A power distribution cabinet lightning protection and overvoltage protection method based on deep learning

Abstract

The application discloses a power distribution cabinet lightning and overvoltage protection method based on deep learning, comprising the following steps: S1, collecting multi-source electrical parameter data; S2, performing preprocessing on the multi-source electrical parameter data; S3, performing time sequence dimension and variable dimension interaction modeling through an improved iTransformer network to generate a lightning and overvoltage risk prediction vector; S4, determining the risk level of lightning and overvoltage through a dynamic CUSUM detection method and outputting an early warning level label; S5, matching corresponding protection measures and performing protection control based on the early warning level label; S6, collecting real-time operation state data of the power distribution cabinet and inputting the data to the improved iTransformer network for incremental updating; and S7, structurally arranging the risk prediction vector, the early warning level label and the operation state data and uploading the data to a cloud platform. The application improves the accuracy and response efficiency of lightning and overvoltage risk identification of the power distribution cabinet.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power systems and industrial safety protection technology, and in particular to a method for lightning strike and overvoltage protection of distribution cabinets based on deep learning. Background Technology

[0002] With the increasing demand for smart grids and digital management of power equipment, surge interference and overvoltage anomalies induced by lightning strikes in the power distribution process have become significant hidden dangers affecting power supply reliability and equipment safety. Existing distribution cabinet protection methods mainly rely on static protection devices and preset threshold rules for control. However, in the face of complex electromagnetic disturbance environments and nonlinear parameter fluctuation scenarios, the following technical problems exist:

[0003] Multi-source electrical signals are complex, encompassing analog, digital, and pulse signals. The frequency distribution and time synchronization of these signals vary significantly, making it difficult for existing methods to provide unified modeling and dynamic feature extraction. During the operation of distribution cabinets, current, voltage, frequency, and electromagnetic interference are highly coupled, making it difficult for traditional rules to accurately capture their interconnected characteristics and early abnormal evolution trends. Furthermore, existing methods generally lack the ability to dynamically identify lightning-induced abrupt changes, failing to achieve real-time prediction and graded response to high-risk overvoltage events, resulting in delayed triggering of protective measures and insufficient system robustness. In addition, most of the few deep learning-based solutions fail to distinguish between time-dependent modeling and variable-coupled modeling, ignoring dynamic correlation information between electrical channels, leading to low model prediction accuracy and difficulty in deployment.

[0004] Therefore, how to provide a method for lightning strike and overvoltage protection of power distribution cabinets based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for lightning strike and overvoltage protection of distribution cabinets. This invention fully integrates an improved iTransformer network modeling method and a dynamic risk identification mechanism. It details the multimodal signal preprocessing process for lightning-induced overvoltage scenarios, the time-series-variable dual-branch deep modeling structure, causal mask fusion and pulse gating adjustment strategy, and combines the dynamic CUSUM method to achieve risk level judgment and protection control strategy matching. It has the advantages of timely risk response, accurate anomaly identification, and high degree of intelligence in protection strategy.

[0006] A method for lightning and overvoltage protection of power distribution cabinets based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect multi-source electrical parameter data;

[0008] S2. Perform preprocessing on the multi-source electrical parameter data to generate a standardized time-series dataset;

[0009] S3. Input the standardized time series dataset into the improved iTransformer network, perform interactive modeling of the time series dimension and the variable dimension, and generate risk prediction vectors for lightning strikes and overvoltages;

[0010] S4. Based on the risk prediction vector, the risk level of lightning strikes and overvoltages is determined by the dynamic CUSUM detection method, and the warning level label is output.

[0011] S5. Based on the warning level label, match the corresponding protective measures and execute protective controls;

[0012] S6. Collect the operating status data of the power distribution cabinet in real time and input it into the improved iTransformer network for incremental updates;

[0013] S7. Structure and organize the risk prediction vector, early warning level labels, and operational status data, and upload them to the cloud platform.

[0014] Optionally, step S1 specifically includes:

[0015] S11. Install voltage sensors, current transformers, frequency monitoring modules, electromagnetic interference sensing modules, and integrated temperature and humidity sensors inside and outside the distribution cabinet to collect multi-source electrical parameter data during the operation of the distribution cabinet.

[0016] S12. The multi-source electrical parameter data includes analog signals, digital signals, and pulse signals. The analog signals include three-phase voltage data, three-phase current data, ambient temperature data, and ambient humidity data. The digital signals include frequency change data. The pulse signals include electromagnetic interference intensity data.

[0017] Optionally, step S2 specifically includes:

[0018] S21. The analog signal in the multi-source electrical parameter data is converted into a digital signal stream by an analog-to-digital converter. The pulse signal is extracted by local extremum analysis and the pulse features are encoded into pulse fixed-length features. The pulse features include pulse amplitude, pulse width, pulse occurrence position and pulse count per unit time.

[0019] S22. Format the digital signal and pulse fixed-length features in the digital signal stream and multi-source electrical parameter data into a time series structure to generate a time series dataset;

[0020] S23. Perform time alignment on the time series dataset and resample it according to a uniform sampling period to form a multi-channel time series dataset;

[0021] S24. The outliers in the multi-channel time series dataset are identified and removed using the sliding window statistical method. The noise is smoothed by the median filtering method and normalized by the Z-score standardization method to obtain the normalized time series dataset.

