Intelligent box-type substation monitoring method based on edge computing

Abstract

The application relates to the technical field of power monitoring, and discloses an intelligent box-type transformer substation monitoring method based on edge computing. The method comprises the following steps: collecting box transformer electrical quantities, heat quantities and environmental quantities to generate a box transformer operation time sequence tensor, obtaining an abnormal score through residual channel and attention channel cross gating fusion, calculating an adaptive alarm threshold in combination with a load rate, and generating an alarm instruction after triggering threshold resetting according to a residual sequence kurtosis value. The application solves the technical problems that the fixed threshold in the existing box transformer monitoring scheme cannot consider the differentiated detection of the difference between the trend abnormality and the sudden abnormality, and the adaptive threshold is invalid due to historical inertia accumulation in the data distribution mutation scene.

Description

Technical Field

[0001] This application relates to the field of power monitoring technology, and in particular to a monitoring method for intelligent prefabricated substations based on edge computing. Background Technology

[0002] As a key device in the power distribution system responsible for voltage reduction and power distribution on the user side, the real-time monitoring of the operating status of prefabricated substations is of great significance to ensuring the power supply security of end users. Existing prefabricated substation monitoring solutions mainly adopt centralized SCADA architecture or IoT cloud platform architecture. Electrical parameters, temperature parameters, and environmental parameters collected by sensors are uniformly uploaded to a remote server through a communication network for threshold comparison and alarm determination. Some solutions deploy edge computing gateways near the prefabricated substation to achieve local data aggregation and linkage control with a simple rule engine. Anomaly detection at the edge mainly relies on pre-set fixed thresholds or static judgment rules based on statistical mean plus a fixed multiple of standard deviation.

[0003] Fixed threshold strategies cannot distinguish between two different types of anomalies in transformer substation operation. Slow degradation processes such as insulation aging and increased contact resistance manifest as a continuous shift in the signal trend baseline, while sudden faults such as overload impacts and sudden increases in partial discharge exhibit short-duration pulse characteristics. A single detection channel cannot simultaneously handle these two types of anomalies with different time scales. Furthermore, the operating baseline of the transformer substation continuously drifts with seasonal temperature differences, load structure changes, and equipment aging. Fixed thresholds cannot dynamically adjust to these baseline changes, leading to an increased false alarm rate when the baseline rises and a increased false negative rate when the baseline falls.

[0004] Even if an adaptive threshold mechanism is introduced to enable the alarm threshold to track changes in the operating baseline, the mechanism itself will cause new problems. When the device characteristics change abruptly or the sensor experiences zero-point drift, the historical smoothing inertia of the adaptive threshold will gradually absorb the abrupt signal into the baseline, causing the threshold to be dragged in one direction and gradually lose its sensitivity to such anomalies. In other words, the cumulative offset of the adaptive process itself becomes a hidden failure mode, and the existing solution lacks a monitoring and correction mechanism for the trend of adaptive threshold offset. Summary of the Invention

[0005] This application provides a smart prefabricated substation monitoring method based on edge computing, which solves the technical problems in existing prefabricated substation monitoring schemes, such as the inability of fixed thresholds to take into account the differentiated detection of trend anomalies and sudden anomalies, and the failure of adaptive thresholds due to the accumulation of historical inertia in scenarios of sudden changes in data distribution.

[0006] This application provides a monitoring method for intelligent prefabricated substations based on edge computing, the monitoring method for intelligent prefabricated substations based on edge computing includes: Step S1: Collect electrical, thermal, and environmental quantities of the prefabricated substation, and generate the prefabricated substation operation sequence tensor by time-frequency alignment and normalization splicing. Step S2: Input the transformer substation operation sequence tensor into the residual channel and the attention channel respectively. The residual channel extracts the trend baseline through one-dimensional depthwise separable convolution and subtracts it from the transformer substation operation sequence tensor to obtain the residual score sequence. The attention channel performs weighted aggregation on the transformer substation operation sequence tensor through single-head self-attention to obtain the attention score sequence. The residual score sequence and the attention score sequence are mapped to gated vectors by the Sigmoid function, then cross-multiplied and added step by step to obtain the cross-gated anomaly score. Step S3: Calculate the adaptive alarm threshold of the transformer based on the mean and standard deviation of the cross-gating anomaly score within the sliding evaluation window, combined with the sensitivity coefficient corresponding to the transformer load rate. Step S4: Based on the comparison result of the kurtosis value of the residual scoring sequence in the sliding window and the preset trigger threshold, reset the transformer substation adaptive alarm threshold, compare the cross-gating abnormal score with the reset transformer substation adaptive alarm threshold, and generate an edge monitoring alarm command.

[0007] In the technical solution provided in this application, the timing tensor of the transformer operation is input into the residual channel and the attention channel for parallel processing. The residual channel uses one-dimensional depthwise separable convolution to extract the trend baseline and calculate the degree of deviation. The attention channel uses single-head self-attention to capture the feature differences between each time step within the timing window. This allows slow deterioration anomalies and sudden impact anomalies to be detected by channels that are good at processing features of the corresponding time scales, avoiding the limitation of a single channel being unable to handle two different types of anomalies. The residual score sequence and the attention score sequence are mapped to gate vectors by the Sigmoid function, then cross-multiplied and added step-by-step to obtain the cross-gated anomaly score. This cross-gating mechanism makes the two channels no longer output scores independently, but form a bidirectional information modulation relationship. When one channel detects an anomalous signal enhancement, its gate vector approaches 1, thereby amplifying the response sensitivity of the other channel at the same time. When the two types of anomalous signals overlap in time, a synergistic amplification effect is produced. This coupling enhancement characteristic cannot be obtained by simple linear weighted fusion, making the response amplitude of the cross-gated anomaly score to complex fault precursors significantly higher than the independent output of any single channel.

