Substation remote fault warning system and method based on deep learning

Abstract

The present invention discloses a substation remote fault warning system and method based on deep learning, comprising the following steps: obtaining a standardized substation operation data set; constructing an improved variational autoencoder model; obtaining an optimized improved variational autoencoder model; using the optimized improved variational autoencoder model to perform multiple rounds of training on the standardized substation operation data set to form a highly sensitive and highly accurate anomaly detection model; collecting substation operation data in real time and inputting it into the anomaly detection model, analyzing the substation operation data in real time, and detecting anomaly patterns that deviate significantly from the normal substation operation state; when the anomaly detection model detects an anomaly pattern, the system automatically triggers a remote fault warning. The present invention further identifies typical patterns of temperature rise anomaly, voltage fluctuation anomaly, equipment degradation trend anomaly, and sudden composite anomaly based on the degree of deviation of the reconstruction error, thereby enhancing the interpretability and guidance of the warning.

Description

Technical Field

[0001] The present invention relates to the technical field of substations, and in particular to a substation remote fault early warning system and method based on deep learning. Background Art

[0002] With the continuous improvement of the intelligence level of power systems, substations, as the core nodes of power grid operation, their safe and stable operation is of great significance to the reliability of the entire power supply system. Substation remote status monitoring and fault warning, as an important part of the automated operation and maintenance of power systems, have gradually become the focus of academic research and engineering practice.

[0003] On the one hand, most existing technologies model substation operation data based on traditional statistical analysis methods or shallow machine learning algorithms. Existing methods usually rely on fixed rules or manually set feature engineering, and it is difficult to extract deep correlation patterns from high-dimensional, complex, and dynamically changing operation data. In scenarios where there is noise in the operation data, the data dimension is high, and the abnormal features are weak, the traditional methods' ability to identify abnormal conditions is significantly reduced, resulting in missed reports and false alarms in the early warning system.

[0004] On the other hand, although some studies have attempted to introduce deep learning models to perform feature modeling and anomaly detection on substation status, the training process of such methods is usually highly dependent on model hyperparameters. In actual deployment, due to the lack of an effective parameter optimization mechanism, the model is prone to fall into local optimality, resulting in poor adaptability in different substations or different working conditions.

[0005] To sum up, there is an urgent need for a substation remote fault warning method with strong feature extraction capability, parameter optimization capability and refined identification capability to improve the timeliness and accuracy of fault response. Summary of the Invention

[0006] One purpose of the present invention is to propose a remote fault warning system and method for substations based on deep learning. The present invention further identifies typical patterns of temperature rise anomalies, voltage fluctuation anomalies, equipment degradation trend anomalies and sudden composite anomalies according to the degree of deviation of the reconstruction error, thereby enhancing the interpretability and guidance of the warning.

[0007] A deep learning-based remote fault warning method for a substation according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect and preprocess substation operation data to obtain a standardized substation operation data set;

[0009] S2. Build an improved variational autoencoder model, which includes a device-state-sensitive encoder and an attention mechanism decoder, and generates a low-dimensional latent space representation and reconstructs the data.

[0010] S3. Calculate the reconstruction error evaluation index of the substation operation data based on the standardized substation operation dataset and the reconstructed data. Using the improved Firefly Optimization Algorithm with the reconstruction error evaluation index as the objective function, perform a global search and adaptive adjustment of the key hyperparameters of the improved variational autoencoder model to obtain the optimized improved variational autoencoder model.

[0011] S4. Utilize the optimized improved variational autoencoder model to conduct multiple rounds of training on a standardized substation operation dataset, forming a highly sensitive and accurate anomaly detection model. This anomaly detection model automatically learns and captures the implicit feature information and abnormal patterns in the substation operation data.

[0012] S5. Real-time collection of substation operation data and input into the anomaly detection model, which analyzes the substation operation data in real time and detects abnormal patterns that deviate significantly from normal substation operation status;

[0013] S6. When the anomaly detection model detects an abnormal pattern, the system automatically triggers a remote fault warning and sends the abnormal feature data and substation status indicator information to the operation and maintenance platform.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Set the substation operation data collection time window, collect the substation's multi-dimensional operation parameter data within the time window, and form the original substation operation data set :

[0016] ;

[0017] in, Indicates the Substation operation data samples, Indicates the total number of substation operation data samples, is the timestamp, is the current parameter, is the voltage parameter, is the temperature parameter, is the oil level parameter;

[0018] S12. Delete missing items from the original substation operation data set, remove abnormal or discontinuous data records with timestamps, merge duplicate records, and reorder them to generate a cleaned substation operation data set;

[0019] S13. Calculate the mean and standard deviation of the current parameters, voltage parameters, temperature parameters, and oil level parameters of the cleaned substation operation data set, calculate the abnormal judgment interval for each parameter, and consider parameters outside the abnormal judgment interval as outliers and remove them;

[0020] S14. Perform noise suppression on the substation operation dataset after outliers have been removed. Smooth each parameter type using a sliding average method, replacing the parameter value at each moment with the mean value within the sliding window, to obtain a noise-reduced substation operation dataset.

[0021] S15. Normalize the substation operation data set after noise reduction, obtain the minimum and maximum values ​​of current parameters, voltage parameters, temperature parameters and oil level parameters, and linearly map each parameter value to the interval , unify the dimensions and scales of different parameters, and output a standardized substation operation data set after normalization .

[0022] Optionally, the S2 includes the following steps:

[0023] S21. Based on the standardized substation operation data set , an improved variational autoencoder model integrating device state sensitivity factors is constructed, wherein the improved variational autoencoder model includes two parts: a device state sensitive encoder and an attention mechanism decoder.

[0024] S22. In the device state sensitive encoder, by introducing the device state sensitivity factor Adaptively weight the latent variable distribution of any sample in the standardized substation operation dataset to obtain a low-dimensional latent space distribution that fits the substation fault characteristics. :

[0025] ;

[0026] in, is the equipment state sensitivity factor, which is determined by the short-term fluctuation variance and long-term trend offset of the standardized substation operation data. When the standardized substation operation data shows an abnormal trend, the equipment state sensitivity factor increases adaptively, enhancing the sensitivity of the improved variational autoencoder model to weak abnormal patterns. and They represent the mean vector and standard deviation vector of the latent space distribution of the standardized substation operation data after being mapped by the equipment state sensitive encoder, For the Standardized substation operation data samples, is a state-sensitive latent distribution used to model the The potential spatial probability of a substation sample after considering the impact of abnormal state, is a Gaussian distribution with a mean and variance after device state modulation, Represents a vector element-wise multiplication operation.

