Transformer loss monitoring method and transformer loss monitoring system

Abstract

The embodiment of the invention provides a transformer loss monitoring method and a transformer loss monitoring system, and the monitoring method comprises the steps: dynamically updating winding resistance, excitation reactance and iron core magnetic conductivity according to an electromagnetic parameter and a temperature parameter in a state vector space through employing a complex frequency domain parameter identification algorithm which introduces a fusion expansion Cole-Cole model; the no-load loss and the load loss are decoupled through an improved all-electric difference method, the dynamically updated excitation reactance and the iron core magnetic conductivity are used for decoupling the no-load loss, and the dynamically updated winding resistor is used for decoupling the load loss; and taking the decoupled loss data and the associated parameters in the state vector space as the input of an LSTM-Transform hybrid model loaded on the edge calculation module, and predicting and outputting the load loss trend in the future time period. According to the transformer loss monitoring method, the real-time performance, the accuracy and the adaptability of transformer loss monitoring are improved.

Description

Technical Field

[0001] This application relates to the field of transformer loss monitoring technology, and more particularly to transformer loss monitoring methods and transformer loss monitoring systems. Background Technology

[0002] As a core piece of equipment in the power system, transformers account for a large proportion of the total power transmission and distribution losses.

[0003] In recent years, the deepening of smart grid construction has accelerated the transformation of transformer monitoring technology from traditional offline periodic testing to online real-time monitoring. Traditional offline testing relies on manual inspections, typically involving oil sample analysis and electrical parameter testing every few months. This not only suffers from data lag but also struggles to capture sudden changes in losses under instantaneous load fluctuations. In contrast, online monitoring technology, through the implantation of sensors to achieve continuous data acquisition, has become an important means of improving the reliability of transformer operation.

[0004] In related technologies, transformer losses consist of no-load losses (iron losses) and load losses (copper losses). The former is related to core hysteresis and eddy current effects, and is significantly affected by core temperature and magnetic field distribution; the latter depends on winding current, resistance, and leakage flux loss, while winding resistance changes nonlinearly with temperature. When calculating losses solely based on voltage and current signals, it is impossible to distinguish additional losses caused by factors such as core saturation and winding skin effect, resulting in significant loss monitoring errors. Therefore, current mainstream online monitoring solutions still have significant technical bottlenecks, most notably the limitations of single-parameter monitoring. For example, by collecting voltage / current signals through electricity meters and current transformers and combining them with load rate correction models to calculate real-time losses, this single-parameter monitoring method cannot fully reflect the transformer's operating status, resulting in large errors in loss monitoring. Summary of the Invention

[0005] The purpose of this application is to provide a transformer loss monitoring method and a transformer loss monitoring system, so as to at least solve the technical problem of large loss monitoring error in existing energy consumption monitoring technologies.

[0006] To achieve the above objectives, the embodiments of this application provide the following technical solutions.

[0007] In a first aspect, according to an embodiment of this application, a transformer loss monitoring method is provided, comprising the following steps: A multiphysics fusion model based on a dynamic Bayesian network is constructed. An online learning mechanism based on a two-factor dynamic update strategy is introduced into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time. The multimodal data preprocessed by the fusion model are coupled and analyzed to generate the transformer state vector space. Based on the electromagnetic and temperature parameters in the state vector space, the winding resistance, excitation reactance and core permeability are dynamically updated using a complex frequency domain parameter identification algorithm that incorporates a fusion extended Cole-Cole model. The no-load loss and load loss are decoupled by an improved all-electric differential method. The dynamically updated excitation reactance and core permeability are used to decouple the no-load loss, and the dynamically updated winding resistance is used to decouple the load loss. The decoupled loss data and the correlation parameters in the state vector space are used as inputs to the LSTM-Transformer hybrid model loaded in the edge computing module to predict and output the load loss trend in the future time period. In the hybrid model, the loss time series features are captured by bidirectional LSTM layers, and the long-range correlation of multi-physics fields is mined by the modular multi-head attention structure Transformer layer.

[0008] Preferably, in the step of constructing a multiphysics fusion model based on a dynamic Bayesian network, the nodes of the dynamic Bayesian network are divided into physical layer nodes, feature layer nodes, and decision layer nodes. The physical layer nodes correspond to multimodal raw data, including winding temperature field data, voltage and current spectrum data, characteristic gas concentration data, vibration acceleration signals, and frequency cepstral coefficients of acoustic fingerprint signals. The feature layer nodes are high-order features extracted from the multimodal data of the physical layer nodes, including temperature difference gradient, electromagnetic harmonic distortion rate, gas component ratio, vibration dominant frequency energy ratio, and acoustic signature feature entropy. The decision-making layer nodes are the fused state assessment parameters, including insulation aging index, winding overheat risk value, and loss anomaly probability.

[0009] Preferably, the step of introducing an online learning mechanism based on a two-factor dynamic update strategy into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time includes: In the conventional time-varying factor, the window length of the sliding window is dynamically adjusted based on the load volatility. The multimodal data within the time window is divided into recent data segments, medium-term data segments, and long-term data segments according to the collection time. Weights are assigned and corrected for each period data segment. Based on the weighted multimodal data, the conditional probability table is periodically updated using the maximum likelihood estimation method. In the abnormal response factor, when the rate of change and duration of the physical layer node parameters meet the preset abnormal threshold conditions, an abnormal mode is triggered. The preset abnormal threshold conditions include temperature change characteristics, characteristic gas concentration change characteristics in oil, and vibration signal change characteristics. In the abnormal mode, the update frequency of the conditional probability table of the corresponding node is increased, and Bayesian estimation is used for iterative update. Through L1 regularization to constrain parameter drift, the KL divergence of the probability table before and after the update is kept within the preset divergence threshold.

[0010] Preferably, the step of performing coupled analysis on the preprocessed multimodal data of the fusion model to generate the transformer state vector space includes: Spatiotemporal correlation processing of multimodal data includes: aligning multimodal data in the time dimension; and generating a continuous field distribution from discretely distributed physical layer node data through spatial interpolation in the spatial dimension, and establishing a spatial correlation matrix with the modal data at the corresponding spatial location. The spatial correlation matrix is ​​used to characterize the correlation strength of multimodal data at different spatial locations. The weight coefficients of each data point in the multimodal data are calculated based on the feature importance evaluation algorithm. The correlation degree of the features is judged by combining the spatial correlation matrix. When the correlation degree of any two features exceeds the preset correlation threshold, the features are selectively retained based on the weight coefficients to form a coupled feature set. Based on the coupled feature set, a time decay factor based on the correlation strength is assigned to each dimension feature to construct an initial state vector; at the same time, a dynamic adaptation mechanism for operating conditions is introduced to increase the corresponding operating condition feature dimensions, forming a transformer state vector space with dynamically adjustable dimensions.