[0008] The adaptive alarm threshold for the transformer substation is calculated based on the mean and standard deviation of the cross-gating anomaly score within the sliding evaluation window, combined with the sensitivity coefficient corresponding to the load rate of the transformer substation. This allows the alarm threshold to be smoothly updated in line with the long-term changes in the operating baseline, and to automatically adjust the detection sensitivity according to the current load level. Under heavy load conditions, the sensitivity coefficient is set to a smaller value to lower the threshold and make alarms easier to trigger, which matches the actual need for earlier warnings as equipment failures develop faster under heavy load conditions. Based on this, the adaptive alarm threshold of the transformer substation is reset according to the comparison between the kurtosis value of the residual scoring sequence within the sliding window and the preset trigger threshold. As a fourth-order statistic reflecting the sharpness of the data distribution, the kurtosis value is much more sensitive to the occurrence of extreme deviations in the residual sequence than the variance. When the kurtosis value of multiple consecutive windows exceeds the trigger threshold, it is determined that the data distribution pattern has changed significantly. At this time, the historical information accumulated by the exponential moving average is cleared all at once and the threshold is reinitialized with the statistical characteristics of the current window. This cuts off the inertia of the adaptive mechanism's memory of the old distribution characteristics, so that the threshold can quickly return to the level that matches the current real data distribution in the event of sudden changes in equipment characteristics or sensor drift. This fundamentally solves the second-order problem derived from the introduction of the adaptive mechanism itself, which is that the adaptive threshold is continuously dragged in one direction and gradually becomes ineffective. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an embodiment of the intelligent prefabricated substation monitoring method based on edge computing in this application. Figure 2 This is a schematic diagram of the dynamic update of the adaptive alarm threshold and the kurtosis-triggered reset process of the transformer substation in the embodiments of this application. Detailed Implementation

[0011] This application provides a monitoring method for intelligent prefabricated substations based on edge computing. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent prefabricated substation monitoring method based on edge computing in this application includes: Step S1: Collect electrical, thermal, and environmental quantities of the prefabricated substation, and generate the prefabricated substation operation sequence tensor by time-frequency alignment and normalization splicing. Specifically, the transformer substation operation sequence tensor is a two-dimensional numerical matrix with fixed time steps as rows and all sensor feature channels as columns. Electrical quantities include nine channels: three-phase current RMS value, three-phase voltage RMS value, total harmonic distortion rate, partial discharge pulse count, and maximum pulse amplitude. Thermal quantities include two channels: cable joint temperature and transformer top oil temperature. Environmental quantities include three channels: substation temperature, relative humidity, and smoke concentration, totaling fourteen feature channels. Time-frequency alignment refers to unifying the signals with different sampling rates onto the same time scale using a one-second step. High-frequency signals are compressed by taking the RMS value within each second window, while low-frequency signals are interpolated to fill in missing time points. Normalization splicing involves independently mapping each feature channel to a zero-to-one interval and arranging them side-by-side along the column direction. A 300-row, fourteen-column transformer substation operation sequence tensor is formed by truncation using a 300-second sliding window. The 300-second window length corresponds to a complete load fluctuation unit in the transformer substation's power distribution cycle, and a new tensor is generated every 60 seconds.

[0013] Step S2: Input the transformer substation operation sequence tensor into the residual channel and the attention channel respectively. The residual channel extracts the trend baseline through one-dimensional depthwise separable convolution and subtracts it from the transformer substation operation sequence tensor to obtain the residual score sequence. The attention channel performs weighted aggregation on the transformer substation operation sequence tensor through single-head self-attention to obtain the attention score sequence. The residual score sequence and the attention score sequence are mapped to gated vectors by the Sigmoid function, then cross-multiplied and added step by step to obtain the cross-gated anomaly score. Specifically, the residual channel and attention channel are designed to address two typical anomaly patterns in transformer substations. The residual channel uses one-dimensional depthwise separable convolution to extract the gradually changing trend lines of electrical and temperature parameters along the time axis. The fluctuations remaining after subtracting the trend lines from the original signal are the residuals. The larger the residual, the further it deviates from the normal trend at the current moment. Such deviations usually correspond to slow degradation processes such as insulation aging or increased contact resistance. The attention channel focuses on which time steps within the time window have significant feature differences from other time steps. It assigns higher weights to time steps with significant differences through a self-attention mechanism. These differences usually correspond to rapid abrupt events such as overload impacts or sudden increases in partial discharge. The meaning of cross-gating is that when the attention channel identifies a sudden change feature at a certain moment, its Sigmoid mapping value approaches one. At this time, the score of the residual channel at the same moment is fully preserved or even amplified, which is equivalent to the sudden change signal activating the sensitivity of the trend channel to anomalies. Conversely, when the residual channel detects a continuous trend deviation, it also amplifies the response of the attention channel to weak local features. The two channels reinforce each other rather than simply adding them together.

[0014] Step S3: Calculate the adaptive alarm threshold of the transformer substation based on the mean and standard deviation of the cross-gating anomaly score within the sliding evaluation window, combined with the sensitivity coefficient corresponding to the transformer substation load rate. Specifically, the core mechanism of the adaptive alarm threshold for transformer substations is an exponential moving average. A smoothing coefficient of 0.9 means that approximately 90% of the weight in the current threshold comes from historical accumulation, and 10% comes from statistics from the latest evaluation window. This setting makes the threshold insensitive to short-term fluctuations but able to track long-term trends. The sensitivity coefficient is linked to the transformer substation's load factor: 3.5 for light load conditions below 0.3, 3.0 for normal conditions between 0.3 and 0.8, and 2.5 for heavy load conditions above 0.8. The higher the load factor, the smaller the sensitivity coefficient, the lower the threshold, and the easier it is to trigger an alarm. This is because during heavy load operation, the thermal and electrical stresses on the equipment increase simultaneously, leading to faster fault development and requiring a more sensitive detection threshold. The load factor is calculated by dividing the average active power of the low-voltage outgoing circuits by the rated capacity indicated on the transformer substation's nameplate.

[0015] Step S4: Based on the comparison result of the kurtosis value of the residual scoring sequence in the sliding window and the preset trigger threshold, reset the adaptive alarm threshold of the transformer substation, compare the cross-gating abnormal score with the reset adaptive alarm threshold of the transformer substation, and generate an edge monitoring alarm command.

