Substation equipment condition monitoring and intelligent fault warning method based on deep learning

Abstract

This invention discloses a method for substation equipment status monitoring and intelligent fault warning based on deep learning. The method includes the following steps: S1. obtaining a preprocessed multi-source status dataset; S2. generating a high-dimensional equipment status feature matrix; S3. forming a preliminary fault status identification result using a fault-sensitive deep belief network model; S4. obtaining an optimized sensitive deep belief network model; S5. using the optimized sensitive deep belief network model to perform online analysis on the multi-source status data collected in real time, generating real-time fault prediction results for the equipment status and classifying and grading the fault risks; and S6. triggering a remote fault warning mechanism based on the real-time fault prediction results, sending fault warning information containing the fault risk level, warning signal, and emergency response recommendations to the substation operation and maintenance center. This invention effectively supports the intelligent alarm linkage and hierarchical control of the dispatching system.

Description

Technical Field

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

[0002] As the level of intelligent power systems continues to increase, substations, as core nodes in the transmission and distribution network, are facing significant challenges. Their operational safety and equipment health have a significant impact on the overall stability of the power grid. As a key component of intelligent operations and maintenance systems, remote fault warning for substations plays a crucial role in monitoring equipment status, identifying potential faults, and predicting risk trends.

[0003] Currently, mainstream substation fault warning methods are mostly based on static threshold settings or empirical models. Their core approach is to set fixed upper and lower limits for monitoring parameters such as voltage, current, and temperature, triggering warnings once the values ​​exceed these limits. While these methods are simple to implement and offer a certain level of early fault response capability, they suffer from significant drawbacks. For one thing, static thresholds struggle to account for the normal fluctuations of different equipment under various operating conditions, leading to frequent false alarms and missed alarms. Furthermore, empirical models are unable to capture nonlinear fault evolution processes, making early warning effectiveness significantly insufficient, particularly when faced with complex system disturbances and hidden fault precursors.

[0004] In addition, some academic and engineering fields have attempted to introduce machine learning-based classification and recognition methods to predict and analyze substation failures. However, such methods generally rely on artificial feature construction and are difficult to adapt to the high-dimensional, multi-source, and strongly time-series coupled data characteristics of actual power systems. They have problems such as weak model generalization ability, complex parameter adjustment, and insufficient sensitivity to abnormal conditions, which limits their promotion and application in engineering practice.

[0005] Current mainstream deep learning applications focus on traditional convolutional neural networks or recurrent neural networks. While these networks perform well in image recognition and sequence prediction, they often encounter the multidimensional, low signal-to-noise ratio, and dynamically evolving data characteristics of substation equipment operation. This often leads to local optima due to inappropriate initial model parameters or rigid training structures, making dynamic adjustments difficult to adapt to the diverse operating environments of substations in different regions. Furthermore, existing research still primarily relies on manual trial-and-error or random initialization to adjust the parameters of deep learning models, lacking intelligent optimization mechanisms that integrate with real-time system operational feedback.

[0006] In summary, there is an urgent need for an innovative intelligent early warning method that can integrate multi-source status information, adaptively optimize learning structures, and take into account global search capabilities and fault-sensitive characteristics, so as to achieve early and accurate identification and graded response to equipment failures in complex power systems, thereby ensuring the safe and stable operation of substations and the overall power grid. Summary of the Invention

[0007] One purpose of the present invention is to propose a substation equipment status monitoring and intelligent fault warning method based on deep learning. The present invention can not only improve the response speed of fault handling, but also support the dynamic evolution tracking of risks across time periods, and effectively support the intelligent alarm linkage and hierarchical control of the dispatching system.

[0008] A method for substation equipment status monitoring and intelligent fault warning based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Real-time acquisition of multi-source state data sets and preprocessing to obtain preprocessed multi-source state data sets;

[0010] S2. Use the preprocessed multi-source state dataset to construct the device state feature vector. Use time series data segmentation and multi-scale sliding window methods to extract key features related to fault evolution and generate a high-dimensional device state feature matrix.

[0011] S3. A fault-sensitive deep belief network model performs multi-level unsupervised feature learning on the high-dimensional device state feature matrix to form preliminary fault state recognition results;

[0012] S4. Apply the jellyfish swarm optimization algorithm to perform global search and dynamic tuning of the weight and bias parameters of the deep belief network model. The deviation between the initial fault state recognition result and the predetermined recognition performance index is used as feedback information during the adjustment process to obtain an optimized sensitive deep belief network model.

[0013] S5. Use the optimized sensitive deep belief network model to perform online analysis on multi-source status data collected in real time, generate real-time fault prediction results for equipment status, and classify and grade fault risks;

[0014] S6. Based on the real-time fault prediction results, trigger the remote fault warning mechanism and send fault warning information including fault risk level, warning signal and emergency response suggestions to the substation operation and maintenance center.

[0015] Optionally, the S1 includes the following steps:

[0016] S11. Set the data collection cycle. In each collection cycle, voltage data is collected by voltage sensors, current sensors, temperature sensors and switch status collection devices deployed at key locations of the substation. , current data , temperature data And switch status data , construct the original multi-source state dataset :

[0017] ;

[0018] in, Indicates the i-th collection record, Indicates the timestamp of the i-th data collection, The number of samples in the set time window;

[0019] S12. Remove noise from the original multi-source state data set, use the moving average method to smooth the continuous variables, and compare the current value of each data point with its previous value. The adjacent data points are averaged as the denoising result of the data point to eliminate the influence of short-term spike noise on the data;

[0020] S13. Perform outlier detection and elimination on the original multi-source state data set after the preliminary smoothing process to determine the degree of deviation of a single point in the overall data distribution. The difference between the value of each data point and the mean of its dimension is divided by the standard deviation of that dimension to obtain a normalized value. When the normalized value is greater than the preset anomaly determination threshold, the point is considered an outlier and is eliminated.

[0021] S14. Perform missing value filling on the multi-source state data set after outliers have been eliminated. For continuous variables such as voltage data, current data, temperature data, and switch state data, use the average of the valid data points before and after the data point as the interpolation result to maintain the continuity and change trend of the data sequence in the multi-source state data set.