[0011] Preferably, the method of introducing a complex frequency domain parameter identification algorithm that integrates and extends the Cole-Cole model, and dynamically updating the winding resistance, excitation reactance, and core permeability based on the electromagnetic and temperature parameters in the state vector space, includes: Electromagnetic and temperature parameters are extracted from the state vector space. The temperature parameters are used for temperature compensation preprocessing of the complex frequency domain impedance spectrum. The complex frequency domain impedance spectrum is obtained by time-frequency conversion based on the acquired transformer signal. The electromagnetic parameters include harmonic components, excitation current, and magnetic field strength. A dynamic model of winding resistance is constructed, and the relaxation characteristics of the extended Cole-Cole model are integrated. The winding resistance is represented as the superposition of DC resistance and additional resistance of each harmonic. The additional resistance is associated with the harmonic components of electromagnetic parameters in the state vector space. A frequency-dependent model of excitation reactance is constructed, and a complex permeability parameter is introduced. The excitation current and core temperature data in the state vector space are correlated to describe the variation law of excitation reactance at different frequencies. A coupled model of iron core permeability is constructed, and the spatial distribution characteristics of permeability are corrected by combining the magnetic field strength in the state vector space. A nonlinear least squares fitting algorithm is adopted, with the preprocessed complex frequency domain impedance spectrum as input. The initial identification values ​​of winding resistance, excitation reactance and core permeability are calculated iteratively based on the fusion extended Cole-Cole model. Electromagnetic parameters and temperature parameters in the state vector space are introduced as constraints to correct the initial identification values. The winding resistance, excitation reactance, and core permeability are updated based on parameter update triggering conditions. The parameter update triggering conditions include changes in electromagnetic parameters or temperature parameters in the state vector space exceeding a preset threshold.

[0012] Preferably, the step of decoupling no-load loss and load loss through the improved all-electric differential method includes: Input the real-time parameters obtained through complex frequency domain parameter identification, including dynamically updated winding resistance, excitation reactance and core permeability, and extract the temperature field gradient, harmonic distortion rate and vibration characteristic parameters in the state vector space as auxiliary decoupling variables. A sub-model for calculating no-load loss is constructed. Based on the excitation reactance and core permeability, the overall proportion of hysteresis loss and eddy current loss is corrected by combining the core average temperature. The no-load time harmonic component in the associated state vector space is also included in the no-load loss calculation sub-model. When the core temperature field gradient exceeds a preset threshold, the no-load loss calculation results are thermally corrected. A load loss calculation sub-model is constructed. Based on the winding resistance and the fundamental and harmonic components of the current, a skin effect correction coefficient is introduced to calculate the winding copper loss and harmonic additional loss. At the same time, the vibration characteristic parameters in the state vector space are associated. When the vibration dominant frequency offset exceeds the set range, the mechanical additional loss in the load loss is compensated. A differential coupling equation is established, and the total loss is decomposed into no-load loss component and load loss component through a separation coefficient matrix. The elements of the separation coefficient matrix are dynamically adjusted based on the transformer's rated parameters and real-time operating conditions.

[0013] Preferably, the structure of the LSTM-Transformer hybrid model includes: The preprocessing layer includes a parallel convolutional module and a temporal module. The convolutional module uses multiple convolutional kernels to extract local features from the input multidimensional temporal data and outputs a local temporal feature vector. The temporal module corrects temporal misalignment of the data through a gated recurrent unit (GRU) and outputs an aligned global temporal feature vector. The outputs of the convolutional module and the temporal module are fused into an input feature matrix through a feature concatenation operation. The feature coding layer adopts a bidirectional LSTM structure, including a forward LSTM unit and a backward LSTM unit. The outputs of the two units are weighted and summed to generate bidirectional temporal features. A cross-layer feature interaction layer is set up, in which a feature mapping matrix and a cross-layer attention mechanism are set. The feature mapping matrix transforms the bidirectional temporal features into intermediate features that match the dimensions of the Transformer layer. The cross-layer attention mechanism calculates the similarity weights between the LSTM features and the input features of the Transformer layer, and enables the features of the two layers to be dynamically fused through residual connections. The Transformer enhanced decoding layer adopts a modular multi-head attention structure, which includes multiple parallel attention sub-modules that perform attention calculations for electrical parameter features, thermal features, and chemical features respectively. A dynamic attention scaling mechanism is introduced, inversely proportional to the time-varying volatility of the feature, to reduce the interference of noisy features on the attention weights. The output fusion layer includes a feature selection gate and an output regression unit. The feature selection gate selects key features from LSTM features and Transformer features based on L1 regularization. The output regression unit adopts a two-layer fully connected network and outputs a standardized load loss prediction value through the Sigmoid activation function. The learnable parameters of the two fully connected network layers in the output regression unit are dynamically calibrated through the online distillation mechanism of the edge computing module.

[0014] Preferably, in the feature encoding layer employing a bidirectional LSTM structure, an adaptive forget gate mechanism is introduced, whereby the gate parameters are modulated in real time by the load volatility in the state vector space, specifically including: Extract the load current sequence within a preset sliding time window from the state vector space. Calculate the load volatility, expressed as: ; In the formula, This represents the average current value within the time window. This represents the load volatility, used to characterize the relative degree of load fluctuation; n represents the total number of load current sequences. This represents the current value of the i-th load current sequence; The initial parameters of the forget gate of the bidirectional LSTM are set to The forget gate parameter is adjusted based on the load volatility, expressed as: ; In the formula, This indicates the preset load stability threshold. , This represents the adjustment coefficient. This indicates severe load fluctuations. This indicates that the load is stable, and f represents the adjusted forget gate parameter; The temporal features output by the forward LSTM unit and the temporal features output by the backward LSTM unit are weighted and fused based on the real-time forget gate parameter f, as follows: In the formula, Indicates the features after fusion. This represents the timing characteristics of the output of the forward LSTM unit. This represents the timing characteristics of the output from the backward LSTM unit.

[0015] Preferably, the step of using the decoupled loss data and the correlation parameters in the state vector space as input to the LSTM-Transformer hybrid model loaded in the edge computing module to predict and output the load loss trend in the future time period includes: Extract the decoupled load loss data and the associated parameters in the state vector space to form a multidimensional input dataset. Divide the dataset into a sample sequence by a sliding time window. Each sample contains historical data of a preset duration and corresponding labels. The edge computing module loads a pre-trained hybrid model, takes the sample sequence as input, and outputs the predicted load loss value, trend slope, and feature importance weights for a preset time period in the future. The Monte Carlo dropout method is used to calculate the confidence interval of the prediction results. When the confidence is lower than the preset threshold, the latest feature data in the state vector space is introduced to re-predict. Establish a prediction error feedback mechanism to periodically compare the deviation between the predicted value and the actual loss data: when the deviation exceeds the set range, trigger the lightweight parameter fine-tuning of the hybrid model, including: correcting the forgetting factor of the LSTM layer and the attention coefficient of the Transformer layer based on the deviation gradient, and calling the online distillation mechanism of the edge computing module during the fine-tuning process. The model training sample pool is synchronized with the state vector space update frequency, and new samples after preprocessing are stored in real time. The system also judges whether the operating conditions have changed abruptly by using the operating condition features in the state vector space. When the change index exceeds the threshold, the weight ratio of new samples in the training pool is automatically increased, and incremental training of the model is triggered, updating only the multi-head attention parameters of the Transformer layer.