[0022] S25. The normalized time series dataset is segmented according to the set sliding window time step and sliding step size to generate a historical sample segment sequence. The historical sample segment sequence is then organized into a three-dimensional tensor structure according to the number of samples, the number of multi-source electrical parameter channels, and the sliding window time step to form a standardized time series dataset.

[0023] Optionally, step S3 specifically includes:

[0024] S31. Input the standardized time series dataset into the improved iTransformer network, which includes a time series modeling branch, a variable modeling branch, a cross-fusion module, an impulse gating adjustment module, and a risk prediction module.

[0025] S32. Extract local dynamic features from the time series segments of each multi-source electrical parameter channel in the standardized time series dataset using a one-dimensional convolutional structure of the time series modeling branch, generate local feature tensors, input the local feature tensors into the dilated convolution stacked structure to extract medium- and long-term time-dependent features, generate global perception feature tensors, and unify the global perception feature tensors into the set feature dimensions through linear mapping to generate time series feature tensors.

[0026] The dilated convolution stack structure includes 3 layers of dilated convolution units. The kernel size of each layer of dilated convolution units is 3, the dilation rate is 1, 2 and 4 respectively, the number of output channels is 64, and the activation function is GELU.

[0027] S33. Perform variable modeling operations on the multi-source electrical parameter channel dimension of the standardized time series dataset at each sliding window time step through the variable modeling branch, specifically:

[0028] At each sliding window time step, the corresponding multi-source electrical parameter channel feature vectors are used to generate channel embedding feature tensors through fully connected embedding mapping;

[0029] The attention weights between multi-source electrical parameter channels are calculated for the channel embedding feature tensor. Based on the attention weights, the channel embedding feature tensor is weighted and aggregated to generate the associated feature tensor. The variable associated feature tensor is then unified in feature dimension through linear mapping to generate the variable feature tensor.

[0030] S34. Input the temporal feature tensor and the variable feature tensor into the cross-fusion module to perform interactive modeling operations and generate a mask fusion feature tensor.

[0031] S35. The pulse gating adjustment module generates a pulse fixed-length feature tensor by using a fully connected embedding mapping to generate a pulse fixed-length feature tensor, converts the pulse fixed-length feature tensor into a gating weight tensor by using a gating mapping function, and performs element-wise weighting on the mask fusion feature tensor and the gating weight tensor to generate a pulse adjustment feature tensor.

[0032] S36. Input the pulse modulation feature tensor into the risk prediction module, generate a latent feature tensor through linear mapping and ReLU activation function, and generate a risk prediction vector through linear mapping and Softmax normalization function. The risk prediction vector represents the predicted probability distribution of the current sample in various lightning strike and overvoltage risk levels.

[0033] Optionally, step S34 specifically includes:

[0034] The time-series feature tensor and the variable feature tensor are linearly mapped and concatenated along the feature dimension to obtain a joint feature tensor. The joint feature tensor is then linearly mapped through three sets of fully connected weight matrices to generate query tensor, key tensor and value tensor.

[0035] Along the sliding window time step dimension, a causal mask matrix is ​​added to the joint feature tensor, retaining only the attention paths of the current and past time steps. This causal mask matrix is ​​a lower triangular two-dimensional mask matrix consistent with the sliding window time step. Let the number of sliding window time steps be... Then the causal mask matrix And satisfy:

[0036] ;

[0037] in, Indicates the current time step of the sliding window. This indicates the time step of the sliding window that is being monitored. Indicates the first The sliding window time step for the first Attention masking state at each sliding window time step, when When, it indicates that the current sliding window time step only focuses on itself or past sliding window time steps, and the mask value is set to 0. If the current sliding window time step wants to focus on a future sliding window time step, violating causal logic, then the mask value is set to... ;

[0038] Causality mask matrix During the attention calculation process, the mask attention weight matrix is ​​obtained, and matrix multiplication is performed between the mask attention weight matrix and the value vector to obtain the mask feature tensor.

[0039] After linearly mapping the joint feature tensor, a residual connection is made with the mask feature tensor, and a mask fusion feature tensor is generated through layer normalization.

[0040] Optionally, step S4 specifically includes:

[0041] S41. Within the set sliding window, calculate the local mean and local standard deviation of the risk prediction vector;

[0042] S42. Based on the current predicted value, local mean, and local standard deviation, construct the positive increment and negative increment for each channel respectively. When the current predicted value is greater than the local mean, calculate the positive increment; when the current predicted value is less than the local mean, calculate the negative increment.

[0043] S43. The positive and negative increments of all channels are weighted and fused according to the channel weights to obtain the global mutation score. The channel weights are obtained by normalizing the local standard deviation.

[0044] S44. Within the set sliding window, calculate the lower and upper thresholds of the mutation level based on the global mutation score sequence.

[0045] S45. Compare the current global mutation score with the lower and upper thresholds of the mutation level. If the global mutation score is lower than the lower threshold, it is determined to be at the normal level, and the warning level label is L0. If the global mutation score is greater than or equal to the lower threshold and less than or equal to the upper threshold, it is determined to be at the warning level, and the warning level label is L1. If the global mutation score is higher than the upper threshold, it is determined to be at the high-risk level, and the warning level label is L2.

[0046] Optionally, step S5 specifically includes:

[0047] S51. When the warning level label is L0, only the risk status information is recorded, and no protective control is implemented.

