of the transformer key monitoring point at the current time point and the temperature reference value, and determine whether the transformer has an overheating risk based on a comparison between the deviation value and a preset threshold.

[0018] Compared with the prior art, the transformer operating state monitoring method and system provided by the application adopts an artificial intelligence algorithm based on deep learning to perform feature engineering and time sequence joint analysis on the load current, cooling mode, cooling medium flow rate and ambient temperature of the transformer, so as to mine the thermal behavior characteristics of the transformer, and on this basis, further utilize a deep learning model trained on a large amount of healthy transformer operating data to perform transformer temperature prediction, so as to obtain the expected temperature of each key monitoring point of the transformer in the normal state, and then perform overheat risk early warning based on the difference between the expected temperature and the actual measured temperature of the transformer in the normal state. The method can realize real-time and accurate early warning of the overheat risk of the transformer, overcome the limitations of the traditional fixed temperature threshold monitoring method, and improve the sensitivity and robustness of overheat risk identification. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application, when taken in conjunction with the accompanying drawings. The drawings provided in the specification and the embodiments of the present application together serve to provide a further understanding that enables others skilled in the art to make or use the present application. The drawings provided are for illustrative purposes and are not intended to limit the present application thereto, and constitute a part of the specification. In the drawings, the same reference numerals refer to the same components or steps throughout the specification.

[0020] Figure 1 A flowchart of the transformer operating state monitoring method according to the embodiment of the present application.

[0021] Figure 2 A data flow diagram of the transformer operating state monitoring method according to the embodiment of the present application.

[0022] Figure 3 A flowchart of substep S2 of the transformer operating state monitoring method according to the embodiment of the present application.

[0023] Figure 4 A flowchart of substep S3 of the transformer operating state monitoring method according to the embodiment of the present application.

[0024] Figure 5 A flowchart of substep S31 of the transformer operating state monitoring method according to the embodiment of the present application.

[0025] Figure 6 A block diagram of the transformer operating state monitoring system according to the embodiment of the present application. DETAILED DESCRIPTION

[0026] As shown in the present application and claims, unless the context clearly indicates otherwise, the words "a," "an," "the," and / or "this" are not limited in scope to the singular, but rather encompass both the singular and the plural. Generally, the terms "comprises" and "comprising" are not intended to be exclusive, but rather to permit inclusion of additional steps and elements, methods or apparatuses.

[0027] Although the present application makes various references to certain modules in the system according to embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are merely illustrative, and different aspects of the system and method can use different modules.

[0028] Flowcharts are used in the present application to illustrate the operations performed by the system according to embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in sequence. Rather, various steps can be processed in reverse order or simultaneously, as desired. Other operations can also be added to or removed from these processes.

[0029] In the following, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It is obvious that the described embodiments are only a part of the embodiments of the present application, and not all embodiments of the present application, and it should be understood that the present application is not limited by the example embodiments described herein.

[0030] It is worth noting that in the present application, all actions of obtaining data are carried out in compliance with the corresponding data protection regulations and policies of the country where the device is located, and with the authorization given by the owner of the corresponding device.

[0031] In view of the technical problems described in the above background art, the present application proposes a transformer operating state monitoring method, which uses an artificial intelligence algorithm based on deep learning to perform feature engineering and time series joint analysis on the load current, cooling method, cooling medium flow rate and environmental temperature of the transformer, to mine the thermal behavior characteristics of the transformer, and on this basis, further uses a deep learning model trained on a large amount of healthy transformer operating data to perform transformer temperature prediction, to obtain the expected temperature of each key monitoring point of the transformer in the normal state, and then, based on the difference between the expected temperature and the actual measured temperature of the transformer in the normal state, to perform overheat risk warning. This method can realize real-time and accurate warning of the overheat risk of the transformer, overcome the limitations of the traditional fixed temperature threshold monitoring method, and improve the sensitivity and robustness of overheat risk identification.

[0032] Figure 1 Flowchart of the transformer operating state monitoring method according to embodiments of the present application. Figure 2A data flow chart of a transformer operating state monitoring method according to an embodiment of the present application. As shown in Figure 1 and Figure 2 The transformer operating state monitoring method comprises the steps of: S1, acquiring transformer operating state data in a predetermined time window in real time, the transformer operating state data comprising a time series of load current, a time series of ambient temperature, a time series of transformer cooling mode and cooling medium flow rate; S2, performing operating state feature engineering on the transformer operating state data to obtain a time series of transformer operating state feature vectors; S3, inputting the time series of transformer operating state feature vectors into a trained transformer normal thermal state prediction model to obtain a temperature reference value of a transformer key monitoring point at a current time point; S4, acquiring a temperature measurement value of the transformer key monitoring point at the current time point in real time; S5, calculating a deviation value between the temperature measurement value of the transformer key monitoring point at the current time point and the temperature reference value, and determining whether the transformer has an overheating risk based on a comparison between the deviation value and a preset threshold value.