[0016] Secondly, according to another embodiment of this application, a transformer loss monitoring system is provided; The transformer loss monitoring system includes the following modules: The fusion model construction module is used to construct a multiphysics fusion model based on a dynamic Bayesian network. An online learning mechanism based on a two-factor dynamic update strategy is introduced into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time. The vector space generation module is used to perform coupled analysis on the multimodal data after the fusion model preprocessing to generate the transformer state vector space; The update module is used to dynamically update the winding resistance, excitation reactance and core permeability based on the electromagnetic parameters and temperature parameters in the state vector space using a complex frequency domain parameter identification algorithm that incorporates a fusion extended Cole-Cole model. The loss decoupling module is used to decouple no-load loss from load loss using an improved all-electric differential method. The dynamically updated excitation reactance and core permeability are used to decouple no-load loss, while the dynamically updated winding resistance is used to decouple load loss. The loss prediction module is used to take the decoupled loss data and the correlation parameters in the state vector space as input to the LSTM-Transformer hybrid model loaded in the edge computing module, and predict and output the load loss trend in the future time period. In the hybrid model, the loss time series features are captured by bidirectional LSTM layers, and the long-range correlation of multi-physics fields is mined by the modular multi-head attention structure Transformer layers.

[0017] Compared with the prior art, the beneficial effects of the transformer loss monitoring method and transformer loss monitoring system of the present application embodiments are: This application constructs a dynamic Bayesian network to fuse multi-physics data and introduces a two-factor dynamic update strategy of conventional time-varying factors and abnormal response factors to update the conditional probability table, which can effectively fuse multi-modal data and solve the problem of heterogeneity of multi-source data; in particular, the two-factor strategy ensures that the model always matches the actual operating state of the transformer. This application eliminates the temporal misalignment and spatial discreteness of multimodal data by constructing a spatiotemporal alignment and spatial correlation matrix, forming a state carrier containing multi-dimensional features such as electromagnetics, temperature, and vibration; the dynamic dimension adjustment mechanism can adapt to different working conditions, ensuring that the state vector space can fully capture the complex operating state of the transformer; This application uses electromagnetic parameters and temperature coefficients in state vector space, and utilizes a complex frequency domain algorithm that integrates extended Cole-Cole models to dynamically update winding resistance, excitation reactance, and core permeability. This enables the dynamic changes of transformer parameters with frequency, temperature, and operating conditions, overcoming the shortcomings of traditional static parameters that cannot adapt to fluctuations in operating conditions. This application addresses the different physical nature of no-load loss and load loss by using corresponding dynamic parameters for decoupling. By combining differential coupling equations with dynamically adjusted separation coefficient matrices, the coupling interference between the two types of loss is effectively eliminated, thereby improving the precision of transformer loss monitoring. The bidirectional LSTM used in the hybrid model of this application can effectively capture the temporal dependence of loss data and adapt to the dynamic changes of loss over time; the modular multi-head attention mechanism can specifically mine the long-range correlation between multiple physical fields such as electrical parameters and thermal characteristics, and solve the problem of trend prediction under multi-factor coupling. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] In the attached diagram: Figure 1 A flowchart illustrating the implementation of the transformer loss monitoring method provided in this application embodiment; Figure 2 This is a structural block diagram of the transformer loss monitoring system provided in an embodiment of this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0021] According to the embodiments of this application, a method embodiment for transformer loss monitoring is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0022] Figure 1 This is a flowchart of a transformer loss monitoring method according to an embodiment of this application; like Figure 1 As shown, the transformer loss monitoring method includes the following steps: Step S101: Construct a multiphysics fusion model based on a dynamic Bayesian network, and introduce an online learning mechanism based on a two-factor dynamic update strategy into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time. This embodiment divides the nodes of the dynamic Bayesian network into physical layer nodes, feature layer nodes, and decision layer nodes, and realizes the fusion analysis of physical field data through a hierarchical structure. In this embodiment, the physical layer node corresponds to the multimodal raw data, that is, the physical layer node corresponds to the transformer multimodal raw monitoring data, which is used to reflect the original state of each physical field, including winding temperature field data, voltage and current spectrum data, characteristic gas concentration data (such as hydrogen, methane, acetylene, etc. collected by gas chromatograph), vibration acceleration signal and frequency cepstral coefficient of acoustic fingerprint signal. In this embodiment, the feature layer node is used to extract features and reduce the dimensionality of the data from the physical layer node, and is used to characterize the higher-order correlation features of the multi-physics field. Specifically, the feature layer node is a higher-order feature extracted from the multi-modal data of the physical layer node, including temperature difference gradient (the rate of change of temperature difference between the winding and the oil temperature), electromagnetic harmonic distortion rate, gas component ratio, vibration dominant frequency energy ratio and acoustic signature entropy. In this embodiment, based on the coupling analysis of the feature layer nodes, transformer condition assessment parameters are output. Among them, the decision layer nodes are the fused condition assessment parameters, including insulation aging index, winding overheating risk value, and loss anomaly probability. The insulation aging index is the insulation paper polymerization degree attenuation coefficient calculated by comprehensively considering gas composition and temperature gradient, with a value ranging from 0 to 1. The closer the coefficient is to 1, the more severe the aging. The winding overheating risk value is a weighted risk value based on winding temperature, current density, and harmonic loss. The loss anomaly probability is the probability that the actual loss deviates from the theoretical value, calculated by combining electromagnetic parameters and vibration characteristics.

[0023] In this embodiment, physical layer nodes and feature layer nodes are associated through conditional probability, and the initial conditional probability table is generated based on historical operation and maintenance data.

[0024] Step S102: Perform coupled analysis on the preprocessed multimodal data of the fusion model to generate the transformer state vector space; Step S103: Based on the electromagnetic parameters and temperature parameters in the state vector space, dynamically update the winding resistance, excitation reactance, and core permeability using a complex frequency domain parameter identification algorithm that incorporates the fusion extended Cole-Cole model. Step S104: Decouple no-load loss and load loss using an improved all-electric differential method. The dynamically updated excitation reactance and core permeability are used to decouple no-load loss, and the dynamically updated winding resistance is used to decouple load loss. Step S105: The decoupled loss data and the correlation parameters in the state vector space are used as inputs to the LSTM-Transformer hybrid model loaded in the edge computing module to predict and output the load loss trend in the future time period. In the hybrid model, the loss time series features are captured by the bidirectional LSTM layer, and the long-range correlation of multi-physics fields is mined by the modular multi-head attention structure Transformer layer.

[0025] Furthermore, in step S101 of this embodiment, the step of introducing an online learning mechanism based on a two-factor dynamic update strategy into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time includes: In the conventional time-varying factor, the window length of the sliding window is dynamically adjusted based on the load volatility. The multimodal data within the time window is divided into recent data segments, medium-term data segments, and long-term data segments according to the collection time. Weights are assigned and corrected for each period data segment. Based on the weighted multimodal data, the conditional probability table is periodically updated using the maximum likelihood estimation method. With the settings of the above embodiments, the conventional time-varying factor is used to adapt to the changes in conventional operating conditions such as slow fluctuations in transformer load and gradual changes in ambient temperature. By dynamically adjusting the data window and weighting strategy, the conditional probability table (CPT) is updated smoothly with the operating conditions. Furthermore, in this embodiment, the load fluctuation rate is used to characterize the degree of drastic change in transformer load. It is used to dynamically adjust the window length of the sliding window. First, three-phase load current data is collected from the physical layer node, including the formed current sequence. Load volatility is expressed as: ; In the formula, This represents the average current value within the time window. This represents the load volatility, used to characterize the relative degree of load fluctuation; n represents the total number of load current sequences. This represents the current value of the i-th load current sequence; In the dynamic adjustment of the sliding window length based on load volatility, the load volatility and the window length of the sliding window are negatively correlated, that is, the greater the load volatility, the shorter the window length. In a preferred implementation, the change in window length between two consecutive intervals should not exceed 50%. For example, it is acceptable to adjust the window length from 30 minutes to 15 minutes. However, if the window length jumps directly from 30 minutes to 5 minutes, it is considered an excessive change. In this case, a smooth transition mechanism needs to be introduced, inserting a 10-minute buffer between 30 minutes and 5 minutes to avoid oscillations in the conditional probability table caused by abrupt changes in data distribution.

