A harmonic detection and early warning system for intelligent transformers

Abstract

The present invention relates to the technical field of transformer harmonic monitoring, and in particular to a harmonic detection and early warning system for intelligent transformers. By acquiring the voltage and current signals of the transformer in real time, the frequency domain decomposition is performed using Fourier transform, and multi-dimensional harmonic features such as amplitude, phase and frequency are extracted, and the features are input into a deep learning model for training to generate a dynamic harmonic feature model. On this basis, a harmonic anomaly classification module analyzes harmonic signals in real time, determines the type and associates the position of abnormal harmonics, and generates a report containing the type and position of the anomaly. Through the dynamic early warning module, the grid load factor is calculated based on the real-time collected grid load state parameters, and the harmonic detection threshold is dynamically adjusted in combination with the initially set amplitude and frequency thresholds. The application of dynamic thresholds enhances the system's adaptability to changes in load states, and significantly improves the sensitivity and accuracy of harmonic detection and early warning.

Description

Technical Field

[0001] The invention relates to the technical field of transformer harmonic monitoring, and in particular to a harmonic detection and early warning system for an intelligent transformer. Background Art

[0002] As an important part of the power grid, the operating environment of the transformer is becoming increasingly complex. The widespread use of load fluctuations and nonlinear loads has led to increasingly serious harmonic problems. Harmonics not only increase the power loss of the transformer, but also cause problems such as equipment heating and insulation aging. In severe cases, it may even cause transformer failure, affecting the safe and stable operation of the power grid. Therefore, how to effectively detect harmonics and provide early warning has become an important research direction in the power industry.

[0003] In the existing technology, harmonic detection mainly relies on the threshold judgment method. The traditional harmonic detection system analyzes the voltage and current signals collected in real time by setting fixed amplitude and frequency thresholds. Once the harmonic signal exceeds the preset threshold, an alarm is triggered. However, the fixed threshold detection method has obvious limitations:

[0004] Insufficient sensitivity: Fixed thresholds cannot dynamically adapt to complex and changing grid load conditions, which may result in missed reporting of harmonic anomalies under light load conditions and false alarms under heavy load conditions.

[0005] Lack of anomaly classification: Fixed thresholds only provide simple anomaly detection and cannot further analyze the type of abnormal harmonics, limiting its application depth in harmonic diagnosis. Summary of the invention

[0006] To solve the above problems, the present invention provides a harmonic detection and early warning system for intelligent transformers, which learns harmonic anomalies through a deep learning model and performs real-time monitoring and classification, adjusts the harmonic early warning threshold through real-time grid load, and performs harmonic anomaly early warning based on the monitored abnormal harmonics and the abnormal warning threshold.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A harmonic detection and early warning system for an intelligent transformer, comprising a harmonic data acquisition module, a harmonic learning module, a harmonic anomaly classification module and a dynamic early warning module connected in sequence;

[0009] The harmonic data acquisition module is used to collect the voltage signal and current signal of the transformer in real time, construct a time series signal set including harmonic current data and harmonic voltage data, and pre-process the time series signal set to obtain a harmonic data set;

[0010] The harmonic learning module is used to decompose the signal using Fourier transform based on the preprocessed harmonic data set to generate a harmonic feature set, including the amplitude, phase and frequency characteristics of each order of harmonics; the harmonic feature set is input into the deep learning model to train and generate a dynamic harmonic feature model;

[0011] The harmonic anomaly classification module is used to analyze the real-time harmonic feature set based on the trained dynamic harmonic feature model, detect abnormal harmonic signals and generate an abnormal harmonic report, wherein the abnormal harmonic report includes the abnormal type and location of the abnormal harmonic;

[0012] The dynamic warning module is used to dynamically update the harmonic warning threshold based on the real-time grid load state parameters, and perform warning operations according to abnormal harmonic signals and the harmonic warning threshold. The warning operation outputs abnormal harmonic reports.