[0027] S23. Low-dimensional latent space distribution based on device state-sensitive encoder output , using the reparameter sampling method to obtain the latent variable sampling value :

[0028] ;

[0029] in, is a random variable modulated by the device state sensitivity factor. Through the sampling process after the device state sensitivity factor is modulated, the latent variable sampling value is Capture and characterize subtle abnormal information in standardized substation operation data.

[0030] S24. Using latent variable sampling values ​​in the attention mechanism decoder Generate and reconstruct standardized substation operation data samples , the attention mechanism decoder adopts the adaptive attention weight matrix Realize the dynamic contribution adjustment of different potential variable sampling value dimensions to the substation operation data feature reconstruction process:

[0031] ;

[0032] in, For the reconstructed standardized substation operation data samples, It is the output function of the attention mechanism decoder, which is used to restore the modulated latent variables to the multi-dimensional parameter prediction values ​​of the substation status, reflecting the ability to actively focus on abnormal changes in specific dimensions.

[0033] S25. Combine the device state sensitive encoder and the attention mechanism decoder for end-to-end training. The training goal is to minimize the loss function of the improved variational autoencoder model that integrates the device state sensitivity factor. :

[0034] ;

[0035] in, For the The mean square error between the standardized data of a substation and its reconstructed value is used to measure the ability of the improved variational autoencoder model to reconstruct the current substation status. It is a direct basis for judging deviation from normal in fault warning. is the KL divergence term, is the balance coefficient, is the number of standardized substation operation data samples.

[0036] Optionally, S3 includes the following steps:

[0037] S31. Based on standardized substation operation data set and reconstruction of standardized substation operation data sets , calculate the reconstruction error evaluation index between each standardized substation operation data sample and its corresponding reconstructed data sample :

[0038] ;

[0039] in, For the The reconstruction error evaluation index of the standardized substation operation data samples is used to characterize the abnormality degree of the substation operation data.

[0040] S32. Construct an overall reconstruction error evaluation index based on the reconstruction error evaluation index of all standardized substation operation data samples , used to comprehensively evaluate the improved variational autoencoder model's ability to perceive abnormalities in substation status:

[0041] .

[0042] S33. An improved firefly optimization algorithm integrating dynamic light intensity modulation and adaptive step size update is proposed. The improved firefly optimization algorithm uses the overall reconstruction error evaluation index As the optimization objective function, the key hyperparameters in the improved variational autoencoder model of the fusion device state sensitivity factor are globally searched and adaptively optimized. The improved firefly optimization algorithm uses the following update rule to update the first Firefly position vector :

[0043] ;

[0044] in, For the The first iteration The position vector of each firefly represents the key hyperparameter set of the improved variational autoencoder model to be optimized. For the The first iteration The position vector of each firefly, For the The first iteration The position vector of each firefly represents the key hyperparameter set of the optimized improved variational autoencoder model. is a random perturbation vector that obeys the standard Gaussian distribution, is the dynamic light intensity attraction factor, reflecting the In the iteration Firefly The attraction strength of a firefly:

[0045] ;

[0046] in, is the initial light intensity attraction factor, is the light intensity attenuation factor, is the abnormal sensitivity enhancement coefficient, For the Firefly 、 The Euclidean distance between them in the hyperparameter space, 、 Firefly 、 In the The overall reconstruction error evaluation index corresponding to the iteration; The adaptive step coefficient driven by the substation status is dynamically adjusted by the overall reconstruction error evaluation index:

[0047] ;

[0048] in, is the basic step length, is the step size adaptive adjustment coefficient, The target overall reconstruction error evaluation index is used. When the overall reconstruction error evaluation index caused by the firefly corresponding position parameters deviates from the set threshold, the step size is adaptively increased to enhance the parameter space search capability.

[0049] S34. Repeatedly execute the improved firefly optimization algorithm, when the overall reconstruction error evaluation index When the set convergence threshold is reached or the maximum number of iterations is reached, the optimal hyperparameter configuration of the improved variational autoencoder model that optimizes the fusion device state sensitivity factor is obtained. .

[0050] Optionally, the S4 includes the following steps:

[0051] S41. Input the standardized substation operation dataset into the optimized improved variational autoencoder model constructed based on the optimal hyperparameter configuration, and perform multiple rounds of training, continuously iterating the model parameters during the training process;

[0052] S42. During each round of training, the improved variational autoencoder model extracts a low-dimensional latent space representation of the standardized substation operating data using a device-state-sensitive encoder. The data is then restored using an attention-based decoder. The reconstruction error (mean squared error) between the original and reconstructed data is calculated for each sample. The Kullback-Leibler divergence between the posterior latent distribution and the prior distribution of each sample is also calculated to constrain the distribution structure of the latent space.

[0053] S43. During multiple rounds of training iterations, record the reconstruction error change sequence of each sample in each round of training, and construct the reconstruction error trajectory from the reconstruction error change sequence to dynamically analyze and improve the stability and convergence of the variational autoencoder model during training;

[0054] S44. After training is complete, calculate the mean and variance of the final reconstruction error for all standardized substation operation data samples and construct a baseline distribution of the reconstruction error. This baseline distribution is used to characterize the range of the model's ability to reconstruct substation data under normal conditions.

[0055] S45. An anomaly detection model is constructed based on the trained optimized improved variational autoencoder model. The anomaly detection model uses whether the reconstruction error of the input sample significantly deviates from the baseline distribution as the judgment standard: when the actual reconstruction error of a standardized substation operation data sample exceeds the mean of the reconstruction error baseline distribution plus a certain multiple of the standard deviation, the sample is judged to be abnormal; among which, the multiple parameter in the anomaly detection threshold is the sensitivity coefficient, which is used to control the tolerance range of the early warning system to the degree of abnormal deviation, and to balance the early warning accuracy and false alarm rate under different substation operating environments.