[0026] Furthermore, in the step of periodically updating the conditional probability table using the maximum likelihood estimation method, the update cycle is synchronized with the load volatility calculation cycle to ensure that the conditional probability table matches the rhythm of operating condition changes; in the weighted fusion of multimodal data, for each parameter of the physical layer node, a weighted value is calculated according to the data segment weight, expressed as: ; In the formula, , and Indicates weight, This represents the average value of the parameter in the recent data segment. This represents the average value of the parameter in the intermediate data segment. This represents the average value of parameters for a data segment from a longer period. This represents the parameter values ​​after weighted fusion; For any two associated nodes in a dynamic Bayesian network, calculate the conditional probability based on the weighted data. , is represented as: ; in, Indicates that X takes the weighted data. And Y takes The frequency of Y, where m represents the number of discretization levels of Y; Furthermore, after the calculation is completed, the new conditional probability value overwrites the old conditional probability table and is stored in the memory database of the edge computing module; maximum likelihood estimation combined with multimodal data fusion ensures the statistical significance of the probability calculation.

[0027] Furthermore, in the abnormal response factor of the two-factor dynamic update strategy, when the rate of change and duration of the physical layer node parameters meet the preset abnormal threshold conditions, an abnormal mode is triggered; the preset abnormal threshold conditions include temperature change characteristics, characteristic gas concentration change characteristics in oil, and vibration signal change characteristics; under the abnormal mode, the update frequency of the conditional probability table of the corresponding node is increased, and Bayesian estimation is used for iterative update, and L1 regularization is used to constrain parameter drift, so that the KL divergence of the probability table before and after the update is within the preset divergence threshold.

[0028] In this embodiment, the anomaly response factor is used to quickly capture sudden abnormal states of transformers, such as local overheating and discharge faults. By triggering high-frequency updates and constraint iterations, the conditional probability table can adapt to abnormal operating conditions in a timely manner while avoiding drastic parameter fluctuations. In the update frequency adjustment, for the physical layer nodes that trigger anomalies and their associated nodes, the update frequency of the conditional probability table (CPT) is increased from the usual 5 minutes / time to 1 minute / time, while non-associated nodes maintain the usual frequency (reducing computational redundancy). In data priority processing, multimodal data during abnormal periods are weighted according to the principle of prioritizing abnormal parameters to ensure that abnormal information dominates the CPT update.

[0029] In the iterative update step using Bayesian estimation, the CPT before the anomalous mode is triggered is used as the prior distribution. Based on the newly acquired anomalous data, a likelihood function is calculated to characterize the probability of observing anomalous data under the current parameters. The likelihood function is expressed as: In the formula, Let N represent the i-th abnormal data point, and N represent the amount of data during the abnormal period. Combining the prior distribution and the likelihood function, the posterior distribution, i.e., the updated conditional probability table CPT, is calculated using Bayes' theorem. The posterior distribution retains the rationality of the historical prior while incorporating new information from the abnormal data.

[0030] Through the above embodiments, this application constructs a dynamic Bayesian network to fuse multi-physics data and introduces a two-factor dynamic update strategy of conventional time-varying factors and abnormal response factors to update the conditional probability table, which can effectively fuse multi-modal data and solve the problem of heterogeneity of multi-source data; wherein, through the set two-factor strategy, the model is always matched with the actual operating state of the transformer.

[0031] Furthermore, in some embodiments of this application, the step of performing coupled analysis on the preprocessed multimodal data of the fusion model to generate the transformer state vector space includes: Spatiotemporal correlation processing is performed on multimodal data to transform discrete, heterogeneous multimodal data into spatiotemporally consistent features. Through temporal alignment and spatial continuity, it eliminates the fragmentation of data across the spatiotemporal dimensions. Specifically, this includes: In the time dimension, multimodal data is aligned, such as using timestamp synchronization to achieve temporal consistency for differences in sampling frequencies; In the spatial dimension, discretely distributed physical layer node data are spatially interpolated to generate a continuous field distribution, and a spatial correlation matrix is ​​established with the modal data at the corresponding spatial locations. The spatial correlation matrix is ​​used to characterize the correlation strength of multimodal data at different spatial locations. In the construction of the spatial correlation matrix, the matrix is ​​used to characterize the correlation strength of multimodal data at different spatial locations. The weight coefficients of each data point in the multimodal data are calculated based on the feature importance assessment algorithm. The correlation degree of features is judged by combining the spatial correlation matrix. When the correlation degree of any two features exceeds the preset correlation threshold, features are selectively retained based on the weight coefficients to form a coupled feature set. In the formation of the coupled feature set, redundant features are eliminated and strongly coupled, high-information features are retained through feature importance assessment and correlation degree screening, thereby improving the effectiveness of the state vector. In feature importance assessment, the random forest feature importance algorithm is used to calculate the weight of each multimodal data point. This weight is used to reflect the contribution of the feature to the state assessment. Based on the coupled feature set, a time decay factor based on the correlation strength is assigned to each dimension feature to construct an initial state vector; at the same time, a dynamic adaptation mechanism for operating conditions is introduced to increase the corresponding operating condition feature dimensions, forming a transformer state vector space with dynamically adjustable dimensions.

[0032] The time decay factor in this embodiment is expressed as: ; In the formula, This represents the decay factor of feature k at time t. This represents the average correlation between feature k and other features. This represents the attenuation coefficient; the higher the correlation, the slower the attenuation. Furthermore, the constructed initial state vector is represented as: ; The embodiments of this application dynamically increase the feature dimension according to the changes in the transformer's operating conditions to ensure the comprehensiveness of the state vector space; wherein, the final generated transformer state vector space has the advantages of spatiotemporal consistency, dynamic dimension and feature effectiveness; Therefore, the embodiments of this application eliminate the temporal misalignment and spatial discreteness of multimodal data by constructing a spatiotemporal alignment and spatial correlation matrix, forming a state carrier containing multi-dimensional features such as electromagnetics, temperature, and vibration; the dynamic dimension adjustment mechanism can adapt to different working conditions, ensuring that the state vector space can fully capture the complex operating state of the transformer.