[0013] Furthermore, the construction of a time series signal set including harmonic current data and harmonic voltage data comprises the following steps:

[0014] The collected voltage signal and current signal are synchronously corrected by a time synchronization algorithm to generate synchronized voltage signal and current signal;

[0015] The synchronized voltage and current signals are filtered by a bandpass filtering algorithm to remove fundamental waves and high-frequency noise, and to extract characteristic signals within the harmonic frequency band;

[0016] The characteristic signal in the harmonic frequency band is segmented and formatted into a time series signal set through segmentation processing.

[0017] Furthermore, the preprocessing is normalization processing.

[0018] Furthermore, the step of inputting the harmonic feature set into the deep learning model and training and generating a dynamic harmonic feature model comprises the following steps:

[0019] Classify and label each feature sample in the harmonic feature set as normal harmonics and abnormal harmonics, and generate a standardized training data set;

[0020] The training data set is input into a convolutional neural network model, which includes several convolutional layers, pooling layers and fully connected layers. The convolutional layers are used to extract local features of harmonic amplitude, phase and frequency, the pooling layers are used to reduce data dimension, and the fully connected layers are used to perform classification output;

[0021] The convolutional neural network model is trained through the back-propagation algorithm, the classification error is evaluated using the cross-entropy loss function, the model parameters are updated through the Adam optimization algorithm, and the network weights are iteratively optimized.

[0022] Furthermore, the formula of the convolutional neural network model is as follows:

[0023] ;

[0024] in, Output probability for the kth class; is the jth input feature in the input feature vector, including the amplitude, phase and frequency information of the harmonic; is the weight parameter from input feature j to hidden unit i in the hidden layer; is the weight parameter from hidden unit i to output category k in the output layer; and are the bias parameters of the hidden layer and the output layer respectively; is the hidden layer activation function; is the activation function of the output layer.

[0025] Furthermore, the analysis of the real-time harmonic feature set, detection of abnormal harmonic signals and generation of an abnormal harmonic report comprises the following steps:

[0026] The harmonic feature set collected in real time is input into the trained dynamic harmonic feature model to extract abnormal harmonic signals and abnormal types that deviate from the normal harmonic pattern;

[0027] For the deviated abnormal harmonic signal, the position of the abnormal harmonic signal is obtained from the pre-associated position table, and an abnormal harmonic report is generated according to the abnormal harmonic signal position and abnormal type.

[0028] Furthermore, the harmonic warning threshold includes an amplitude threshold and a frequency threshold.

[0029] Furthermore, the dynamically updating the harmonic warning threshold based on the real-time grid load state parameter includes:

[0030] Normalizing the real-time grid load state parameter into a grid load factor, wherein the grid load factor ranges from 0 to 1;

[0031] The grid load factor is multiplied by the initial amplitude threshold and the initial frequency threshold respectively to obtain the amplitude threshold and the frequency threshold.

[0032] Furthermore, the warning operation performs warning notification via voice broadcast and text message.

[0033] The beneficial effects of the present invention are as follows: the present invention obtains the transformer voltage and current signals in real time, uses Fourier transform to perform frequency domain decomposition on the harmonic signals, extracts multidimensional harmonic features including amplitude, phase and frequency, and inputs them into the deep learning model for training to generate a dynamic harmonic feature model. On this basis, the harmonic anomaly classification module analyzes the harmonic signals in real time, determines the type and associates the position of the abnormal harmonics, and generates an abnormal harmonic report containing the abnormal type and position. Further, through the dynamic early warning module, based on the real-time collected grid load state parameters, it is standardized to the grid load factor, combined with the initially set amplitude threshold and frequency threshold, and dynamically calculates the harmonic detection threshold. The introduction of dynamic thresholds improves the adaptability of the harmonic detection system to changes in load states, thereby enhancing the sensitivity and accuracy of early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a structural schematic diagram of a harmonic detection and early warning system for an intelligent transformer in the present invention.