[0033] In the above transformer operating state monitoring method, the step S1 of acquiring transformer operating state data in a predetermined time window in real time comprises a time series of load current, a time series of ambient temperature, a time series of transformer cooling mode and a time series of cooling medium flow rate. It can be understood that the thermal state of a transformer is the result of dynamic balance between its internal loss (mainly caused by load current) and external heat dissipation conditions (affected by ambient temperature, cooling mode and cooling medium flow rate), and the traditional fixed temperature threshold alarm cannot adapt to such dynamic working condition changes, which is prone to misjudgment. Therefore, in order to comprehensively grasp the key factors affecting the temperature of the transformer, the present application deploys corresponding sensors to collect time series data of ambient temperature, load current of the transformer, cooling mode (such as oil-immersed self-cooling, forced oil circulation air cooling, etc.) and cooling medium flow rate (such as oil pump speed, fan speed, etc.) in real time, so as to realize continuous monitoring of the internal and external operating parameters of the transformer and capture the thermal state evolution law of the transformer under dynamic working conditions.

[0034] In the process of implementation, the load current is one of the important factors affecting the thermal state of the transformer, and its fluctuations directly reflect the working load situation of the transformer. To accurately measure the time series data of the load current, high-precision current transformers or Rogowski coils are usually used. Such sensors can quickly respond to changes in current and convert electrical signals into digital signals that are easy to process. By installing such sensors at key nodes of the transformer, data of the load current can be continuously recorded. Considering that the load current may fluctuate greatly due to the instantaneous changes in the power grid load, the selected sensor needs to have a high sampling rate and resolution to capture all subtle changes. At the same time, electrical isolation between the sensor and the transformer needs to be ensured to avoid interfering with the normal operation of the transformer.

[0035] For monitoring the ambient temperature, temperature sensors such as platinum resistance thermometers (RTD) or thermocouples can be installed at different locations around the transformer. These sensors can provide stable and accurate temperature readings, helping to understand the actual conditions of the environment in which the transformer is located. Since the ambient temperature may be affected by factors such as seasonal changes, diurnal temperature differences, etc., when selecting sensors, their measurement range and accuracy must be considered to suit the actual application scenario. In addition, to ensure the comprehensiveness and accuracy of the data, multiple temperature sensors need to be reasonably laid out to cover all key areas around the transformer, ensuring that the most accurate temperature information is obtained regardless of changes in ambient temperature.

[0036] The cooling method and flow rate of the cooling medium of the transformer directly affect the effective dissipation of heat inside the transformer. According to different cooling methods, such as oil-immersed self-cooling, forced oil circulation air cooling, etc., corresponding sensors are selected for monitoring. For example, in oil-immersed self-cooled transformers, the flow rate of the cooling medium can be indirectly reflected by detecting the speed of the oil pump; while in forced oil circulation air-cooled transformers, both fan speed and oil pump speed need to be monitored. For the measurement of such parameters, rotary encoders or Hall effect-based sensors can be used, which can accurately track the speed changes of motors or pumps and convert them into data for further analysis. It is worth noting that the choice of cooling method is often based on the design of the transformer and the specific application scenario, so when designing the monitoring system, the characteristics of various cooling methods need to be fully considered, and the most suitable technical means for parameter acquisition need to be selected.

[0037] To ensure that the data collected by various sensors can be effectively utilized, a high-efficiency data transmission network needs to be established. This network not only needs to support the simultaneous access of a large number of sensors, but also needs to ensure the real-time and stability of data transmission. The common approach is to use industrial-grade wireless communication technology or wired Ethernet connection to build a data transmission architecture covering the entire substation. In this way, not only can the real-time collection of transformer operating state data be realized, but also a solid foundation can be provided for subsequent data storage and analysis. In addition, with the development of Internet of Things technology, more and more intelligent sensors begin to integrate data preprocessing functions, which can complete preliminary data filtering and compression locally, thereby reducing the burden on the central processing system and improving overall efficiency.

[0038] In the entire data collection process, in addition to the deployment of hardware facilities, software support is also crucial. This includes developing dedicated drivers and interface protocols for different types of sensors to ensure seamless data connection to the central control system. At the same time, a complete data management system needs to be designed to classify and store the collected data, manage indexes, and perform regular backups, ensuring that historical data for any time period can be easily retrieved even after a long period of operation. In addition, considering that there may be multiple devices of different brands and models in the power system, compatibility issues cannot be ignored. Therefore, during implementation, open standards and industry specifications should be followed as much as possible to promote good intercommunication between components.