[0048] S52. When the warning level label is L1, execute the first-level protection strategy, which includes limiting non-critical loads, activating the overvoltage mitigation circuit, and generating a local warning log.

[0049] S53. When the warning level label is L2, the secondary protection strategy is executed. The secondary protection strategy includes issuing trip control commands and executing zone power outages, and uploading risk information and response status to the cloud platform.

[0050] Optionally, step S6 specifically includes:

[0051] S61, The operating status data of the distribution cabinet includes three-phase voltage operating trajectory, three-phase current operating trajectory, power grid frequency change sequence, electromagnetic interference intensity change sequence, ambient temperature and humidity change sequence, protection control action execution record and historical early warning level label sequence;

[0052] S62. Preprocess the power distribution cabinet operation status data to generate a multi-source feedback dataset;

[0053] S63. Input the multi-source feedback dataset into the improved iTransformer network, and update the model parameters of the improved iTransformer network through incremental training. During the incremental training process, calculate the cross-entropy loss function based on the historical warning level label sequence and risk prediction vector, and perform gradient update and optimization of the model parameters with the cross-entropy loss function as the target.

[0054] Optionally, step S7 specifically includes:

[0055] S71. Integrate the risk prediction vector, early warning level label and distribution cabinet operation status data in a structured manner to generate a distribution cabinet operation status data package.

[0056] S72. Upload the power distribution cabinet operation status data packet to the remote cloud monitoring platform. The remote cloud monitoring platform supports remote visualization display, which includes a risk prediction trend chart, a warning level heat map, and a protection response execution timeline chart.

[0057] The beneficial effects of this invention are:

[0058] First, this invention achieves two-dimensional deep modeling of multi-source electrical parameter data during the operation of a distribution cabinet by constructing an improved iTransformer network that includes temporal modeling and variable modeling branches, effectively improving the ability to extract features related to lightning strikes and overvoltage risks. The temporal modeling branch introduces a dilated convolution stacking structure to enhance the modeling ability for medium- and long-term time-dependent features; the variable modeling branch employs a channel attention mechanism to strengthen the accuracy of correlation modeling between multi-source electrical parameter channels, ultimately forming more discriminative temporal and variable feature tensors.

[0059] Secondly, this invention introduces a causal mask into the cross-fusion module to ensure that the temporal modeling during risk prediction conforms to causal logic, avoiding interference with model accuracy from future information leakage. Furthermore, by introducing pulse fixed-length features into the pulse gating adjustment module, dynamic response and feature correction to sudden electromagnetic interference factors are achieved, enhancing the model's robustness and sensitivity to abnormal disturbances.

[0060] Furthermore, this invention uses a dynamic CUSUM detection method to determine the risk levels of lightning strikes and overvoltages, constructs multi-level early warning level labels, and combines the label levels with corresponding protection and control strategies, realizing a closed-loop response mechanism from prediction to control. After the protection and control are executed, real-time data on the operating status of the distribution cabinet is collected, a multi-source feedback dataset is constructed, and incremental updates of the model parameters are performed, further improving the model's adaptability and prediction accuracy in dynamic environments.

[0061] In summary, this invention significantly improves feature modeling depth, risk identification accuracy, protection response effectiveness, and model sustainable optimization capabilities. It enables efficient prediction and accurate protection of distribution cabinets in sudden risk scenarios such as lightning strikes and overvoltages, and has good engineering application value and promotion prospects. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a schematic diagram of a deep learning-based method for lightning and overvoltage protection of power distribution cabinets proposed in this invention.

[0064] Figure 2 This is a flowchart of the improved iTransformer structure in a deep learning-based method for lightning and overvoltage protection of power distribution cabinets proposed in this invention.

[0065] Figure 3 This invention presents a dynamic CUSUM detection method and protection control flowchart in a deep learning-based method for lightning and overvoltage protection of distribution cabinets. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1-3 A method for lightning and overvoltage protection of power distribution cabinets based on deep learning includes the following steps:

[0068] S1. Collect multi-source electrical parameter data;

[0069] S2. Perform preprocessing on the multi-source electrical parameter data to generate a standardized time-series dataset;

[0070] S3. Input the standardized time series dataset into the improved iTransformer network, perform interactive modeling of the time series dimension and the variable dimension, and generate risk prediction vectors for lightning strikes and overvoltages;

[0071] S4. Based on the risk prediction vector, the risk level of lightning strikes and overvoltages is determined by the dynamic CUSUM detection method, and the warning level label is output.

[0072] S5. Based on the warning level label, match the corresponding protective measures and execute protective controls;

[0073] S6. Collect the operating status data of the power distribution cabinet in real time and input it into the improved iTransformer network for incremental updates;

[0074] S7. Structure and organize the risk prediction vector, early warning level labels, and operational status data, and upload them to the cloud platform.

[0075] In this embodiment, step S1 specifically includes:

[0076] S11. Install voltage sensors, current transformers, frequency monitoring modules, electromagnetic interference sensing modules, and integrated temperature and humidity sensors inside and outside the distribution cabinet to collect multi-source electrical parameter data during the operation of the distribution cabinet.

[0077] S12. The multi-source electrical parameter data includes analog signals, digital signals, and pulse signals. The analog signals include three-phase voltage data, three-phase current data, ambient temperature data, and ambient humidity data. The digital signals include frequency change data. The pulse signals include electromagnetic interference intensity data.