[0035] Figure 2 It is a flow chart of the steps of training and generating a dynamic harmonic feature model in the present invention. DETAILED DESCRIPTION

[0036] See also Figure 1-Figure 2 As shown, the present invention relates to a harmonic detection and early warning system for an intelligent transformer, comprising a harmonic data acquisition module, a harmonic learning module, a harmonic anomaly classification module and a dynamic early warning module connected in sequence;

[0037] The harmonic data acquisition module is used to collect the voltage signal and current signal of the transformer in real time, construct a time series signal set including harmonic current data and harmonic voltage data, and pre-process the time series signal set to obtain a harmonic data set;

[0038] The harmonic learning module is used to decompose the signal using Fourier transform based on the preprocessed harmonic data set to generate a harmonic feature set, including the amplitude, phase and frequency characteristics of each order of harmonics; the harmonic feature set is input into the deep learning model to train and generate a dynamic harmonic feature model;

[0039] The harmonic anomaly classification module is used to analyze the real-time harmonic feature set based on the trained dynamic harmonic feature model, detect abnormal harmonic signals and generate an abnormal harmonic report, wherein the abnormal harmonic report includes the abnormal type and location of the abnormal harmonic;

[0040] The dynamic warning module is used to dynamically update the harmonic warning threshold based on the real-time grid load state parameters, and perform warning operations according to abnormal harmonic signals and the harmonic warning threshold. The warning operation outputs abnormal harmonic reports.

[0041] Specifically, first, the harmonic data acquisition module uses current transformers and voltage transformers to obtain the current and voltage signals of the transformer in real time. The collected data is time-aligned through a time synchronization algorithm to ensure the synchronization of the voltage and current signals. Subsequently, the signal is filtered using a bandpass filtering algorithm to filter out the fundamental component and high-frequency noise, retaining only the effective characteristic signals within the harmonic frequency band. The processed signal is time-series processed by a segmentation algorithm to construct a time series signal set containing harmonic current data and harmonic voltage data, forming a standardized harmonic data set, laying the foundation for subsequent analysis. In the harmonic learning module, the harmonic data set is decomposed in the frequency domain by Fourier transform, and the amplitude, phase and frequency characteristics of each order of harmonics are extracted to generate a harmonic feature set. For example, for a 50 Hz power system, the decomposed harmonic features include specific parameters such as the second harmonic (100 Hz) and the third harmonic (150 Hz). The extracted feature set is input into the deep learning model for training. The deep learning model adopts a multi-layer convolutional neural network architecture. The convolution layer is used to extract local features, the pooling layer is used for dimensionality reduction and noise suppression, and the fully connected layer realizes comprehensive analysis of features. The model weights are optimized by the back propagation algorithm and the cross entropy loss function, and finally a dynamic harmonic feature model is generated, which can adapt to the changes in harmonic patterns under different power grid operation conditions. The harmonic anomaly classification module analyzes the real-time input harmonic features based on the dynamic harmonic feature model and identifies abnormal harmonic signals that deviate significantly from the normal mode. For example, when the amplitude of the third harmonic exceeds the historical average by more than twice, the system identifies it as an abnormality and determines the specific location of the abnormal harmonic in combination with the pre-set position association table. The generated abnormal harmonic report records the abnormal type (such as amplitude abnormality, frequency offset, etc.) and the corresponding location in detail, providing accurate diagnostic basis for operation and maintenance personnel. The dynamic early warning module obtains the load of the power grid in real time, standardizes these parameters into the power grid load factor, and calculates the dynamic harmonic detection threshold in combination with the initially set amplitude threshold and frequency threshold. For example, when the load power fluctuates significantly, the system adjusts the harmonic detection threshold according to the real-time load factor to adapt to the current operating conditions. If the detected abnormal harmonic signal exceeds the dynamic threshold range, the system immediately triggers the early warning operation and sends an early warning notification containing the abnormal type and location to the operation and maintenance personnel in the form of voice broadcast and text message, so that timely response measures can be taken.