[0039] In the above transformer operating state monitoring method, the step S2, the transformer operating state data is subjected to operating state feature engineering to obtain a time series of transformer operating state feature vectors. Specifically, since the original monitoring data contains parameters of different dimensions and different dynamic characteristics (such as discrete cooling mode encoding and continuous flow rate time series), directly inputting them into a deep learning model may cause mismatch between features or information loss. Therefore, in order to extract more representative transformer thermal state features, based on feature engineering technology, the application integrates heterogeneous data in the transformer operating state data through multi-modal integration and time series modeling, thereby converting the original monitoring data into a unified dimensional feature representation to obtain a time series of transformer operating state feature vectors. Among them, Figure 3 The flowchart of sub-step S2 of the transformer operating state monitoring method according to the embodiment of the application. As shown in FIG. 2, the transformer operating state data is subjected to operating state feature engineering to obtain a time series of transformer operating state feature vectors. Figure 3As shown, the step S2 includes steps of: S21, one-hot encoding the transformer cooling mode to obtain a transformer cooling mode one-hot encoding feature vector; S22, combining the time series of the load current, the time series of the ambient temperature and the time series of the cooling medium flow rate after standardization processing and time series alignment to obtain a time series of transformer thermal behavior feature vectors; S23, based on the time stamp of each transformer thermal behavior feature vector in the time series of transformer thermal behavior feature vectors, embedding a transformer running time length mark and the transformer cooling mode one-hot encoding feature vector in the respective transformer thermal behavior feature vector to obtain a time series of enhanced transformer thermal behavior feature vectors; S24, performing time series modeling analysis on the time series of enhanced transformer thermal behavior feature vectors to obtain a time series of transformer operating state feature vectors.

[0040] Specifically, the step S21, one-hot encoding the transformer cooling mode to obtain a transformer cooling mode one-hot encoding feature vector. It should be understood that the present application considers that the cooling mode of the transformer belongs to a discrete classification variable and cannot be directly used for numerical calculation, while a machine learning model needs structured numerical input for effective training and inference. Therefore, in order to convert the cooling mode information into a mathematical representation that can be processed by the model, the present application is based on the principle of one-hot encoding, that is, by assigning a unique binary vector to each cooling mode, to realize the numerical conversion of discrete variables. Specifically, if there are N possible cooling modes of the transformer, an N-dimensional vector is constructed, in which the dimension corresponding to the current cooling mode is 1 and the rest is 0. For example, if the cooling mode is forced oil circulation air cooling (OFAF) and it is the second in the encoding order, the vector [0, 1, 0, …] is generated. In this way, without introducing numerical size misleading, the category information of the cooling mode can be completely preserved, providing standardized input for subsequent feature fusion.

[0041] Specifically, the step S22, the time series of the load current, the time series of the ambient temperature and the time series of the cooling medium flow rate are standardized and time series aligned to obtain the time series of the transformer thermal behavior feature vector. Specifically, since the original data of the load current, the ambient temperature and the cooling medium flow rate have different dimensions (such as ampere, degree Celsius, meter / second) and numerical ranges, direct combination will cause some feature factors to dominate the gradient update due to the large value, while other features are weakened in the model training process. In addition, the communication delay of different sensors may cause the time series data to be not strictly synchronized. Therefore, in order to eliminate the dimensional difference and ensure the time consistency, the present application is based on the standardization (Z-Score normalization) and time alignment technology, by converting the time series data of the load current, the ambient temperature and the cooling medium flow rate into the standard normal distribution with the mean of 0 and the standard deviation of 1, the influence of the dimensional and numerical range between different data dimensions is eliminated. At the same time, according to the time stamp of data acquisition, the alignment operation is carried out to ensure that the load current, the ambient temperature and the cooling medium flow rate data at the same time can be accurately corresponded. Then, the standardized load current, ambient temperature and cooling medium flow rate data at each time point are combined into a feature vector in a certain order, for example [load current standardized value, ambient temperature standardized value, cooling medium flow rate standardized value], and arranged in time sequence to form a time series, thereby obtaining the time series of the transformer thermal behavior feature vector. In this way, the transformer thermal behavior feature with comparability and time synchronization is effectively constructed, which lays a foundation for subsequent data analysis.

[0042] Specifically, the step S23, based on the time stamp of each transformer thermal behavior feature vector in the time sequence of the transformer thermal behavior feature vectors, respectively embeds a transformer running time length label and the transformer cooling mode one-hot encoding feature vector in the respective transformer thermal behavior feature vector to obtain a time sequence of enhanced transformer thermal behavior feature vectors. It should be understood that since the thermal characteristics of the transformer are not only affected by the instantaneous working condition (such as the current load, flow rate), but also closely related to the running time length and the active control strategy of the cooling system. Therefore, in order to enhance the representation ability of the features, the present application based on the feature embedding technology, by embedding the running time length (such as the cumulative hours) and the one-hot encoding of the cooling mode into the transformer thermal behavior feature vector of each time step, to integrate the global context information of the transformer thermal behavior characteristics. Specifically, first, according to the time stamp, calculate the transformer running time length corresponding to each transformer thermal behavior feature vector, and convert it into a suitable numerical form, for example, a numerical value in hours. Then, add the running time length label and the above-mentioned transformer cooling mode one-hot encoding feature vector to each transformer thermal behavior feature vector in a certain order to form an expanded high-dimensional feature. For example, [load current standardized value, ambient temperature standardized value, cooling medium flow rate standardized value, running time length, cooling mode one-hot encoding], thereby obtaining a time sequence of enhanced transformer thermal behavior feature vectors. In this way, the thermal cumulative effect of the transformer and the long-term influence of the cooling strategy on the thermal state can be explicitly modeled, thereby improving the adaptability of the model to complex working conditions.