[0078] In this embodiment, step S2 specifically includes:

[0079] S21. The analog signal in the multi-source electrical parameter data is converted into a digital signal stream by an analog-to-digital converter. The pulse signal is extracted by local extremum analysis and the pulse features are encoded into pulse fixed-length features. The pulse features include pulse amplitude, pulse width, pulse occurrence position and pulse count per unit time.

[0080] S22. Format the digital signal and pulse fixed-length features in the digital signal stream and multi-source electrical parameter data into a time series structure to generate a time series dataset;

[0081] S23. Perform time alignment on the time series dataset and resample it according to a uniform sampling period to form a multi-channel time series dataset;

[0082] S24. The outliers in the multi-channel time series dataset are identified and removed using the sliding window statistical method. The noise is smoothed by the median filtering method and normalized by the Z-score standardization method to obtain the normalized time series dataset.

[0083] S25. The normalized time series dataset is segmented according to the set sliding window time step and sliding step size to generate a historical sample segment sequence. The historical sample segment sequence is then organized into a three-dimensional tensor structure according to the number of samples, the number of multi-source electrical parameter channels, and the sliding window time step to form a standardized time series dataset.

[0084] In this embodiment, step S3 specifically includes:

[0085] S31. Input the standardized time series dataset into the improved iTransformer network, which includes a time series modeling branch, a variable modeling branch, a cross-fusion module, an impulse gating adjustment module, and a risk prediction module.

[0086] S32. Extract local dynamic features from the time series segments of each multi-source electrical parameter channel in the standardized time series dataset using a one-dimensional convolutional structure of the time series modeling branch, generate local feature tensors, input the local feature tensors into the dilated convolution stacked structure to extract medium- and long-term time-dependent features, generate global perception feature tensors, and unify the global perception feature tensors into the set feature dimensions through linear mapping to generate time series feature tensors.

[0087] The one-dimensional convolutional structure has a kernel size of 3, an output channel count of 32, and an activation function of ReLU.

[0088] The dilated convolution stack structure includes 3 layers of dilated convolution units. The kernel size of each layer of dilated convolution units is 3, the dilation rate is 1, 2 and 4 respectively, the number of output channels is 64, and the activation function is GELU.

[0089] The temporal feature tensor is a four-dimensional tensor structure, including the number of samples, the number of multi-source electrical parameter channels, the number of sliding window time steps, and the feature dimension. The feature dimension is consistent with the number of output channels of the dilated convolution stack structure.

[0090] S33. Perform variable modeling operations on the multi-source electrical parameter channel dimension of the standardized time series dataset at each sliding window time step through the variable modeling branch, specifically:

[0091] At each sliding window time step, the corresponding multi-source electrical parameter channel feature vectors are used to generate channel embedding feature tensors through fully connected embedding mapping;

[0092] The attention weights between multi-source electrical parameter channels are calculated for the channel embedding feature tensor. Based on the attention weights, the channel embedding feature tensor is weighted and aggregated to generate the associated feature tensor. The variable associated feature tensor is then unified in feature dimension through linear mapping to generate the variable feature tensor.

[0093] For example, during the operation of a certain distribution cabinet, the collected multi-source electrical parameter data includes: phase A voltage 230V, phase B voltage 228V, phase C voltage 231V; phase A current 32A, phase B current 30A, phase C current 31A; grid frequency 49.7Hz; peak electromagnetic interference intensity 7.8V; ambient temperature 36℃; ambient humidity 84%. This data was continuously collected once per second over the past 60 seconds, forming a standardized time-series data segment of a sample. The time step is 60 steps, and there are 9 multi-source electrical parameter channels, including 3-phase voltage, 3-phase current, 1 frequency, 1 electromagnetic disturbance, and 1 temperature and humidity integrated channel.

[0094] In the time-series modeling branch, each electrical parameter channel of the sample corresponds to a time series segment. Local trends are extracted through one-dimensional convolution, and then dilated convolution is used to capture longer-term trend patterns. Finally, these are uniformly mapped to the same feature dimension, forming a time-series feature tensor that represents the dynamic evolution of each channel along the time axis. For example, in a three-phase current signal, the A-phase current exhibits a slow upward trend from 30A to 38A within 60 time steps, and a local pulse with an amplitude higher than 40A appears between the 48th and 52nd time steps. This time-series change can be analyzed by extracting the local pulse pattern features through one-dimensional convolution, and then further identifying its upward trend through dilated convolution stacking. This indicates that the change is not an isolated anomaly but may be a precursor to a sudden change in system current caused by a lightning strike.

[0095] In the variable modeling branch, firstly, at each sliding window time step, the features of all channels at the current moment are aggregated, and the semantics of each parameter are extracted through channel embedding. For example, at the 45th time step of the sliding window, the features of the nine channels of the distribution cabinet at the current moment are mapped to a unified embedding space dimension through channel embedding, forming a channel embedding feature tensor with consistent dimensions. Next, an attention mechanism is used to identify important coupling relationships between channels. For example, a significant increase in the attention coefficient between the electromagnetic interference channel and the B-phase current channel is found, indicating a significant dynamic coupling change; at the same time, the attention of the frequency channel and the three-phase voltage also shows a synchronous fluctuation trend, which may indicate a linkage between voltage fluctuations and grid frequency disturbances in the system. Finally, all channel embedding feature vectors are weighted and aggregated according to attention weights to obtain an associated feature vector representing the coupling structure of multi-source electrical parameters at the current moment, and mapped to a unified feature dimension through linear transformation to form a variable feature tensor.