[0042] Furthermore, the construction of a time series signal set including harmonic current data and harmonic voltage data comprises the following steps:

[0043] The collected voltage signal and current signal are synchronously corrected by a time synchronization algorithm to generate synchronized voltage signal and current signal;

[0044] The synchronized voltage and current signals are filtered by a bandpass filtering algorithm to remove fundamental waves and high-frequency noise, and to extract characteristic signals within the harmonic frequency band;

[0045] The characteristic signal in the harmonic frequency band is segmented and formatted into a time series signal set through segmentation processing.

[0046] In some embodiments, in the time synchronization correction stage, a time synchronization algorithm based on cross-correlation analysis is used to synchronize and correct the collected voltage signal and current signal. Specifically, the time offset is calculated using the phase relationship between the collected voltage and current signals, and the time alignment of the signal is achieved by optimizing the peak position of the cross-correlation function of the calculated offset. This process can eliminate the time error caused by the delay of the acquisition device or the clock drift, and ensure that the voltage and current signals are consistent under the same time reference, thereby providing high-precision input for subsequent signal processing. In the filtering processing stage, a bandpass filtering algorithm is used to process the synchronized voltage and current signals. In the segmentation processing stage, the filtered signal is segmented. The algorithm dynamically determines the starting point and end point of the segmentation based on the periodic characteristics and frequency domain characteristics of the signal to avoid the information loss that may be caused by the fixed segmentation method. For example, by analyzing the periodic characteristics of the harmonic signal, each complete fundamental cycle is marked and the signal is divided into independent time sequence segments. In addition, in order to ensure the accuracy of the segmentation, a threshold detection mechanism is introduced to screen the signal amplitude and eliminate the peaks or abnormal values ​​in the signal. The harmonic signals after segmented processing are stored in a time-series form, and a timestamp is attached to mark the acquisition time, so as to construct a standardized time-series signal set. The final time-series signal set is stored in a matrix form, the rows of the matrix represent the time step, and the columns represent the amplitude and frequency information of each order of harmonics. Such a structure is not only convenient for subsequent Fourier transform and deep learning model input, but also can improve the versatility and adaptability of the signal through standardized processing. In this embodiment, the synergy of time synchronization, bandpass filtering and segmented processing algorithms provides high-precision basic data for harmonic detection of intelligent transformers.

[0047] Furthermore, the preprocessing is normalization processing.

[0048] Specifically, the normalization process is based on the linear mapping method to adjust the amplitude of the signal to a fixed value range, which can be [0,1]. Specifically, for each harmonic signal segment, the normalization parameters are generated by calculating the maximum and minimum values ​​of the signal segment. Then, these parameters are used to adjust the signal point by point to ensure that the processed data not only retains the relative characteristics of the original signal, but also has a unified value range. In the normalization process, the voltage and current signals are processed separately. This separate processing method can retain the independent characteristics between voltage and current, while providing more accurate input data for the subsequent feature extraction steps. After normalization, the value range of the data is strictly limited to the specified interval, thereby avoiding the problem of unstable model training caused by too large or too small values. The harmonic data after normalization is further organized into a structured data set, which can be directly used as the input of Fourier transform. This standardized data form not only makes the algorithm more adaptable in a changing signal environment, but also effectively improves the robustness of the model in the detection and classification process.