[0043] Specifically, in one specific example of the present application, the step S24 comprises: inputting the time series of the enhanced transformer thermal behavior feature vector into a time series encoder based on an LSTM model to obtain the time series of the transformer operating state feature vector. It should be understood that, since the change of the transformer thermal state has strong time sequence dependence (such as the current temperature being affected by the cumulative effect of historical load and cooling process). Therefore, in order to extract the high-order time sequence pattern in the time series of the enhanced transformer thermal behavior feature vector, the present application is based on the time series modeling capability of the long short-term memory network (LSTM), through its gating mechanism (input gate, forget gate, output gate) to selectively remember the historical thermal state information of the transformer, while ignoring irrelevant noise data, to achieve accurate prediction of the future thermal state of the transformer. Specifically, the time series of the enhanced transformer thermal behavior feature vector is input into a multi-layer LSTM network, and each time step of the LSTM unit receives the enhanced transformer thermal behavior feature vector at the current time step and updates its internal state according to the historical information. Through layer-by-layer transmission and processing, the LSTM network can capture the long-term dependence in the time series of the enhanced transformer thermal behavior feature vector, that is, the complex pattern of the change of the transformer thermal state over time and operating conditions. Finally, the output layer of the LSTM network generates the hidden state at each time step to obtain the time series of the transformer operating state feature vector. Based on this, each time step feature in the time series of the transformer operating state feature vector not only contains the instantaneous feature of the transformer thermal behavior, but also incorporates historical operating conditions and long-term running effect information, providing high discriminative features for subsequent baseline temperature prediction.

[0044] In the above transformer operating state monitoring method, the step S3 inputs the time series of the transformer operating state feature vector into the trained transformer normal thermal state prediction model to obtain the temperature baseline value of the transformer key monitoring point at the current time point, and the transformer key monitoring point includes the winding, the core and the oil top layer. That is, based on the current operating state of the transformer, in order to determine its expected temperature under normal conditions, in order to provide a baseline for overheating risk warning, the present application is based on the powerful feature learning and prediction capability of the deep learning model, by using the transformer normal thermal state prediction model trained on a large number of healthy transformer operating data to analyze the time series of the transformer operating state feature vector, to utilize the correlation mapping relationship between the thermal behavior pattern of the normal fault-free transformer under different operating conditions and the temperature of each key monitoring point learned by the transformer normal thermal state prediction model in the training process, to predict the expected temperature of each key monitoring point of the transformer under normal state under the current operating condition, and to reveal the expected thermal state of the transformer under fault-free state. Wherein, Figure 4 The flow chart of the sub-step S3 of the transformer operating state monitoring method according to the embodiment of the present application. As Figure 4As shown, the step S3 includes steps of: S31, performing local semantic enhancement based double-layer message passing coding on the time sequence of the transformer operating state feature vectors to obtain a transformer operating state feature time sequence propagation coding vector; and S32, performing feature decoding on the transformer operating state feature time sequence propagation coding vector to obtain the temperature reference value of the current time point transformer key monitoring point.

[0045] Specifically, the step S31 performs local semantic enhancement based double-layer message passing coding on the time sequence of the transformer operating state feature vectors to obtain a transformer operating state feature time sequence propagation coding vector. Specifically, although the time sequence modeling method based on the LSTM model can effectively capture the long-term time sequence dependence relationship of the transformer operating state, there is a problem of insufficient capture of local time domain feature interaction and fine structure information, and the transformer operating states of adjacent time points often contain stronger time sequence correlation and transient feature change law. Therefore, in order to further enhance the feature representation, the application proposes a local semantic enhancement based double-layer message passing coding method, which enhances the feature representation ability of each transformer operating state feature vector by using the local neighborhood context information of the transformer operating state feature vector, so as to realize effective transmission of transformer thermal state information between adjacent time points, and then aggregate through global feature interaction to obtain a more comprehensive and fine transformer operating state feature time sequence propagation coding vector, so as to more accurately reflect the dynamic change process of the transformer thermal state. Wherein, Figure 5 The flowchart of the sub-step S31 of the transformer operating state monitoring method according to the embodiment of the application is shown in FIG. 4. Figure 5 As shown, the step S31 includes steps of: S311, determining the window size of the local semantic enhancement perception window of each transformer operating state feature vector based on the feature distribution of each transformer operating state feature vector in the time sequence of the transformer operating state feature vectors; S312, performing local semantic embedding enhancement on each transformer operating state feature vector based on all transformer operating state feature vectors in the local semantic enhancement perception window of the transformer operating state feature vector to obtain a time sequence of transformer operating state feature local semantic enhancement coding vectors; and S313, performing global time domain transmission coding on the time sequence of the transformer operating state feature local semantic enhancement coding vectors to obtain the transformer operating state feature time sequence propagation coding vector.

[0046] More specifically, the step S311 is expressed by a formula as follows:

[0047]

[0048] Wherein, exp(·) represents an exponential function operation with e as the base, h i and hj respectively represent the i-th and j-th transformer operating state feature vectors in the time series of transformer operating state feature vectors, ||·|| represents the calculation of the Euclidean norm, τ represents the temperature coefficient, s ij represents the feature correlation degree between h i and h j , is a preset neighborhood, ∈ represents a first smoothing term, γ represents a second smoothing term, k represents the index of the vector, represents the correlation weight coefficient of the j-th transformer operating state feature vector in the preset neighborhood with respect to h i , log2(·) represents the logarithmic function with base 2, ε i represents the near-neighbor feature distribution entropy of the vector h i , w max represents the preset maximum window size, w i represents the window size of the local semantic enhancement perception window corresponding to the transformer operating state feature vector h i .

