A transformer state prediction method based on big data analysis

Abstract

The application provides a mutual inductor state prediction method based on big data analysis, and relates to the field of mutual inductor state prediction, and the specific steps include: constructing a mutual inductor state data set, segmenting tensorization time series data in the data set, generating a mutual inductor state feature vector by using a sine transformation, a tanh function, combining linear projection and cross-dimension difference items, introducing a disturbance correction kernel and combining softmax weighted fusion to generate a cross-time feature vector, calculating total disturbance energy after extracting global dependence and local fluctuation information to form a disturbance feature representation, fusing the disturbance feature representation to form a fusion feature vector, constructing an overall curvature representation, extracting long-term trends and local peak information by using average pooling and maximum pooling, splicing the fusion feature vector to form a multi-scale fusion feature, and performing linear transformation and Sigmoid output on the spliced and flattened multi-scale fusion features of all segments to obtain mutual inductor state prediction values, so that the accuracy and robustness of mutual inductor state prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of transformer state prediction, and in particular to a transformer state prediction method based on big data analysis. BACKGROUND

[0002] As a core measurement and protection device in the power system, the operation state of the transformer is directly related to the collection accuracy of voltage and current signals and the reliability of system protection action. Transformer state monitoring plays an important role in fault diagnosis, operation evaluation and equipment maintenance. With the expansion of the power grid scale and the increase of the complexity of the operating environment, the operation data of the transformer under the influence of factors such as temperature change, load fluctuation and insulation aging show nonlinear and multi-scale characteristics. The traditional method of relying on manual maintenance or single signal monitoring cannot fully reflect the health state of the transformer, and it is necessary to introduce a method based on big data analysis to model and predict the operation state of the transformer to improve the accuracy of fault warning and state evaluation.

[0003] At present, transformer state prediction mainly relies on single physical quantity monitoring or static modeling method based on statistical characteristics, usually only considers limited dimensional information such as current, voltage or environmental temperature and humidity, lacks comprehensive integration of electrical quantities, insulation quantities and environmental quantities, and is difficult to fully describe the operation characteristics of the transformer under complex working conditions. At the same time, although some methods introduce time series analysis or traditional machine learning models, they generally have the problems of insufficient feature extraction ability and poor sensitivity to cross-time dependence, resulting in limited prediction accuracy and insufficient adaptability to sudden disturbances or periodic changes.

[0004] The operation state of the transformer is jointly affected by multi-source data such as electrical quantities, insulation quantities and environmental quantities. Under different time scales, it not only contains long-term evolution trend, but also has transient abnormality caused by local discharge pulse or current impact, and its cross-time dependence and local disturbance show high complexity. Therefore, a prediction method is proposed, which fuses segmented representation, cross-time disturbance modeling and global-local fusion, can realize efficient extraction and dynamic modeling of multi-source features, and thus accurately predict the state of the transformer under complex operating conditions. SUMMARY

[0005] The application provides a mutual inductor state prediction method based on big data analysis, aiming at the problems of large state data quantity, unstable cross-time dependence and frequent local disturbance of the mutual inductor in long-term operation, a multi-stage prediction mechanism is provided by fusing segmented representation, cross-time dependence modeling and multi-scale curvature fusion, including obtaining mutual inductor state data and constructing a mutual inductor state data set, calculating segmented representation through sinusoidal transformation and tanh function, combining linear projection and cross-dimension difference item to generate a mutual inductor state feature vector, introducing a disturbance correction kernel to obtain attention weight through softmax transformation and weighting fusion of main trend and periodic disturbance information to generate a cross-time feature vector, forming a disturbance feature representation based on global dependence relationship and local fluctuation information and calculating total disturbance energy to obtain a fusion feature vector, combining overall curvature representation and average pooling and maximum pooling to extract long-term trend and local peak value information to form a multi-scale fusion feature, and finally flattening through linear transformation and Sigmoid mapping to output the mutual inductor state prediction value, so as to realize unified modeling and accurate prediction of global evolution and local disturbance characteristics of the mutual inductor operation state.

[0006] A mutual inductor state prediction method based on big data analysis, the specific method is:

[0007] S1, obtaining the state data of the mutual inductor, and constructing a mutual inductor state data set after preprocessing;

[0008] S2, dividing the tensorized time series data in the data set into segments, calculating the segmented representation, generating the feature representation value by using the sinusoidal transformation and tanh function, combining the linear projection and the cross-dimension difference item, and generating the mutual inductor state feature vector;

[0009] S3, introducing the disturbance correction kernel between different segments, obtaining the attention weight through weighted aggregation and softmax transformation, extracting the main trend and the periodic disturbance information based on the mutual inductor state feature vector, and generating the cross-time feature vector by using the attention weight for weighted fusion;

[0010] S4, extracting the global dependence relationship and the local fluctuation information of the cross-time feature vector, obtaining the disturbance feature representation through layer normalization and ReLU after weighted fusion, obtaining the total disturbance energy through squaring and absolute value correction, and forming the fusion feature vector after fusion with the disturbance feature representation;

[0011] S5, constructing the overall curvature representation based on the adjacent segment fusion feature vectors, extracting the long-term trend and the local peak value information by using the average and maximum pooling, and splicing the current segment fusion feature vector to form the multi-scale fusion feature;

[0012] S6, flattening all segmented multi-scale fusion features, outputting the mutual inductor state prediction value after linear transformation and Sigmoid mapping;

[0013] S7, a mutual inductor state prediction model based on big data analysis is constructed, a mutual inductor state data set is input, a weighted binary cross entropy is taken as a loss function, and iterative training is performed in sequence through steps S2 to S6 until convergence is reached, and the optimization training of the model is completed.

