Substation hidden danger defect trend prediction method based on cloud edge cooperation

Abstract

The embodiment of the invention provides a substation hidden danger defect trend prediction method based on cloud-side cooperation, and belongs to the technical field of intelligent operation and maintenance of an electric power system. Comprising the following steps: acquiring inspection data of a substation through an inspection host, and deploying edge computing nodes on the substation side; preprocessing the inspection data; performing preliminary fault judgment at an edge computing node according to the obtained inspection data; sending the inspection data subjected to fault judgment into a lightweight prediction model in an edge computing node, and performing preliminary defect trend prediction on the inspection data to obtain structured prediction data; taking the inspection data and the structured prediction data as state data and uploading the state data to a provincial side cloud platform computing center; constructing a depth prediction model in a provincial side cloud platform computing center according to the historical data; and predicting the state data uploaded by each edge computing node, outputting a defect prediction result, and giving a corresponding operation and maintenance strategy recommendation according to the defect prediction result.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, specifically to a method for predicting the trend of hidden dangers and defects in substations based on cloud-edge collaboration. Background Technology

[0002] Currently, with the development of new power systems, the level of intelligence in substation equipment is constantly improving. Traditional substation defect management relies heavily on manual inspections and post-event analysis, which suffers from problems such as slow response, low data utilization, and weak predictive capabilities. To address this, companies such as the State Grid have introduced various intelligent inspection systems and equipment monitoring terminals. However, these systems often suffer from severe data silos, making it difficult to form an effective closed loop. Predictive models are deployed in the cloud, resulting in high latency and difficulty in meeting real-time requirements. Furthermore, they lack effective cloud-edge collaboration mechanisms, and data collected at the edge cannot participate in model optimization in real time.

[0003] Therefore, there is an urgent need for an intelligent prediction method that can quickly analyze and process data from edge devices and collaboratively train and deploy models with cloud platforms. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the trend of potential hazards and defects in substations based on cloud-edge collaboration. By combining edge computing with deep modeling in the cloud, it achieves hierarchical processing and distributed intelligent prediction of multi-source equipment data. By deploying a lightweight prediction model on the substation-side edge computing nodes, rapid assessment and real-time early warning of critical equipment status can be achieved, improving system response speed and operational efficiency. Furthermore, by constructing a deep prediction model integrating Transformer and graph neural networks at the provincial cloud platform computing center, the medium- and long-term prediction capability of equipment operating trends is enhanced, improving the accuracy and stability of the prediction results.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for predicting the trend of hidden dangers and defects in substations based on cloud-edge collaboration, including: The inspection host acquires inspection data of the substation and deploys edge computing nodes on the substation side; The inspection data is preprocessed; Based on the obtained inspection data, a preliminary fault judgment is made at the edge computing node; The inspection data after fault diagnosis is sent to the lightweight prediction model in the edge computing node to perform preliminary defect trend prediction on the inspection data and obtain structured prediction data. The inspection data and the structured prediction data are uploaded as status data to the provincial cloud platform computing center; A deep prediction model was built based on historical data at the provincial cloud platform computing center. The status data uploaded by each edge computing node is predicted, the defect prediction result is output, and the corresponding operation and maintenance strategy recommendation is given based on the defect prediction result. The defect prediction results and corresponding operation and maintenance strategies are recommended and sent to the corresponding edge computing nodes. The equipment is inspected based on the defect prediction results and the corresponding operation and maintenance strategies, and the monitoring results are fed back to the edge computing node and the provincial cloud platform computing center to optimize the lightweight prediction model deployed on the edge computing node and the deep prediction model of the provincial cloud platform computing center.