[0049] That is, according to the feature distribution characteristics of each vector in the time series of transformer operating state feature vectors, a local semantic enhancement perception window that can best capture the semantically relevant local environment is dynamically adapted for each transformer operating state feature vector, thereby avoiding the problem of insufficient information capture or redundancy caused by the inability to match the feature distribution difference when a fixed window is used. Each local semantic enhancement perception window can be adaptively focused on a semantically relevant local feature set to accurately cover the local context with high semantic correlation to the transformer operating state feature vector, thereby laying a more accurate feature representation foundation for global time domain transfer coding and temperature reference value prediction.

[0050] More specifically, in one specific example of the present application, the step S312 comprises: inputting all transformer operating state feature vectors in the local semantic enhancement perception window of the transformer operating state feature vector into a local neighborhood enhancement fusion network based on an attention mechanism to obtain the transformer operating state feature local semantic enhancement coding vector, which is represented by the formula:

[0051]

[0052] wherein softmax(·) represents a normalized exponential function, W q , W k and W g represent different weight parameter matrices in the local neighborhood enhancement fusion network, (·)T represents the transpose of the vector, d represents the feature dimension of h i , and α ij represents h jcorresponding attention weight factor, Sigmoid(·) represents a sigmoid activation function, h i corresponding transformer operating state feature local semantic enhancement encoding vector.

[0053] That is, by local neighborhood enhancement fusion based on attention mechanism, the context features with high semantic correlation degree with the transformer operating state feature vector within the local semantic enhancement perception window can be dynamically focused, and the extraction of key local features is strengthened and irrelevant noise is suppressed by using the characteristics of adaptive weight allocation of attention mechanism. Based on this, the generated transformer operating state feature local semantic enhancement encoding vector not only retains the feature details at the current time, but also integrates the associated information at the adjacent time, significantly improves the description ability of the feature to the local dynamic change of the transformer operating state, and further makes the subsequent temperature benchmark value prediction based on the feature more accurately reflect the normal thermal state of the transformer under the current working condition, provides a more reliable benchmark for the overheating risk assessment, enhances the capture sensitivity of the system to the local abnormal features in the transformer operation, and reduces the misjudgment or omission caused by insufficient representation of local features.

[0054] In particular, considering that the temperature coefficient τ and the first smoothing term ∈ and the second smoothing term γ are introduced in the determination process of the window size of the local semantic enhancement perception window, the deformation measure quantization needs to be considered synchronously for the weight parameter matrix W g , W q and W k in the local neighborhood enhancement fusion process to improve the context semantic embedding under the predetermined local semantic enhancement perception window. Based on this, in one preferred example of the present application, all transformer operating state feature vectors in the local semantic enhancement perception window of the transformer operating state feature vector are input into the local neighborhood enhancement fusion network based on attention mechanism to obtain the transformer operating state feature local semantic enhancement encoding vector, including: first, performing local neighborhood semantic dynamic adaptation correction on the weight parameter matrix of the local neighborhood enhancement fusion network based on attention mechanism to obtain an updated weight parameter matrix; and then based on the updated weight parameter matrix, performing local neighborhood enhancement fusion based on attention mechanism on all transformer operating state feature vectors in the local semantic enhancement perception window of the transformer operating state feature vector to obtain the transformer operating state feature local semantic enhancement encoding vector.

[0055] Specifically, first, for the vector (τ, ∈, γ), linear interpolation operation is performed to make its length consistent with the transformer operating state feature vector, and then one-dimensional convolution is performed to complete nonlinear curvature transformation, so as to obtain the strain coupling feature vector V α , and the weight parameter matrix Wg , W q and W k As the mapping reference, the curvature strain adaptive feature projection vectors of the weight parameter matrix are obtained:

[0056] V α1 = V α W g + V α

[0057] V α2 = V α W q + V α

[0058] V α3 = V α W k + V α

[0059] wherein V α represents the strain coupling feature vector, V α1 , V α2 and V α3 respectively represent the curvature strain adaptive feature projection vectors corresponding to the weight parameter matrices W g , W q and W k .

[0060] Then, considering the correlation constraint for the curvature strain, the core curvature guiding factor of the parameter dynamic modulation can be simplified, that is, the exogenous topological disturbance excitation can be ignored, and thus the above curvature strain adaptive feature projection vectors are respectively correlated with the eigenvectors of the weight parameter matrix, that is, V eg , V eq and V ek in the following formula, to modify the weight parameter matrix, which is expressed as:

[0061] W ′g = (V α1 T V eg )⊙W g

[0062] W ′q = (V α2 T V eq )⊙W q

[0063] W ′k = (V α3 T V ek )⊙W k

[0064] wherein V eg , V eq and V ek respectively represent the eigenvectors of the weight parameter matrices W g , W q and W k , represents the dot product operation, W ′g , W ′q and W ′k respectively represent the updated weight parameter matrices corresponding to W g , W q and W k .