[0022] S15. Normalize the multi-source state data set with missing values, subtract the minimum value in each dimension from the data, and then divide it by the difference between the maximum and minimum values ​​of the dimension, and map the data of each dimension to the interval , eliminating the dimensional differences between different physical quantities;

[0023] S16. The multi-source state dataset after noise removal, outlier removal, missing value filling and normalization is defined as the preprocessed multi-source state dataset .

[0024] Optionally, the S2 includes the following steps:

[0025] S21. Preprocessed multi-source state dataset According to the collection time series Divide the segment into equal intervals and set the segment length to , divide the preprocessed multi-source state data set into several time series data sub-segments :

[0026] ;

[0027] in, , is the total time span of the preprocessed dataset, Indicates the Data subsegments within a time window;

[0028] S22. For each time series data segment Construct a multi-scale feature extraction structure based on sliding window, set the sliding window length set, and for each window length Perform sliding traversal operations to extract statistical features of voltage data, current data, temperature data, and switch status data in each sliding window, including the mean , standard deviation , slope and range ; Mean:

[0029] ;

[0030] Standard Deviation:

[0031] ;

[0032] Slope:

[0033] ;

[0034] Very bad:

[0035] ;

[0036] in, Indicates the Section 1 The first Single-dimensional data at a moment, is the mean of the sliding window index;

[0037] S23. For all time series data sub-segments and window length The extracted statistical features are integrated to generate a time period index and scale index A set of feature vectors for combined indexes ; Each eigenvector is represented as:

[0038] ;

[0039] in, 、 、 、 Represents the voltage at Paragraph, Section The mean, standard deviation, slope and range characteristics under the scale, and the current, temperature and switch status data of other dimensions are extracted in the same way;

[0040] S24. Collect all the feature vectors Splicing to form a high-dimensional device state feature matrix ,in is the number of time windows, is the sliding window scale number, The number of single-segment feature dimensions extracted at each scale.

[0041] Optionally, S3 includes the following steps:

[0042] S31. Calculate the fault sensitivity coefficient of each characteristic component based on the safety margin of each key parameter during substation operation to form a fault sensitivity coefficient matrix :

[0043] ;

[0044] in, express Middle Time period, The eigenvector of the scale, To reflect the preset threshold value of the safe operation status of the substation, is the failure risk amplification factor;

[0045] Calculated fault sensitivity coefficient matrix and high-dimensional equipment state characteristic matrix Multiply element by element to form a fault-sensitive input matrix ;

[0046] S32. Using the Fault-Sensitive Input Matrix Construct a fault-sensitive deep belief network model. The fault-sensitive deep belief network model adopts a multi-layer improved restricted Boltzmann machine stacking structure and is expressed as:

[0047] ;

[0048] in, Indicates the Layer-modified restricted Boltzmann machine, is the total number of layers of the fault-sensitive deep belief network model. In each layer of the fault-sensitive deep belief network model, the input vector is recorded as , for layer 1, , the hidden layer is represented as ;

[0049] S33. Improved restricted Boltzmann machine at each layer Inside, for the input vector and the latent vector The energy function integrating the fault sensitivity coefficient is used for modeling:

[0050] ;

[0051] in, Indicates the Tier input nodes, For the Layer The input node and The connection weights between hidden nodes, and are the input node and hidden node bias of this layer respectively, For the The fault sensitivity coefficient corresponding to the input node is taken from the fault sensitive input matrix After corresponding mapping, we get: is the number of neurons in the visible layer of layer l, is the number of neurons in the lth hidden layer, Represents the hidden layer vector of the lth layer;

[0052] S34. Using the contrast divergence algorithm, under the premise of ensuring that the fault anomaly characteristics in the multi-source status data of the substation are fully enhanced, the improved restricted Boltzmann machine of each layer is constructed. Parameters Perform unsupervised pre-training layer by layer and output each hidden layer As the input of the next layer, during the training process, the fault sensitivity coefficient Adjust the weight of each input node in the gradient update so that the fault-sensitive deep belief network model focuses on the characteristics of abnormal deviation from the safety threshold;

[0053] S35. After completing the unsupervised pre-training of all modified restricted Boltzmann machine layers, supervised fine-tuning of the entire fault-sensitive deep belief network model is performed, using the fault label vector generated by historical fault records. ,in, is the number of fault categories, and the cross entropy loss function is used to jointly optimize the parameters of the entire network to generate a fault feature expression vector ;

[0054] S36. Express the fault feature vector Pass it to the output layer of the fault-sensitive deep belief network model, use the fault mode classifier to perform forward reasoning, and output the preliminary fault state recognition result .

[0055] Optionally, the S4 includes the following steps:

[0056] S41. Set the connection weight parameters of each layer in the trained fault-sensitive deep belief network model and bias parameters , uniformly expanded into parameter vectors ,in, Indicates the jellyfish individual number, is the size of the jellyfish group, each jellyfish individual represents a set of potential optimal configurations of the fault-sensitive deep belief network model under the current working conditions of the substation;

[0057] S42. Perform forward reasoning on the multi-source state dataset using the fault-sensitive deep belief network model corresponding to the individual jellyfish positions to obtain preliminary fault identification output. and compare the initial fault identification output with the historical label Comparison, the introduction of substation key equipment weight coefficient Forming a device-weighted cross entropy loss function that integrates prediction performance and device importance :

[0058] ;

[0059] Among them, Q represents the number of devices participating in the evaluation, represents the critical weight of the qth device, c represents the total number of fault types that can be predicted by the model, represents the true i-th fault label of the q-th device in the training sample, Indicates individual jellyfish The predicted probability output of the deep belief network model represented by the representation for the i-th type of fault on the q-th device;

[0060] Device-weighted cross entropy loss function It not only reflects the accuracy of the model in identifying various types of faults, but also strengthens the focus on key equipment through the operation critical weights, ensuring that the optimization process does not ignore the components in the substation that are critical to system stability while improving the overall accuracy. It is more in line with the power engineering requirements for safety redundancy, equipment-level priority response, and fault criticality.