[0056] Optionally, the S5 includes the following steps:

[0057] S51. Based on the deployed real-time monitoring system, substation operation data is collected and preprocessed. The preprocessed real-time standardized substation operation data is input into the trained anomaly detection model. The anomaly detection model compares the reconstruction error of the current sample with the historical normal state error baseline to determine whether the current substation operation status has significantly deviated.

[0058] S52. If the reconstruction error of the current sample is greater than the upper limit of the normal interval of the error baseline, the abnormal state recognition mechanism is triggered and the abnormal mode and abnormal category judgment results are output;

[0059] S53. The substation operation data sample number, reconstruction error value, abnormal parameter type and corresponding classification result judged as abnormal mode are combined into an abnormal detection report and cached in the result module. At the same time, the abnormal detection result is transmitted to the remote operation and maintenance platform.

[0060] Optionally, the abnormal pattern classification rules are as follows:

[0061] Temperature rise abnormality mode: When the real-time temperature parameter reconstruction error exceeds twice the error baseline mean, and the current parameter is simultaneously higher than the rising trend threshold, it is judged as a potential heating abnormality or local load abnormality during substation operation;

[0062] Abnormal voltage fluctuation mode: When the voltage parameter reconstruction error exceeds the upper limit of the error baseline for three consecutive time windows and the oil level parameter remains stable, it is judged as line fluctuation, voltage instability or bus oscillation abnormality;

[0063] Abnormal equipment degradation trend mode: When the reconstruction errors of all parameters are higher than the floating state threshold and remain higher than the floating state threshold for more than five consecutive time windows, it is judged that the performance of key substation equipment is degrading or hidden aging;

[0064] Sudden composite abnormality mode: When the reconstruction errors of current, temperature, and oil level parameters simultaneously exceed three times the baseline standard deviation, it is judged as a sudden composite abnormality event in the substation system.

[0065] Optionally, the sudden composite abnormal event of the substation system includes a local fault, an abnormal load shock or a cooling system failure.

[0066] A substation remote fault warning system based on deep learning is used to implement a substation remote fault warning method based on deep learning, including the following modules:

[0067] A data acquisition module is used to collect substation operation data in real time based on a set time window. The substation operation data includes current parameters, voltage parameters, temperature parameters and oil level parameters, and upload the collected results to the data preprocessing module;

[0068] A data preprocessing module is used to perform missing value elimination, timestamp correction, duplicate record merging, sliding average noise reduction and minimum-maximum normalization on the substation operation data input by the data acquisition module, and output a standardized substation operation data set;

[0069] An improved variational autoencoder modeling module is used to build an improved variational autoencoder model that incorporates device state sensitivity factors based on a standardized substation operation dataset. The model consists of an equipment state-sensitive encoder and an attention mechanism decoder, supporting nonlinear dimensionality reduction, latent space modeling, and output reconstruction of input data.

[0070] The reconstruction error evaluation module is used to calculate the error between the reconstructed samples generated by the improved variational autoencoder model and the input samples, and to construct an overall reconstruction error evaluation index based on the error results of all samples, which is used as the basis for optimization;

[0071] An improved Firefly Optimization module uses the overall reconstruction error evaluation index as the optimization target. It uses an improved Firefly Optimization algorithm that integrates dynamic light intensity modulation and a substation state-driven step size mechanism to perform global search and adaptive adjustment of key hyperparameters of the improved variational autoencoder model to obtain the optimal hyperparameter configuration.

[0072] The anomaly detection model training module is used to perform multiple rounds of training on the improved variational autoencoder model under the optimal hyperparameter configuration, build an anomaly detection model with high sensitivity and high generalization ability, and establish an anomaly recognition baseline distribution based on the model's final reconstruction error;

[0073] The real-time fault detection module receives standardized substation operation data collected in real time, calls the anomaly detection model to determine whether the current data deviates significantly from the normal operating state, and outputs the anomaly label, anomaly parameter type, and anomaly degree.

[0074] The abnormal result output module is used to generate a structured abnormality report based on the abnormal data, abnormality type and judgment basis identified by the fault detection module, and transmit it to the substation operation and maintenance center through remote communication.

[0075] The beneficial effects of the present invention are:

[0076] (1) The present invention introduces a device state sensitivity factor into the encoder, dynamically adjusts the distribution parameters of the latent variables according to the short-term fluctuations and long-term trend changes of the substation operation data, further regulates the projection method of the sample in the latent space during the re-parameter sampling stage, and adaptively adjusts the feature response weights of each dimension in the reconstruction process through the attention mechanism decoder. Compared with the traditional fixed-structure VAE model, the present invention enables the model to respond to small but critical state changes at the early stage of anomalies, thereby enhancing the model's ability to recognize local anomalies.

[0077] (2) The present invention designs a dynamic light intensity function combined with the reconstruction error evaluation index, introduces the substation state perception enhancement factor to regulate the attraction relationship between different individuals, and dynamically adjusts the individual search step size according to the degree of deviation between the overall reconstruction error and the target error in model training, thereby realizing flexible switching between global search and local refinement, and effectively avoiding falling into the local optimum.

[0078] (3) The present invention no longer relies on a single threshold judgment in the anomaly detection process. Instead, it constructs a complete reconstruction error baseline distribution through the training stage, and constructs a multi-level anomaly detection strategy in the form of standard deviation weighting in real-time data analysis. According to the degree of deviation of the reconstruction error, it further identifies the typical patterns of temperature rise anomaly, voltage fluctuation anomaly, equipment degradation trend anomaly and sudden composite anomaly, thereby enhancing the interpretability and guidance of the early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0080] Figure 1 This is a flowchart of a substation remote fault warning method based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0081] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0082] refer to Figure 1 , a remote fault warning method for substations based on deep learning, comprising the following steps:

[0083] S1. Collect and preprocess substation operation data to obtain a standardized substation operation data set;

[0084] S2. Build an improved variational autoencoder model, which includes a device-state-sensitive encoder and an attention mechanism decoder, and generates a low-dimensional latent space representation and reconstructs the data.

[0085] S3. Calculate the reconstruction error evaluation index of the substation operation data based on the standardized substation operation dataset and the reconstructed data. Using the improved Firefly Optimization Algorithm with the reconstruction error evaluation index as the objective function, perform a global search and adaptive adjustment of the key hyperparameters of the improved variational autoencoder model to obtain the optimized improved variational autoencoder model.