[0033] Furthermore, in some embodiments of this application, step S103, which introduces a complex frequency domain parameter identification algorithm that integrates the extended Cole-Cole model, and dynamically updates the winding resistance, excitation reactance, and core permeability based on the electromagnetic parameters and temperature parameters in the state vector space, includes: Electromagnetic and temperature parameters are extracted from the state vector space. The temperature parameters are used for temperature compensation preprocessing of the complex frequency domain impedance spectrum. The complex frequency domain impedance spectrum is obtained by time-frequency conversion based on the acquired transformer signal. The time-domain signal (the transformer signal that varies with time) is converted into a frequency-domain signal by Fast Fourier Transform (FFT). The electromagnetic parameters include harmonic components, excitation current, and magnetic field strength. A dynamic model of winding resistance is constructed, and the relaxation characteristics of the extended Cole-Cole model are integrated. The winding resistance is represented as the superposition of DC resistance and additional resistance of each harmonic. The additional resistance is associated with the harmonic components of electromagnetic parameters in the state vector space. The dynamic characteristics of winding resistance are based on the combined effect of fundamental and harmonic currents. The fundamental current affects the DC resistance, while the harmonic current generates additional resistance due to the skin effect and proximity effect. The winding resistance dynamic model in this embodiment decomposes the total resistance into DC resistance and additional resistance of each harmonic by integrating the relaxation characteristics of the extended Cole-Cole model, thereby achieving an accurate description of the resistance change at different frequencies. Among them, the influencing factor of DC resistance is the winding temperature. For example, an increase in temperature will lead to an increase in resistance. The dynamic model of winding resistance is directly related to the temperature characteristics of the winding hot spot in the state vector space. The actual operating temperature of the winding can be calculated. Furthermore, through the temperature compensation mechanism, the standard DC resistance at the reference temperature is corrected to the real-time DC resistance at the current temperature. For the additional resistance of each harmonic, this embodiment introduces the relaxation characteristics of the extended Cole-Cole model. The higher the frequency of the high-frequency harmonic, the stronger the skin effect and proximity effect, and the larger the additional resistance. The calculation of the additional resistance is directly related to the characteristics of the harmonic components in the state vector, including the frequency, current amplitude and harmonic distortion rate of each harmonic.

[0034] Furthermore, step S103 also includes: A frequency-dependent model of excitation reactance is constructed, and a complex permeability parameter is introduced. The excitation current and core temperature data in the state vector space are correlated to describe the variation law of excitation reactance at different frequencies. The excitation reactance is related to the core permeability. The permeability has frequency dependence (increased frequency leads to increased eddy current loss and decreased permeability) and temperature dependence (temperature changes change the resistance to core domain movement and affect hysteresis characteristics). In this embodiment, the frequency dependence model of the excitation reactance introduces a complex permeability parameter to correlate frequency, excitation current and core temperature, which can represent the variation law of excitation reactance at different frequencies. A coupled model of iron core permeability is constructed, and the spatial distribution characteristics of permeability are corrected by combining the magnetic field strength in the state vector space. In this embodiment, the permeability of the iron core is not uniformly distributed. It is affected by the ampere-turn distribution of the winding and the edge effect. The magnetic field strength at different locations is different, which causes the permeability to change nonlinearly with spatial location and magnetic field strength. The iron core permeability coupling model corrects the spatial distribution characteristics of the permeability by fusing the spatial magnetic field data in the state vector, thereby achieving an accurate representation of the magnetic properties of the iron core. Furthermore, a nonlinear least squares fitting algorithm is adopted, with the preprocessed complex frequency domain impedance spectrum as input, and the initial identification values ​​of winding resistance, excitation reactance and core permeability are calculated iteratively based on the fusion extended Cole-Cole model. Electromagnetic parameters and temperature parameters in the state vector space are introduced as constraints to correct the initial identification values. In the step of iteratively calculating the initial identification value, the sum of squared residuals between the calculated and measured values ​​of the complex frequency domain impedance spectrum is used as the objective function to initialize the initial values ​​of winding resistance, excitation reactance and core permeability. These values ​​are then substituted into the objective function for iterative solution. When the rate of change of the sum of squared residuals between two consecutive iterations is less than 0.1%, the iteration is stopped and the initial identification value is output. Furthermore, constraints are introduced to correct the initial identification values, including: Based on the harmonic components in the state vector space, a high-frequency additional resistance constraint is applied to the initial identification value of the winding resistance, so that the corrected winding resistance increases with the increase of the harmonic distortion rate. Based on the excitation current and magnetic field strength in the state vector space, a saturation constraint is applied to the initial identification value of the core permeability, so that the corrected permeability decreases as the magnetic field strength increases. Based on the winding hot spot temperature and the core average temperature in the state vector space, temperature coefficient constraints are applied to the initial identification values ​​of winding resistance and permeability to make the correction results conform to the influence of temperature on electromagnetic parameters. Finally, this application updates the winding resistance, excitation reactance, and core permeability based on parameter update triggering conditions. The parameter update triggering conditions include changes in electromagnetic parameters or temperature parameters in the state vector space exceeding a preset threshold.

[0035] This application uses electromagnetic parameters and temperature coefficients in state vector space, and employs a complex frequency domain algorithm that integrates and extends the Cole-Cole model to dynamically update winding resistance, excitation reactance, and core permeability. This enables the dynamic changes of transformer parameters with frequency, temperature, and operating conditions, overcoming the shortcomings of traditional static parameters that cannot adapt to fluctuations in operating conditions.

[0036] Furthermore, in some embodiments of this application, step S104, which involves decoupling no-load loss from load loss using an improved all-electric differential method, includes: Input the real-time parameters obtained through complex frequency domain parameter identification, including dynamically updated winding resistance, excitation reactance and core permeability, and extract the temperature field gradient, harmonic distortion rate and vibration characteristic parameters in the state vector space as auxiliary decoupling variables. A sub-model for calculating no-load loss is constructed. Based on the excitation reactance and core permeability, the overall proportion of hysteresis loss and eddy current loss is corrected by combining the average core temperature. For example, when the temperature rises, the resistivity of the core material increases, and the eddy current loss decreases, i.e., its proportion decreases, while the hysteresis loss is more sensitive to temperature, and its proportion will relatively increase. The no-load loss calculation sub-model is associated with the no-load harmonic components in the state vector space. When the core temperature field gradient exceeds a preset threshold, the no-load loss calculation result is thermally corrected. For example, when the core temperature field gradient exceeds the preset threshold, it indicates that there is local overheating in the core. Therefore, this application introduces a thermal correction coefficient to adjust the no-load loss calculation result upward to compensate for local overheating loss. A sub-model for calculating load losses is constructed. Based on the winding resistance and the fundamental and harmonic components of the current, a skin effect correction coefficient (proportional to the square root of the current frequency) is introduced to calculate the winding copper loss and harmonic additional loss. Simultaneously, vibration characteristic parameters in the state vector space are correlated. When the vibration dominant frequency offset exceeds the set range, mechanical additional loss in the load loss is compensated. The vibration dominant frequency offset in the state vector space is used to characterize mechanical faults such as winding loosening and cooling fan malfunction. When the offset exceeds the set range, it is determined that the mechanical additional loss has increased. Through the vibration loss compensation coefficient, the mechanical additional loss is superimposed on the load loss to ensure that the loss calculation includes non-electromagnetic factors. A differential coupling equation is established, and the total loss is decomposed into no-load loss components and load loss components through a separation coefficient matrix. The elements of the separation coefficient matrix are dynamically adjusted based on the transformer's rated parameters and real-time operating conditions. The separation coefficient matrix is ​​a 2×2 square matrix, where the row dimension corresponds to the loss type: the first row is the no-load loss correlation coefficient, and the second row is the load loss correlation coefficient; the column dimension corresponds to the coupling relationship: the first column reflects the weight of its own loss (self-coupling), and the second column reflects the cross-influence with another type of loss (mutual coupling); all elements of the matrix take values ​​in the range (0,1), and satisfy the physical constraint that the sum of the matrix's rows and columns is 1, ensuring the integrity of the coupling relationship. In addition, the matrix elements of the separation coefficient matrix can be dynamically adjusted, such as by adjusting and updating according to the transformer's rated parameters and real-time operating conditions, so that the separation weights match the actual loss characteristics.