[0014] Preferably, the construction process of the mutual inductor state data set includes three stages of data acquisition, preprocessing and structuring, the acquisition of the mutual inductor state data includes electrical quantity data, environmental quantity data and insulation quantity data, the electrical quantity data is the current and voltage signals output by the secondary side of the mutual inductor, the environmental quantity data includes the surface temperature of the mutual inductor body and the temperature and humidity of the operating environment, and the insulation quantity data includes partial discharge pulses, leakage currents and dielectric loss factors, in the preprocessing stage, the time alignment is first performed on the collected multiple types of data, the time stamps of the electrical quantity data, the environmental quantity data and the insulation quantity data are corrected through interpolation based on the sampling time of the voltage signal, in the structuring stage, the cutting is performed according to a fixed time sequence length, the electrical quantity data, the environmental quantity data and the insulation quantity data are stacked in the form of channels within the time sequence length, the tensorized time sequence data is formed, and finally the mutual inductor state data set is constructed.

[0015] Preferably, in the S2 step, the tensorized time sequence data in the data set is divided into segments, the segment representation is calculated, the feature representation value is generated by using the sine transform and the tanh function, and the mutual inductor state feature vector is generated by combining the linear projection and the cross-dimension difference term.

[0016] Each dimension of the tensorized time sequence data is divided into segments with a length of , each value in the segment is squared and then subtracted by a term proportional to the absolute value of the value, and then normalized and averaged to obtain the segment representation of each dimension.

[0017] The linear projection is performed on the feature representation value of the first segment and the first dimension, and then the cross-dimension difference term is introduced, the squared difference is performed between the feature representation value of the first dimension and the feature representation values of all other dimensions, and the projection result is added to obtain the fused cross-dimension relationship representation of the first segment and the first dimension.

[0018] The fused cross-dimension relationship representations of all dimensions of the first segment are spliced to obtain the mutual inductor state feature vector.

[0019] Further, in view of the problems of large numerical scale difference, frequent local impact interference and difficult to depict cross-dimension coupling relationship of time series data of electrical quantity, environmental quantity and insulation quantity formed by the transformer during long-term operation, the application proposes an embedding mechanism combined with segmented representation construction and cross-dimension difference item enhancement, first, the tensorized time series data is divided according to the preset segmentation length, the square minus absolute value proportional item modification operation is performed on each dimension sequence in the segment, and the segmented representation is obtained by normalization and averaging, so as to weaken the abnormal peak value caused by current impact, local discharge pulse and temperature and humidity fluctuation and the like, and ensure the robustness and stability of the segmented representation, then, the linear projection is performed on the segmented representation of the first dimension of the first segment, the feature is mapped to a unified space, the influence of different physical quantity dimensions and inconsistent numerical distribution on the modeling accuracy is eliminated, on this basis, the cross-dimension difference item is introduced, the square difference calculation is performed on the feature representation value of the first dimension and other dimensions, and is fused with the linear projection result, to obtain the fusion cross-dimension relationship representation of the first segment of the first dimension, so that the fusion cross-dimension relationship representation retains the single-dimension feature trend and strengthens the multi-source coupling effect, finally, the fusion cross-dimension relationship representations of all dimensions in the same segment are spliced to generate the transformer state feature vector, which provides high-quality input for subsequent modeling.

[0020] Preferably, in the S3 step, the disturbance correction kernel between different segments is introduced, the attention weight is obtained by weighted aggregation and softmax transformation, the main trend and periodic disturbance information are extracted based on the transformer state feature vector, and the cross-time feature vector is generated by weighted fusion using the attention weight;

[0021] For the fusion cross-dimension relationship representations of the first segment and the first segment on the first dimension, the square difference of the two is calculated, a penalty term proportional to the absolute value of the difference of the two is introduced, and the disturbance correction kernel of the first segment and the first segment on the first dimension is generated.

[0022] The disturbance correction kernels of all dimensions of the first segment and the first segment are weighted aggregated, and the cross-time disturbance weight of the first segment and the first segment is obtained after averaging.

[0023] The cross-time disturbance weight is transformed into the attention weight of the first segment and the first segment by using softmax.

[0024] Based on the state feature vector of the mutual inductor, the main trend information is extracted using a linear projection matrix, and the periodic perturbation information is obtained through a sine function. The main trend information and the periodic perturbation information are then fused using attention weights to obtain the [missing information]. Segment-time feature vector.