[0061] S43. Initialize the jellyfish population And set the maximum number of iterations , in each round t, combined with the dynamic operating weight of the current substation operating status , use the ocean current driven update formula to achieve global optimization:

[0062] ;

[0063] in, represents the position vector of the sth jellyfish individual in the tth iteration, Represents the updated position vector of the sth jellyfish individual in the t+1th iteration It represents the dynamic operating condition weight introduced for the substation operating status in the tth iteration, which is used to adjust the responsiveness of the jellyfish swarm’s global search behavior to environmental disturbances at different times. It represents the ocean current drift factor, which is used to control the extent to which the jellyfish individual approaches the current optimal individual and reflects the cluster movement speed of the jellyfish group. represents the position vector of the individual with the best fitness in the entire jellyfish group in the tth iteration;

[0064] S44. Introducing an active foraging behavior update mechanism that combines recognition error and fault level response to accelerate the adjustment of fault-sensitive deep belief network models with errors greater than a preset value:

[0065] ;

[0066] in, is the active search coefficient, For randomly selected individuals, For the The response weight of each jellyfish in the current round is used to guide the convergence direction to correct the fault misjudgment area;

[0067] S45. In each iteration, the current-driven and active foraging mechanisms are dynamically switched based on the individual fitness change trend and the performance difference of the global optimal model, and the current optimal model parameter vector is recorded. ;

[0068] S46. If the maximum number of iterations is reached or continuous The global optimal fitness improvement in the iteration is less than the convergence threshold , then terminate the iteration and obtain the optimal parameter vector Decoded into a set of fault-sensitive deep belief network model parameters , forming an optimized fault-sensitive deep belief network model .

[0069] Optionally, the S5 includes the following steps:

[0070] S51. Using the optimized fault-sensitive deep belief network model , the pre-processed multi-source state data set collected in real time Input into the model, perform forward reasoning calculations, and obtain each time point The equipment status prediction results under ,in, Indicates the Class failure at time point predicted probability of occurrence;

[0071] S52. Prediction results The predicted probabilities of various types of faults are sorted and the maximum value is determined, and the fault type corresponding to the maximum probability is extracted , the fault type As the main failure mode of the current equipment operation status, and record the predicted probability of the main failure mode ;

[0072] S53. Combine typical substation operating conditions, equipment risk levels, and historical fault impact data to establish fault risk classification rules;

[0073] S54. According to the fault type and predicted probability , determine the current time point The equipment failure risk level under the condition is calculated and the real-time failure prediction result label triple is output, including the prediction category, prediction probability and corresponding risk level. ,in, ;

[0074] S55. Sequence the prediction result labels in multiple consecutive time periods Perform time series fusion analysis to generate fault trend evolution curve.

[0075] Optionally, the fault risk classification rules are established according to the following fault risk level classification standards:

[0076] Low risk level: predicted probability , and the corresponding fault type Slight fluctuations include short-term current anomalies, low-amplitude temperature fluctuations, and single-point transient interference.

[0077] Medium risk level: predicted probability , and the corresponding fault type Belong to the medium warning category, including continuous equipment overtemperature, unplanned switching signals, and long-term voltage deviation from steady state;

[0078] High risk level: predicted probability , and the corresponding fault type Key anomalies involved include failure of circuit breaker operation, severe current fluctuations, and continuous temperature exceeding the limit.

[0079] The beneficial effects of the present invention are:

[0080] (1) The present invention introduces a fault sensitivity coefficient matrix in the feature learning stage. By constructing a fault risk amplification factor that is proportional to the degree of deviation of the voltage, current, temperature, and switch state operating parameters, the weight enhancement processing of key features is achieved. It is embedded in the energy function of the improved restricted Boltzmann machine. During the unsupervised pre-training process, the deep network structure is guided to focus on the abnormal signal area with the greatest potential for failure in the operation of substation equipment. Compared with the traditional deep belief network structure, it performs better in the extraction of low-amplitude fluctuations, multi-dimensional data coupling, and time-series mixed faults. Under the condition of multi-source information fusion, the accuracy of fault pattern recognition and the ability to extract precursors are greatly improved.

[0081] (2) The present invention introduces the jellyfish swarm optimization algorithm into the deep model parameter tuning process, and introduces a dynamic operating condition weight function and a behavior selection function based on the prediction error response in view of the characteristics of the variable operating conditions of the substation and the large differences in the operating areas. It constructs a hybrid optimization update strategy that integrates ocean current drive and active foraging mechanism. It can dynamically adjust the optimization direction and amplitude according to the real-time status, prediction deviation and response speed of the equipment, so that the model parameters continue to converge to the optimal solution under the current operating state, significantly reducing the risk of overfitting and improving the model stability and generalization ability.

[0082] (3) In the real-time reasoning stage, the present invention combines the output probability of the optimized deep model with the criticality of equipment operation to construct a risk classification standard based on the weighting of equipment importance, and realizes the multi-level fault risk division of low risk, medium risk and high risk. By serializing the prediction results of multiple consecutive moments, the fault evolution trend and level change trend curve are further output, providing power operation and maintenance personnel with a more forward-looking and executable decision-making basis. It can not only improve the response speed of fault handling, but also support the dynamic evolution tracking of risks across time periods, and effectively support the intelligent alarm linkage and hierarchical control of the dispatching system. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] 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:

[0084] Figure 1 This is a flow chart of a substation equipment status monitoring and intelligent fault warning method based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0085] 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.

[0086] refer to Figure 1, a substation equipment status monitoring and intelligent fault early warning method based on deep learning, including the following steps:

[0087] S1. Real-time acquisition of multi-source state data sets and preprocessing to obtain preprocessed multi-source state data sets;

[0088] S2. Use the preprocessed multi-source state dataset to construct the device state feature vector. Use time series data segmentation and multi-scale sliding window methods to extract key features related to fault evolution and generate a high-dimensional device state feature matrix.