[0049] Furthermore, the step of inputting the harmonic feature set into the deep learning model and training and generating a dynamic harmonic feature model comprises the following steps:

[0050] Classify and label each feature sample in the harmonic feature set as normal harmonics and abnormal harmonics, and generate a standardized training data set;

[0051] The training data set is input into a convolutional neural network model, which includes several convolutional layers, pooling layers and fully connected layers. The convolutional layers are used to extract local features of harmonic amplitude, phase and frequency, the pooling layers are used to reduce data dimension, and the fully connected layers are used to perform classification output;

[0052] The convolutional neural network model is trained through the back-propagation algorithm, the classification error is evaluated using the cross-entropy loss function, the model parameters are updated through the Adam optimization algorithm, and the network weights are iteratively optimized.

[0053] It should be noted that, first, each feature sample in the harmonic feature set is classified and labeled. Specifically, based on historical operating data and expert experience, each set of harmonic features (including amplitude, phase and frequency) is classified and labeled, and normal harmonic signals are labeled as "normal", while signals that deviate from the normal range are labeled as "abnormal". For example, when the amplitude of a certain order of harmonic is significantly higher than the historical mean and the frequency deviates from the reference frequency by more than the set threshold, it is labeled as an abnormal harmonic. This labeling process is carried out by combining automated algorithms with manual correction to ensure the accuracy and consistency of the labeling. After the labeling is completed, a standardized training data set is generated, in which each sample includes both feature information and the corresponding classification label. Then, the training data set is input into the convolutional neural network (CNN) model. The model structure consists of several convolutional layers, pooling layers and fully connected layers, and each module plays a specific role in feature extraction and classification. The convolutional layer extracts local features in the harmonic signal through local receptive field operations, such as the amplitude and frequency pattern of harmonics of a specific order. The pooling layer reduces the data dimension by downsampling, retaining key features while reducing computational complexity. For example, the maximum pooling operation selects the local maximum value in the convolution output, which strengthens the response to key features. The fully connected layer maps the extracted high-dimensional features to the classification results, realizing the accurate distinction between normal harmonics and abnormal harmonics. During the model training process, the back propagation algorithm is used to optimize the parameters of the CNN model. Specifically, the classification output is first calculated through forward propagation, and the difference between the predicted result and the actual label is evaluated based on the cross entropy loss function. Then, the back propagation algorithm is used to calculate the gradient information to guide the update of the model weights and bias parameters. The optimization algorithm selects the Adam optimizer, which has the ability to adjust the learning rate adaptively, can converge quickly during the training process, and avoid falling into the local optimum. In order to improve the generalization ability of the model, data enhancement technology is introduced during the training process to randomly perturb the training data set, such as adding Gaussian noise or adjusting the harmonic amplitude. This method can simulate the harmonic changes under different operating conditions and improve the model's adaptability to unseen data. In addition, by setting a validation set, the model performance during the training process is monitored in real time to avoid overfitting problems. Finally, after multiple rounds of iterative optimization, the generated dynamic harmonic feature model can accurately capture the key features of the harmonic signal and achieve efficient classification of normal harmonics and abnormal harmonics.

[0054] Furthermore, the formula of the convolutional neural network model is as follows:

[0055] ;

[0056] in, Output probability for the kth class; is the jth input feature in the input feature vector, including the amplitude, phase and frequency information of the harmonic; is the weight parameter from input feature j to hidden unit i in the hidden layer; is the weight parameter from hidden unit i to output category k in the output layer; and are the bias parameters of the hidden layer and the output layer respectively; is the hidden layer activation function; is the activation function of the output layer.