[0025] Furthermore, addressing the issues of time-dependent instability and frequent transient interference in the collected transformer state data, including electrical quantities, insulation quantities, and environmental quantities, this invention proposes a time-dependent disturbance modeling method based on a disturbance correction kernel and an attention mechanism. Firstly, for the... Section and the The section in The dimensional fusion of cross-dimensional relationships involves calculating the squared difference between the two dimensions and introducing a penalty term proportional to the absolute value of the difference to form a perturbation correction kernel. This kernel weakens the impact of abnormal peaks such as current surges, partial discharge pulses, and sudden temperature and humidity changes on the determination of cross-time differences. Then, the perturbation correction kernels of each dimension are weighted, aggregated, and averaged to obtain the cross-time perturbation weight, thus comprehensively characterizing the overall differences between electrical quantities, insulation quantities, and environmental quantities at different time periods. Next, softmax is used to transform the perturbation weight into attention weights, enabling the model to adaptively highlight time-series segments highly relevant to the current segment and suppress noise segments with low correlation to the actual operating state. Finally, based on the transformer state feature vector, a linear projection matrix is ​​used to extract the main trend information reflecting long-term evolution. Simultaneously, a sine function is combined to generate periodic perturbation information, and the two types of features are weighted and fused through attention weights to form the second... The segment-time feature vector mechanism not only maintains the continuity of the operating trend, but also accurately captures periodic fluctuations and sudden disturbances in electrical, insulation and environmental signals, thereby achieving robust modeling and accurate characterization of the time-dependent relationship of the transformer.

[0026] Preferably, in step S4, the global dependency and local fluctuation information of the cross-time feature vector are extracted, and after weighted fusion, the perturbation feature representation is obtained by layer normalization and ReLU. The total perturbation energy is obtained by square and absolute value correction, and then fused with the perturbation feature representation to form a fused feature vector.

[0027] Using the weight matrix to assess the first Perform a linear transformation on the feature vectors across time segments to extract the global dependencies between data, and calculate the first... Segment cross-time feature vector and the first The absolute value of the difference between the means of the feature vectors across time segments is used to extract local fluctuation information. The local fluctuation information, weighted by the global dependency and the perturbation enhancement factor, is then added and processed through layer normalization and ReLU activation to obtain the first... The perturbation characteristics of the segment are represented;

[0028] The disturbance feature of the first dimension is squared, and the absolute value of the disturbance suppression factor weighted first dimensional feature value is subtracted to obtain the disturbance energy, and the disturbance energy of all dimensions in the first segment is accumulated to obtain the total disturbance energy in the first segment;

[0029] The total disturbance energy of the first segment is multiplied by the disturbance feature representation, compressed by the tanh nonlinearity, and finally multiplied element by element with the original disturbance feature representation to output the fusion feature vector that fuses global and local information.

[0030] Further, for the state data of the transformer during operation, both the long-term evolution global trend and the local disturbance caused by factors such as current impact, partial discharge and sudden change of temperature and humidity are included, and the application proposes a global and local disturbance fusion mechanism. First, the weight matrix is used to linearly transform the cross-time feature vector of the first segment to extract the global dependence between data, and the absolute value of the difference between the vector and its mean value is calculated to extract local fluctuation information that can reflect electrical quantity mutation and environmental fluctuations. The weighted disturbance enhancement factor is added to the global dependence, normalized by layer and activated by ReLU to obtain the disturbance feature representation of the first segment, so as to balance the global stability and local sensitivity. Then, the eigenvalues of the disturbance feature representation in the first dimension are squared, and the disturbance suppression factor weighted absolute value term is subtracted to obtain the disturbance energy, and the disturbance energy of all dimensions is accumulated to obtain the total disturbance energy in the first segment, so as to highlight significant changes while suppressing abnormal amplification caused by random noise. Finally, the total disturbance energy is multiplied by the disturbance feature representation, and the tanh nonlinearity is used to compress extreme energy values, and then the original disturbance feature representation is fused element by element to obtain the fusion feature vector that integrates global dependence and local disturbance, which ensures the ability to describe the long-term state evolution trend of the transformer, and can effectively capture and constrain transient interference caused by current peak, partial discharge pulse and environmental mutation, thereby improving the robustness and prediction accuracy of the model.

[0031] Preferably, in the S5 step, the overall curvature representation is constructed based on the adjacent segment fusion feature vector, the long-term trend and local peak information are extracted using average and maximum pooling, and the multi-scale fusion feature is formed by splicing the current segment fusion feature vector.

[0032] The first segment and the first ​The square of the two-norm of the difference of the fusion feature vectors of the segments is obtained, a forward change amount is further calculated, and the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, so as to construct the overall curvature representation of the third segment The square of the two-norm of the difference of the fusion feature vectors of the segments is obtained, a forward change amount is further calculated, and the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, so as to construct the overall curvature representation of the third segment The square of the two-norm of the difference of the fusion feature vectors of the segments is obtained, a forward change amount is further calculated, and the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, so as to construct the overall curvature representation of the third segment 、 、 The second-order difference vector of the three segments is calculated, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment

[0033] On the basis of the first segment, the overall curvature representation sequence of the past length is taken, average pooling and maximum pooling are performed, smooth long-term trend information and local peak information under extreme disturbance are obtained respectively, and the two types of pooling results are spliced with the fusion feature vector of the first segment to form the multi-scale fusion feature of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment.