[0086] S4. Utilize the optimized improved variational autoencoder model to conduct multiple rounds of training on a standardized substation operation dataset, forming a highly sensitive and accurate anomaly detection model. This anomaly detection model automatically learns and captures the implicit feature information and abnormal patterns in the substation operation data.

[0087] S5. Real-time collection of substation operation data and input into the anomaly detection model, which analyzes the substation operation data in real time and detects abnormal patterns that deviate significantly from normal substation operation status;

[0088] S6. When the anomaly detection model detects an abnormal pattern, the system automatically triggers a remote fault warning and sends the abnormal feature data and substation status indicator information to the operation and maintenance platform.

[0089] In this embodiment, S1 includes the following steps:

[0090] S11. Set the substation operation data collection time window, collect the substation's multi-dimensional operation parameter data within the time window, and form the original substation operation data set :

[0091] ;

[0092] in, Indicates the Substation operation data samples, Indicates the total number of substation operation data samples, is the timestamp, is the current parameter, is the voltage parameter, is the temperature parameter, is the oil level parameter;

[0093] S12. Delete missing items from the original substation operation data set, remove abnormal or discontinuous data records with timestamps, merge duplicate records, and reorder them to generate a cleaned substation operation data set;

[0094] S13. Calculate the mean and standard deviation of the current parameters, voltage parameters, temperature parameters, and oil level parameters of the cleaned substation operation data set, calculate the abnormal judgment interval for each parameter, and consider parameters outside the abnormal judgment interval as outliers and remove them;

[0095] S14. Perform noise suppression on the substation operation dataset after outliers have been removed. Smooth each parameter type using a sliding average method, replacing the parameter value at each moment with the mean value within the sliding window, to obtain a noise-reduced substation operation dataset.

[0096] S15. Normalize the substation operation data set after noise reduction, obtain the minimum and maximum values ​​of current parameters, voltage parameters, temperature parameters and oil level parameters, and linearly map each parameter value to the interval , unify the dimensions and scales of different parameters, and output a standardized substation operation data set after normalization .

[0097] In this embodiment, S2 includes the following steps:

[0098] S21. Based on the standardized substation operation data set , an improved variational autoencoder model that integrates device state sensitivity factors is constructed. The improved variational autoencoder model includes two parts: a device state sensitive encoder and an attention mechanism decoder.

[0099] S22. In the device state sensitive encoder, by introducing the device state sensitivity factor Adaptively weight the latent variable distribution of any sample in the standardized substation operation dataset to obtain a low-dimensional latent space distribution that fits the substation fault characteristics. :

[0100] ;

[0101] in, is the equipment state sensitivity factor, which is determined by the short-term fluctuation variance and long-term trend offset of the standardized substation operation data. When the standardized substation operation data shows an abnormal trend, the equipment state sensitivity factor increases adaptively, enhancing the sensitivity of the improved variational autoencoder model to weak abnormal patterns. and They represent the mean vector and standard deviation vector of the latent space distribution of the standardized substation operation data after being mapped by the equipment state sensitive encoder, For the Standardized substation operation data samples, is a state-sensitive latent distribution used to model the The potential spatial probability of a substation sample after considering the impact of abnormal state, is a Gaussian distribution with a mean and variance after device state modulation, Represents a vector element-wise multiplication operation.

[0102] S23. Low-dimensional latent space distribution based on device state-sensitive encoder output , using the reparameter sampling method to obtain the latent variable sampling value :

[0103] ;

[0104] in, is a random variable modulated by the device state sensitivity factor. Through the sampling process after the device state sensitivity factor is modulated, the latent variable sampling value is Capture and characterize subtle abnormal information in standardized substation operation data.

[0105] S24. Using latent variable sampling values ​​in the attention mechanism decoder Generate and reconstruct standardized substation operation data samples , the attention mechanism decoder adopts the adaptive attention weight matrix Realize the dynamic contribution adjustment of different potential variable sampling value dimensions to the substation operation data feature reconstruction process:

[0106] ;

[0107] in, For the reconstructed standardized substation operation data samples, It is the output function of the attention mechanism decoder, which is used to restore the modulated latent variables to the multi-dimensional parameter prediction values ​​of the substation status, reflecting the ability to actively focus on abnormal changes in specific dimensions.

[0108] S25. Combine the device state sensitive encoder and the attention mechanism decoder for end-to-end training. The training goal is to minimize the loss function of the improved variational autoencoder model that integrates the device state sensitivity factor. :

[0109] ;

[0110] in, For the The mean square error between the standardized data of a substation and its reconstructed value is used to measure the ability of the improved variational autoencoder model to reconstruct the current substation status. It is a direct basis for judging deviation from normal in fault warning. is the KL divergence term, is the balance coefficient, is the number of standardized substation operation data samples.

[0111] In this embodiment, S3 includes the following steps:

[0112] S31. Based on standardized substation operation data set and reconstruction of standardized substation operation data sets , calculate the reconstruction error evaluation index between each standardized substation operation data sample and its corresponding reconstructed data sample :

[0113] ;

[0114] in, For the The reconstruction error evaluation index of the standardized substation operation data samples is used to characterize the abnormality degree of the substation operation data.

[0115] S32. Construct an overall reconstruction error evaluation index based on the reconstruction error evaluation index of all standardized substation operation data samples , used to comprehensively evaluate the improved variational autoencoder model's ability to perceive abnormalities in substation status:

[0116] .