[0057] First, the input feature vector contains key information such as the amplitude, phase, and frequency of the harmonic signal. These features are processed by the hidden layer, and each hidden unit generates an intermediate value based on the weighted summation of the input features. Specifically, the weight parameter is used to adjust the influence of the input features, and the bias parameter is used for correction. In order to introduce nonlinear modeling capabilities, each intermediate value is nonlinearly transformed through an activation function (such as the ReLU function), thereby enhancing the model's ability to capture complex feature patterns. The output of the hidden layer is further processed and enters the output layer. In the output layer, the activation values ​​of the hidden units are weighted and summed again, and combined with the bias parameters of the output layer to generate a score value corresponding to the classification category. Subsequently, these score values ​​are normalized by the softmax function and converted into a probability distribution for each category. The softmax function is characterized by ensuring that the output value is non-negative and the sum is 1, so that the possibility of the input feature belonging to each category can be intuitively represented. For example, for a harmonic feature set, the model processing process will extract key feature patterns in the input signal, such as abnormal amplitudes or frequency offsets of specific order harmonics, and judge whether the signal is an abnormal harmonic based on these patterns. The output results not only provide classification labels, but also indicate the confidence level of the classification through probability values, providing a basis for subsequent abnormal report generation and early warning. Through the above mechanism, the convolutional neural network model can learn the complex patterns of harmonic signals in the multi-dimensional feature space, adapt to the harmonic changes under different load conditions, and thus achieve efficient classification of harmonic signals.

[0058] Furthermore, the analysis of the real-time harmonic feature set, detection of abnormal harmonic signals and generation of an abnormal harmonic report comprises the following steps:

[0059] The harmonic feature set collected in real time is input into the trained dynamic harmonic feature model to extract abnormal harmonic signals and abnormal types that deviate from the normal harmonic pattern;

[0060] For the deviated abnormal harmonic signal, the position of the abnormal harmonic signal is obtained from the pre-associated position table, and an abnormal harmonic report is generated according to the abnormal harmonic signal position and abnormal type.

[0061] Specifically, first, the harmonic feature set collected in real time is input into the dynamic harmonic feature model. This model is based on a convolutional neural network model, and extracts signals that deviate from the normal harmonic pattern by pattern matching of the input features. The model analyzes the differences between multidimensional features such as amplitude, frequency, and phase and historical normal harmonic features. For example, when the amplitude of a certain order harmonic exceeds the normal range and the frequency shifts, the model determines it as an abnormal harmonic signal. The results of anomaly detection include not only the abnormal type of the signal (such as abnormal amplitude, frequency drift, etc.), but also the confidence of the abnormal feature, which is used to evaluate the reliability of the detection results.

[0062] After detecting an abnormal harmonic signal, the system will locate it in combination with the preset position signal association table. The position signal association table records the sources of harmonic signals in different parts of the transformer, such as the device location corresponding to each voltage and current sampling point. When the model detects an abnormal signal, the system matches the location information according to the sampling point of the abnormal signal to determine the physical source of the abnormal signal. For example, if the abnormal signal comes from the input current transformer, the device location will be marked in the report so that subsequent operation and maintenance personnel can quickly locate the problem. Finally, an abnormal harmonic report is generated based on the extracted abnormal type and location information. The report content includes the type, characteristic value, occurrence location and timestamp of the abnormal harmonic. For example, a report may record the abnormal amplitude of the third harmonic, the amplitude exceeds the normal range by 30%, the frequency offset is 2 Hz, it occurs at the current sampling point at the input of the main transformer, and the detection time is a specific timestamp. The generated report is stored in a standardized format to ensure that it can be directly called in the early warning module and notified to relevant personnel through voice or text messages.

[0063] Furthermore, the harmonic warning threshold includes an amplitude threshold and a frequency threshold.

[0064] Specifically, the setting of the amplitude threshold is based on historical operating data and equipment characteristics. Specifically, by statistically analyzing the amplitude of the harmonic signal during the normal operation of the transformer, the upper and lower limits of the amplitude of each order of harmonics are determined as the reference threshold. For example, for the third harmonic signal, its amplitude threshold may be set to 10% above and below the normal operating range. In real-time detection, when the amplitude of the harmonic signal exceeds the threshold, the system will determine it as an amplitude abnormality. The setting of the frequency threshold is based on the fundamental frequency of the system, and the reasonable upper and lower limits are determined by analyzing the frequency fluctuation range under different load conditions. The dynamic update of the harmonic warning threshold is achieved by combining the real-time grid load state parameters. For example, when the load increases, the system will automatically adjust the upper and lower limits of the amplitude threshold and the frequency threshold to adapt to more complex operating conditions. This dynamic adjustment mechanism effectively avoids false alarms or missed alarms caused by fixed thresholds.