[0034] Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment. Further, in view of the fact that the operating data of the transformer potential transformer contains both smooth long-term trend and severe disturbance caused by current impact, local discharge or environmental mutation, the application proposes a multi-scale fusion mechanism based on overall curvature representation, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a forward change amount, the square of the two-norm of the difference of the fusion feature vectors of the adjacent segments is calculated to obtain a backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three segments is constructed, the square of the two-norm thereof is calculated, and a constant 1 is added as a denominator to obtain the overall curvature representation of the third segment.

[0035] ​​​​​​​​Preferably, in the S6 step, all segmented multi-scale fusion features are flattened, linearly transformed and outputted after Sigmoid mapping to obtain the mutual inductor state prediction value;

[0036] All segmented multi-scale fusion features are sequentially stacked to form a comprehensive feature matrix, the comprehensive feature matrix is flattened, multiplied by a learnable weight matrix and added with a bias to obtain a linear output result, and then a mutual inductor state prediction value is obtained through a Sigmoid function.

[0037] Further, in view of the problem that the final discriminant output needs to be realized on the basis of all segments for mutual inductor state prediction, the application proposes a prediction output mechanism based on linear mapping and Sigmoid compression, all segmented multi-scale fusion features are sequentially stacked to form a comprehensive feature matrix, and a flattening operation is performed on the comprehensive feature matrix to unify the feature representation of different segments in the time dimension, then the flattened comprehensive feature vector is multiplied by a learnable weight matrix and added with a bias to obtain a linear output result, and the linear result is mapped to a normalized mutual inductor state prediction value through a Sigmoid function, which can compress the complex segment representation into an intuitive probability output, thereby realizing accurate prediction of the mutual inductor state.

[0038] Preferably, in the S7 step, a mutual inductor state prediction model based on big data analysis is constructed, a mutual inductor state data set is inputted, a weighted binary cross-entropy is used as a loss function, and the model is iteratively trained to convergence through steps S2 to S6 to complete the optimization training of the model.

[0039] Further, the model is realized by using Python programming language and PyTorch deep learning framework, an NVIDIA 3090 24GB GPU is used in the training process, a weighted binary cross-entropy is used as the loss function to balance the imbalance between normal and abnormal samples, an Adam optimizer is selected for parameter updating, and the learning rate is dynamically adjusted by using an exponential decay strategy.

[0040] In the above technical solution, the application provides the following technical effects and advantages:

[0041] 1、The application introduces a segmented representation in the tensorized time series in the mutual inductor state data set, generates a feature representation value by combining a square term and an absolute value term with a sine transformation and a tanh function, and then generates a mutual inductor state feature vector by combining linear projection and cross-dimension difference items, thereby effectively weakening the extreme peak interference caused by current impact, partial discharge pulse and temperature and humidity mutation, ensuring the stability of the segmented feature representation, explicitly depicting the coupling relationship between electrical quantities, insulation quantities and environmental quantities, and thereby improving the robust modeling capability and cross-dimension correlation expression capability of the mutual inductor multi-source features.

[0042] 2、The application obtains attention weights by constructing a disturbance correction kernel, weighted aggregation and combining a softmax attention mechanism, and extracts main trend information and periodic disturbance information based on the mutual inductor state feature vector respectively, generates a cross-time feature vector by using the attention weights for weighted fusion, suppresses transient noise in the operation data, and adaptively highlights key time sequence segments with high correlation to the current segment, thereby ensuring the continuity of the state trend expression and accurately reflecting periodic disturbance characteristics such as current and voltage signal fluctuations, temperature and humidity changes, thereby enhancing the stability and accuracy of cross-time dependent modeling;

[0043] 3、The application utilizes linear transformation and the absolute value of mean difference to construct global dependence and local fluctuation information, obtains disturbance feature representation through layer normalization and ReLU after weighted fusion, and further calculates total disturbance energy through squaring and absolute value correction, and obtains a robust fusion feature vector by fusing the original disturbance feature representation; subsequently, the overall curvature representation is constructed based on adjacent segments, and long-term trend and local peak information are extracted by combining average pooling and maximum pooling, finally, multi-scale fusion features are formed, and the mutual inductor state prediction value is output through linear mapping and Sigmoid function, which takes into account global evolution trend and local disturbance features, effectively captures the smooth degradation law in the aging process and the extreme peak value problem of the mutual inductor state data, and thereby significantly improves the state prediction accuracy and reliability of the mutual inductor under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a flowchart of the mutual inductor state prediction method based on big data analysis provided by the application.

[0045] Figure 2 It is a mutual inductor state feature vector generation structure diagram provided by the application.

[0046] Figure 3 It is a cross-time feature vector generation structure diagram provided by the application.

[0047] Figure 4 It is a multi-scale fusion feature generation structure diagram provided by the application.

[0048] Figure 5 It is a model training loss curve diagram provided by the application.