[0065] In this way, the local semantic embedding enhancement based on the weight parameter matrix can consider the introduction of the deformation field coupling component in the window size determination of the local semantic enhancement perception window, thereby improving the context semantic embedding expression effect of the sequence of transformer operating state feature local semantic enhancement coding vectors based on the local semantic enhancement perception window.

[0066] More specifically, in one specific example of the present application, the step S313 comprises inputting the time sequence of transformer operating state feature local semantic enhancement coding vectors into a message global transmission coding network based on a Transformer architecture to obtain the transformer operating state feature time sequence propagation coding vector, which is represented by the formula:

[0067]

[0068] wherein, represents the time sequence of transformer operating state feature local semantic enhancement coding vectors, and respectively represent the first, second and i-th transformer operating state feature local semantic enhancement coding vectors in , and Transformer(·) represents the Transformer architecture, v f represents the transformer operating state feature time sequence propagation coding vector.

[0069] That is, the Transformer architecture is used to model the time series of the transformer operating state feature local semantic enhanced encoding vector in the global range, so that the transformer operating state feature local semantic enhanced encoding vector at each time point can dynamically integrate the feature information of all historical time points in the entire time sequence. Based on this, the generated transformer operating state feature time series propagation encoding vector can effectively capture the evolution law and coupling relationship of the load current, environmental temperature and other parameters in the long time sequence dimension, thereby improving the prediction accuracy of the temperature reference value of the transformer key monitoring point, providing more comprehensive time sequence feature support for the deviation value based overheating risk warning, and enhancing the ability to capture potential thermal fault signs of the transformer in long-term operation.

[0070] Specifically, the step S32 decodes the transformer operating state feature time series propagation encoding vector to obtain the temperature reference value of the transformer key monitoring point at the current time point. That is, in order to convert the transformer operating state feature time series propagation encoding vector into a temperature prediction value with clear physical meaning, the application further decodes the transformer operating state feature time series propagation encoding vector using the decoding layer of the transformer normal thermal state prediction model. Specifically, the decoding layer is based on a multi-task learning framework and predicts the temperature reference value of each key monitoring point of the transformer through multiple decoding sub-networks in parallel. Each decoding sub-network includes a fully connected layer and an activation function, which is used to map a high-dimensional feature vector to a temperature prediction value. In the decoding process, each decoding sub-network shares the same transformer operating state feature time series propagation encoding vector as input, but has independent network parameters and output, so as to simultaneously predict the temperature reference values of multiple transformer key monitoring points such as transformer windings, cores and oil top layers. In this way, not only the accuracy and efficiency of temperature prediction are improved, but also the sensitivity and adaptability of the model to temperature changes of different monitoring points are enhanced.

[0071] In the above transformer operating state monitoring method, the step S4 acquires the temperature measurement value of the transformer key monitoring point at the current time point in real time. It should be understood that the temperature reference value is the expected temperature that the transformer key monitoring point should reach under the ideal fault-free state. Therefore, in order to judge whether the transformer has an overheating risk, the actual thermal state of the transformer needs to be further compared with the expected normal thermal state. Based on this, the application acquires the real temperature condition of the transformer at the current time point through the temperature sensor installed on each key monitoring point of the transformer, as a direct basis for comparison with the temperature reference value predicted by the model.

[0072] In the transformer operating state monitoring method, in step S5, a deviation value between the temperature measurement value of the transformer key monitoring point at the current time point and the temperature reference value is calculated, and based on comparison between the deviation value and a preset threshold value, it is determined whether the transformer has an overheating risk. That is, in order to quantify the difference between the actual thermal state of the transformer and the healthy thermal state of the transformer under the current working condition, and to judge whether there is a potential overheating fault risk based on the degree of the difference, the application calculates the difference between the measured temperature and the reference temperature, and compares it with the preset allowed deviation range, to realize accurate identification of the overheating risk. Specifically, for each key monitoring point (winding, core, oil top layer), the arithmetic difference between the current time point temperature measurement value and the corresponding temperature reference value predicted is calculated to reveal the degree of actual temperature exceeding the normal expected level. Then, the calculated deviation value of each key monitoring point is compared with the preset threshold value of each key monitoring point. Here, the preset threshold value is the allowed deviation upper limit set according to the type of transformer, design margin, historical operation experience and statistical analysis of model prediction error. For example, the deviation threshold value of the winding may be set to 5℃. If the temperature deviation value of any key monitoring point exceeds the corresponding preset threshold value, it is determined that the transformer has an overheating risk at the monitoring point, and a corresponding alarm signal or prompt information may be triggered to instruct the operation and maintenance personnel to further check or take intervention measures. In this way, the sensitivity and accuracy of the overheating risk monitoring can be significantly improved, and the temperature fluctuations caused by normal working condition changes are effectively filtered out, providing more reliable technical support for ensuring the safe operation of the transformer.