[0089] S3. A fault-sensitive deep belief network model performs multi-level unsupervised feature learning on the high-dimensional device state feature matrix to form preliminary fault state recognition results;

[0090] S4. Apply the jellyfish swarm optimization algorithm to perform global search and dynamic tuning of the weight and bias parameters of the deep belief network model. The deviation between the initial fault state recognition result and the predetermined recognition performance index is used as feedback information during the adjustment process to obtain an optimized sensitive deep belief network model.

[0091] S5. Use the optimized sensitive deep belief network model to perform online analysis on multi-source status data collected in real time, generate real-time fault prediction results for equipment status, and classify and grade fault risks;

[0092] S6. Based on the real-time fault prediction results, trigger the remote fault warning mechanism and send fault warning information including fault risk level, warning signal and emergency response suggestions to the substation operation and maintenance center.

[0093] In this embodiment, S1 includes the following steps:

[0094] S11. Set the data collection cycle. In each collection cycle, voltage data is collected by voltage sensors, current sensors, temperature sensors and switch status collection devices deployed at key locations of the substation. , current data , temperature data And switch status data , construct the original multi-source state dataset :

[0095]

[0096] in, Indicates the i-th collection record, Indicates the timestamp of the i-th data collection, The number of samples in the set time window;

[0097] S12. Remove noise from the original multi-source state data set, use the moving average method to smooth the continuous variables, and compare the current value of each data point with its previous value. The adjacent data points are averaged as the denoising result of the data point to eliminate the influence of short-time spike noise on the data;

[0098] S13. Perform outlier detection and elimination on the original multi-source state data set after the preliminary smoothing process to determine the degree of deviation of a single point in the overall data distribution. The difference between the value of each data point and the mean of its dimension is divided by the standard deviation of that dimension to obtain a normalized value. When the normalized value is greater than the preset anomaly determination threshold, the point is considered an outlier and is eliminated.

[0099] S14. Perform missing value filling on the multi-source state data set after outliers have been eliminated. For continuous variables such as voltage data, current data, temperature data, and switch state data, use the average of the valid data points before and after the data point as the interpolation result to maintain the continuity and change trend of the data sequence in the multi-source state data set.

[0100] S15. Normalize the multi-source state data set with missing values, subtract the minimum value in each dimension from the data, and then divide it by the difference between the maximum and minimum values ​​of the dimension, and map the data of each dimension to the interval , eliminating the dimensional differences between different physical quantities;

[0101] S16. The multi-source state dataset after noise removal, outlier removal, missing value filling and normalization is defined as the preprocessed multi-source state dataset .

[0102] In this embodiment, S2 includes the following steps:

[0103] S21. Preprocessed multi-source state dataset According to the collection time series Divide the segment into equal intervals and set the segment length to , divide the preprocessed multi-source state data set into several time series data sub-segments :

[0104] ;

[0105] in, , is the total time span of the preprocessed dataset, Indicates the Data subsegments within a time window;

[0106] S22. For each time series data segment Construct a multi-scale feature extraction structure based on sliding window, set the sliding window length set, and for each window length Perform sliding traversal operations to extract statistical features of voltage data, current data, temperature data, and switch status data in each sliding window, including the mean , standard deviation , slope and range ; Mean:

[0107] ;

[0108] Standard Deviation:

[0109] ;

[0110] Slope:

[0111] ;

[0112] Very bad:

[0113] ;

[0114] in, Indicates the Section 1 The first Single-dimensional data at a moment, is the mean of the sliding window index;

[0115] S23. For all time series data sub-segments and window length The extracted statistical features are integrated to generate a time period index and scale index A set of feature vectors for combined indexes ; Each eigenvector is represented as:

[0116] ;

[0117] in, 、 、 、 Respectively represent the voltage in Paragraph, Section The mean, standard deviation, slope and range characteristics under the scale, and the current, temperature and switch status data of other dimensions are extracted in the same way;

[0118] S24. Collect all the feature vectors Splicing to form a high-dimensional device state feature matrix ,in is the number of time windows, is the sliding window scale number, The number of single-segment feature dimensions extracted at each scale.

[0119] In this embodiment, S3 includes the following steps:

[0120] S31. Calculate the fault sensitivity coefficient of each characteristic component based on the safety margin of each key parameter during substation operation to form a fault sensitivity coefficient matrix :

[0121] ;

[0122] in, express Middle Time period, The eigenvector of the scale, To reflect the preset threshold value of the safe operation status of the substation, is the failure risk amplification factor;

[0123] Calculated fault sensitivity coefficient matrix and high-dimensional equipment state characteristic matrix Multiply element by element to form a fault-sensitive input matrix ;

[0124] S32. Using the Fault-Sensitive Input Matrix Construct a fault-sensitive deep belief network model. The fault-sensitive deep belief network model adopts a multi-layer improved restricted Boltzmann machine stacking structure and is expressed as:

[0125] ;

[0126] in, Indicates the Layer-modified restricted Boltzmann machine, is the total number of layers of the fault-sensitive deep belief network model. In each layer of the fault-sensitive deep belief network model, the input vector is recorded as , for layer 1, , the hidden layer is represented as ;

[0127] The construction process of the improved restricted Boltzmann machine includes:

[0128] At the structural level, the improved restricted Boltzmann machine retains the basic structure of the traditional restricted Boltzmann machine, that is, a two-layer symmetric network consisting of a visible layer and a hidden layer, with full bidirectional connections between the two layers and no connections between nodes within the layer;

[0129] At the input feature processing level, the improved restricted Boltzmann machine introduces a fault sensitivity coefficient control mechanism in the visible layer node, that is, assigning weights to the features of each dimension in the input vector. The weights come from the fault sensitivity matrix , used to amplify characteristic signals significantly related to abnormal equipment conditions and suppress weak response inputs under background noise or normal fluctuations;

[0130] At the energy function modeling level, the improved restricted Boltzmann machine adjusts the energy function of the traditional restricted Boltzmann machine and embeds the fault sensitivity coefficient , forming a fault-guided energy expression, which enables the model to automatically focus on the data components that deviate from the normal operating state during the learning process;

[0131] In terms of training mechanism, the improved restricted Boltzmann machine retains contrastive divergence as a training method. However, during the gradient update process, the fault sensitivity coefficient is introduced into the gradient expression of error backpropagation, so that the model parameter adjustment prioritizes the feature dimensions that are amplified when the device is in a potential fault state.