[0049] Figure 6 It is a mutual inductor state prediction model prediction effect diagram based on big data analysis provided by the application. DETAILED DESCRIPTION

[0050] The application provides a mutual inductor state prediction method based on big data analysis, aiming at the problems of large state data quantity, unstable cross-time dependence and frequent local disturbance of the mutual inductor in the long-term operation process, a prediction mechanism is provided by fusing segmented representation, cross-time feature modeling and multi-scale curvature fusion, including constructing a mutual inductor state feature vector based on segmented representation and cross-dimension difference, generating a cross-time feature vector by combining a disturbance correction kernel and an attention mechanism, forming a disturbance feature representation by fusing global dependence relationship and local fluctuation information and introducing overall curvature representation to generate multi-scale fusion features, and finally outputting a mutual inductor state prediction value through linear transformation and Sigmoid mapping, so as to realize unified modeling and accurate prediction of global evolution and local disturbance characteristics of the mutual inductor.

[0051] S1, acquire state data of the mutual inductor, and construct a mutual inductor state data set after preprocessing.

[0052] Please refer to Figure 1 The mutual inductor state prediction method based on big data analysis in the embodiment of the application has the following specific steps.

[0053] The construction process of the mutual inductor state data set includes three stages of data acquisition, preprocessing and structuring, the mutual inductor state data includes electrical quantity data, environmental quantity data and insulation quantity data, wherein the electrical quantity data is the current and voltage signals output from the secondary side of the mutual inductor, the environmental quantity data includes the surface temperature of the mutual inductor body and the temperature and humidity of the operating environment, and the insulation quantity data includes local discharge pulse, leakage current and dielectric loss factor, in the data acquisition stage, the sampling rate of the current and voltage signals is set to 5 kHz to ensure complete recording of harmonic and dynamic components, the surface temperature of the body and the temperature and humidity of the operating environment are sampled at a period of 1 minute to reflect the slow change process, the sampling rate of the local discharge pulse signal is set to 100 kHz to capture the transient impact characteristics, and the leakage current and dielectric loss factor are obtained in units of hours as the key reference for long-term insulation degradation, in the preprocessing stage, first, the collected multiple types of data are time-aligned, the sampling time of the voltage signal is taken as the reference, and the time stamps of the electrical quantity data, the environmental quantity data and the insulation quantity data are corrected by interpolation, in the structuring stage, the data is cut according to a fixed time sequence length, the time sequence length is set to 60 minutes, the electrical quantity data, the environmental quantity data and the insulation quantity data are stacked in the form of channels within the time sequence length to form tensorized time sequence data, and finally the mutual inductor state data set is constructed.

[0054] S2, divide the tensorized time sequence data in the data set into segments, calculate the segmented representation, generate feature representation values by using sine transformation and tanh function, and generate a mutual inductor state feature vector by combining linear projection and cross-dimension difference items.

[0055] Furthermore, in step S2, the transformer state feature vector is generated, and the process is as follows: Figure 2 As shown, the specific steps include:

[0056] S21, the tensor-quantized time series data is denoted as , ,in The length of the time series. The number of data types for the current transformer status. For the first The data collected at each time step, Within each dimension, the sequence is divided into segments of length [length missing]. The segments were obtained in total. The data is divided into segments. Each value within a segment is squared, and a term proportional to the absolute value of that value is subtracted. The result is then normalized and averaged to reduce the impact of extreme anomalies caused by current surges, partial discharge pulses, and sudden temperature and humidity changes, thus obtaining a robust segmented representation. The specific mathematical model is as follows:

[0057] ;

[0058] in, For the first Duan Di Piecewise representation of dimensions The segment length is... For the first Time step The numerical value of dimension, For learnable dimension-dependent perturbation factors, This is for the absolute value operation;

[0059] In this embodiment, Set to 60. Set to 5, and get The system is divided into segments to ensure stable segmentation while highlighting local variations. Initialized to 0.5 to suppress extreme outliers caused by transformer current surges and temperature / humidity fluctuations during large-scale data acquisition. The value is set to 8, which represents the number of data types for the current transformer status data.

[0060] S22. Segmentation representation With sine transformation The summation is used to simulate the dynamic disturbance of the transformer state signal under periodic operating conditions. This disturbance is mapped to a finite interval using tanh. The specific mathematical model is as follows:

[0061] ;

[0062] in, For the first segment dimensional feature representation value;

[0063] S23, performing linear projection on the feature representation , and then introducing a cross-dimensional difference term, calculating the square difference between the feature representation value of the first dimension and all other dimensions, adding the projection result, explicitly encoding the coupling relationship between different mutual inductor state data, obtaining the representation after fusing the cross-dimensional relationship, and the specific mathematical model is:

[0064] ;

[0065] wherein, is the fused cross-dimensional relationship representation of the first segment dimension, is a learnable dimension-specific projection weight, is a learnable correlation coefficient between the first dimension and the first dimension, is the feature representation value of the first segment dimension;

[0066] S24, splicing the fused cross-dimensional relationship representations of all dimensions of the first segment to obtain a mutual inductor state feature vector .

[0067] S3, introducing a disturbance correction kernel between different segments, obtaining attention weights through weighted aggregation and softmax transformation, extracting main trend and periodic disturbance information based on the mutual inductor state feature vector, and generating a cross-time feature vector by weighted fusion using the attention weights.