[0006] Optionally, the inspection data is preprocessed, including: The infrared images in the inspection data are normalized to obtain a first infrared image with uniform size; Determine whether the first infrared image is a single-channel image; If it is determined that the first infrared image is not a single-channel image, each pixel of the first infrared image is converted into an integer gray value according to formula (1). (1) in, For image Medium pixel grayscale value, For pixels Temperature value, The minimum temperature value of the image. The maximum temperature value of the image, the target range of grayscale values ​​is [0, ... An 8-bit image corresponds to 0 to 255, therefore = 256; If the first infrared image is determined to be a single-channel image, then according to formula (2), each pixel of the first infrared image is traversed, the frequency of grayscale values ​​appearing in the target range is counted, and a histogram is generated. (2) in, This is a histogram of grayscale values ​​within the target range. Image width, Image height, For the target range; Based on the histogram, the total number of pixels with gray values ​​less than or equal to the target range is obtained according to formula (3). (3) in, The cumulative distribution function represents the total number of pixels whose gray values ​​are less than or equal to the target range. , This represents the total number of pixels. A histogram showing grayscale values ​​less than or equal to the target range; The total number of pixels is normalized according to formula (4) to generate mapped pixels. (4) in, This represents the mapping function generated by normalizing the cumulative distribution function to the target grayscale range. This represents the smallest non-zero value in the cumulative distribution function; Traverse each pixel of the first infrared image, map the grayscale value of each pixel to a mapped pixel, and generate an enhanced image; Determine whether the grayscale value of each pixel in the enhanced image is within the target range; If the grayscale value of each pixel in the enhanced image is determined to be within the target range, the enhanced image is output.

[0007] Optionally, the inspection data is preprocessed, including: Obtain the total time series data of the inspection data; The moving average of the time series data at the current time is obtained according to formula (5). (5) in, Indicates at time The moving average, To adjust the sliding window size, For the first time series data Data points.

[0008] Optionally, the edge computing node performs a preliminary fault judgment based on the acquired inspection data, including: Obtain the two-dimensional temperature matrix of the enhanced image; Based on the two-dimensional temperature matrix, the temperature of all pixels in the enhanced image is traversed to obtain the highest temperature value; Obtain the temperature difference between the highest temperature value and the reference temperature; When the temperature difference is detected to be greater than the first threshold, there is a potential risk of abnormal heating, and an early warning is issued. The partial discharge sequence at the current moment is obtained according to formula (6). (6) in, For the entire partial discharge signal sequence, For the first The amplitude of each pulse, This represents the number of discharge pulses detected within the current acquisition period. When the number and amplitude of discharge pulses in the partial discharge sequence exceed a preset value, there is a tendency for insulation breakdown, and an early warning is issued. The gas concentration value of the substation at the current moment is obtained according to formula (7). (7) in, In time The gas concentration vector collected by the oil chromatograph at the substation at any given time. It's the hydrogen concentration. methane concentration, Acetylene concentration, and These are the concentrations of ethylene and ethane; According to formula (8), the gas concentration value is used to diagnose the gas concentration and obtain the gas concentration ratio. (8) in, This is the ratio of acetylene concentration to methane concentration. This refers to the acetylene concentration. This refers to the methane concentration. When the gas concentration ratio is greater than the second threshold, there is a risk of high-energy discharge, and an early warning is issued. When the gas concentration ratio meets the third threshold, a warning is issued indicating a potential thermal malfunction or common aging problem.

[0009] Optionally, the inspection data after fault diagnosis is fed into a lightweight prediction model in the edge computing node to perform preliminary defect trend prediction on the inspection data and obtain structured prediction data, including: The inspection data is acquired, and multimodal feature fusion is performed on the inspection data. The fused inspection data is constructed into a time-series feature sequence and used as input to the trained lightweight prediction model to obtain the corresponding predicted value. The comprehensive risk score is obtained based on the predicted value using formula (9). (9) in, This indicates the overall risk score of the current equipment. This indicates the number of key indicators being monitored. Indicates the first The weight of each indicator Segmented scoring is used to represent the predicted value. The rating; Defect evaluation is performed based on the comprehensive risk score.