[0096] S34. Input the temporal feature tensor and the variable feature tensor into the cross-fusion module to perform interactive modeling operations, specifically including:

[0097] The time-series feature tensor and the variable feature tensor are linearly mapped and concatenated along the feature dimension to obtain a joint feature tensor. The joint feature tensor is then linearly mapped through three sets of fully connected weight matrices to generate query tensor, key tensor and value tensor.

[0098] Along the sliding window time step dimension, a causal mask matrix is ​​added to the joint feature tensor, retaining only the attention paths of the current and past time steps. This causal mask matrix is ​​a lower triangular two-dimensional mask matrix consistent with the sliding window time step. Let the number of sliding window time steps be... Then the causal mask matrix And satisfy:

[0099] ;

[0100] in, Indicates the current time step of the sliding window. This indicates the time step of the sliding window that is being monitored. Indicates the first The sliding window time step for the first Attention masking state at each sliding window time step, when When, it indicates that the current sliding window time step only focuses on itself or past sliding window time steps, and the mask value is set to 0. If the current sliding window time step wants to focus on a future sliding window time step, violating causal logic, then the mask value is set to... ;

[0101] Causality mask matrix During the attention calculation process, the mask attention weight matrix is ​​obtained. :

[0102] ;

[0103] in, This indicates a query tensor. Represents the key tensor. The embedding dimension of the attention mapping space is represented; and matrix multiplication is performed between the mask attention weight matrix and the value vector to obtain the mask feature tensor;

[0104] After linearly mapping the joint feature tensor, a residual connection is made with the mask feature tensor, and a mask fusion feature tensor is generated through layer normalization.

[0105] S35. The pulse gating adjustment module generates a pulse fixed-length feature tensor by using a fully connected embedding mapping to generate a pulse fixed-length feature tensor, converts the pulse fixed-length feature tensor into a gating weight tensor by using a gating mapping function, and performs element-wise weighting on the mask fusion feature tensor and the gating weight tensor to generate a pulse adjustment feature tensor.

[0106] S36. Input the pulse modulation feature tensor into the risk prediction module, generate a latent feature tensor through linear mapping and ReLU activation function, and generate a risk prediction vector through linear mapping and Softmax normalization function. The risk prediction vector represents the predicted probability distribution of the current sample in various lightning strike and overvoltage risk levels.

[0107] In this embodiment, step S4 specifically includes:

[0108] S41. Within the set sliding window, calculate the local mean and local standard deviation of the risk prediction vector;

[0109] S42, Based on the current predicted value Local mean and local standard deviation Construct the positive increment for each channel respectively. and negative increment :

[0110] ;

[0111] ;

[0112] in, This represents the positive increment of the previous time step. This represents the negative increment of the previous time step. This is an adjustable sensitivity parameter; when the current predicted value is greater than the local mean, a positive increment is calculated; when the current predicted value is less than the local mean, a negative increment is calculated.

[0113] S43. Weight the positive and negative increments of all channels according to their channel weights to obtain the global mutation score. The channel weights Obtained by normalization from local standard deviation;

[0114] ;

[0115] ;

[0116] in, Represents the maximum value function;

[0117] S44. Within the set sliding window, calculate the lower threshold of the mutation level based on the global mutation score sequence. With threshold upper limit :

[0118]

[0119] ;

[0120] in, and These represent the mean and standard deviation of the global mutation score sequence within the defined sliding window. The adjustment coefficient is , and satisfies . ;

[0121] S45. Compare the current global mutation score with the lower and upper thresholds of the mutation level. If the global mutation score is lower than the lower threshold, it is determined to be at the normal level, and the warning level label is L0. If the global mutation score is greater than or equal to the lower threshold and less than or equal to the upper threshold, it is determined to be at the warning level, and the warning level label is L1. If the global mutation score is higher than the upper threshold, it is determined to be at the high-risk level, and the warning level label is L2.

[0122] For example, within a certain sliding window, the predicted risk values ​​for the three-phase voltage (phase A), three-phase voltage (phase B), and electromagnetic interference intensity in the current sliding window, calculated from historical sample segments, are 0.87, 0.54, and 0.36, respectively. The local means are 0.62, 0.67, and 0.24, respectively, and the local standard deviation is 0.10 for all three. Therefore, the positive increment of the three-phase voltage (phase A) is... The negative increment of the three-phase voltage B is The positive increment of electromagnetic interference intensity is The above increments are then weighted and fused with channel weights of 0.4, 0.35, and 0.25 to obtain the global mutation score for the current sliding window:

[0123] ;

[0124] The moving mean of the global mutation score sequence within the current sliding window is 1.2, and the standard deviation is 0.2. Therefore, the lower threshold for the mutation level is 1.0, and the upper threshold range is 1.4. Since the current score of 1.755 is higher than the upper threshold of 1.4, it is judged to be of high risk, and the warning level label is output as L2.