[0073] In summary, the transformer operating state monitoring method based on the embodiments of the application is illustrated, which uses an artificial intelligence algorithm based on deep learning to perform feature engineering and time series joint analysis on the load current, cooling method, cooling medium flow rate and environmental temperature of the transformer to mine the thermal behavior characteristics of the transformer. On this basis, a deep learning model trained on a large amount of healthy transformer operating data is further used for transformer temperature prediction to obtain the expected temperature of each key monitoring point of the transformer under normal state, and then, based on the difference between the expected temperature of the transformer under normal state and the actual measured temperature, an overheating risk warning is performed. This method can realize real-time and accurate warning of the overheating risk of the transformer, overcome the limitations of the traditional fixed temperature threshold monitoring method, and improve the sensitivity and robustness of the overheating risk identification.

[0074] Further, a transformer operating state monitoring system is also provided.

[0075] Figure 6 A block diagram of the transformer operating state monitoring system according to the embodiments of the application is shown in FIG. 1. As shown in FIG. 1, the transformer operating state monitoring system comprises a transformer operating state monitoring method according to the embodiments of the application, a transformer operating state monitoring device, a transformer operating state monitoring platform and a transformer operating state monitoring database. Figure 6As shown, the transformer operating state monitoring system 100 according to the embodiment of the present application comprises: an operating state data acquisition module 110, configured to acquire transformer operating state data in a predetermined time window in real time, wherein the transformer operating state data comprises a time sequence of load current, a time sequence of ambient temperature, a transformer cooling mode, and a time sequence of cooling medium flow rate; an operating state feature engineering module 120, configured to perform operating state feature engineering on the transformer operating state data to obtain a time sequence of transformer operating state feature vectors; a normal thermal state prediction module 130, configured to input the time sequence of transformer operating state feature vectors into a trained transformer normal thermal state prediction model to obtain a temperature reference value of a transformer key monitoring point at a current time point; a real-time temperature measurement module 140, configured to acquire a temperature measurement value of the transformer key monitoring point at the current time point in real time; and an overheating risk assessment module 150, configured to calculate a deviation value between the temperature measurement value of the transformer key monitoring point at the current time point and the temperature reference value, and determine whether the transformer has an overheating risk based on a comparison between the deviation value and a preset threshold value.

[0076] Here, those skilled in the art can understand that the specific operations of each module in the above transformer operating state monitoring system have been described in detail above with reference to the description of the transformer operating state monitoring method Figures 1 to 5 of the above embodiment, and therefore, the repeated description thereof will be omitted.

[0077] The above describes the basic principles of the present application in combination with specific embodiments, but it should be pointed out that the advantages, advantages, effects, etc. mentioned in the present application are only examples and are not limited, and these advantages, advantages, effects, etc. cannot be considered as the must-have of each embodiment of the present application. In addition, the specific details of the above embodiments are only for the purpose of example and for the purpose of understanding, and the above details do not limit the present application to the must-use of the above specific details.

[0078] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in a certain embodiment can be referred to the related description of other embodiments. In the several embodiments provided by the present application, it should be understood that the disclosed system and method can be implemented by other means. For example, the system embodiments described above are only illustrative, for example, the unit division is only a logical function division, and actual implementation can have another division manner. The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment scheme.

[0079] It will be obvious to a person skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments and can be implemented in other embodiments without departing from the scope of the application. The foregoing embodiments are to be considered in all respects only as illustrative and not restrictive, and all changes coming within the meaning and equivalency range of the appended claims are intended to be embraced therein. Any reference signs in the claims should not be construed as limiting the scope of the claims.

[0080] Furthermore, it is clear that the word "comprising" does not exclude other elements or steps, and the singular does not exclude the plural. A single processor or other unit can fulfil the functions of several units recited in the system claims.

[0081] Finally, it should be noted that the above- described description is merely intended to illustrate and not to limit the present application. Some of the embodiments have been described above purely by way of example and without any intention that they should be construed as limiting the present application. Modifications or equivalent arrangements to those described above are possible and within the scope of the present application.

Claims

1. A method for monitoring transformer operating status, characterized in that: include: Real-time acquisition of transformer operating status data within a predetermined time window, the transformer operating status data including a time series of load current, a time series of ambient temperature, a time series of transformer cooling mode, and a time series of cooling medium flow rate; Performing operation state feature engineering on the transformer operation state data to obtain a time series of transformer operation state feature vectors; Inputting the time series of the transformer operating state feature vector into the trained transformer normal thermal state prediction model to obtain the temperature reference value of the transformer key monitoring point at the current time point; Real-time acquisition of temperature measurement values ​​of key monitoring points of the transformer at the current time point; The deviation between the temperature measurement value of the key monitoring point of the transformer at the current time point and the temperature reference value is calculated, and based on the comparison between the deviation value and a preset threshold, it is determined whether the transformer has an overheating risk.

2. The transformer operating status monitoring method according to claim 1, characterized in that: The key monitoring points of the transformer include windings, iron core and oil top layer.