[0010] Optionally, acquiring the inspection data and performing multimodal feature fusion on the inspection data includes: A feature vector is constructed based on the current maximum temperature, temperature difference, peak value of the local signal, and gas concentration value. The feature vector is input into the fully connected layer, and the attention weight of each sub-vector in the feature vector is obtained according to formulas (10) to (13). (10) (11) (12) (13) in, For bias terms, , , and These are the attention weights for the component vectors representing the highest temperature value, temperature difference, peak value of the local signal, and gas concentration value, respectively. , , This is the weight matrix. This is the highest temperature value. This is the temperature difference value. The peak value of the local signal. This represents the gas concentration value. The feature layer fusion result is obtained according to formula (14). (14) in, This is the result of feature layer fusion; According to formula (15), the temporal feature sequence constructed based on the feature layer fusion result is obtained. (15) in, It is a time-series characteristic sequence. For a point in time.

[0011] Optionally, the feature layer fusion result is obtained, including: The weight matrix is ​​trained and optimized according to the loss function of formula (16), and the gradient of each weight matrix is ​​obtained. (16) in, It is the output of feature layer fusion. This is the number of samples input this time. The loss function; Update the gradient according to formula (17). (17) in, It's the learning rate, which controls the step size for each update. For gradient; When the loss function converges to its minimum value, each weight value is determined.

[0012] Optionally, the fused inspection data is constructed into a time-series feature sequence and used as input to a trained lightweight prediction model to obtain the corresponding predicted value, including: The temporal feature sequence is used as the prediction input gate and input into the network of the LSTM model; According to formula (18), the temporal feature sequence at the current time and the hidden state at the previous time are obtained. The forgetting ratio that needs to be retained after entering the forget gate is obtained. (18) in, For the forgetting rate, It is the hidden state from the previous moment. It is the input feature from the previous time step. The weight matrix of the forget gate, Bias term, yes The activation function outputs a value in the range [0,1], which controls the proportion of information forgotten. Based on the current time-series feature sequence and the forgetting ratio, output the predicted state time-series feature sequence for the next time step.

[0013] Optionally, training the lightweight prediction model includes: Acquire training data and feed it into the LSTM model to obtain lightweight prediction data; The root mean square error value over a time period is obtained according to formula (19) and used as an evaluation of the prediction results of the LSTM model. (19) in, Predict feature sequence values ​​for the LSTM model. The actual feature sequence value, For a period of time, This is the root mean square error; Based on the evaluation results, the parameter model of the LSTM model is optimized to complete the training of the LSTM model.

[0014] Optionally, the status data uploaded by each edge computing node is predicted, a defect prediction result is output, and a corresponding operation and maintenance strategy recommendation is given based on the defect prediction result, including: Obtain the device status data at the current moment according to formula (20). (20) in, It is equipment At any moment state, It is equipment Other dimensions of data, It is equipment At any moment Input features, It is an aggregate function used for synthesizing devices. Status information in various dimensions; The state data is classified using a fully connected layer and a softmax activation function, and the prediction result of the device is obtained according to formula (21). ,(twenty one) in, and These are the weights and biases of the fully connected layer. For the predicted results; Based on the prediction results, recommended operation and maintenance strategies will be output.

[0015] The purpose of this invention, based on the above technical solutions, is to provide a method for predicting substation hidden dangers and defects based on cloud-edge collaboration. By combining edge computing with deep modeling in the cloud, it achieves hierarchical processing and distributed intelligent prediction of multi-source equipment data. Deploying a lightweight prediction model on the substation-side edge computing nodes enables rapid assessment and real-time early warning of critical equipment status, improving system response speed and operational efficiency. Constructing a deep prediction model integrating Transformer and graph neural networks at the provincial cloud platform computing center enhances the medium- and long-term prediction capabilities of equipment operating trends, improving the accuracy and stability of prediction results. This invention effectively improves the prediction level of substation equipment defects and intelligent operation and maintenance capabilities, and is applicable to power system status monitoring and risk early warning in multiple scenarios, improving operation and maintenance efficiency.