[0125] This invention employs a dynamic CUSUM detection method to perform abrupt change point detection on the risk prediction vector, achieving accurate differentiation of lightning strike and overvoltage risk levels. By calculating local mean and local standard deviation within a set sliding window, constructing positive and negative increments, and fusing channel weights to generate a global abrupt change score, it effectively identifies potential risk fluctuation trends. Combining the abrupt change level discrimination interval with multi-level classification of the current abrupt change score, it can output warning level labels L0, L1, or L2, significantly enhancing the real-time performance and graded response capability of risk warnings, and improving the accuracy of response to sudden lightning strikes and overvoltage events, as well as the level of safety protection.

[0126] In this embodiment, step S5 specifically includes:

[0127] S51. When the warning level label is L0, only the risk status information is recorded, and no protective control is implemented.

[0128] S52. When the warning level label is L1, execute the first-level protection strategy, which includes limiting non-critical loads, activating the overvoltage mitigation circuit, and generating a local warning log.

[0129] S53. When the warning level label is L2, the secondary protection strategy is executed. The secondary protection strategy includes issuing trip control commands and executing zone power outages, and uploading risk information and response status to the cloud platform.

[0130] In this embodiment, step S6 specifically includes:

[0131] S61, The operating status data of the distribution cabinet includes three-phase voltage operating trajectory, three-phase current operating trajectory, power grid frequency change sequence, electromagnetic interference intensity change sequence, ambient temperature and humidity change sequence, protection control action execution record and historical early warning level label sequence;

[0132] S62. Preprocess the power distribution cabinet operation status data to generate a multi-source feedback dataset;

[0133] S63. Input the multi-source feedback dataset into the improved iTransformer network, and update the model parameters of the improved iTransformer network through incremental training. During the incremental training process, calculate the cross-entropy loss function based on the historical warning level label sequence and risk prediction vector, and perform gradient update and optimization of the model parameters with the cross-entropy loss function as the target.

[0134] The model parameters include the convolution weights of the temporal modeling branch, the embedding mapping matrix of the variable modeling branch, the attention mapping parameters and normalization coefficients of the cross-fusion module, the gating weight parameters of the pulse gating adjustment module, and the linear transformation weights and normalization function parameters in the risk prediction module.

[0135] In this embodiment, step S7 specifically includes:

[0136] S71. Integrate the risk prediction vector, early warning level label and distribution cabinet operation status data in a structured manner to generate a distribution cabinet operation status data package.

[0137] S72. Upload the power distribution cabinet operation status data packet to the remote cloud monitoring platform. The remote cloud monitoring platform supports remote visualization display, which includes a risk prediction trend chart, a warning level heat map, and a protection response execution timeline chart.

[0138] Example 1:

[0139] To verify the feasibility of this invention in practice, it was applied to a smart distribution room in a 220kV substation. The distribution room is primarily responsible for power distribution to the substation's office area, relay protection zone, communication area, and backup power supply area. This distribution cabinet operates year-round in the southern coastal region, characterized by frequent thunderstorms and high humidity, and frequently faces risks such as lightning-induced overvoltage, electromagnetic interference, and voltage instability. Previously, protection relied mainly on fixed threshold alarm mechanisms and manual intervention, which resulted in problems such as untimely response, frequent false alarms and missed alarms, and outdated protective measures.

[0140] In implementation, based on the method of this invention, voltage sensors, current transformers, frequency monitoring modules, electromagnetic interference sensing modules, and integrated temperature and humidity sensors are deployed in the input and output buses, three-phase circuits, power modules, and environmental chamber of the distribution cabinet to collect operating electrical parameters such as three-phase voltage, three-phase current, grid frequency, electromagnetic interference intensity, and ambient temperature and humidity. The analog signals are digitized using the A / D module in the edge AI computing module, and features such as amplitude, width, location, and count per unit time of pulse interference are extracted. All parameters are constructed into a standardized multi-channel time-series dataset with a unified sampling period, and segmentation is performed using a sliding window.

[0141] Subsequently, the standardized time-series dataset is input into an improved iTransformer network structure. By introducing time-series modeling and variable modeling branches, electrical signals are simultaneously modeled in both the time and variable dimensions. Precursor features of lightning strikes and overvoltage abrupt changes are extracted, and pulse gating is incorporated to enhance the response to transient disturbances. The resulting risk prediction vector represents the predicted probability distribution across three risk levels: normal, warning, and high-risk. Furthermore, a dynamic CUSUM detection method is used to identify abrupt changes in the predicted score sequence, promptly outputting risk level labels.

[0142] To quantitatively verify the performance of the method of this invention, it was compared with fixed threshold methods, LSTM-based methods, and Transformer-based methods. The test data covered three months of historical operating data from a substation and 20 real records of thunderstorms. The accuracy, recall, F1 score, response latency, false alarm rate, and protection success rate were evaluated. The experimental results are shown in Table 1.