[0016] Specifically, the kurtosis value reflects the sharpness of the distribution of the residual score sequence within a sliding window. A normal distribution has zero kurtosis. When the kurtosis is significantly greater than zero (i.e., peaked), it indicates that most residual values ​​within the window are concentrated near the mean, but a small number of extreme deviations occur. In a transformer substation scenario, this distribution pattern corresponds to a sudden jump in sensor performance or a step change in equipment characteristics. When the kurtosis is significantly less than zero (i.e., flat-peaked), it indicates that the residual values ​​are abnormally evenly distributed over a wide range. In a transformer substation scenario, this corresponds to signal truncation caused by sensor range saturation. The preset trigger threshold is 2.0, an empirical value determined based on the 99th percentile of the kurtosis distribution in historical normal operation data. Only when three consecutive windows exceed the trigger threshold is it considered a distribution pattern change event rather than an occasional disturbance. Once triggered, the historical cumulative amount of the exponential moving average is cleared, and the threshold is reinitialized with the statistical characteristics of the current window to avoid historical inertia masking the true changes in equipment status. Edge monitoring alarm commands are generated when the maximum value of the cross-gating anomaly score window reaches or exceeds the reset threshold. The command content carries an alarm timestamp and a dominant channel identifier. The dominant channel identifier is determined by comparing the gating residual score and the gating attention score. If the former is larger, it is marked as trend-dominant; if the latter is larger, it is marked as abrupt change-dominant. Based on this, maintenance personnel can determine whether to investigate slow degradation faults or sudden impact faults.

[0017] In one specific embodiment, step S1 includes: Electrical parameters of the high and low voltage circuits of the prefabricated substation are collected, temperature parameters of the cable joints and transformers are collected, and environmental parameters of the interior of the prefabricated substation are collected to obtain a multi-source heterogeneous sampling sequence containing different sampling frequencies. Using a fixed time step as a reference, the root mean square value of the high-frequency signal in the multi-source heterogeneous sampling sequence is calculated by windowing, and the low-frequency signal is filled to the fixed time step by linear interpolation to obtain the equally spaced resampling sequence. The 1st percentile and 99th percentile of each feature channel in the equally spaced resampled sequence within a preset historical period are taken as the normalization boundary. Each feature channel is mapped point by point to the zero-one interval to obtain the normalized feature sequence. All feature channels of the normalized feature sequence are concatenated along the feature dimension, and continuous time steps are truncated according to the preset sliding window length to generate the order tensor of the transformer during operation.

[0018] Specifically, the electrical parameters on the high-voltage side include instantaneous values ​​of three-phase current and three-phase voltage, sampled at a frequency of 4000Hz. On the low-voltage side, they include active power, total harmonic distortion (THD), and three-phase voltage imbalance, sampled at a frequency of 1Hz. Inside the high-voltage switchgear, a UHF partial discharge pulse signal is also included, sampled at a frequency of 100MHz. Temperature parameters include cable joint temperature and transformer top oil temperature, both collected by wireless temperature sensors at a sampling frequency of 0.1Hz (once every 10 seconds). Environmental parameters include internal temperature, relative humidity, and smoke concentration, also sampled at a frequency of 0.1Hz. All of these signals constitute a multi-source heterogeneous sampling sequence. The 4000Hz and 100MHz signals are classified as high-frequency signals, the 0.1Hz signal as a low-frequency signal, and the 1Hz signal, which coincides with a fixed time step, requires no additional processing. With a fixed time step of 1 second, for a 4000Hz three-phase current and voltage signal, each 1-second window contains 4000 sampling points. The squares of these 4000 points are averaged and then squared using the root mean square formula to compress them into a single root mean square value for that second. For a 100MHz partial discharge signal, the number of pulse occurrences and the amplitude of the pulse with the largest amplitude are counted within each 1-second window, forming a pulse counting channel and a maximum pulse amplitude channel, respectively. For a 0.1Hz temperature and environmental signal, there is a 10-second interval between two adjacent original sampling points. The nine missing 1-second intervals are filled in using linear interpolation, i.e., interpolating the two preceding and following real sampling values ​​proportionally to the time distance. After the above processing, all signals are unified into an equally spaced resampling sequence of 1 point per second, totaling 14 characteristic channels.

[0019] The 1st and 99th percentiles were chosen as the normalization boundaries instead of the minimum and maximum values. This is because occasional sensor spikes or communication packet loss during transformer operation can produce extreme outliers. If extreme values ​​were used as the normalization boundaries, the normal operation data would be compressed into an extremely narrow range, losing its discriminative power. The 1st and 99th percentiles can exclude the extreme samples at each end of the range, which is 1%. A 30-day historical period was preset, meaning that the normalization upper and lower bounds for each channel were determined using data from 1 point per second, totaling 2,592,000 samples per channel, over the past 30 days. The 30-day period covers the load differences between weekdays and weekends, daytime and nighttime, and the temperature fluctuation range within a month. The mapping method involves taking the original value of each characteristic channel at each time point, subtracting the 1st percentile of that channel, and then dividing by the difference between the 99th and 1st percentiles. The mapping result falls between 0 and 1. If the original value is below the 1st percentile, it is truncated to 0; if it is above the 99th percentile, it is truncated to 1. After normalization, the 14 channels are arranged sequentially along the feature dimension to form a sequence of 14-dimensional vectors corresponding to 1 second. Then, 300 consecutive time steps are extracted according to the preset sliding window length of 300 seconds to form a 300-row, 14-column box-type variable runtime sequence tensor. The window advances every 60 seconds to generate a new tensor, and the 60-second advance step is consistent with the length of the subsequent sliding evaluation window.

[0020] In one specific embodiment, in step S2, the residual channel extracts the trend baseline through one-dimensional depthwise separable convolution and subtracts it from the operating sequence tensor of the transformer substation to obtain the residual scoring sequence, including: The timing tensor of the transformer substation is input into a trend extractor consisting of multiple layers of cascaded one-dimensional depthwise separable convolutions. Each layer of one-dimensional depthwise separable convolution consists of a channel-wise convolution along the time dimension and a 1x1 pointwise convolution. Each layer is followed by batch normalization and ReLU activation, and the output is a trend baseline with the same dimension as the timing tensor of the transformer substation. The residual tensor is obtained by subtracting the operating sequence tensor of the transformer substation from the trend baseline element by element. The L2 norm of the residual tensor at each time step is calculated along the feature dimension to obtain the residual score sequence.