[0016] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a substation hidden danger and defect trend prediction method based on cloud-edge collaboration, which is one embodiment of the present invention. Figure 2 This is a flowchart of a preprocessing procedure for inspection data according to one embodiment of the present invention; Figure 3 This is a flowchart of a preprocessing procedure for inspection data according to one embodiment of the present invention; Figure 4 This is a flowchart illustrating the preliminary fault diagnosis process according to one embodiment of the present invention. Figure 5 This is a flowchart of a preliminary defect trend prediction method according to one embodiment of the present invention; Figure 6 This is a flowchart of a multimodal feature fusion embodiment of the present invention; Figure 7 This is a flowchart illustrating the process of obtaining feature layer fusion results according to one embodiment of the present invention; Figure 8 This is a flowchart illustrating the process of obtaining the corresponding predicted value according to one embodiment of the present invention; Figure 9 This is a flowchart of training a lightweight prediction model according to one embodiment of the present invention. Figure 10 This is a flowchart illustrating the corresponding operation and maintenance strategy recommendation process in one embodiment of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0019] In the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.

[0020] like Figure 1 The diagram shows a flowchart of a substation hidden danger and defect trend prediction method based on cloud-edge collaboration, according to one embodiment of the present invention. Figure 1 In this context, the prediction method may include: In step S1, the inspection data of the substation is obtained through the inspection host, and edge computing nodes are deployed on the substation side; In step S2, the inspection data is preprocessed; In step S3, a preliminary fault judgment is made at the edge computing node based on the acquired inspection data; In step S4, the inspection data after fault judgment is sent to the lightweight prediction model in the edge computing node to perform preliminary defect trend prediction on the inspection data and obtain structured prediction data. In step S5, the inspection data and structured prediction data are uploaded as status data to the provincial cloud platform computing center; In step S6, a deep prediction model is constructed at the provincial cloud platform computing center based on historical data; In step S7, the status data uploaded by each edge computing node is predicted, the defect prediction result is output, and the corresponding operation and maintenance strategy recommendation is given based on the defect prediction result. In step S8, the defect prediction results and corresponding operation and maintenance strategy recommendations are sent to the corresponding edge computing nodes; In step S9, the equipment is inspected based on the defect prediction results and the corresponding operation and maintenance strategies, and the monitoring results are fed back to the edge computing nodes and the provincial cloud platform computing center to optimize the lightweight prediction model deployed on the edge computing nodes and the deep prediction model of the provincial cloud platform computing center.