[0037] This application addresses the different physical nature of no-load loss and load loss by using corresponding dynamic parameters for decoupling. By combining differential coupling equations with dynamically adjusted separation coefficient matrices, the coupling interference between the two types of loss is effectively eliminated, thereby improving the precision of transformer loss monitoring. Furthermore, in some embodiments of this application, the structure of the LSTM-Transformer hybrid model includes: A preprocessing layer, which includes parallel convolutional modules and temporal modules; In this embodiment, the convolution module uses multiple convolution kernels to extract local features from the input multidimensional temporal data and outputs a local temporal feature vector. The convolution module includes three parallel convolution kernels with sizes of 3, 5, and 7 (corresponding to the time window length), and each convolution kernel outputs 16-dimensional features. Each convolution operation is followed by a ReLU activation function and max pooling, and the final output is a local temporal feature vector with a dimension of 48. In this embodiment, the timing module corrects data timing misalignment through a gated cyclic unit (GRU) and outputs an aligned global timing feature vector. In this embodiment, the outputs of the convolutional module and the temporal module are fused into the input feature matrix through a feature concatenation operation. The feature encoding layer adopts a bidirectional LSTM structure, including a forward LSTM unit and a backward LSTM unit. The outputs of the two units are weighted and summed to generate bidirectional temporal features. In the weighted summation, the weights are dynamically assigned based on the temporal importance of the features. The weight parameters are learned through a single-layer perceptron. Compared with simple concatenation, weighted summation can highlight the bidirectional dependence of key time steps and improve the discriminative ability of features. A cross-layer feature interaction layer is set up, in which a feature mapping matrix and a cross-layer attention mechanism are set. The feature mapping matrix transforms the bidirectional temporal features into intermediate features that match the dimensions of the Transformer layer. The cross-layer attention mechanism is used to avoid the loss of key information in LSTM features after mapping. The cross-layer attention mechanism calculates the similarity weight between LSTM features and input features of the Transformer layer, and enables the two layers of features to be dynamically fused through residual connections, so as to ensure that the temporal modeling advantages of LSTM and the global correlation advantages of Transformer are preserved at the same time. The Transformer-enhanced decoding layer employs a modular multi-head attention structure, comprising multiple parallel attention sub-modules that perform attention calculations for electrical, thermal, and chemical features, respectively. A dynamic attention scaling mechanism is introduced, where the scaling factor is inversely proportional to the time-varying volatility of the feature, used to reduce the interference of noisy features on the attention weights. Specifically, for each feature, the variance within the sliding window is used as the time-varying volatility; the smaller the scaling factor, the higher the time-varying volatility. The output fusion layer includes a feature selection gate and an output regression unit. The feature selection gate is used to filter key features. Based on L1 regularization, the feature selection gate selects key features from LSTM features and Transformer features to reduce the interference of redundant features and reduce the computational cost of the model. The output regression unit adopts a two-layer fully connected network and outputs standardized load loss prediction values ​​through the Sigmoid activation function. The learnable parameters of the two fully connected network layers in the output regression unit are dynamically calibrated through the online distillation mechanism of the edge computing module. The online distillation mechanism is used for dynamic parameter calibration at the edge. In order to adapt to the dynamic changes in the transformer operating conditions, the learnable parameters of the fully connected network are calibrated in real time through the online distillation of the edge computing module. For example, a lightweight teacher model based on XGBoost is deployed at the edge, and its prediction results are used as soft labels to guide the student model (i.e., the fully connected network) to update its parameters.

[0038] Furthermore, in the feature encoding layer employing a bidirectional LSTM structure, an adaptive forget gate mechanism is introduced. The gate parameters are modulated in real time by the load volatility in the state vector space, specifically including: Extract the load current sequence within a preset sliding time window from the state vector space. Calculate load volatility; set the initial parameters of the forget gate of the bidirectional LSTM to... The forget gate parameter is adjusted based on the load volatility, expressed as: ; In the formula, This indicates the preset load stability threshold. , This represents the adjustment coefficient. This indicates severe load fluctuations. This indicates that the load is stable, and f represents the adjusted forget gate parameter; The temporal features output by the forward LSTM unit and the temporal features output by the backward LSTM unit are weighted and fused based on the real-time forget gate parameter f, as follows: In the formula, Indicates the features after fusion. This represents the timing characteristics of the output of the forward LSTM unit. This represents the timing characteristics of the output from the backward LSTM unit.

[0039] Furthermore, in some embodiments of this application, the step of using the decoupled loss data and the correlation parameters in the state vector space as input to the LSTM-Transformer hybrid model loaded in the edge computing module to predict and output the load loss trend in the future time period includes: Extract the decoupled load loss data and the associated parameters in the state vector space to form a multidimensional input dataset. Divide the dataset into a sample sequence by a sliding time window. Each sample contains historical data of a preset duration and corresponding labels. The edge computing module loads a pre-trained hybrid model, takes the sample sequence as input, and outputs the predicted load loss value, trend slope, and feature importance weights for a preset time period in the future. To quantify the uncertainty of the prediction, this embodiment uses the Monte Carlo dropout method to calculate the confidence interval of the prediction results. When the confidence level is lower than a preset threshold, the latest feature data in the state vector space is introduced for re-prediction until the confidence level is ≤10%. By monitoring errors in real time and adjusting parameters accordingly, the model dynamically adapts to changes in data distribution, avoiding the accumulation of prediction biases over long-term operation. To this end, this embodiment establishes a prediction error feedback mechanism to periodically compare the deviation between predicted values ​​and actual loss data. When the deviation exceeds a set range, it triggers lightweight parameter fine-tuning of the hybrid model, including: The forgetting factor of the LSTM layer and the attention coefficient of the Transformer layer are corrected based on the bias gradient. The fine-tuning process calls the online distillation mechanism of the edge computing module. Specifically, in correcting the LSTM forgetting factor, the forgetting gate parameter is adjusted based on the bias gradient: if the predicted value is consistently lower than the measured value, the forgetting factor is increased to enhance the model's memory of recent data; otherwise, the forgetting factor is decreased. In correcting the attention coefficient of the Transformer layer, the correlation between the bias and the importance weight of each feature is calculated. If the contribution of the prediction bias is relatively high, the attention coefficient corresponding to the feature with the prediction bias is reduced to reduce its impact on the prediction. With the above settings, the model training sample pool and the state vector space update frequency are kept synchronized, and the new samples after preprocessing are stored in real time. In this embodiment, the model can quickly adapt to the sudden changes in transformer operating conditions through real-time updates of the sample pool and incremental training. Furthermore, this embodiment determines whether a sudden change in the operating condition occurs by using the operating condition features in the state vector space: when the change index exceeds the threshold, the weight ratio of new samples in the training pool is automatically increased, and incremental training of the model is triggered, updating only the multi-head attention parameters of the Transformer layer.