[0132] At the inter-layer transmission mechanism level, the improved restricted Boltzmann machine supports its output implicit vector As the next layer Input ,ensure that the layer-by-layer feature expression is multi-level abstracted and aggregated towards the fault-related direction;

[0133] The improved restricted Boltzmann machine introduces a fault-sensitive control mechanism at the input end, embeds feature weighted expression in energy modeling, and guides parameters to focus on optimizing key feature dimensions during training. This enables the entire network to have stronger perception and representation capabilities for fault precursor data in substations, forming a core feature learning unit for remote fault warning tasks.

[0134] S33. Improved restricted Boltzmann machine at each layer Inside, for the input vector and the latent vector The energy function integrating the fault sensitivity coefficient is used for modeling:

[0135] ;

[0136] in, Indicates the Tier input nodes, For the Layer The input node and The connection weights between hidden nodes, and are the input node and hidden node bias of this layer respectively, For the The fault sensitivity coefficient corresponding to the input node is taken from the fault sensitive input matrix After corresponding mapping, we get: is the number of neurons in the visible layer of layer l, is the number of neurons in the lth hidden layer, Represents the hidden layer vector of the lth layer;

[0137] S34. Using the contrast divergence algorithm, under the premise of ensuring that the fault anomaly characteristics in the multi-source status data of the substation are fully enhanced, the improved restricted Boltzmann machine of each layer is constructed. Parameters Perform unsupervised pre-training layer by layer and output each hidden layer As the input of the next layer, during the training process, the fault sensitivity coefficient Adjust the weight of each input node in the gradient update so that the fault-sensitive deep belief network model focuses on the characteristics of abnormal deviation from the safety threshold;

[0138] S35. After completing the unsupervised pre-training of all modified restricted Boltzmann machine layers, supervised fine-tuning of the entire fault-sensitive deep belief network model is performed, using the fault label vector generated by historical fault records. ,in, is the number of fault categories, and the cross entropy loss function is used to jointly optimize the parameters of the entire network to generate a fault feature expression vector ;

[0139] S36. Express the fault feature vector Pass it to the output layer of the fault-sensitive deep belief network model, use the fault mode classifier to perform forward reasoning, and output the preliminary fault state recognition result .

[0140] In this embodiment, S4 includes the following steps:

[0141] S41. Set the connection weight parameters of each layer in the trained fault-sensitive deep belief network model and bias parameters , uniformly expanded into parameter vectors ,in, Indicates the jellyfish individual number, is the size of the jellyfish group, each jellyfish individual represents a set of potential optimal configurations of the fault-sensitive deep belief network model under the current working conditions of the substation;

[0142] S42. Perform forward reasoning on the multi-source state dataset using the fault-sensitive deep belief network model corresponding to the individual jellyfish positions to obtain preliminary fault identification output. and compare the initial fault identification output with the historical label Comparison, the introduction of substation key equipment weight coefficient Forming a device-weighted cross entropy loss function that integrates prediction performance and device importance :

[0143] ;

[0144] Among them, Q represents the number of devices participating in the evaluation, represents the critical weight of the qth device, c represents the total number of fault types that can be predicted by the model, represents the true i-th fault label of the q-th device in the training sample, Indicates individual jellyfish The predicted probability output of the deep belief network model represented by the representation for the i-th type of fault on the q-th device;

[0145] Device-weighted cross entropy loss function It not only reflects the accuracy of the model in identifying various types of faults, but also strengthens the focus on key equipment through the operation critical weights, ensuring that the optimization process does not ignore the components in the substation that are critical to system stability while improving the overall accuracy. It is more in line with the power engineering requirements for safety redundancy, equipment-level priority response, and fault criticality.

[0146] S43. Initialize the jellyfish population And set the maximum number of iterations , in each round t, combined with the dynamic operating weight of the current substation operating status , use the ocean current driven update formula to achieve global optimization:

[0147] ;

[0148] in, represents the position vector of the sth jellyfish individual in the tth iteration, Represents the updated position vector of the sth jellyfish individual in the t+1th iteration It represents the dynamic operating condition weight introduced for the substation operating status in the tth iteration, which is used to adjust the responsiveness of the jellyfish swarm’s global search behavior to environmental disturbances at different times. It represents the ocean current drift factor, which is used to control the extent to which the jellyfish individual approaches the current optimal individual and reflects the cluster movement speed of the jellyfish group. represents the position vector of the individual with the best fitness in the entire jellyfish group in the tth iteration;

[0149] S44. Introducing an active foraging behavior update mechanism that combines recognition error and fault level response to accelerate the adjustment of fault-sensitive deep belief network models with errors greater than a preset value:

[0150] ;

[0151] in, is the active search coefficient, For randomly selected individuals, For the The response weight of each jellyfish in the current round is used to guide the convergence direction to correct the fault misjudgment area;

[0152] S45. In each iteration, the current-driven and active foraging mechanisms are dynamically switched based on the individual fitness change trend and the performance difference of the global optimal model, and the current optimal model parameter vector is recorded. ;

[0153] S46. If the maximum number of iterations is reached or continuous The global optimal fitness improvement in the iteration is less than the convergence threshold , then terminate the iteration and obtain the optimal parameter vector Decoded into a set of fault-sensitive deep belief network model parameters , forming an optimized fault-sensitive deep belief network model .

[0154] In this embodiment, S5 includes the following steps:

[0155] S51. Using the optimized fault-sensitive deep belief network model , the pre-processed multi-source state data set collected in real time Input into the model, perform forward reasoning calculations, and obtain each time point The equipment status prediction results under ,in, Indicates the Class failure at time point predicted probability of occurrence;

[0156] S52. Prediction results The predicted probabilities of various types of faults are sorted and the maximum value is determined, and the fault type corresponding to the maximum probability is extracted , the fault type As the main failure mode of the current equipment operation status, and record the predicted probability of the main failure mode ;

[0157] S53. Combine typical substation operating conditions, equipment risk levels, and historical fault impact data to establish fault risk classification rules;

[0158] S54. According to the fault type and predicted probability , determine the current time point The equipment failure risk level under the condition is calculated and the real-time failure prediction result label triple is output, including the prediction category, prediction probability and corresponding risk level. ,in, ;

[0159] S55. Sequence the prediction result labels in multiple consecutive time periods Perform time series fusion analysis to generate fault trend evolution curve.