[0143] Table 1. Performance Comparison of Different Methods under Lightning and Overvoltage Protection Tasks

[0144]

[0145] As shown in Table 1, the method of this invention exhibits significant advantages over the comparative methods in lightning strike and overvoltage protection tasks. In the three classification performance indicators of accuracy, recall, and F1 score, the method of this invention achieves 96.7%, 95.2%, and 95.9% respectively, significantly higher than the other three methods, indicating that it possesses higher overall judgment ability and classification stability in multi-category risk level identification. Regarding response latency, the average response time of the method of this invention is only 73ms, far lower than the 326ms of the fixed threshold method, the 168ms of the LSTM-based method, and the 124ms of the Transformer-based method, demonstrating that this method has a faster risk identification and protection response speed, meeting the real-time protection requirements of substations for sudden lightning strikes or induced overvoltages. In terms of false alarm rate and protection success rate, which directly affect the reliability of engineering deployment, the false alarm rate of the method of this invention is controlled at 2.1%, which is significantly lower than the 11.4% of the fixed threshold method and the 6.8% and 4.3% of other deep learning methods, respectively. At the same time, its protection success rate reaches 97.8%, which is more than 10% higher than the traditional solution, fully demonstrating the engineering practicality of this invention in accurately identifying risk levels and reasonably triggering protection mechanisms.

[0146] In practical applications, the method of this invention significantly improves the accuracy of lightning strike and overvoltage risk identification, effectively reduces the false alarm rate, shortens the protection response time, and significantly enhances the reliability of the protection mechanism. It demonstrates good engineering adaptability and field deployment value, and has broad prospects for promotion and application.

[0147] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for lightning strike and overvoltage protection of power distribution cabinets based on deep learning, characterized in that, Includes the following steps: S1. Collect multi-source electrical parameter data; S2. Perform preprocessing on the multi-source electrical parameter data to generate a standardized time-series dataset; S3. Input the standardized time series dataset into the improved iTransformer network, perform interactive modeling of the time series dimension and the variable dimension, and generate risk prediction vectors for lightning strikes and overvoltages; Specifically, step S3 includes: S31. Input the standardized time series dataset into the improved iTransformer network, which includes a time series modeling branch, a variable modeling branch, a cross-fusion module, an impulse gating adjustment module, and a risk prediction module. S32. Extract local dynamic features from the time series segments of each multi-source electrical parameter channel in the standardized time series dataset using a one-dimensional convolutional structure of the time series modeling branch, generate local feature tensors, input the local feature tensors into the dilated convolution stacked structure to extract medium- and long-term time-dependent features, generate global perception feature tensors, and unify the global perception feature tensors into the set feature dimensions through linear mapping to generate time series feature tensors. S33. Perform variable modeling operations on the multi-source electrical parameter channel dimension of the standardized time series dataset at each sliding window time step through the variable modeling branch, specifically: At each sliding window time step, the corresponding multi-source electrical parameter channel feature vectors are used to generate channel embedding feature tensors through fully connected embedding mapping; The attention weights between multi-source electrical parameter channels are calculated for the channel embedding feature tensor. Based on the attention weights, the channel embedding feature tensor is weighted and aggregated to generate the associated feature tensor. The variable associated feature tensor is then unified in feature dimension through linear mapping to generate the variable feature tensor. S34. Input the temporal feature tensor and the variable feature tensor into the cross-fusion module to perform interactive modeling operations and generate a mask fusion feature tensor. S35. The pulse gating adjustment module generates a pulse fixed-length feature tensor by using a fully connected embedding mapping to generate a pulse fixed-length feature tensor, converts the pulse fixed-length feature tensor into a gating weight tensor by using a gating mapping function, and performs element-wise weighting on the mask fusion feature tensor and the gating weight tensor to generate a pulse adjustment feature tensor. S36. Input the pulse modulation feature tensor into the risk prediction module, generate a latent feature tensor through linear mapping and ReLU activation function, and generate a risk prediction vector from the latent feature tensor through linear mapping and Softmax normalization function. The risk prediction vector represents the predicted probability distribution of the current sample in various lightning strike and overvoltage risk levels. S4. Based on the risk prediction vector, the risk level of lightning strikes and overvoltages is determined by the dynamic CUSUM detection method, and the warning level label is output. S5. Based on the warning level label, match the corresponding protective measures and execute protective controls; S6. Collect the operating status data of the power distribution cabinet in real time and input it into the improved iTransformer network for incremental updates; S7. Structure and organize the risk prediction vector, early warning level labels, and operational status data, and upload them to the cloud platform.

2. The method for lightning and overvoltage protection of power distribution cabinets based on deep learning according to claim 1, characterized in that, Step S1 specifically includes: S11. Install voltage sensors, current transformers, frequency monitoring modules, electromagnetic interference sensing modules, and integrated temperature and humidity sensors inside and outside the distribution cabinet to collect multi-source electrical parameter data during the operation of the distribution cabinet. S12. The multi-source electrical parameter data includes analog signals, digital signals, and pulse signals. The analog signals include three-phase voltage data, three-phase current data, ambient temperature data, and ambient humidity data. The digital signals include frequency change data. The pulse signals include electromagnetic interference intensity data.

3. The method for lightning and overvoltage protection of power distribution cabinets based on deep learning according to claim 1, characterized in that, Step S2 specifically includes: S21. The analog signal in the multi-source electrical parameter data is converted into a digital signal stream by an analog-to-digital converter. The pulse signal is extracted by local extremum analysis and the pulse features are encoded into pulse fixed-length features. The pulse features include pulse amplitude, pulse width, pulse occurrence position and pulse count per unit time. S22. Format the digital signal and pulse fixed-length features in the digital signal stream and multi-source electrical parameter data into a time series structure to generate a time series dataset; S23. Perform time alignment on the time series dataset and resample it according to a uniform sampling period to form a multi-channel time series dataset; S24. The outliers in the multi-channel time series dataset are identified and removed using the sliding window statistical method. The noise is smoothed by the median filtering method and normalized by the Z-score standardization method to obtain the normalized time series dataset. S25. The normalized time series dataset is segmented according to the set sliding window time step and sliding step size to generate a historical sample segment sequence. The historical sample segment sequence is then organized into a three-dimensional tensor structure according to the number of samples, the number of multi-source electrical parameter channels, and the sliding window time step to form a standardized time series dataset.