[0065] Furthermore, the dynamically updating the harmonic warning threshold based on the real-time grid load state parameter includes:

[0066] Normalizing the real-time grid load state parameter into a grid load factor, wherein the grid load factor ranges from 0 to 1;

[0067] The grid load factor is multiplied by the initial amplitude threshold and the initial frequency threshold respectively to obtain the amplitude threshold and the frequency threshold.

[0068] Specifically, first, the load state parameters of the power grid are collected in real time. These parameters usually include key indicators such as load power and load fluctuation rate that reflect the operation status of the power grid. In order to facilitate subsequent calculations, these original parameters are standardized into a power grid load factor, the value of which is between [0,1]. The standardization process maps the current parameter value to the standardized range based on the maximum and minimum values ​​in the historical data. For example, when the power grid load power is close to the historical maximum value, the load factor is close to 1; when the load power is low, the load factor is close to 0. The purpose of standardization is to provide a unified, dimensionless indicator for dynamically adjusting the warning threshold. Then, according to the calculated power grid load factor, the initially set amplitude threshold and frequency threshold are dynamically adjusted. The amplitude threshold is used to measure whether the strength of the harmonic signal exceeds the normal range, while the frequency threshold is used to judge the degree of frequency deviation of the harmonic signal. The dynamic adjustment process is achieved by multiplying the load factor with the initial threshold. For example, assuming that the initial threshold of the amplitude of a certain order harmonic is 5% of the normal range, and the initial threshold of the frequency is ±0.5Hz, when the grid load factor is 0.8, the amplitude threshold will be adjusted to 4%, and the frequency threshold will be adjusted to ±0.4 Hz. This adjustment ensures that under high load conditions, the threshold range is appropriately expanded to reduce false alarms; under low load conditions, the threshold range is appropriately shrunk to enhance the sensitivity of detection. Through the above dynamic adjustment, the system can flexibly change the detection standard according to the real-time operation status of the grid, so as to more accurately capture abnormal harmonic signals. When the amplitude or frequency of the real-time detected harmonic signal exceeds the dynamically adjusted threshold range, the system immediately determines that the signal is an abnormal signal, classifies it as a specific abnormal type, and triggers an early warning operation.

[0069] Furthermore, the warning operation performs warning notification via voice broadcast and text message.

[0070] Specifically, when the system detects an abnormal harmonic signal and generates an abnormal harmonic report, the early warning operation is immediately triggered. The early warning notification content is based on the core information of the abnormal harmonic report, including the type of abnormal signal (such as abnormal amplitude or frequency), the degree of abnormality (deviation amplitude or frequency value), the abnormal location (such as a specific transformer sampling point), and the time of occurrence. After structured processing, these information are generated into two forms of notification: voice and text. The voice broadcast converts the report content into real-time voice reminders through text-to-speech technology, which is suitable for rapid response by on-site duty personnel. Text messages send complete abnormal information to designated recipients in the form of SMS, email or instant messaging for detailed records and remote notifications.

[0071] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A harmonic detection and early warning system for intelligent transformers, characterized in that: It includes a harmonic data acquisition module, a harmonic learning module, a harmonic anomaly classification module and a dynamic early warning module which are connected in sequence; The harmonic data acquisition module is used to collect the voltage signal and current signal of the transformer in real time, construct a time series signal set including harmonic current data and harmonic voltage data, and pre-process the time series signal set to obtain a harmonic data set; The harmonic learning module is used to decompose the signal using Fourier transform based on the preprocessed harmonic data set to generate a harmonic feature set, including the amplitude, phase and frequency characteristics of each order of harmonics; the harmonic feature set is input into the deep learning model to train and generate a dynamic harmonic feature model; The harmonic anomaly classification module is used to analyze the real-time harmonic feature set based on the trained dynamic harmonic feature model, detect abnormal harmonic signals and generate an abnormal harmonic report, wherein the abnormal harmonic report includes the abnormal type and location of the abnormal harmonic; The dynamic warning module is used to dynamically update the harmonic warning threshold based on the real-time grid load state parameters, and perform warning operations according to abnormal harmonic signals and the harmonic warning threshold. The warning operation outputs abnormal harmonic reports.