[0160] In this embodiment, the fault risk classification rules are established according to the following fault risk level classification standards:

[0161] Low risk level: predicted probability , and the corresponding fault type Slight fluctuations include short-term current anomalies, low-amplitude temperature fluctuations, and single-point transient interference.

[0162] Medium risk level: predicted probability , and the corresponding fault type Belong to the medium warning category, including continuous equipment overtemperature, unplanned switching signals, and long-term voltage deviation from steady state;

[0163] High risk level: predicted probability , and the corresponding fault type Key anomalies involved include failure of circuit breaker operation, severe current fluctuations, and continuous temperature exceeding the limit.

[0164] Example 1: In June 2024, a 220kV substation was located in the western industrial zone. The substation is an important hub of the regional power transmission and distribution network, responsible for the power supply dispatching task of 8 large corporate users in the surrounding area and the urban power grid. The substation contains 3 sets of main transformers, 6 sets of circuit breakers, 16 sets of disconnectors and a large number of current, voltage and temperature monitoring equipment. As the summer industrial electricity consumption peaks, the long-term high-load operation of the equipment significantly increases the probability of failure. To ensure the safe and stable operation of the substation, the operation and maintenance unit decided to deploy the present invention at the station for actual verification, and conduct a performance comparison analysis with the traditional static threshold warning system and the intelligent recognition method based on the SVM model currently in operation.

[0165] The project testing period lasted 30 days, from June 10, 2024, to July 9, 2024. During the deployment process, synchronous sensing devices were installed in key areas, including the main transformer outgoing line, the circuit breaker contact temperature rise location, the primary current busbar, and the ambient temperature and humidity monitoring points. The system collected the following four types of real-time status data:

[0166] Voltage data : The sampling frequency is 10Hz, and the monitoring range is 0-250kV;

[0167] Current data : The sampling frequency is 10Hz, and the monitoring range is 0-3000A;

[0168] Temperature data : The sampling frequency is 1Hz, and the monitoring range is -20°C-150°C;

[0169] Switch status data : Event-triggered acquisition, recording the circuit breaker or disconnector action time, duration and feedback status code.

[0170] In the data processing phase, the original data is first subjected to standardized preprocessing, including noise filtering, outlier removal, missing value interpolation and normalization mapping, to generate a cleaned state data set. The data was then segmented into 10-minute time periods. A sliding window (with sizes of 30 seconds, 60 seconds, and 120 seconds) was used to extract the statistical characteristics of the device status (mean, standard deviation, slope, and range). The corresponding state feature vector for each segment was calculated.

[0171] By introducing prior empirical rules for equipment failure, we set voltage fluctuations exceeding ±5%, current continuous deviation from the rated value by more than 10%, abnormal temperature rise rate exceeding 5°C / min, and abnormal switch tripping frequency higher than 3 times / hour as reference thresholds. Combined with the degree of sample deviation, a fault sensitivity coefficient matrix is ​​constructed and embedded into the multi-layer deep belief network training structure.

[0172] During the model training phase, actual annotated operational data from March to May 2024 was collected as the training set, including:

[0173] Table 1 Training set parameter data

[0174]

[0175] After preliminary training of the deep belief network using the data in Table 1 above, the jellyfish swarm optimization algorithm is introduced to perform global dynamic tuning of the network parameters. The optimization process uses the cross entropy loss function and the weighted error of the equipment risk weight as evaluation indicators. 100 rounds of iterations are used, and the swarm size is 30 jellyfish individuals. Finally, the deep belief network model with the optimal parameter configuration is obtained. .

[0176] During the 30-day deployment period, a total of approximately 1.224 million monitoring data points were collected. Table 2 shows the performance comparison results of the early warning system output:

[0177] Table 2 Performance comparison results of the present invention and the traditional method

[0178]

[0179] According to operational feedback, at 02:47 AM on June 28, 2024, the early warning system successfully identified for the first time an unplanned tripping of circuit breaker zone A3, classifying it as a "high-risk" event. The system completed deduction within just 0.47 seconds of identifying the abnormal condition and automatically sent a remote warning message to the Jining Power Grid Dispatching System. An on-site investigation confirmed the cause of the fault to be an arc discharge triggering the protective switch, thus preventing equipment burnout caused by excessive contact temperature rise. The traditional system, however, did not generate an alarm during this period and only recorded an overtemperature event one hour later, after the equipment was shut down.

[0180] In the other four scenarios of on-site equipment anomalies, the system completed risk warnings more than 30 seconds in advance and accurately classified them as "medium risk" or "low risk" levels, providing a scientific basis for on-duty operation and maintenance personnel to formulate inspection sequences and scheduling plans.

[0181] In addition, during the model operation, the jellyfish swarm optimization algorithm automatically and dynamically adjusts the network's weights and bias parameters based on the current sampling data distribution. From June 15 to June 18, when the equipment load showed a continuous growth trend, the model achieved faster adaptive convergence through optimization, effectively avoiding the occurrence of fault identification lags and misjudgments.

[0182] In summary, the proposed substation remote fault warning method, based on the combination of the jellyfish swarm optimization algorithm and a deep belief network, offers the advantages of high precision, low latency, and strong adaptability. It can implement real-time warnings and dynamic risk assessments under complex operating conditions using multi-source state data, effectively enhancing the automation and safety assurance capabilities of the substation intelligent operation and maintenance system. Example 1 demonstrates the feasibility and effectiveness of the proposed method in practical engineering applications.