[0117] S33. An improved firefly optimization algorithm integrating dynamic light intensity modulation and adaptive step size update is proposed. The improved firefly optimization algorithm uses the overall reconstruction error evaluation index As the optimization objective function, the key hyperparameters in the improved variational autoencoder model of the fusion device state sensitivity factor are globally searched and adaptively optimized. The improved firefly optimization algorithm uses the following update rule to update the first Firefly position vector :

[0118] ;

[0119] in, For the The first iteration The position vector of each firefly represents the key hyperparameter set of the improved variational autoencoder model to be optimized. For the The first iteration The position vector of each firefly, For the The first iteration The position vector of each firefly represents the key hyperparameter set of the optimized improved variational autoencoder model. is a random perturbation vector that obeys the standard Gaussian distribution, is the dynamic light intensity attraction factor, reflecting the In the iteration Firefly The attraction strength of a firefly:

[0120] ;

[0121] in, is the initial light intensity attraction factor, is the light intensity attenuation factor, is the abnormal sensitivity enhancement coefficient, For the Firefly 、 The Euclidean distance between them in the hyperparameter space, 、 Firefly 、 In the The overall reconstruction error evaluation index corresponding to the iteration; The adaptive step coefficient driven by the substation status is dynamically adjusted by the overall reconstruction error evaluation index:

[0122] ;

[0123] in, is the basic step length, is the step size adaptive adjustment coefficient, The target overall reconstruction error evaluation index is used. When the overall reconstruction error evaluation index caused by the firefly corresponding position parameters deviates from the set threshold, the step size is adaptively increased to enhance the parameter space search capability.

[0124] S34. Repeatedly execute the improved firefly optimization algorithm, when the overall reconstruction error evaluation index When the set convergence threshold is reached or the maximum number of iterations is reached, the optimal hyperparameter configuration of the improved variational autoencoder model that optimizes the fusion device state sensitivity factor is obtained. .

[0125] In this embodiment, S4 includes the following steps:

[0126] S41. Input the standardized substation operation dataset into the optimized improved variational autoencoder model constructed based on the optimal hyperparameter configuration, and perform multiple rounds of training, continuously iterating the model parameters during the training process;

[0127] S42. During each round of training, the improved variational autoencoder model extracts a low-dimensional latent space representation of the standardized substation operating data using a device-state-sensitive encoder. The data is then restored using an attention-based decoder. The reconstruction error (mean squared error) between the original and reconstructed data is calculated for each sample. The Kullback-Leibler divergence between the posterior latent distribution and the prior distribution of each sample is also calculated to constrain the distribution structure of the latent space.

[0128] S43. During multiple rounds of training iterations, record the reconstruction error change sequence of each sample in each round of training, and construct the reconstruction error trajectory from the reconstruction error change sequence to dynamically analyze and improve the stability and convergence of the variational autoencoder model during training;

[0129] S44. After training is complete, calculate the mean and variance of the final reconstruction error for all standardized substation operation data samples and construct a baseline distribution of the reconstruction error. This baseline distribution is used to characterize the range of the model's ability to reconstruct substation data under normal conditions.

[0130] S45. An anomaly detection model is constructed based on the trained optimized improved variational autoencoder model. The anomaly detection model uses whether the reconstruction error of the input sample significantly deviates from the baseline distribution as the judgment standard: when the actual reconstruction error of a standardized substation operation data sample exceeds the mean of the reconstruction error baseline distribution plus a certain multiple of the standard deviation, the sample is judged to be abnormal; among which, the multiple parameter in the anomaly detection threshold is the sensitivity coefficient, which is used to control the tolerance range of the early warning system to the degree of abnormal deviation, and to balance the early warning accuracy and false alarm rate under different substation operating environments.

[0131] In this embodiment, S5 includes the following steps:

[0132] S51. Based on the deployed real-time monitoring system, substation operation data is collected and preprocessed. The preprocessed real-time standardized substation operation data is input into the trained anomaly detection model. The anomaly detection model compares the reconstruction error of the current sample with the historical normal state error baseline to determine whether the current substation operation status has significantly deviated.

[0133] S52. If the reconstruction error of the current sample is greater than the upper limit of the normal interval of the error baseline, the abnormal state recognition mechanism is triggered and the abnormal mode and abnormal category judgment results are output;

[0134] S53. The substation operation data sample number, reconstruction error value, abnormal parameter type and corresponding classification result judged as abnormal mode are combined into an abnormal detection report and cached in the result module. At the same time, the abnormal detection result is transmitted to the remote operation and maintenance platform.

[0135] In this embodiment, the abnormal pattern classification rules are as follows:

[0136] Temperature rise abnormality mode: When the real-time temperature parameter reconstruction error exceeds twice the error baseline mean, and the current parameter is simultaneously higher than the rising trend threshold, it is judged as a potential heating abnormality or local load abnormality during substation operation;

[0137] Abnormal voltage fluctuation mode: When the voltage parameter reconstruction error exceeds the upper limit of the error baseline for three consecutive time windows and the oil level parameter remains stable, it is judged as line fluctuation, voltage instability or bus oscillation abnormality;

[0138] Abnormal equipment degradation trend mode: When the reconstruction errors of all parameters are higher than the floating state threshold and remain higher than the floating state threshold for more than five consecutive time windows, it is judged that the performance of key substation equipment is degrading or hidden aging;

[0139] Sudden composite abnormality mode: When the reconstruction errors of current, temperature, and oil level parameters simultaneously exceed three times the baseline standard deviation, it is judged as a sudden composite abnormality event in the substation system.

[0140] In this embodiment, the sudden composite abnormal event of the substation system includes a local fault, an abnormal load shock, or a cooling system failure.

[0141] A substation remote fault warning system based on deep learning is used to implement a substation remote fault warning method based on deep learning, including the following modules:

[0142] The data acquisition module is used to collect substation operation data in real time based on a set time window. The substation operation data includes current parameters, voltage parameters, temperature parameters and oil level parameters, and upload the collection results to the data preprocessing module;

[0143] The data preprocessing module is used to perform missing value elimination, timestamp correction, duplicate record merging, sliding average noise reduction, and minimum-maximum normalization on the substation operation data input by the data acquisition module, and output a standardized substation operation data set;

[0144] An improved variational autoencoder modeling module is used to build an improved variational autoencoder model that incorporates equipment state sensitivity factors based on a standardized substation operation dataset. The model consists of an equipment state-sensitive encoder and an attention mechanism decoder, supporting nonlinear dimensionality reduction, latent space modeling, and reconstructed output of the input data.

[0145] The reconstruction error evaluation module is used to calculate the error between the reconstructed samples generated by the improved variational autoencoder model and the input samples, and to construct an overall reconstruction error evaluation index based on the error results of all samples, which is used as the basis for optimization;

[0146] An improved Firefly Optimization module uses the overall reconstruction error evaluation index as the optimization target. It uses an improved Firefly Optimization algorithm that integrates dynamic light intensity modulation and a substation state-driven step size mechanism to perform global search and adaptive adjustment of key hyperparameters of the improved variational autoencoder model to obtain the optimal hyperparameter configuration.