2. A harmonic detection and early warning system for intelligent transformers according to claim 1, characterized in that: The construction of a time series signal set containing harmonic current data and harmonic voltage data comprises the following steps: The collected voltage signal and current signal are synchronously corrected by a time synchronization algorithm to generate synchronized voltage signal and current signal; The synchronized voltage and current signals are filtered by a bandpass filtering algorithm to remove fundamental waves and high-frequency noise, and to extract characteristic signals within the harmonic frequency band; The characteristic signal in the harmonic frequency band is segmented and formatted into a time series signal set through segmentation processing.

3. The harmonic detection and early warning system for intelligent transformer according to claim 1 is characterized in that: The preprocessing is a normalization process.

4. The harmonic detection and early warning system for intelligent transformer according to claim 1 is characterized in that: Inputting the harmonic feature set into the deep learning model and training and generating a dynamic harmonic feature model comprises the following steps: Classify and label each feature sample in the harmonic feature set as normal harmonics and abnormal harmonics, and generate a standardized training data set; The training data set is input into a convolutional neural network model, which includes several convolutional layers, pooling layers and fully connected layers. The convolutional layers are used to extract local features of harmonic amplitude, phase and frequency, the pooling layers are used to reduce data dimension, and the fully connected layers are used to perform classification output; The convolutional neural network model is trained through the back-propagation algorithm, the classification error is evaluated using the cross-entropy loss function, the model parameters are updated through the Adam optimization algorithm, and the network weights are iteratively optimized.

5. A harmonic detection and early warning system for intelligent transformers according to claim 4, characterized in that: The formula of the convolutional neural network model is as follows: ； in, Output probability for the kth class; is the jth input feature in the input feature vector, including the amplitude, phase and frequency information of the harmonic; is the weight parameter from input feature j to hidden unit i in the hidden layer; is the weight parameter from hidden unit i to output category k in the output layer; and are the bias parameters of the hidden layer and the output layer respectively; is the hidden layer activation function; is the activation function of the output layer.

6. A harmonic detection and early warning system for intelligent transformers according to claim 4, characterized in that: The analysis of the real-time harmonic feature set, detection of abnormal harmonic signals and generation of an abnormal harmonic report comprises the following steps: The harmonic feature set collected in real time is input into the trained dynamic harmonic feature model to extract abnormal harmonic signals and abnormal types that deviate from the normal harmonic pattern; For the deviated abnormal harmonic signal, the position of the abnormal harmonic signal is obtained from the pre-associated position table, and an abnormal harmonic report is generated according to the abnormal harmonic signal position and abnormal type.

7. A harmonic detection and early warning system for intelligent transformers according to claim 6, characterized in that: The harmonic warning threshold includes an amplitude threshold and a frequency threshold.

8. The harmonic detection and early warning system for intelligent transformer according to claim 7, characterized in that: The dynamically updating the harmonic warning threshold based on the real-time grid load state parameter includes: Normalizing the real-time grid load state parameter into a grid load factor, wherein the grid load factor ranges from 0 to 1; The grid load factor is multiplied by the initial amplitude threshold and the initial frequency threshold respectively to obtain the amplitude threshold and the frequency threshold.

9. The harmonic detection and early warning system for intelligent transformer according to claim 1, characterized in that: The warning operation is performed through voice broadcast and text message.

Images