[0021] Specifically, the trend extractor consists of three layers of one-dimensional depthwise separable convolutions connected in series. Each layer's channel-wise convolution operates along the time dimension, with a kernel size of 15. This means that each time, data from 15 consecutive time steps is used to perform a sliding weighted average within a single feature channel. The 15-second receptive field corresponds to the typical duration of a load switch or reactive power compensation switching action in a transformer substation. After the three layers are connected in series, the total receptive field reaches 43 time steps, covering medium- to long-term trend changes such as transformer oil temperature response hysteresis and cable temperature rise inertia. Channel-wise convolution independently extracts the smooth trend in the time direction for each of the 14 channels, without exchanging information between different channels. The subsequent 1x1 pointwise convolution linearly mixes the features of the 14 channels at each time step, encoding the coupling relationship between electrical quantity channels and temperature channels. For example, when the root mean square value of the three-phase current continuously increases, the cable joint temperature usually also increases; pointwise convolution can capture this cross-channel synchronous trend. Each layer's output undergoes batch normalization to stabilize the numerical distribution of each channel within a range with a mean of 0 and a variance of 1, preventing deep gradient vanishing. The ReLU activation function sets negative values ​​to zero, retaining only positive trend features. After processing through three layers, the output is a 300-row, 14-column matrix with the exact same dimensions as the input box-variable runtime sequence tensor. This matrix is ​​the trend baseline, representing the normal operating trend that each channel should exhibit within the current time window.

[0022] The residual tensor obtained by subtracting the trend baseline element-wise from the timing tensor of the transformer substation is also a 300-row, 14-column matrix. Each element reflects the degree to which the actual value at the corresponding time step and channel deviates from the normal trend. Positive values ​​indicate that the actual value is higher than the expected trend, negative values ​​indicate that it is lower than the expected trend, and the larger the absolute value, the more severe the deviation. The residual tensor retains the deviation information of each of the 14 channels, but anomaly detection requires a scalar to measure the overall deviation at each time step. Therefore, the L2 norm is calculated along the feature dimension, that is, the residual values ​​of the 14 channels at each time step are squared, summed, and then squared, compressing the 14-dimensional vector into a non-negative scalar. The reason for choosing the L2 norm instead of simple summation or the L1 norm is that the L2 norm is more sensitive to large deviations in a single channel. Suppose that at a certain moment only the partial discharge pulse count channel is abnormal while the other 13 channels are normal, the L2 norm will amplify the contribution of that channel due to the squaring operation, making the anomaly more prominent, while the L1 norm or mean will be diluted by the other normal channels. After L2 norm calculation, each of the 300 time steps corresponds to a scalar value, forming a residual score sequence of length 300.

[0023] In one specific embodiment, in step S2, the attention channel performs weighted aggregation of the transformer substation's operating sequence tensor through single-head self-attention to obtain an attention score sequence, including: The timing tensor of the transformer substation is mapped through the Query linear projection layer and the Key-Value shared linear projection layer to obtain the query matrix and the key-value matrix, respectively. Multiply the query matrix and the transpose of the key matrix, divide by the square root of the feature dimension, and normalize using the Softmax function to obtain the inter-step attention weight matrix. Multiply the attention weight matrix between time steps with the key matrix to obtain the attention output tensor. Calculate the L2 norm of the attention output tensor for each time step along the feature dimension to obtain the attention score sequence.

[0024] Specifically, both the Query linear projection layer and the Key-Value shared linear projection layer are essentially matrix multiplication operations without bias terms. The transformer substation's runtime sequence tensor has a dimension of 300 rows and 14 columns. The Query linear projection layer internally maintains a 14-row, 14-column learnable weight matrix. The 14-dimensional feature vector of the input tensor at each time step is right-multiplied by this weight matrix to obtain a 14-dimensional query vector. This operation is performed at each of the 300 time steps, resulting in a 300-row, 14-column query matrix. The Key-Value shared linear projection layer also maintains a 14-row, 14-column weight matrix, mapping the transformer substation's runtime sequence tensor to a 300-row, 14-column key-value matrix in the same way. The reason for sharing the same projection layer for both Key and Value instead of using two separate projection layers is that edge computing nodes have limited memory and computing power. Sharing the projection layer reduces the number of parameters from three 14x14 matrices to two, or from 588 parameters to 392. Furthermore, experiments show that in scenarios with low feature dimensionality, such as transformer substation time-series data, the accuracy gain from separating Key and Value is insufficient to offset the cost of doubling the number of parameters. Multiplying the query matrix by the transpose of the key-value matrix produces a 300x300 matrix, where the element in the i-th row and j-th column represents the inner product between the query vector at time step i and the key vector at time step j. A larger inner product indicates a more similar feature pattern between the two time steps. Dividing by the square root of the feature dimension 14, or approximately 3.74, is a scaling operation designed to prevent the inner product value from becoming too large as the feature dimension increases, causing the output of the Softmax function to approach a one-hot distribution. Without scaling, the sum of the inner products of the 14 components would be too large, and Softmax would concentrate almost all the weights on the time step with the largest inner product while ignoring information from other time steps. After scaling, the weight distribution between time steps is smoother.

[0025] After Softmax normalization, the sum of the elements in each row of the 300x300 matrix is ​​1. The 300 elements in the i-th row represent the attention weights of the i-th time step to all other time steps. A higher weight means that the i-th time step considers the j-th time step to be more closely related to itself. In the context of transformer substation monitoring, when a partial discharge pulse increase or a sudden change in three-phase current occurs at a certain time step, the characteristic pattern of that time step differs significantly from the surrounding normal time steps. Its corresponding attention weights will exhibit a unique distribution: the weights are dispersed and uniform among normal time steps, while the weights of abnormal time steps are extremely low among normal time steps and extremely high among themselves or other abnormal time steps. Multiplying the attention weight matrix between time steps by the key-value matrix is ​​equivalent to performing a weighted summation and recombination of the feature vectors of the 300 time steps. Each time step obtains a new 14-dimensional vector. Abnormal time steps will gather more feature energy from themselves and other similar abnormal time steps in the recombined vector, thus widening the gap with normal time steps. The L2 norm of the 300-row, 14-column attention output tensor is calculated along the feature dimension for each time step. The processing method is exactly the same as that of the residual channel: the 14 components in each row are squared, summed, and then squared to compress into a single scalar. These 300 scalars constitute the attention scoring sequence. Because anomalous time steps accumulate more feature energy, their L2 norm is naturally higher than that of normal time steps. Therefore, the positions with higher values ​​in the attention scoring sequence correspond to the moments when mutation-type anomalous events occur.