[0040] Therefore, the bidirectional LSTM used in the hybrid model of this application can effectively capture the temporal dependence of loss data and adapt to the dynamic changes of loss over time; the modular multi-head attention mechanism can specifically mine the long-range correlation between multiple physical fields such as electrical parameters and thermal characteristics, and solve the problem of trend prediction under multi-factor coupling.

[0041] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0042] Please refer to Figure 2 In another embodiment of the present invention, a transformer loss monitoring system is provided; The transformer loss monitoring system includes the following modules: The fusion model construction module 201 is used to construct a multi-physics fusion model based on a dynamic Bayesian network. An online learning mechanism based on a two-factor dynamic update strategy is introduced into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time. The vector space generation module 202 is used to perform coupled analysis on the multimodal data after the fusion model preprocessing to generate the transformer state vector space. The update module 203 is used to dynamically update the winding resistance, excitation reactance and core permeability based on the electromagnetic parameters and temperature parameters in the state vector space using a complex frequency domain parameter identification algorithm that incorporates a fusion extended Cole-Cole model. The loss decoupling module 204 is used to decouple no-load loss and load loss through an improved all-electric differential method. The dynamically updated excitation reactance and core permeability are used to decouple no-load loss, and the dynamically updated winding resistance is used to decouple load loss. The loss prediction module 205 is used to take the decoupled loss data and the correlation parameters in the state vector space as input to the LSTM-Transformer hybrid model loaded in the edge computing module, and predict and output the load loss trend in the future time period. In the hybrid model, the loss time series features are captured by the bidirectional LSTM layer, and the long-range correlation of multi-physics fields is mined by the modular multi-head attention structure Transformer layer.

[0043] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0044] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0045] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for monitoring transformer losses, characterized in that, Includes the following steps: A multiphysics fusion model based on a dynamic Bayesian network is constructed. An online learning mechanism based on a two-factor dynamic update strategy is introduced into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time. The multimodal data preprocessed by the fusion model are coupled and analyzed to generate the transformer state vector space. Based on the electromagnetic and temperature parameters in the state vector space, the winding resistance, excitation reactance and core permeability are dynamically updated using a complex frequency domain parameter identification algorithm that incorporates a fusion extended Cole-Cole model. The no-load loss and load loss are decoupled by an improved all-electric differential method. The dynamically updated excitation reactance and core permeability are used to decouple the no-load loss, and the dynamically updated winding resistance is used to decouple the load loss. The decoupled loss data and the correlation parameters in the state vector space are used as inputs to the LSTM-Transformer hybrid model loaded in the edge computing module to predict and output the load loss trend in the future time period. In the hybrid model, the loss time series features are captured by bidirectional LSTM layers, and the long-range correlation of multi-physics fields is mined by the modular multi-head attention structure Transformer layer.

2. The transformer loss monitoring method according to claim 1, characterized in that, In the steps of constructing a multiphysics fusion model based on a dynamic Bayesian network, the nodes of the dynamic Bayesian network are divided into physical layer nodes, feature layer nodes, and decision layer nodes. The physical layer nodes correspond to multimodal raw data, including winding temperature field data, voltage and current spectrum data, characteristic gas concentration data, vibration acceleration signals, and frequency cepstral coefficients of acoustic fingerprint signals. The feature layer nodes are high-order features extracted from the multimodal data of the physical layer nodes, including temperature difference gradient, electromagnetic harmonic distortion rate, gas component ratio, vibration dominant frequency energy ratio, and acoustic signature feature entropy. The decision-making layer nodes are the fused state assessment parameters, including insulation aging index, winding overheat risk value, and loss anomaly probability.

3. The transformer loss monitoring method according to claim 2, characterized in that, The step of introducing an online learning mechanism based on a two-factor dynamic update strategy into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time includes: In the conventional time-varying factor, the window length of the sliding window is dynamically adjusted based on the load volatility. The multimodal data within the time window is divided into recent data segments, medium-term data segments, and long-term data segments according to the collection time. Weights are assigned and corrected for each period data segment. Based on the weighted multimodal data, the conditional probability table is periodically updated using the maximum likelihood estimation method. In the abnormal response factor, when the rate of change and duration of the physical layer node parameters meet the preset abnormal threshold conditions, an abnormal mode is triggered. The preset abnormal threshold conditions include temperature change characteristics, characteristic gas concentration change characteristics in oil, and vibration signal change characteristics. In the abnormal mode, the update frequency of the conditional probability table of the corresponding node is increased, and Bayesian estimation is used for iterative update. Through L1 regularization to constrain parameter drift, the KL divergence of the probability table before and after the update is kept within the preset divergence threshold.

4. The transformer loss monitoring method according to claim 3, characterized in that, The steps for coupling analysis of the preprocessed multimodal data from the fusion model to generate the transformer state vector space include: Spatiotemporal correlation processing of multimodal data includes: aligning multimodal data in the time dimension; and generating a continuous field distribution from discretely distributed physical layer node data through spatial interpolation in the spatial dimension, and establishing a spatial correlation matrix with the modal data at the corresponding spatial location. The spatial correlation matrix is ​​used to characterize the correlation strength of multimodal data at different spatial locations. The weight coefficients of each data point in the multimodal data are calculated based on the feature importance evaluation algorithm. The correlation degree of the features is judged by combining the spatial correlation matrix. When the correlation degree of any two features exceeds the preset correlation threshold, the features are selectively retained based on the weight coefficients to form a coupled feature set. Based on the coupled feature set, a time decay factor based on the correlation strength is assigned to each dimension feature to construct an initial state vector; at the same time, a dynamic adaptation mechanism for operating conditions is introduced to increase the corresponding operating condition feature dimensions, forming a transformer state vector space with dynamically adjustable dimensions.

5. The transformer loss monitoring method according to claim 4, characterized in that, A complex frequency domain parameter identification algorithm based on the extended Cole-Cole model is introduced. This algorithm dynamically updates the winding resistance, excitation reactance, and core permeability based on the electromagnetic and temperature parameters in the state vector space. The steps include: Electromagnetic and temperature parameters are extracted from the state vector space. The temperature parameters are used for temperature compensation preprocessing of the complex frequency domain impedance spectrum. The complex frequency domain impedance spectrum is obtained by time-frequency conversion based on the acquired transformer signal. The electromagnetic parameters include harmonic components, excitation current, and magnetic field strength. A dynamic model of winding resistance is constructed, and the relaxation characteristics of the extended Cole-Cole model are integrated. The winding resistance is represented as the superposition of DC resistance and additional resistance of each harmonic. The additional resistance is associated with the harmonic components of electromagnetic parameters in the state vector space. A frequency-dependent model of excitation reactance is constructed, and a complex permeability parameter is introduced. The excitation current and core temperature data in the state vector space are correlated to describe the variation law of excitation reactance at different frequencies. A coupled model of iron core permeability is constructed, and the spatial distribution characteristics of permeability are corrected by combining the magnetic field strength in the state vector space. A nonlinear least squares fitting algorithm is adopted, with the preprocessed complex frequency domain impedance spectrum as input. The initial identification values ​​of winding resistance, excitation reactance and core permeability are calculated iteratively based on the fusion extended Cole-Cole model. Electromagnetic parameters and temperature parameters in the state vector space are introduced as constraints to correct the initial identification values. The winding resistance, excitation reactance, and core permeability are updated based on parameter update triggering conditions. The parameter update triggering conditions include changes in electromagnetic parameters or temperature parameters in the state vector space exceeding a preset threshold.