3. The transformer operating status monitoring method according to claim 1, characterized in that: Performing operation state feature engineering on the transformer operation state data to obtain a time series of transformer operation state feature vectors, including: Performing one-hot encoding on the transformer cooling mode to obtain a one-hot encoding feature vector of the transformer cooling mode; Normalizing and aligning the time series of the load current, the ambient temperature, and the cooling medium flow rate to obtain a time series of a transformer thermal behavior characteristic vector; Based on the timestamp of each transformer thermal behavior feature vector in the time series of the transformer thermal behavior feature vector, embedding the transformer operating time mark and the transformer cooling mode one-hot encoding feature vector in each transformer thermal behavior feature vector to obtain an enhanced time series of the transformer thermal behavior feature vector; A time series modeling analysis is performed on the time series of the enhanced transformer thermal behavior characteristic vector to obtain the time series of the transformer operation state characteristic vector.

4. The transformer operating status monitoring method according to claim 3, characterized in that: Performing time series modeling analysis on the time series of the enhanced transformer thermal behavior characteristic vector to obtain the time series of the transformer operating state characteristic vector includes: The time series of the enhanced transformer thermal behavior feature vector is input into a time series encoder based on an LSTM model to obtain the time series of the transformer operating state feature vector.

5. The transformer operating status monitoring method according to claim 1, characterized in that: Inputting the time series of the transformer operating state feature vector into the trained transformer normal thermal state prediction model to obtain the temperature reference value of the transformer key monitoring point at the current time point, including: Performing double-layer message passing encoding based on local semantic enhancement on the time series of the transformer operating state feature vector to obtain a transformer operating state feature time series propagation coding vector; The transformer operating state characteristic time series propagation coding vector is feature decoded to obtain the temperature reference value of the transformer key monitoring point at the current time point.

6. The transformer operating status monitoring method according to claim 5, characterized in that: The time series of the transformer operating state feature vector is subjected to double-layer message passing encoding based on local semantic enhancement to obtain a transformer operating state feature time series propagation encoding vector, including: Determining the window size of the local semantic enhancement perception window of each transformer operating state feature vector based on the feature distribution of each transformer operating state feature vector in the time series of the transformer operating state feature vector; Based on all transformer operating state feature vectors in the local semantic enhancement perception window of each transformer operating state feature vector, each transformer operating state feature vector is locally semantically embedded and enhanced to obtain a time series of transformer operating state feature local semantic enhancement encoding vectors; Global time domain transfer coding is performed on the time series of the local semantic enhancement coding vector of the transformer operating state feature to obtain the transformer operating state feature time series propagation coding vector.

7. The transformer operating status monitoring method according to claim 6, characterized in that: Based on all transformer operating state feature vectors in the local semantic enhancement perception window of each transformer operating state feature vector, each transformer operating state feature vector is locally semantically embedded and enhanced to obtain a time series of transformer operating state feature local semantic enhancement encoding vectors, including: All transformer operating state feature vectors in the local semantic enhancement perception window of the transformer operating state feature vector are input into the local neighborhood enhancement fusion network based on the attention mechanism to obtain the transformer operating state feature local semantic enhancement encoding vector.

8. The transformer operating status monitoring method according to claim 7, characterized in that: Inputting all transformer operating state feature vectors in the local semantic enhancement perception window of the transformer operating state feature vector into a local neighborhood enhancement fusion network based on an attention mechanism to obtain a local semantic enhancement encoding vector of the transformer operating state feature, including: Performing local neighborhood semantic dynamic adaptation correction on the weight parameter matrix of the local neighborhood enhancement fusion network based on the attention mechanism to obtain an updated weight parameter matrix; Based on the updated weight parameter matrix, all transformer operating state feature vectors in the local semantic enhancement perception window of the transformer operating state feature vector are subjected to local neighborhood enhancement fusion based on the attention mechanism to obtain the transformer operating state feature local semantic enhancement encoding vector.

9. The transformer operating status monitoring method according to claim 8, characterized in that: Performing global time domain transfer coding on the time series of the local semantic enhancement coding vector of the transformer operating state feature to obtain the transformer operating state feature time series propagation coding vector, including: The time series of the local semantic enhancement coding vector of the transformer operating state feature is input into the message global domain transmission coding network based on the Transformer architecture to obtain the transformer operating state feature time series propagation coding vector.

10. A transformer operating status monitoring system, characterized in that: include: An operating status data acquisition module is used to acquire the transformer operating status data in real time within a predetermined time window, wherein the transformer operating status data includes a time series of load current, a time series of ambient temperature, a time series of transformer cooling mode, and a time series of cooling medium flow rate; An operating state feature engineering module, configured to perform operating state feature engineering on the transformer operating state data to obtain a time series of transformer operating state feature vectors; A normal thermal state prediction module is used to input the time series of the transformer operating state feature vector into the trained transformer normal thermal state prediction model to obtain the temperature reference value of the transformer key monitoring point at the current time point; A real-time temperature measurement module is used to obtain the temperature measurement value of the key monitoring point of the transformer at the current time point in real time; The overheating risk assessment module is used to calculate the deviation between the temperature measurement value of the key monitoring point of the transformer at the current time point and the temperature reference value, and determine whether the transformer has an overheating risk based on the comparison between the deviation value and a preset threshold.

Images