[0026] In one specific embodiment, in step S2, the residual score sequence and the attention score sequence are mapped to gating vectors by the Sigmoid function, then cross-multiplied and added step-by-step to obtain the cross-gated anomaly score, including: The attention score sequence is mapped by the Sigmoid function to obtain the first gating vector. The first gating vector is then multiplied element by element by the residual score sequence to obtain the gating residual score. The residual score sequence is mapped by the Sigmoid function to obtain the second gating vector. The second gating vector is then multiplied element by element with the attention score sequence to obtain the gated attention score. The gating residual score and the gating attention score are added together step by step to obtain the cross-gating anomaly score.

[0027] Specifically, the output range of the Sigmoid function strictly falls within the open interval of 0 to 1. The larger the input value, the closer the output is to 1; the smaller the input value, the closer the output is to 0. When the input is 0, the output is exactly 0.5. The first gating vector, obtained by mapping the attention score sequence through the Sigmoid function, has 300 elements, each with a value between 0 and 1. Physically, it represents an openness signal provided by the attention channel at each time step. At time steps with higher attention scores (i.e., when local abrupt changes are detected), the first gating vector approaches 1, meaning that the score of the residual channel at that moment is almost completely preserved. At time steps with lower attention scores (i.e., when the feature is stable without abrupt changes), the first gating vector falls back to around 0.5, and the score of the residual channel is attenuated by about half. The gated residual score is obtained by multiplying the first gating vector element-wise with the residual score sequence. This operation makes the output of the residual channel no longer independent but dynamically adjusted by the attention channel. Taking the actual operation scenario of a transformer substation as an example, when the transformer winding insulation slowly ages, causing the partial discharge baseline to gradually rise, the residual channel can capture this long-term trend deviation. However, if a current spike caused by a sudden load shear is superimposed within the same time window, the attention channel will give an extremely high score at the time step when the spike occurs. The first gating vector approaches 1 at this time step, so that the residual score is completely preserved at this moment, or even reaches a higher value due to the superposition of baseline deviation and sudden change. This superposition amplification is the key difference between cross-gating and simple addition. In simple addition, the contribution of the two channels is always a fixed proportion, while cross-gating allows one channel to dynamically adjust its output intensity according to the detection result of the other channel.

[0028] The reverse modulation process follows the same logic but applies to different fault modes. The residual score sequence is mapped using a Sigmoid function to obtain the second gating vector. Time steps with high residual scores correspond to second gating vectors approaching 1, while time steps with low residual scores correspond to second gating vectors around 0.5. The second gating vector is then multiplied element-wise with the attention score sequence to obtain the gated attention score. This amplifies the sensitivity of the attention channel to detecting sudden changes at the same time when the residual channel detects a significant trend deviation. In the context of a prefabricated substation, a sustained abnormal increase in cable joint temperature represents a trend deviation. The residual channel responds first and gives a high score. If the attention channel detects an increase in partial discharge pulses with a relatively small amplitude at the same time, this weak signal is easily ignored without gating. However, after modulation with the second gating vector, its score is amplified and it is not missed. This precisely matches the fault evolution pattern of prefabricated substations. Temperature anomalies and increased partial discharges are often different manifestations of the same insulation degradation process. The existence of a trend deviation itself means that weak sudden change signals deserve more attention. The gated residual score and the gated attention score are added together at the corresponding positions of the same time step to obtain the cross-gated anomaly score. The sequence length is still 300. The value at each time step comprehensively reflects the coupling result of the severity of trend deviation and the significance of mutation characteristics. The two types of anomalous signals form a mutually reinforcing relationship under the cross-gating mechanism rather than each contributing a fixed component independently.

[0029] In one specific embodiment, step S3 includes: The cross-gated anomaly score is divided into continuous non-overlapping sliding evaluation windows according to a preset window length. The window mean and window standard deviation are calculated for the cross-gated anomaly score in each sliding evaluation window. The window mean and window standard deviation are recursively calculated by performing exponential moving averages on the window mean and window standard deviation respectively using a preset smoothing coefficient to obtain the smoothed mean and smoothed standard deviation. Read the average active power on the low-voltage side of the box-type substation within the current sliding evaluation window, divide the average value by the rated capacity of the box-type substation to obtain the load factor of the box-type substation, and determine the corresponding sensitivity coefficient based on the preset load range in which the load factor of the box-type substation is located. The adaptive alarm threshold of the transformer substation is obtained by adding the smoothed mean and the product of the sensitivity coefficient and the smoothed standard deviation.

[0030] Specifically, the preset window length is 60 seconds, meaning that data from 60 consecutive time steps in the cross-gated anomaly score sequence are grouped into the same sliding evaluation window, with no overlap between windows and the windows being contiguous. The 60-second window length is consistent with the step size of the transformer substation operation sequence tensor; each generation of a new transformer substation operation sequence tensor corresponds to a new sliding evaluation window for evaluation. Within each window, the arithmetic mean of the 60 cross-gated anomaly scores is calculated, and the sample standard deviation is calculated to obtain the window standard deviation. The denominator of the sample standard deviation is 59 instead of 60 to eliminate finite sample bias. The preset smoothing coefficient for the exponential moving average recursion is 0.9. In the recursive formula, the current smoothed mean is equal to 0.9 multiplied by the smoothed mean of the previous window plus 0.1 multiplied by the current window mean. The recursive method for the smoothed standard deviation is exactly the same. A smoothing coefficient of 0.9 means that approximately 90% of the weight in the current threshold comes from historical accumulation, while only 10% comes from the statistics of the latest window. This setting makes the threshold insensitive to short-term load fluctuations and sensor random noise common in transformer operation, but can track long-term baseline changes caused by seasonal changes or equipment aging. Initial values ​​are required during recursive startup. The initial values ​​for the smoothing mean and smoothing standard deviation are taken as the arithmetic mean and the arithmetic mean of the window mean and the window standard deviation of all sliding evaluation windows within 24 hours of continuous operation after the edge computing node is deployed and online, respectively. The 24-hour initialization period can cover a complete day-night load cycle of the transformer.