6. The transformer loss monitoring method according to claim 5, characterized in that, The steps to decouple no-load loss and load loss using the improved all-electric differential method include: Input the real-time parameters obtained through complex frequency domain parameter identification, including dynamically updated winding resistance, excitation reactance and core permeability, and extract the temperature field gradient, harmonic distortion rate and vibration characteristic parameters in the state vector space as auxiliary decoupling variables. A sub-model for calculating no-load loss is constructed. Based on the excitation reactance and core permeability, the overall proportion of hysteresis loss and eddy current loss is corrected by combining the core average temperature. The no-load time harmonic component in the associated state vector space is also included in the no-load loss calculation sub-model. When the core temperature field gradient exceeds a preset threshold, the no-load loss calculation results are thermally corrected. A load loss calculation sub-model is constructed. Based on the winding resistance and the fundamental and harmonic components of the current, a skin effect correction coefficient is introduced to calculate the winding copper loss and harmonic additional loss. At the same time, the vibration characteristic parameters in the state vector space are associated. When the vibration dominant frequency offset exceeds the set range, the mechanical additional loss in the load loss is compensated. A differential coupling equation is established, and the total loss is decomposed into no-load loss component and load loss component through a separation coefficient matrix. The elements of the separation coefficient matrix are dynamically adjusted based on the transformer's rated parameters and real-time operating conditions.

7. The transformer loss monitoring method according to claim 6, characterized in that, The structure of the LSTM-Transformer hybrid model includes: The preprocessing layer includes a parallel convolutional module and a temporal module. The convolutional module uses multiple convolutional kernels to extract local features from the input multidimensional temporal data and outputs a local temporal feature vector. The temporal module corrects temporal misalignment of the data through a gated recurrent unit (GRU) and outputs an aligned global temporal feature vector. The outputs of the convolutional module and the temporal module are fused into an input feature matrix through a feature concatenation operation. The feature coding layer adopts a bidirectional LSTM structure, including a forward LSTM unit and a backward LSTM unit. The outputs of the two units are weighted and summed to generate bidirectional temporal features. A cross-layer feature interaction layer is set up, in which a feature mapping matrix and a cross-layer attention mechanism are set. The feature mapping matrix transforms the bidirectional temporal features into intermediate features that match the dimensions of the Transformer layer. The cross-layer attention mechanism calculates the similarity weights between the LSTM features and the input features of the Transformer layer, and enables the features of the two layers to be dynamically fused through residual connections. The Transformer enhanced decoding layer adopts a modular multi-head attention structure, which includes multiple parallel attention sub-modules that perform attention calculations for electrical parameter features, thermal features, and chemical features respectively. A dynamic attention scaling mechanism is introduced, inversely proportional to the time-varying volatility of the feature, to reduce the interference of noisy features on the attention weights. The output fusion layer includes a feature selection gate and an output regression unit. The feature selection gate selects key features from LSTM features and Transformer features based on L1 regularization. The output regression unit adopts a two-layer fully connected network and outputs a standardized load loss prediction value through the Sigmoid activation function. The learnable parameters of the two fully connected network layers in the output regression unit are dynamically calibrated through the online distillation mechanism of the edge computing module.

8. The transformer loss monitoring method according to claim 7, characterized in that, In the feature encoding layer employing a bidirectional LSTM structure, an adaptive forget gate mechanism is introduced. The gate parameters are modulated in real time by the load volatility in the state vector space, specifically including: Extract the load current sequence within a preset sliding time window from the state vector space. Calculate the load volatility, expressed as: ； In the formula, This represents the average current value within the time window. This represents the load volatility, used to characterize the relative degree of load fluctuation; n represents the total number of load current sequences. This represents the current value of the i-th load current sequence; The initial parameters of the forget gate of the bidirectional LSTM are set to The forget gate parameter is adjusted based on the load volatility, expressed as: ； In the formula, This indicates the preset load stability threshold. This represents the adjustment coefficient. This indicates severe load fluctuations. This indicates that the load is stable, and f represents the adjusted forget gate parameter; The temporal features output by the forward LSTM unit and the temporal features output by the backward LSTM unit are weighted and fused based on the real-time forget gate parameter f, as follows: In the formula, Indicates the features after fusion. This represents the timing characteristics of the output of the forward LSTM unit. This represents the timing characteristics of the output from the backward LSTM unit.

9. The transformer loss monitoring method according to claim 8, characterized in that, The steps of using the decoupled loss data and the correlation parameters in the state vector space as input to the LSTM-Transformer hybrid model loaded in the edge computing module to predict and output the load loss trend in the future time period include: Extract the decoupled load loss data and the associated parameters in the state vector space to form a multidimensional input dataset. Divide the dataset into a sample sequence by a sliding time window. Each sample contains historical data of a preset duration and corresponding labels. The edge computing module loads a pre-trained hybrid model, takes the sample sequence as input, and outputs the predicted load loss value, trend slope, and feature importance weights for a preset time period in the future. The Monte Carlo dropout method is used to calculate the confidence interval of the prediction results. When the confidence is lower than the preset threshold, the latest feature data in the state vector space is introduced to re-predict. Establish a prediction error feedback mechanism to periodically compare the deviation between the predicted value and the actual loss data: when the deviation exceeds the set range, trigger the lightweight parameter fine-tuning of the hybrid model, including: correcting the forgetting factor of the LSTM layer and the attention coefficient of the Transformer layer based on the deviation gradient, and calling the online distillation mechanism of the edge computing module during the fine-tuning process. The model training sample pool is synchronized with the state vector space update frequency, and new samples after preprocessing are stored in real time. The system also judges whether the operating conditions have changed abruptly by using the operating condition features in the state vector space. When the change index exceeds the threshold, the weight ratio of new samples in the training pool is automatically increased, and incremental training of the model is triggered, updating only the multi-head attention parameters of the Transformer layer.

10. A transformer loss monitoring system for implementing the transformer loss monitoring method as described in any one of claims 1 to 9, characterized in that, The transformer loss monitoring system includes the following modules: The fusion model construction module is used to construct a multiphysics fusion model based on a dynamic Bayesian network. An online learning mechanism based on a two-factor dynamic update strategy is introduced into the fusion model to update the conditional probability table of the dynamic Bayesian network in real time. The vector space generation module is used to perform coupled analysis on the multimodal data after the fusion model preprocessing to generate the transformer state vector space; The update module is used to dynamically update the winding resistance, excitation reactance and core permeability based on the electromagnetic parameters and temperature parameters in the state vector space using a complex frequency domain parameter identification algorithm that incorporates a fusion extended Cole-Cole model. The loss decoupling module is used to decouple no-load loss from load loss using an improved all-electric differential method. The dynamically updated excitation reactance and core permeability are used to decouple no-load loss, while the dynamically updated winding resistance is used to decouple load loss. The loss prediction module is used to take the decoupled loss data and the correlation parameters in the state vector space as input to the LSTM-Transformer hybrid model loaded in the edge computing module, and predict and output the load loss trend in the future time period. In the hybrid model, the loss time series features are captured by bidirectional LSTM layers, and the long-range correlation of multi-physics fields is mined by the modular multi-head attention structure Transformer layers.

Images