[0147] The anomaly detection model training module is used to perform multiple rounds of training on the improved variational autoencoder model under the optimal hyperparameter configuration, build an anomaly detection model with high sensitivity and high generalization ability, and establish an anomaly recognition baseline distribution based on the model's final reconstruction error;

[0148] The real-time fault detection module receives standardized substation operation data collected in real time, calls the anomaly detection model to determine whether the current data deviates significantly from the normal operating state, and outputs the anomaly label, anomaly parameter type, and anomaly degree.

[0149] The abnormal result output module is used to generate a structured abnormality report based on the abnormal data, abnormality type and judgment basis identified by the fault detection module, and transmit it to the substation operation and maintenance center through remote communication.

[0150] Example 1: The online operation monitoring platform of a substation received a remote warning signal for a "sudden temperature rise" from the anomaly detection model proposed in this invention. The system detected that the monitoring data of a main transformer (number: 220kV-TR2) showed a continuous slight upward fluctuation over the past three time periods (a total of 45 seconds). Its temperature reconstruction error jumped from 0.0032 to 0.0079, exceeding the upper limit of two standard deviations of the mean of the reconstruction error baseline distribution, and was accompanied by a slight increase in current. After automatic tracing and analysis, the system found that the abnormal data came from the transformer temperature monitoring sensor channel. The relevant real-time data is as follows:

[0151] Data collection time: 2023-09-18 02:11:30; Current: 295.3 A (within normal range); Voltage: 220.1 kV (stable); Temperature: 56.8°C (4.2°C higher than 10 minutes ago); Oil level: 648 mm (normal); System identification tag: Abnormal temperature rise; Abnormality score: 0.92 (maximum 1.0);

[0152] After the system identifies an anomaly, it immediately creates an anomaly report, which includes:

[0153] Abnormal type: Abnormal temperature rise; Abnormal location: 220kV-TR2 main transformer; Abnormal occurrence time: 2023-09-18 02:11:30; Abnormal duration: 45 seconds before the warning; Corresponding reconstruction error trend: Significant increase; Analysis conclusion: Suspected local load overheating or cooling system efficiency reduction;

[0154] The report is automatically pushed to the on-duty personnel terminal through the power grid dispatching center operation and maintenance platform within 5 seconds, and a text message alarm is simultaneously sent to the on-duty person's mobile phone. After receiving the alarm, the on-duty personnel remotely retrieve the video and infrared imaging system to confirm the temperature rise, and quickly notify the emergency repair team to check the oil pump cooling circuit.

[0155] On-site handling records show that operations and maintenance personnel arrived at the substation at 3:05 AM. On-site inspections confirmed an abnormality in one of the main transformer cooling fan circuits, causing a localized increase in oil temperature. The fan failed to respond to start and stop signals before 2:00 AM due to oxidation of the temperature control relay contacts, indicating an early hardware aging issue. Thanks to the model's high sensitivity in identifying early temperature anomalies and low-latency alarms, operations and maintenance personnel were able to complete the forced switching operation before the oil temperature continued to rise, preventing an overheating trip of the main transformer.

[0156] To verify the advantages of the system, the operation and maintenance center subsequently added this abnormal data to the historical data sample set and used the anomaly detection model proposed in this invention and traditional methods to perform recurrence detection. The total number of retest samples was 30,000, including 60 abnormal samples. The results are shown in Table 1 below:

[0157] Table 1 Data comparison between the anomaly detection model proposed in this paper and the traditional method

[0158]

[0159] The present invention introduces an equipment state sensitivity factor into the encoder, dynamically adjusts the distribution parameters of the latent variables according to the short-term fluctuations and long-term trend changes of the substation operation data, further regulates the projection method of the sample in the latent space during the re-parameter sampling stage, and adaptively adjusts the feature response weights of each dimension in the reconstruction process through the attention mechanism decoder. Compared with the traditional fixed-structure VAE model, the present invention enables the model to respond to small but critical state changes at the early stage of the anomaly, thereby enhancing the model's ability to recognize local anomalies.

[0160] The present invention designs a dynamic light intensity function combined with the reconstruction error evaluation index, introduces the substation state perception enhancement factor to regulate the attraction relationship between different individuals, and dynamically adjusts the individual search step size according to the degree of deviation between the overall reconstruction error and the target error in model training, thereby realizing flexible switching between global search and local refinement, and effectively avoiding falling into local optimality.

[0161] The present invention no longer relies on a single threshold judgment in the anomaly detection process. Instead, it constructs a complete reconstruction error baseline distribution through the training phase, and constructs a multi-level anomaly detection strategy in the form of standard deviation weighting in real-time data analysis. According to the degree of deviation of the reconstruction error, it further identifies typical patterns of temperature rise anomalies, voltage fluctuation anomalies, equipment degradation trend anomalies and sudden composite anomalies, thereby enhancing the interpretability and guidance of the early warning.