[0031] The load factor of the transformer substation is calculated by taking the arithmetic mean of 60 sampled active power values ​​of the low-voltage outgoing circuits within the current 60-second sliding evaluation window, and then dividing the result by the rated capacity indicated on the transformer substation nameplate, yielding a dimensionless ratio between 0 and 1. The preset load range is divided into three segments: a load factor below 0.3 is considered a light load range, with a corresponding sensitivity coefficient of 3.5; a load factor between 0.3 and 0.8 (including both boundaries) is considered a normal range, with a corresponding sensitivity coefficient of 3.0; and a load factor above 0.8 is considered a heavy load range, with a corresponding sensitivity coefficient of 2.5. The logic behind setting the sensitivity coefficient to decrease with increasing load factor is that under heavy load conditions, transformer core losses increase, winding temperature rise intensifies, and cable current carrying capacity approaches its limit. The time window for equipment to transition from a normal state to a fault state is significantly shortened, requiring a lower alarm threshold to achieve earlier warning time. The final calculation of the adaptive alarm threshold for the transformer substation uses the smoothed mean as the baseline level, and then adds the product of the sensitivity coefficient and the smoothed standard deviation to it as the allowable fluctuation range. For example, under normal operating conditions, the sensitivity coefficient is 3.0, and the threshold equals the smoothed mean plus 3.0 times the smoothed standard deviation. This means that an alarm will only be triggered if the cross-gating anomaly score exceeds the recent average level by more than 3 standard deviations, which conforms to the conventional criteria for judging significant anomalies in statistics. The adaptive alarm threshold for the transformer substation is updated once at the end of each sliding evaluation window, and the update frequency is synchronized with the window advancement frequency, i.e., refreshed every 60 seconds.

[0032] In one specific embodiment, step S4 includes: For the residual score sequence, zero-mean unit variance standardization is first performed in each sliding window. The fourth moment mean of the standardized sequence is calculated and then the normal distribution baseline value of 3 is subtracted to obtain the window kurtosis value. Determine whether the absolute value of the window kurtosis of a consecutive preset number of sliding windows is greater than a preset trigger threshold. If so, it is determined to be a distribution pattern change event, triggering the reset of the adaptive alarm threshold for the transformer substation. In response to events that change the distribution pattern, the smoothed mean in the adaptive alarm threshold of the transformer substation is reset to the window mean of the cross-gating abnormal score within the current sliding window, and the smoothed standard deviation is reset to the window standard deviation of the cross-gating abnormal score within the current sliding window. The reset adaptive alarm threshold of the transformer substation is then recalculated based on the reset smoothed mean, smoothed standard deviation and sensitivity coefficient. The maximum value of the cross-gating abnormal score in the current sliding window is compared with the reset adaptive alarm threshold of the transformer substation. When the maximum value is greater than or equal to the reset adaptive alarm threshold of the transformer substation, an edge monitoring alarm command is generated.

[0033] Specifically, the calculation of the window kurtosis value is completed in two steps. First, 60 values ​​from the residual score sequence are taken within each 60-second sliding window. These 60 values ​​are then standardized to zero mean and unit variance. This is done by subtracting the arithmetic mean of these 60 values ​​from each value and then dividing by the sample standard deviation of these 60 values. The standardized sequence has a mean of 0 and a variance of 1. Next, the arithmetic mean of each of the 60 standardized values ​​is calculated by raising it to the fourth power. The result is the fourth moment mean. Finally, 3 is subtracted to obtain the window kurtosis value. The reason for subtracting 3 is that the fourth moment mean of a normal distribution is exactly equal to 3. Subtracting this value makes the kurtosis value corresponding to a normal distribution zero. A positive value indicates a leptokurtic distribution, meaning there are a few extreme deviations in the data, while a negative value indicates a flat distribution, meaning the data distribution is abnormally flat. In the context of transformer substation monitoring, the residual score sequence reflects the degree of deviation from the trend baseline at each time step. During normal operation, the distribution of the residual score is close to normal, with the kurtosis value fluctuating slightly around 0. When the sensor experiences zero-point drift or the equipment characteristics undergo a step change, the residual score will contain a dense cluster of extremely large or small values. The fourth moment is much more sensitive to such extreme values ​​than the second moment (variance). A value that deviates from the mean by 3 times the standard deviation is amplified to 81 times the weight under the fourth power operation, causing the kurtosis value to deviate sharply from 0. The preset trigger threshold is set to 2.0. This value comes from the statistical analysis of the kurtosis value distribution in the historical data of transformer substation operation. Under normal operating conditions, the 99th percentile of the kurtosis value is approximately 1.8. Setting it to 2.0 leaves a certain margin to avoid frequent triggering of resets due to normal fluctuations. The preset number of consecutive windows is 3, which means that the absolute value of the kurtosis value must exceed 2.0 for three consecutive sliding windows (a total of 180 seconds) before it is judged as a distribution pattern change event. The purpose of this continuity requirement is to filter out occasional single-window kurtosis jumps. In transient events such as load shedding or lightning overvoltage, the transformer may have a single window kurtosis abnormality but then recover immediately. Such transient disturbances should not trigger threshold reset.

[0034] After a distribution pattern change event is triggered, all historical information accumulated in the exponential moving average recursion is cleared at once. The smoothed mean is directly replaced with the window mean of the cross-gated anomaly scores within the current sliding window, and the smoothed standard deviation is directly replaced with the window standard deviation of the current window. After the replacement, the reset value is used as a new starting point to continue the recursion with a smoothing coefficient of 0.9. The exponential moving average calculation for subsequent windows is exactly the same as before the reset, the only difference being that the historical accumulation starts from the reset value again. The essence of the reset is to cut off the threshold's memory of the old data distribution characteristics, forcing the threshold to immediately reflect the current true data distribution level. Based on the reset smoothed mean and smoothed standard deviation, combined with the sensitivity coefficient corresponding to the current window, the transformer adaptive alarm threshold is recalculated by adding the smoothed mean and sensitivity coefficient multiplied by the smoothed standard deviation. Within the current window, the maximum value among the 60 cross-gated anomaly scores is compared with the reset threshold. If the maximum value is greater than or equal to the threshold, an edge monitoring alarm command is generated. The edge monitoring alarm command includes an alarm timestamp, the time step position of the maximum cross-gating anomaly score within the current window, the difference in thresholds before and after reset, and the dominant channel identifier. The dominant channel identifier is determined by comparing the gating residual score and the gating attention score at the alarm trigger time step. If the former is larger, it is marked as trend-dominant, indicating that the anomaly source is biased towards slowly deteriorating faults; if the latter is larger, it is marked as sudden change-dominant, indicating that the anomaly source is biased towards sudden impact faults. Maintenance personnel use this identifier to choose whether to investigate progressive hidden dangers such as aging cable joints and deteriorating transformer oil, or acute faults such as overload and short circuits. If no distribution pattern change event is triggered in the current window, the transformer substation adaptive alarm threshold is updated recursively according to the normal exponential moving average without resetting. The alarm judgment process is the same as described above.