[0183] The present invention introduces a fault sensitivity coefficient matrix in the feature learning stage. By constructing a fault risk amplification factor that is proportional to the degree of deviation of the voltage, current, temperature, and switch state operating parameters, the weight enhancement processing of key features is achieved. It is embedded in the energy function of the improved restricted Boltzmann machine. During the unsupervised pre-training process, the deep network structure is guided to focus on the abnormal signal areas with the greatest fault potential in the operation of substation equipment. Compared with the traditional deep confidence network structure, it performs better in the extraction of low-amplitude fluctuations, multi-dimensional data coupling and time-series mixed faults, and greatly improves the accuracy of fault pattern recognition and precursor extraction capabilities under the conditions of multi-source information fusion.

[0184] This paper introduces the jellyfish swarm optimization algorithm into the deep model parameter tuning process. In view of the characteristics of changeable substation operating conditions and large differences in operating areas, it introduces a dynamic operating condition weight function and a behavior selection function based on prediction error response, and constructs a hybrid optimization update strategy that integrates ocean current drive and active foraging mechanism. It can dynamically adjust the optimization direction and amplitude according to the real-time status, prediction deviation and response speed of the equipment, so that the model parameters continue to converge to the optimal solution under the current operating state, significantly reducing the risk of overfitting and improving the model stability and generalization ability.

[0185] In the real-time reasoning stage, the present invention combines the output probability of the optimized deep model with the criticality of equipment operation to construct a risk classification standard based on the weight of equipment importance, realizing a multi-level fault risk division of low risk, medium risk and high risk. By serializing the prediction results of multiple consecutive moments, the fault evolution trend and level change trend curve are further output, providing power operation and maintenance personnel with a more forward-looking and executable decision-making basis, which can not only improve the response speed of fault handling, but also support the dynamic evolution tracking of risks across time periods, and effectively support the intelligent alarm linkage and hierarchical control of the dispatching system.

[0186] 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 substation equipment status monitoring and intelligent fault warning method based on deep learning, characterized in that: The steps include: S1. Real-time acquisition of multi-source state data sets and preprocessing to obtain preprocessed multi-source state data sets; S2. Use the preprocessed multi-source state dataset to construct the device state feature vector. Use time series data segmentation and multi-scale sliding window methods to extract key features related to fault evolution and generate a high-dimensional device state feature matrix. S3. A fault-sensitive deep belief network model performs multi-level unsupervised feature learning on the high-dimensional device state feature matrix to form preliminary fault state recognition results; S4. Apply the jellyfish swarm optimization algorithm to perform global search and dynamic tuning of the weight and bias parameters of the deep belief network model. The deviation between the initial fault state recognition result and the predetermined recognition performance index is used as feedback information during the adjustment process to obtain an optimized sensitive deep belief network model. S5. Use the optimized sensitive deep belief network model to perform online analysis on multi-source status data collected in real time, generate real-time fault prediction results for equipment status, and classify and grade fault risks; S6. Based on the real-time fault prediction results, trigger the remote fault warning mechanism and send fault warning information including the fault risk level, warning signal, and emergency response suggestions to the substation operation and maintenance center; The S3 includes the following steps: S31. Calculate the fault sensitivity coefficient of each characteristic component based on the safety margin of each key parameter during substation operation to form a fault sensitivity coefficient matrix : ； in, express Middle Time period, The eigenvectors under scale, To reflect the preset threshold value of the safe operation status of the substation, is the failure risk amplification factor; Calculated fault sensitivity coefficient matrix and high-dimensional equipment state characteristic matrix Multiply element by element to form a fault-sensitive input matrix ; S32. Using the Fault-Sensitive Input Matrix Building a Fault-Sensitive Deep Belief Network Model , the fault-sensitive deep belief network model adopts a multi-layer improved restricted Boltzmann machine stacking structure and is expressed as: ； in, Indicates the Layer-modified restricted Boltzmann machine, is the total number of layers of the fault-sensitive deep belief network model. In each layer of the fault-sensitive deep belief network model, the input vector is recorded as , for layer 1, , the hidden layer is represented as ; S33. Improved restricted Boltzmann machine at each layer Inside, for the input vector and the latent vector The energy function integrating the fault sensitivity coefficient is used for modeling: ； in, Indicates the Tier input nodes, For the Layer The input node and The connection weights between hidden nodes, and are the input node and hidden node bias of this layer respectively, For the The fault sensitivity coefficient corresponding to the input node is taken from the fault sensitive input matrix After corresponding mapping, we get: is the number of neurons in the visible layer of layer l, is the number of neurons in the lth hidden layer, represents the hidden layer vector of the lth layer; S34. Using the contrast divergence algorithm, under the premise of ensuring that the fault anomaly characteristics in the multi-source status data of the substation are fully enhanced, the improved restricted Boltzmann machine of each layer is constructed. Parameters Perform unsupervised pre-training layer by layer and output each hidden layer As the input of the next layer, during the training process, the fault sensitivity coefficient Adjust the weight of each input node in the gradient update so that the fault-sensitive deep belief network model focuses on the characteristics of abnormal deviation from the safety threshold; S35. After completing the unsupervised pre-training of all modified restricted Boltzmann machine layers, supervised fine-tuning of the entire fault-sensitive deep belief network model is performed, using the fault label vector generated by historical fault records. ,in, is the number of fault categories, and the cross entropy loss function is used to jointly optimize the parameters of the entire network to generate a fault feature expression vector ; S36. Express the fault feature vector Pass it to the output layer of the fault-sensitive deep belief network model, use the fault mode classifier to perform forward reasoning, and output the preliminary fault state recognition result .

2. A method for substation equipment status monitoring and intelligent fault early warning based on deep learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Set the data collection cycle. In each collection cycle, voltage data is collected by voltage sensors, current sensors, temperature sensors and switch status collection devices deployed at key locations of the substation. , current data , temperature data And switch status data , construct the original multi-source state dataset ; S12. Remove noise from the original multi-source state data set, use the moving average method to smooth the continuous variables, and compare the current value of each data point with its previous value. The adjacent data points are averaged as the denoising result of the data point to eliminate the influence of short-term spike noise on the data; S13. Perform outlier detection and elimination on the original multi-source state data set after the preliminary smoothing process to determine the degree of deviation of a single point in the overall data distribution. The difference between the value of each data point and the mean of its dimension is divided by the standard deviation of that dimension to obtain a normalized value. When the normalized value is greater than the preset anomaly determination threshold, the point is considered an outlier and is eliminated. S14. Perform missing value filling on the multi-source state data set after outliers have been eliminated. For continuous variables such as voltage data, current data, temperature data, and switch state data, use the average of the valid data points before and after the data point as the interpolation result to maintain the continuity and change trend of the data sequence in the multi-source state data set. S15. Normalize the multi-source state data set with missing values, subtract the minimum value in each dimension from the data, and then divide it by the difference between the maximum and minimum values ​​of the dimension, and map the data of each dimension to the interval , eliminating the dimensional differences between different physical quantities; S16. The multi-source state dataset after noise removal, outlier removal, missing value filling and normalization is defined as the preprocessed multi-source state dataset .