[0162] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A remote fault warning method for substations based on deep learning, characterized in that: The steps include: S1. Collect and preprocess substation operation data to obtain a standardized substation operation data set; S2. Build an improved variational autoencoder model, which includes a device-state-sensitive encoder and an attention mechanism decoder, and generates a low-dimensional latent space representation and reconstructs the data. S3. Calculate the reconstruction error evaluation index of the substation operation data based on the standardized substation operation dataset and the reconstructed data. Using the improved Firefly Optimization Algorithm with the reconstruction error evaluation index as the objective function, perform a global search and adaptive adjustment of the key hyperparameters of the improved variational autoencoder model to obtain the optimized improved variational autoencoder model. S4. Utilize the optimized improved variational autoencoder model to conduct multiple rounds of training on a standardized substation operation dataset, forming a highly sensitive and accurate anomaly detection model. This anomaly detection model automatically learns and captures the implicit feature information and abnormal patterns in the substation operation data. S5. Real-time collection of substation operation data and input into the anomaly detection model, which analyzes the substation operation data in real time and detects abnormal patterns that deviate significantly from normal substation operation status; S6. When the anomaly detection model detects an abnormal pattern, the system automatically triggers a remote fault warning and sends the abnormal characteristic data and substation status indicator information to the operation and maintenance platform; The S3 includes the following steps: S31. Based on the standardized substation operation data set D standard and reconstruction of standardized substation operation data sets Calculate the reconstruction error evaluation index e between each standardized substation operation data sample and its corresponding reconstructed data sample i ; S32. Based on the reconstruction error evaluation index of all standardized substation operation data samples, construct the overall reconstruction error evaluation index E total , used to comprehensively evaluate the improved variational autoencoder model’s ability to perceive abnormalities in substation status; S33. An improved firefly optimization algorithm integrating dynamic light intensity modulation and adaptive step size update is proposed. The improved firefly optimization algorithm uses the overall reconstruction error evaluation index E total As the optimization objective function, a global search and adaptive optimization adjustment are performed on the key hyperparameters in the improved variational autoencoder model that integrates the device state sensitivity factor. The improved firefly optimization algorithm uses the following update rule to update the position vector of the mth firefly in, is the position vector of the mth firefly at the tth iteration, representing the key hyperparameter set of the improved variational autoencoder model to be optimized, is the position vector of the nth firefly at the tth iteration, is the position vector of the mth firefly at the t+1th iteration, which represents the key hyperparameter set of the optimized improved variational autoencoder model. is a random perturbation vector that obeys the standard Gaussian distribution, is the dynamic light intensity attraction factor, which reflects the attraction intensity of the nth firefly to the mth firefly in the tth iteration: Among them, β0 is the initial light intensity attraction factor, λ is the light intensity attenuation factor, η is the abnormal sensitivity enhancement coefficient, is the Euclidean distance between fireflies m and n in the hyperparameter space at the tth iteration, are the overall reconstruction error evaluation indicators corresponding to fireflies m and n at the tth iteration; The adaptive step coefficient driven by the substation status is dynamically adjusted by the overall reconstruction error evaluation index: Among them, γ0 is the basic step size, ξ is the step size adaptive adjustment coefficient, E target As the target overall reconstruction error evaluation index, when the overall reconstruction error evaluation index caused by the firefly corresponding position parameters deviates from the set threshold, the step size is adaptively increased to enhance the parameter space search capability; S34. Repeatedly execute the improved firefly optimization algorithm, when the overall reconstruction error evaluation index E total When the set convergence threshold is reached or the maximum number of iterations is reached, the optimal hyperparameter configuration x of the improved variational autoencoder model of the optimized fusion device state sensitivity factor is obtained best .

2. A substation remote fault warning method based on deep learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Set the substation operation data collection time window, collect the substation's multi-dimensional operation parameter data within the time window, and form the original substation operation data set D raw ; S12. Delete missing items from the original substation operation data set, remove abnormal or discontinuous data records with timestamps, merge duplicate records, and reorder them to generate a cleaned substation operation data set; S13. Calculate the mean and standard deviation of the current parameters, voltage parameters, temperature parameters, and oil level parameters of the cleaned substation operation data set, calculate the abnormal judgment interval for each parameter, and consider parameters outside the abnormal judgment interval as outliers and remove them; S14. Perform noise suppression on the substation operation dataset after outliers have been removed. Smooth each parameter type using a sliding average method, replacing the parameter value at each moment with the mean value within the sliding window, to obtain a noise-reduced substation operation dataset. S15. Normalize the noise-reduced substation operation dataset to obtain the minimum and maximum values ​​of the current, voltage, temperature, and oil level parameters, and linearly map each parameter value to the interval [0, 1] to unify the dimensions and scales of different parameters. Output the normalized substation operation dataset D after normalization. standard .

3. A substation remote fault warning method based on deep learning according to claim 2, characterized in that: The S2 comprises the following steps: S21. Based on the standardized substation operation data set D standard , construct an improved variational autoencoder model that integrates device state sensitivity factors, the improved variational autoencoder model includes two parts: a device state sensitive encoder and an attention mechanism decoder; S22. In the device state sensitive encoder, by introducing the device state sensitivity factor α i Adaptively weight the latent variable distribution of any sample in the standardized substation operation dataset to obtain a low-dimensional latent space distribution that fits the substation fault characteristics. Among them, α i is the equipment state sensitivity factor, which is determined by the short-term fluctuation variance and long-term trend offset of the standardized substation operation data. When the standardized substation operation data shows an abnormal trend, the equipment state sensitivity factor increases adaptively, enhancing the sensitivity of the improved variational autoencoder model to weak abnormal patterns. i and σ i They represent the mean vector and standard deviation vector of the latent space distribution of the standardized substation operation data after being mapped by the equipment state sensitive encoder, is the i-th standardized substation operation data sample, is a state-sensitive latent distribution, which is used to model the potential spatial probability of the i-th substation sample after considering the influence of state anomalies. is a Gaussian distribution with the mean and variance after device state modulation, ⊙ represents the vector element-wise multiplication operation; S23. Low-dimensional latent space distribution based on device state-sensitive encoder output The latent variable sampling value z is obtained by using the reparameter sampling method i : z i =a i ⊙μ i +a i ⊙s i ⊙∈ i ; Among them, ∈ i is a random variable modulated by the device state sensitivity factor. Through the sampling process modulated by the device state sensitivity factor, the latent variable sampling value z i Capture and characterize subtle abnormal information in standardized substation operation data; S24. In the attention mechanism decoder, use the latent variable sample value z i Generate and reconstruct standardized substation operation data samples The attention mechanism decoder uses the adaptive attention weight matrix W att Realize the dynamic contribution adjustment of different potential variable sampling value dimensions to the substation operation data feature reconstruction process: in, is the i-th reconstructed standardized substation operation data sample, f θ (W att ⊙z i ) is the attention mechanism decoder output function, which is used to restore the modulated latent variables to the multi-dimensional parameter prediction values ​​of the substation status, reflecting the ability to actively focus on abnormal changes in specific dimensions; S25. Combine the device state sensitive encoder and the attention mechanism decoder for end-to-end training. The training goal is to minimize the loss function of the improved variational autoencoder model that integrates the device state sensitivity factor. in, is the mean square error between the standardized data of the i-th substation and its reconstructed value, which is used to measure the reconstruction ability of the improved variational autoencoder model on the current substation state and is the direct basis for judging the deviation from the normal state in fault warning. KL (·) is the KL divergence term, β is the balance coefficient, N ′ is the number of standardized substation operation data samples, and p(z) is the prior distribution of the latent variable z.