[0068] Further, in the S3 step, a cross-time feature vector is generated, and the process is as shown in Figure 3 , and specifically includes the following steps:

[0069] S31, for the fused cross-dimensional relationship representations of the first segment and the first segment on the same dimension , calculating the square difference between the two, introducing a penalty term proportional to the absolute value of the difference between the two, to weaken the influence of transient peaks in the mutual inductor state data caused by current shock, temperature and humidity mutation, etc., thereby generating a disturbance correction kernel, and the specific mathematical model is:

[0070] ;

[0071] wherein, is the fused cross-dimensional relationship representation of the first segment and the first segment in the first dimensional perturbation correction kernel, for the first segment in the first dimensional fusion cross-dimensional relationship representation, dimensional perturbation factor is learnable;

[0072] In the embodiment, initialized to 0.3 to weaken the interference of burst noise while retaining the difference signal of the real working condition change of the transformer;

[0073] S32, the first segment and the first segment of all dimensions are weighted and aggregated, and the average cross-time perturbation weight is obtained, so that different transformer state data contributes differently in the cross-time relationship, and the specific mathematical model is:

[0074] ;

[0075] wherein, is the cross-time perturbation weight of the first segment and the first segment, dimensional weight is learnable;

[0076] S33, using softmax, the cross-time perturbation weight is converted into attention weight to measure the time sequence dependency between segments, and the specific mathematical model is:

[0077] ;

[0078] wherein, is the attention weight of the first segment and the first segment, the greater the attention weight, the higher the cross-time correlation degree of the two segments in the transformer state data, is the cross-time perturbation weight of the first segment and the first segment;

[0079] S34, based on the transformer state feature vector, the main trend information extracted by the linear projection matrix, and the periodic disturbance information obtained by the sine function, the attention weight weighted fusion of main trend information and periodic disturbance information obtains the cross-time feature vector, and the specific mathematical model is:

[0080] ;

[0081] wherein, is the cross-time feature vector of the first segment, , For learnable projection matrices, For the first State feature vector of a segment mutual inductor.

[0082] S4. Extract the global dependency and local fluctuation information of the cross-time feature vectors, and after weighted fusion, obtain the perturbation feature representation through layer normalization and ReLU. After square and absolute value correction, obtain the total perturbation energy, and fuse it with the perturbation feature representation to form a fused feature vector.

[0083] Furthermore, in step S4, the fused feature vector is generated, which specifically includes the following steps:

[0084] S41. Using the weight matrix to analyze cross-time eigenvectors Perform linear transformations to extract global dependencies between data and calculate... and The absolute value of the difference between means is used to extract local fluctuation information, which is used to capture transient disturbances such as current surges, partial discharge pulses, and sudden changes in temperature and humidity, and to combine global dependencies with disturbance enhancement factors. The weighted local fluctuation information is summed, and the perturbation feature representation is obtained through layer normalization and ReLU activation. The specific mathematical model is as follows:

[0085] ;

[0086] in, For the first The perturbation characteristics of the segment are represented. The weight matrix is ​​a learnable matrix. As a learnable perturbation enhancement factor, for The mean calculated from all dimensions For layer normalization, This is for the absolute value operation;

[0087] In this embodiment, Initialize it to 1 to keep local fluctuation information and global dependencies on the same order of magnitude, so that the model can maintain a stable convergence speed at the beginning;

[0088] S42, will No. The eigenvalues ​​of the dimension are squared to highlight large changes, and the weighted eigenvalues ​​of the perturbation suppression factor are subtracted. The absolute value of the dimensional eigenvalues ​​is used to obtain the perturbation energy, which is used to suppress excessive amplification caused by environmental noise or abnormal peaks. The total perturbation energy is obtained by summing the perturbation energies of all dimensions. The specific mathematical model is as follows:

[0089] ;

[0090] wherein, is the total disturbance energy of the th segment, is the disturbance eigenvalue of the th segment, th dimension, is the learnable disturbance suppression factor;

[0091] In the present embodiment, is initialized as 0.2, used to adjust the influence of the absolute value term;

[0092] S43, multiply the total disturbance energy with the disturbance eigenvalue representation, compress through the tanh nonlinearity, avoid abnormal energy dominating the feature distribution, and finally multiply element by element with the original disturbance eigenvalue representation, output the fusion feature vector that fuses global and local information, and the specific mathematical model is:

[0093] ;

[0094] wherein, is the fusion feature vector of the th segment, is element by element multiplication.

[0095] S5, based on the adjacent segment fusion feature vector, construct the overall curvature representation, use average and maximum pooling to extract long-term trend and local peak information, and splice with the current segment fusion feature vector to form a multi-scale fusion feature.

[0096] Further, in the S5 step, the process is as shown in Figure 4 , the two norm square of the difference value of the fusion feature vectors of the th segment and the th segment is calculated to obtain the forward change amount, and the two norm square of the difference value of the fusion feature vectors of the th segment and the th segment is further calculated to obtain the backward change amount, the difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount, and the second-order difference vector of the three , , segments is constructed, and after calculating the two norm square, 1 is added as the denominator to constrain the numerical range, and the overall curvature representation is obtained, and the specific mathematical model is:

[0097] ;

[0098] wherein, is the overall curvature representation of the th segment, , respectively, and segment fusion feature vector, is a two-norm;

[0099] On the basis of the first segment, the past length of the overall curvature representation sequence is taken, where is the overall curvature representation of the first segment, average pooling and maximum pooling are performed, respectively, to obtain smooth long-term trend information and local peak information of extreme disturbance, and the two types of pooling results are spliced with the first segment fusion feature vector to form a multi-scale fusion feature, and the specific mathematical model is:

[0100] ;

[0101] wherein, is the multi-scale fusion feature of the first segment, is the last dimension splicing operation, is the average pooling operation, is the maximum pooling operation.