4. The method for lightning and overvoltage protection of power distribution cabinets based on deep learning according to claim 1, characterized in that, The dilated convolution stack structure includes three layers of dilated convolution units. The kernel size of each layer is 3, the dilation rates are 1, 2, and 4, the number of output channels is 64, and the activation function is GELU.

5. The method for lightning and overvoltage protection of distribution cabinets based on deep learning according to claim 1, characterized in that, Step S34 specifically includes: The time-series feature tensor and the variable feature tensor are linearly mapped and concatenated along the feature dimension to obtain a joint feature tensor. The joint feature tensor is then linearly mapped through three sets of fully connected weight matrices to generate query tensor, key tensor and value tensor. Along the sliding window time step dimension, a causal mask matrix is ​​added to the joint feature tensor, retaining only the attention paths of the current and past time steps. This causal mask matrix is ​​a lower triangular two-dimensional mask matrix consistent with the sliding window time step. Let the number of sliding window time steps be... Then the causal mask matrix And satisfy: ； in, Indicates the current time step of the sliding window. This indicates the time step of the sliding window that is being monitored. Indicates the first The sliding window time step for the first Attention masking state at each sliding window time step, when When, it indicates that the current sliding window time step only focuses on itself or past sliding window time steps, and the mask value is set to 0. If the current sliding window time step wants to focus on a future sliding window time step, violating causal logic, then the mask value is set to... ; Causality mask matrix During the attention calculation process, the mask attention weight matrix is ​​obtained, and matrix multiplication is performed between the mask attention weight matrix and the value vector to obtain the mask feature tensor. After linearly mapping the joint feature tensor, a residual connection is made with the mask feature tensor, and a mask fusion feature tensor is generated through layer normalization.

6. The method for lightning and overvoltage protection of power distribution cabinets based on deep learning according to claim 1, characterized in that, Step S4 specifically includes: S41. Within the set sliding window, calculate the local mean and local standard deviation of the risk prediction vector; S42. Based on the current predicted value, local mean, and local standard deviation, construct the positive increment and negative increment for each channel respectively. When the current predicted value is greater than the local mean, calculate the positive increment; when the current predicted value is less than the local mean, calculate the negative increment. S43. The positive and negative increments of all channels are weighted and fused according to the channel weights to obtain the global mutation score. The channel weights are obtained by normalizing the local standard deviation. S44. Within the set sliding window, calculate the lower and upper thresholds of the mutation level based on the global mutation score sequence. S45. Compare the current global mutation score with the lower and upper thresholds of the mutation level. If the global mutation score is lower than the lower threshold, it is determined to be at the normal level, and the warning level label is L0. If the global mutation score is greater than or equal to the lower threshold and less than or equal to the upper threshold, it is determined to be at the warning level, and the warning level label is L1. If the global mutation score is higher than the upper threshold, it is determined to be at the high-risk level, and the warning level label is L2.

7. The method for lightning and overvoltage protection of power distribution cabinets based on deep learning according to claim 1, characterized in that, Step S5 specifically includes: S51. When the warning level label is L0, only the risk status information is recorded, and no protective control is implemented. S52. When the warning level label is L1, execute the first-level protection strategy, which includes limiting non-critical loads, activating the overvoltage mitigation circuit, and generating a local warning log. S53. When the warning level label is L2, the secondary protection strategy is executed. The secondary protection strategy includes issuing trip control commands and executing zone power outages, and uploading risk information and response status to the cloud platform.

8. The method for lightning and overvoltage protection of power distribution cabinets based on deep learning according to claim 1, characterized in that, Step S6 specifically includes: S61, The operating status data of the distribution cabinet includes three-phase voltage operating trajectory, three-phase current operating trajectory, power grid frequency change sequence, electromagnetic interference intensity change sequence, ambient temperature and humidity change sequence, protection control action execution record and historical early warning level label sequence; S62. Preprocess the power distribution cabinet operation status data to generate a multi-source feedback dataset; S63. Input the multi-source feedback dataset into the improved iTransformer network, and update the model parameters of the improved iTransformer network through incremental training. During the incremental training process, calculate the cross-entropy loss function based on the historical warning level label sequence and risk prediction vector, and perform gradient update and optimization of the model parameters with the cross-entropy loss function as the target.

9. A method for lightning and overvoltage protection of a power distribution cabinet based on deep learning according to claim 1, characterized in that, Step S7 specifically includes: S71. Integrate the risk prediction vector, early warning level label and distribution cabinet operation status data in a structured manner to generate a distribution cabinet operation status data package. S72. Upload the power distribution cabinet operation status data packet to the remote cloud monitoring platform. The remote cloud monitoring platform supports remote visualization display, which includes a risk prediction trend chart, a warning level heat map, and a protection response execution timeline chart.

Images