4. A substation remote fault warning method based on deep learning according to claim 3, characterized in that: The S4 comprises the following steps: S41. Input the standardized substation operation dataset into the optimized improved variational autoencoder model constructed based on the optimal hyperparameter configuration, and perform multiple rounds of training, continuously iterating the model parameters during the training process; S42. During each round of training, the improved variational autoencoder model extracts a low-dimensional latent space representation of the standardized substation operating data using a device-state-sensitive encoder. The data is then restored using an attention-based decoder. The reconstruction error (mean squared error) between the original and reconstructed data is calculated for each sample. The Kullback-Leibler divergence between the posterior latent distribution and the prior distribution of each sample is also calculated to constrain the distribution structure of the latent space. S43. During multiple rounds of training iterations, record the reconstruction error change sequence of each sample in each round of training, and construct the reconstruction error trajectory from the reconstruction error change sequence to dynamically analyze and improve the stability and convergence of the variational autoencoder model during training; S44. After training is complete, calculate the mean and variance of the final reconstruction error for all standardized substation operation data samples and construct a baseline distribution of the reconstruction error. This baseline distribution is used to characterize the range of the model's ability to reconstruct substation data under normal conditions. S45. An anomaly detection model is constructed based on the trained optimized improved variational autoencoder model. The anomaly detection model uses whether the reconstruction error of the input sample significantly deviates from the baseline distribution as the judgment standard: when the actual reconstruction error of a standardized substation operation data sample exceeds the mean of the reconstruction error baseline distribution plus a certain multiple of the standard deviation, the sample is judged to be abnormal; among which, the multiple parameter in the anomaly detection threshold is the sensitivity coefficient, which is used to control the tolerance range of the early warning system to the degree of abnormal deviation, and to balance the early warning accuracy and false alarm rate under different substation operating environments.

5. A substation remote fault warning method based on deep learning according to claim 4, characterized in that: The S5 comprises the following steps: S51. Based on the deployed real-time monitoring system, substation operation data is collected and preprocessed. The preprocessed real-time standardized substation operation data is input into the trained anomaly detection model. The anomaly detection model compares the reconstruction error of the current sample with the historical normal state error baseline to determine whether the current substation operation status has significantly deviated. S54. If the reconstruction error of the current sample is greater than the upper limit of the normal interval of the error baseline, the abnormal state recognition mechanism is triggered and the abnormal mode and abnormal category judgment results are output; S55. The substation operation data sample number, reconstruction error value, abnormal parameter type and corresponding classification result judged as abnormal mode are combined into an abnormal detection report and cached in the result module. At the same time, the abnormal detection result is transmitted to the remote operation and maintenance platform.

6. A substation remote fault warning method based on deep learning according to claim 5, characterized in that: The abnormal pattern classification rules are as follows: Temperature rise abnormality mode: When the real-time temperature parameter reconstruction error exceeds twice the error baseline mean, and the current parameter is simultaneously higher than the rising trend threshold, it is judged as a potential heating abnormality or local load abnormality during substation operation; Abnormal voltage fluctuation mode: When the voltage parameter reconstruction error exceeds the upper limit of the error baseline for three consecutive time windows and the oil level parameter remains stable, it is judged as line fluctuation, voltage instability or bus oscillation abnormality; Abnormal equipment degradation trend mode: When the reconstruction errors of all parameters are higher than the floating state threshold and remain higher than the floating state threshold for more than five consecutive time windows, it is judged that the performance of key substation equipment is degrading or hidden aging; Sudden composite abnormality mode: When the reconstruction errors of current, temperature, and oil level parameters simultaneously exceed three times the baseline standard deviation, it is judged as a sudden composite abnormality event in the substation system.

7. A substation remote fault warning method based on deep learning according to claim 6, characterized in that: The sudden composite abnormal events of the substation system include local failure, abnormal load shock or cooling system failure.

8. A substation remote fault early warning system based on deep learning, used to execute the substation remote fault early warning method based on deep learning according to any one of claims 1 to 7, characterized in that: Includes the following modules: A data acquisition module is used to collect substation operation data in real time based on a set time window. The substation operation data includes current parameters, voltage parameters, temperature parameters and oil level parameters, and upload the collected results to the data preprocessing module; A data preprocessing module is used to perform missing value elimination, timestamp correction, duplicate record merging, sliding average noise reduction and minimum-maximum normalization on the substation operation data input by the data acquisition module, and output a standardized substation operation data set; An improved variational autoencoder modeling module is used to build an improved variational autoencoder model that incorporates device state sensitivity factors based on a standardized substation operation dataset. The model consists of an equipment state-sensitive encoder and an attention mechanism decoder, supporting nonlinear dimensionality reduction, latent space modeling, and output reconstruction of input data. The reconstruction error evaluation module is used to calculate the error between the reconstructed samples generated by the improved variational autoencoder model and the input samples, and to construct an overall reconstruction error evaluation index based on the error results of all samples, which is used as the basis for optimization; An improved Firefly Optimization module uses the overall reconstruction error evaluation index as the optimization target. It uses an improved Firefly Optimization algorithm that integrates dynamic light intensity modulation and a substation state-driven step size mechanism to perform global search and adaptive adjustment of key hyperparameters of the improved variational autoencoder model to obtain the optimal hyperparameter configuration. The anomaly detection model training module is used to perform multiple rounds of training on the improved variational autoencoder model under the optimal hyperparameter configuration, build an anomaly detection model with high sensitivity and high generalization ability, and establish an anomaly recognition baseline distribution based on the model's final reconstruction error; The real-time fault detection module receives standardized substation operation data collected in real time, calls the anomaly detection model to determine whether the current data deviates significantly from the normal operating state, and outputs the anomaly label, anomaly parameter type, and anomaly degree. The abnormal result output module is used to generate a structured abnormality report based on the abnormal data, abnormality type and judgment basis identified by the fault detection module, and transmit it to the substation operation and maintenance center through remote communication.

Images