[0035] Figure 2 This is a schematic diagram of the dynamic update of the adaptive alarm threshold and the kurtosis-triggered reset process of the transformer substation in the embodiments of this application. Figure 2The chart contains two sub-charts sharing the same horizontal axis: the sliding evaluation window number. In the upper sub-chart, the solid line represents the trajectory of the transformer substation adaptive alarm threshold as the window number progresses. The scatter plots represent the maximum cross-gating anomaly score within each sliding evaluation window, and the vertical dashed line marks the window position at the threshold reset time. As seen in the upper sub-chart, within the window number range of 20 to 50, due to the continuous shift in the data distribution of the residual score sequence, the exponential moving average mechanism slowly absorbs the changes, gradually raising the transformer substation adaptive alarm threshold. Although the maximum anomaly score shows a significant increase within the window range of 38 to 43, it never reaches the raised threshold line, posing a risk of missed alarms. The lower sub-chart shows the bar chart representing the window kurtosis value of the residual score sequence within each sliding window, with horizontal dotted lines marking the preset trigger thresholds of ±2.0. Within the window sequence 49 to 53, the kurtosis values ​​of multiple consecutive windows exceed the preset trigger threshold of 2.0, satisfying the judgment condition that three consecutive windows exceed the limit, triggering a distribution pattern change event. The transformer adaptive alarm threshold is reset to the statistical characteristic level of the current window at the vertical dashed line mark. The threshold curve in the upper subgraph shows a significant step drop at the reset moment. After the reset, the threshold falls back from the dragged-up level to a reasonable position that matches the current data distribution, and the scatter points of subsequent windows restore the normal spacing relationship with the threshold line.

[0036] In one specific embodiment, the preset trigger threshold is set to 2.0, and the number of consecutive preset thresholds is 3. When the maximum value is greater than or equal to the reset adaptive alarm threshold of the transformer substation, the edge monitoring alarm command carries the difference between the window kurtosis value corresponding to the current sliding window and the adaptive alarm threshold of the transformer substation before and after the reset. When the maximum value is less than the reset adaptive alarm threshold of the transformer substation but greater than or equal to a preset ratio multiple of the reset adaptive alarm threshold of the transformer substation, an edge monitoring early warning command is generated.

[0037] The numerical basis for setting the preset trigger threshold to 2.0 and the consecutive preset number to 3 has been explained in detail in the aforementioned kurtosis calculation process. In addition to the alarm timestamp and the dominant channel identifier, the edge monitoring alarm command also carries the window kurtosis value corresponding to the current sliding window and the difference between the adaptive alarm threshold of the transformer substation before and after the reset. The window kurtosis value reflects the degree to which the distribution of the residual score sequence deviates from normality when the alarm is triggered. Based on this, maintenance personnel can determine whether the anomaly is caused by an extreme spike or by a change in the overall signal pattern. The threshold difference reflects the adjustment range of the alarm threshold by the reset operation. A positive difference indicates that the threshold has increased after the reset, that is, the baseline level of the current data distribution is higher than the historical cumulative level. A negative difference indicates that the threshold has decreased after the reset, that is, the baseline level has dropped. The two values, together with the dominant channel identifier, form a three-dimensional fault characteristic description. The preset ratio is 0.8, meaning that when the maximum value of the cross-gating abnormal score window has reached 80% of the threshold but not the threshold, an edge monitoring early warning command is generated. The value of 0.8 defines an early warning buffer zone between normal and alarm states; this zone corresponds to a situation where the transformer substation's status has deviated from normal but has not yet reached a severity requiring immediate action. The edge monitoring early warning command differs from the alarm command in that the early warning command does not trigger automatic control actions at the edge; it is only sent to the maintenance terminal via the communication link to provide textual notification. The alarm command, on the other hand, determines whether to simultaneously issue local control commands such as fan start-up or load current limiting based on auxiliary conditions such as the substation temperature. The three-level judgment logic is normal, early warning, and alarm, discretizing the continuous numerical domain of the cross-gating abnormal score into three operational levels, avoiding frequent false alarms caused by repeated jumps near the threshold in binary judgment.

[0038] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A monitoring method for intelligent prefabricated substations based on edge computing, characterized in that, The method includes: Step S1: Collect electrical, thermal, and environmental quantities of the prefabricated substation, and generate the prefabricated substation operation sequence tensor by time-frequency alignment and normalization splicing. Step S2: Input the transformer substation operation sequence tensor into the residual channel and the attention channel respectively. The residual channel extracts the trend baseline through one-dimensional depthwise separable convolution and subtracts it from the transformer substation operation sequence tensor to obtain the residual score sequence. The attention channel performs weighted aggregation on the transformer substation operation sequence tensor through single-head self-attention to obtain the attention score sequence. The residual score sequence and the attention score sequence are mapped to gated vectors by the Sigmoid function, then cross-multiplied and added step by step to obtain the cross-gated anomaly score. Step S3: Calculate the adaptive alarm threshold of the transformer based on the mean and standard deviation of the cross-gating anomaly score within the sliding evaluation window, combined with the sensitivity coefficient corresponding to the transformer load rate. Step S4: Based on the comparison result of the kurtosis value of the residual scoring sequence in the sliding window and the preset trigger threshold, reset the transformer substation adaptive alarm threshold, compare the cross-gating abnormal score with the reset transformer substation adaptive alarm threshold, and generate an edge monitoring alarm command.