3. The method for substation equipment status monitoring and intelligent fault early warning based on deep learning according to claim 1 is characterized in that: The S2 comprises the following steps: S21. Preprocessed multi-source state dataset According to the collection time series Divide the segment into equal intervals and set the segment length to , divide the preprocessed multi-source state data set into several time series data sub-segments ,in, , is the total time span of the preprocessed dataset, Indicates the Data subsegments within a time window; S22. For each time series data segment Construct a multi-scale feature extraction structure based on sliding window, set the sliding window length set, and for each window length Perform sliding traversal operations to extract statistical features of voltage data, current data, temperature data, and switch status data in each sliding window, including the mean , standard deviation , slope and range ; S23. For all time series data sub-segments and window length The extracted statistical features are integrated to generate a time period index and scale index A set of feature vectors for combined indexes ; S24. Collect all the feature vectors Splicing to form a high-dimensional device state feature matrix ,in is the number of time windows, is the sliding window scale number, The number of single-segment feature dimensions extracted at each scale.

4. A method for substation equipment status monitoring and intelligent fault early warning based on deep learning according to claim 3, characterized in that: The S4 comprises the following steps: S41. Set the connection weight parameters of each layer in the trained fault-sensitive deep belief network model and bias parameters , uniformly expanded into parameter vectors ,in, Indicates the jellyfish individual number, is the size of the jellyfish group, each jellyfish individual represents a set of potential optimal configurations of the fault-sensitive deep belief network model under the current working conditions of the substation; S42. Perform forward reasoning on the multi-source state dataset using the fault-sensitive deep belief network model corresponding to the individual jellyfish positions to obtain preliminary fault identification output. and compare the initial fault identification output with the historical label Comparison, the introduction of substation key equipment weight coefficient Forming a device-weighted cross entropy loss function that integrates prediction performance and device importance : ； Among them, Q represents the number of devices participating in the evaluation, represents the critical weight of the qth device, c represents the total number of fault types that can be predicted by the model, represents the true i-th fault label of the q-th device in the training sample, Represents individual jellyfish The predicted probability output of the deep belief network model represented by the representation for the i-th type of fault on the q-th device; S43. Initialize the jellyfish population And set the maximum number of iterations , in each round In the dynamic working condition weight of the current substation operation status , use the ocean current driven update formula to achieve global optimization: ； in, represents the position vector of the sth jellyfish individual in the tth iteration, Represents the updated position vector of the sth jellyfish individual in the t+1th iteration It represents the dynamic operating condition weight introduced for the substation operation status in the tth iteration, which is used to adjust the responsiveness of the jellyfish swarm's global search behavior to environmental disturbances at different times. It represents the ocean current drift factor, which is used to control the extent to which the jellyfish individual approaches the current optimal individual and reflects the cluster movement speed of the jellyfish group. represents the position vector of the individual with the best fitness in the entire jellyfish group in the tth iteration; S44. Introducing an active foraging behavior update mechanism that combines recognition error and fault level response to accelerate the adjustment of fault-sensitive deep belief network models with errors greater than a preset value: ； in, is the active search coefficient, For randomly selected individuals, For the The response weight of each jellyfish in the current round is used to guide the convergence direction to correct the fault misjudgment area; S45. In each iteration, the current-driven and active foraging mechanisms are dynamically switched based on the individual fitness change trend and the performance difference of the global optimal model, and the current optimal model parameter vector is recorded. ; S46. If the maximum number of iterations is reached or continuous The global optimal fitness improvement in the iteration is less than the convergence threshold , then terminate the iteration and obtain the optimal parameter vector Decoded into a set of fault-sensitive deep belief network model parameters , forming an optimized fault-sensitive deep belief network model .

5. A method for substation equipment status monitoring and intelligent fault early warning based on deep learning according to claim 4, characterized in that: The S5 comprises the following steps: S51. Using the optimized fault-sensitive deep belief network model , the pre-processed multi-source state data set collected in real time Input into the model, perform forward reasoning calculations, and obtain each time point The equipment status prediction results under ; S52. Prediction results The predicted probabilities of various types of faults are sorted and the maximum value is determined, and the fault type corresponding to the maximum probability is extracted , the fault type As the main failure mode of the current equipment operation status, and record the predicted probability of the main failure mode ; S53. Combine typical substation operating conditions, equipment risk levels, and historical fault impact data to establish fault risk classification rules; S54. According to the fault type and predicted probability , determine the current time point The equipment failure risk level is as follows:

6. A method for substation equipment status monitoring and intelligent fault early warning based on deep learning according to claim 5, characterized in that: The fault risk classification rules are established according to the following fault risk level classification standards: Low risk level: predicted probability , and the corresponding fault type Slight fluctuations include short-term current anomalies, low-amplitude temperature fluctuations, and single-point transient interference. Medium risk level: predicted probability , and the corresponding fault type Belong to the medium warning category, including continuous equipment overtemperature, unplanned switching signals, and long-term voltage deviation from steady state; High risk level: predicted probability , and the corresponding fault type Key anomalies involved include failure of circuit breaker operation, severe current fluctuations, and continuous temperature exceeding the limit.

7. The method for substation equipment status monitoring and intelligent fault early warning based on deep learning according to claim 5 is characterized in that: The S5 also includes outputting a real-time fault prediction result label triple including the prediction category, prediction probability and corresponding equipment failure risk level. ,in, .

Images