[0102] S6, flatten all segmented multi-scale fusion features, perform linear transformation and Sigmoid mapping, and output the transformer state prediction value.

[0103] Further, in the S6 step, all segmented multi-scale fusion features are stacked in turn to form a comprehensive feature matrix , , the comprehensive feature matrix is flattened, multiplied by a learnable weight matrix and added with a bias to obtain a linear output result, and then the output is mapped to a normalized probability distribution through a Sigmoid function. The specific mathematical model is:

[0104] ;

[0105] wherein, is the transformer state prediction value, the larger the prediction value, the higher the abnormal risk of the transformer, is a learnable weight matrix, wherein is the dimension of the weight matrix, , is a bias, is a Sigmoid function, is a flattening operation.

[0106] Aiming at the characteristics of large amount of state data, unstable cross-time dependence and frequent local disturbance of the mutual inductor in long-term operation, the application constructs a mutual inductor state prediction model based on big data analysis, inputs the mutual inductor state data set, sequentially executes tensorization time series data segmentation and generates mutual inductor state feature vectors, introduces disturbance correction kernel and attention mechanism to generate cross-time feature vectors, fuses global dependence relationship and local fluctuation information and calculates total disturbance energy, combines overall curvature representation and multi-scale pooling to generate multi-scale fusion features, finally outputs mutual inductor state prediction values through linear transformation and Sigmoid mapping, and adopts a weighted binary cross-entropy loss function in the training process to update the model parameters end-to-end, realizes convergence by constraining the difference between the prediction results and the actual state labels, improves the stable response capability of the model under extreme working conditions such as current impact, local discharge pulse and temperature and humidity mutation, and finally forms a mutual inductor state prediction model with high state recognition accuracy.

[0107] Further, in the S7 step, the mutual inductor state prediction model based on big data analysis proposed by the application is realized by using Python programming language and PyTorch deep learning framework, NVIDIA 309024GB GPU is used in the training process, the loss function adopts weighted binary cross-entropy, which is used to balance the imbalance between normal and abnormal samples, the optimizer selects Adam for parameter update, the initial learning rate is set to , and an exponential decay strategy is used for dynamic adjustment, and the batch size is set to 12.

[0108] Further, in the S7 step, the mutual inductor state data set is input into the mutual inductor state prediction model based on big data analysis constructed for processing, and the training loss curve of the model is as shown in Figure 5 From the figure, it can be seen that with the increase of training rounds, the loss value gradually decreases and finally tends to be stable, indicating that the model can be effectively optimized in the training process; the prediction effect is as shown in Figure 6 The horizontal coordinate in the figure is time, in minutes, and the vertical coordinate is the mutual inductor state value, the gray dotted line is the actual evaluation value, and the black solid line is the model prediction value, from the figure, it can be seen that the prediction value and the actual evaluation value are highly consistent in the overall trend, and the local fluctuation error is small, which fully verifies the accuracy and reliability of the method in the mutual inductor state prediction task.

[0109] The above is only the preferred embodiment of the application, and it should be noted that for ordinary skilled persons in the art, without departing from the inventive concept, several modifications and improvements can be made, which are all within the protection scope of the application.

Claims

1. A method for predicting the state of a transformer based on big data analysis, characterized in that, The method comprises the following steps: S1, acquiring state data of the mutual inductor, and constructing a mutual inductor state data set after preprocessing; S2, dividing tensorized time series data in the data set into segments, calculating segment representations, generating feature representation values by using a sine transform and a tanh function, combining linear projection and cross-dimension difference items, and generating a mutual inductor state feature vector; S3, introducing a disturbance correction kernel between different segments, obtaining attention weights through weighted aggregation and softmax transformation, extracting main trends and periodic disturbance information based on the mutual inductor state feature vector, and generating a cross-time feature vector by weighted fusion using the attention weights; S4, extracting global dependency relationships and local fluctuation information of the cross-time feature vector, obtaining disturbance feature representations through layer normalization and ReLU after weighted fusion, obtaining total disturbance energy through squaring and absolute value correction, and forming a fusion feature vector by fusing the disturbance feature representations; S5, constructing an overall curvature representation based on adjacent segment fusion feature vectors, extracting long-term trends and local peak information by using average and maximum pooling, and splicing the current segment fusion feature vector to form a multi-scale fusion feature; S6, flattening all segmented multi-scale fusion features, performing linear transformation and Sigmoid mapping, and outputting a mutual inductor state prediction value; S7, constructing a mutual inductor state prediction model based on big data analysis, inputting the mutual inductor state data set, taking weighted binary cross-entropy as a loss function, sequentially passing through steps S2 to S6, and iteratively training until convergence to complete optimization training of the model.