2. The intelligent prefabricated substation monitoring method based on edge computing according to claim 1, characterized in that, Step S1 includes: Electrical parameters of the high and low voltage circuits of the prefabricated substation are collected, temperature parameters of the cable joints and transformers are collected, and environmental parameters of the interior of the prefabricated substation are collected to obtain a multi-source heterogeneous sampling sequence containing different sampling frequencies. Using a fixed time step as a reference, the root mean square value of the high-frequency signal in the multi-source heterogeneous sampling sequence is calculated by windowing, and the low-frequency signal is filled to the fixed time step by linear interpolation to obtain an equally spaced resampling sequence. The 1st percentile and 99th percentile of each feature channel in the equally spaced resampled sequence within a preset historical period are taken as normalization boundaries. Each feature channel is mapped point by point to the zero-one interval to obtain a normalized feature sequence. All feature channels of the normalized feature sequence are concatenated along the feature dimension, and continuous time steps are truncated according to a preset sliding window length to generate the running sequence tensor of the transformer substation.

3. The intelligent prefabricated substation monitoring method based on edge computing according to claim 1, characterized in that, In step S2, the residual channel extracts the trend baseline through one-dimensional depthwise separable convolution and subtracts it from the operating sequence tensor of the transformer substation to obtain the residual scoring sequence, including: The operating sequence tensor of the transformer substation is input into a trend extractor consisting of multiple layers of cascaded one-dimensional depthwise separable convolutions. Each layer of one-dimensional depthwise separable convolution consists of a channel-wise convolution along the time dimension and a 1x1 pointwise convolution. Each layer is followed by batch normalization and ReLU activation, and the output is a trend baseline with the same dimension as the operating sequence tensor of the transformer substation. The residual tensor is obtained by subtracting the trend baseline element by element from the operating sequence tensor of the transformer substation. The L2 norm of the residual tensor at each time step is calculated along the feature dimension to obtain the residual score sequence.

4. The intelligent prefabricated substation monitoring method based on edge computing according to claim 3, characterized in that, In step S2, the attention channel performs weighted aggregation of the transformer substation's operating sequence tensor through single-head self-attention to obtain an attention score sequence, including: The transformer substation operation sequence tensor is mapped through a Query linear projection layer and a Key-Value shared linear projection layer to obtain a query matrix and a key-value matrix; Multiply the query matrix by the transpose of the key matrix and divide by the square root of the feature dimension, then normalize using the Softmax function to obtain the time-step attention weight matrix. The attention output tensor is obtained by multiplying the attention weight matrix between time steps with the key value matrix. The L2 norm of the attention output tensor at each time step is calculated along the feature dimension to obtain the attention score sequence.

5. The intelligent prefabricated substation monitoring method based on edge computing according to claim 4, characterized in that, In step S2, the residual score sequence and the attention score sequence are mapped to gating vectors by the Sigmoid function, then cross-multiplied and added step-by-step to obtain the cross-gated anomaly score, including: The attention score sequence is mapped using the Sigmoid function to obtain a first gating vector. The first gating vector is then multiplied element-wise with the residual score sequence to obtain the gating residual score. The residual score sequence is mapped using the Sigmoid function to obtain a second gating vector. The second gating vector is then multiplied element-wise with the attention score sequence to obtain the gated attention score. The gating residual score and the gating attention score are added together step by step over time to obtain the cross-gating anomaly score.

6. The intelligent prefabricated substation monitoring method based on edge computing according to claim 1, characterized in that, Step S3 includes: The cross-gated anomaly score is divided into continuous non-overlapping sliding evaluation windows according to a preset window length. The window mean and window standard deviation are calculated for the cross-gated anomaly score in each sliding evaluation window. The window mean and the window standard deviation are recursively calculated by performing exponential moving averages on the window mean and the window standard deviation respectively with a preset smoothing coefficient to obtain the smoothed mean and smoothed standard deviation. Read the average value of the active power on the low-voltage side of the box-type substation within the current sliding evaluation window, divide the average value by the rated capacity of the box-type substation to obtain the load rate of the box-type substation, and determine the corresponding sensitivity coefficient according to the preset load range in which the load rate of the box-type substation is located. The adaptive alarm threshold of the transformer substation is obtained by adding the product of the smoothed mean, the sensitivity coefficient, and the smoothed standard deviation.

7. The intelligent prefabricated substation monitoring method based on edge computing according to claim 1, characterized in that, Step S4 includes: The residual score sequence is first standardized with zero mean and unit variance within each sliding window. The fourth moment mean of the standardized sequence is calculated and the normal distribution baseline value of 3 is subtracted to obtain the window kurtosis value. Determine whether the absolute value of the window kurtosis value of a consecutive preset number of sliding windows is greater than a preset trigger threshold. If so, it is determined to be a distribution pattern change event, triggering the reset of the transformer substation adaptive alarm threshold. In response to the distribution pattern change event, the smoothed mean in the transformer substation adaptive alarm threshold is reset to the window mean of the cross-gating abnormal score within the current sliding window, and the smoothed standard deviation is reset to the window standard deviation of the cross-gating abnormal score within the current sliding window. The reset transformer substation adaptive alarm threshold is then recalculated based on the reset smoothed mean, smoothed standard deviation and sensitivity coefficient. The maximum value of the cross-gating abnormal score in the current sliding window is compared with the reset adaptive alarm threshold of the transformer substation. When the maximum value is greater than or equal to the reset adaptive alarm threshold of the transformer substation, the edge monitoring alarm command is generated.

8. The intelligent prefabricated substation monitoring method based on edge computing according to claim 7, characterized in that, The preset trigger threshold is set to 2.0, and the number of consecutive preset triggers is 3.

9. The intelligent prefabricated substation monitoring method based on edge computing according to claim 7, characterized in that, When the maximum value is greater than or equal to the reset adaptive alarm threshold of the transformer substation, the edge monitoring alarm command carries the difference between the window kurtosis value corresponding to the current sliding window and the adaptive alarm threshold of the transformer substation before and after the reset.

10. The intelligent prefabricated substation monitoring method based on edge computing according to claim 7, characterized in that, When the maximum value is less than the reset adaptive alarm threshold of the transformer substation and greater than or equal to a preset ratio multiple of the reset adaptive alarm threshold of the transformer substation, an edge monitoring early warning command is generated.

Images