2. The method of claim 1, wherein, The mutual inductor state data includes electrical quantity data, environmental quantity data and insulation quantity data, wherein the electrical quantity data is the current and voltage signals output by the secondary side of the mutual inductor; the environmental quantity data includes the surface temperature of the mutual inductor and the temperature and humidity of the operating environment; the insulation quantity data includes partial discharge pulses, leakage current and dielectric loss factor, and forms tensorized time series data after preprocessing and structuring, thereby constructing a mutual inductor state data set.

3. The method of claim 2, wherein the method is characterized by, each dimension of the tensorized time series data is divided into segments of length each value within a segment is squared and subtracted by a term proportional to the absolute value of the value, followed by a normalized average to obtain a segment representation for each dimension; to the first segment the first dimensional feature representation value is linearly projected, and then a cross-dimensional difference term is introduced, the square difference between the first dimensional feature representation value and the feature representation values of all other dimensions is added to the projection result to obtain a fused cross-dimensional relationship representation of the first segment the first dimensional feature representation value is linearly projected, and then a cross-dimensional difference term is introduced, the square difference between the first dimensional feature representation value and the feature representation values of all other dimensions is added to the projection result to obtain a fused cross-dimensional relationship representation of the first segment the first dimensional feature representation value is linearly projected, and then a cross-dimensional difference term is introduced, The first The fusion of all dimensions across the dimension relationship representation splicing is obtained.

4. The method of claim 3, wherein the method is characterized by, For the first segment and the second segment, the fusion of the first dimensional relationship representation is calculated, the square difference of the two is calculated, a penalty term proportional to the absolute value of the difference of the two is introduced, and the first segment and the second segment in the first dimensional disturbance correction kernel are generated; The cross-time perturbation weight of the first segment and the second segment is obtained by weighting and aggregating all dimensions of the perturbation correction kernels of the first segment and the second segment, and taking the average. The cross-time perturbation weights are converted to attention weights for the segment and the segment using softmax. Based on the state feature vector of the mutual inductor, the main trend information extracted by the linear projection matrix is obtained, and the periodic disturbance information is obtained by the sine function. The main trend information and the periodic disturbance information are weighted and fused by using the attention weight to obtain the first Segment cross-time feature vector.

5. The method of claim 4, wherein the method is characterized by, The weight matrix is used to linearly transform the first segment cross-time feature vector, extract global dependence between data, and calculate the difference between the first segment cross-time feature vector and the mean of the first segment cross-time feature vector; the absolute value of the difference is extracted as local fluctuation information; the global dependence and the local fluctuation information weighted by the disturbance enhancement factor are added; and the disturbance feature representation of the first segment is obtained through layer normalization and ReLU activation. The first The disturbance characteristics of segment 1 represent the first Squaring the eigenvalues ​​of dimension 1, and subtracting the weighted eigenvalues ​​of the perturbation suppression factor of dimension 2. The absolute value of the eigenvalues ​​of the dimension is used to obtain the perturbation energy, which is then used to determine the eigenvalues ​​of the dimensional features. The perturbation characteristics of a segment represent the sum of the perturbation energies in all dimensions to obtain the first segment. Total disturbance energy of the segment; The total perturbation energy of the first segment is multiplied by the perturbation feature representation, compressed by a tanh nonlinearity, and finally element-wise multiplied by the original perturbation feature representation, outputting a fused feature vector that fuses global and local information.

6. The method of claim 5, wherein the method is characterized by, The second-order difference vector of the third, fourth and fifth segments is calculated, and a constant 1 is added to the square of the second norm of the second-order difference vector as a denominator to obtain the overall curvature representation of the sixth segment. The square of the second norm of the difference between the fusion feature vectors of the second and third segments is calculated to obtain a forward change amount. The square of the second norm of the difference between the fusion feature vectors of the fourth and fifth segments is calculated to obtain a backward change amount. The difference in the overall change direction is obtained by subtracting the backward change amount from the forward change amount. The second-order difference vector of the third, fourth and fifth segments is calculated, and a constant 1 is added to the square of the second norm of the second-order difference vector as a denominator to obtain the overall curvature representation of the sixth segment. , , The second-order difference vector of the third, fourth and fifth segments is calculated, and a constant 1 is added to the square of the second norm of the second-order difference vector as a denominator to obtain the overall curvature representation of the sixth segment. The second-order difference vector of the third, fourth and fifth segments is calculated, and a constant 1 is added to the square of the second norm of the second-order difference vector as a denominator to obtain the overall curvature representation of the sixth segment. On the basis of the first segment, the overall curvature representation sequence of the past length is subjected to average pooling and maximum pooling, respectively obtaining smooth long-term trend information and local peak information of extreme disturbance, and the two types of pooling results are spliced with the first segment fusion feature vector to form the second segment multi-scale fusion feature.

7. The method of claim 6, wherein the method is based on big data analysis. All segmented multi-scale fusion features are stacked in sequence to form a comprehensive feature matrix, the comprehensive feature matrix is flattened, multiplied by a learnable weight matrix and added with a bias to obtain a linear output result, and then a mutual inductor state prediction value is obtained through a Sigmoid function.

Images