[0021] In Figure 1 In the method shown, step S1 can be used to acquire substation inspection data through the inspection host and deploy edge computing nodes on the substation side. Two devices, the edge computing node and the inspection host, are deployed on the substation side. The inspection host, deployed in the substation, is a device that enables unified access, control, and processing of inspection results from cameras, partial discharge, and oil chromatography online devices, and interacts with the upper-level system. It controls infrared cameras to conduct joint indoor and outdoor equipment inspections, receives inspection data collection files, performs intelligent analysis on the collected data, generates inspection results and reports, and sends alarms promptly. It also has functions such as real-time monitoring and intelligent linkage with main and auxiliary equipment monitoring. The edge computing node has basic model reasoning and data processing capabilities. It is generally deployed on-site as an industrial-grade X86 host, receives video image data collected by the inspection host, performs image recognition, discrimination, and equipment defect analysis based on video streams, images, and specific numerical results, and outputs the analysis results to the inspection host. The inspection host acquires substation inspection data including multi-source data such as infrared, partial discharge, and oil chromatography. Step S2 can be used to preprocess the inspection data. The preprocessing method can take many forms known to those skilled in the art. In one example of the present invention, the preprocessing method may include, for example... Figure 2 The steps are shown. Specifically: In step S11, the infrared images in the inspection data are normalized to obtain a first infrared image with uniform size; In step S12, it is determined whether the first infrared image is a single-channel image; In step S13, if it is determined that the first infrared image is not a single-channel image, each pixel of the first infrared image is converted into an integer grayscale value according to formula (1). (1) in, For image Medium pixel grayscale value, For pixels Temperature value, The minimum temperature value of the image. The maximum temperature value of the image, the target range of grayscale values ​​is [0, ... An 8-bit image corresponds to 0 to 255, therefore = 256; In step S14, if it is determined that the first infrared image is a single-channel image, each pixel of the first infrared image is traversed according to formula (2), the frequency of grayscale values ​​appearing in the target range is counted, and a histogram is generated. (2) in, This is a histogram of grayscale values ​​within the target range. Image width, Image height, For the target range; In step S15, based on the histogram, the total number of pixels with gray values ​​less than or equal to the target range is obtained according to formula (3). (3) in, The cumulative distribution function represents the total number of pixels whose gray values ​​are less than or equal to the target range. , This represents the total number of pixels. A histogram showing grayscale values ​​less than or equal to the target range; In step S16, the total number of pixels is normalized according to formula (4) to generate mapped pixels. (4) in, This represents the mapping function generated by normalizing the cumulative distribution function to the target grayscale range. This represents the smallest non-zero value in the cumulative distribution function; In step S17, each pixel of the first infrared image is traversed, and the gray value of each pixel is mapped to a mapped pixel to generate an enhanced image; In step S18, it is determined whether the grayscale value of each pixel in the enhanced image is within the target range; In step S19, if it is determined that the grayscale value of each pixel in the enhanced image is within the target range, the enhanced image is output.

[0022] In Figure 2 In the method shown, step S11 can be used to normalize the infrared images in the inspection data to obtain a first infrared image with uniform size. The received infrared image data is uniformly scaled to the same size to ensure that all images transmitted to the edge computing nodes have uniform image size, so that the model structure is fixed during inference or training; the image resolution is reduced, the model computational burden is reduced, it is suitable for edge computing devices to perform depth calculation, and it also avoids errors in the model convolutional layer / fully connected layer due to inconsistent sizes. Step S12 can be used to determine whether the first infrared image is a single-channel image. Check the image format to ensure that it is a single-channel image (grayscale image) to avoid processing multi-channel images. Step S13 can be used to convert each pixel of the first infrared image into an integer grayscale value according to formula (1) if the first infrared image is not a single-channel image. Step S14 can be used to traverse each pixel of the first infrared image according to formula (2) if the first infrared image is a single-channel image, count the frequency of grayscale values ​​in the target range, and generate a histogram. Step S15 can be used to obtain the total number of pixels with gray values ​​less than or equal to the target range based on the histogram and according to formula (3). Step S16 can be used to normalize the total number of pixels according to formula (4) to generate mapped pixels for redistributing gray values. Step S17 can be used to traverse each pixel of the first infrared image, map the gray value of each pixel to a mapped pixel, and generate an enhanced image. Step S18 can be used to determine whether the gray value of each pixel in the enhanced image is within the target range. Step S19 can be used to output the enhanced image if the gray value of each pixel in the enhanced image is within the target range, resulting in a more uniform gray distribution and significantly enhanced contrast between the high-temperature area and the background.

[0023] Regarding the preprocessing method, in one example of the present invention, the preprocessing method may further include, for example, Figure 3 The steps are shown. Specifically: In step S21, the total time series data of the inspection data is obtained; In step S22, the moving average of the time series data at the current time is obtained according to formula (5). (5) in, Indicates at time The moving average, To adjust the sliding window size, For the first time series data Data points.

[0024] Step S21 can be used to obtain the total time series data of the inspection data. Step S22 can be used to obtain the moving average of the time series data at the current moment according to formula (5). In the prediction of substation hidden dangers, the moving average is needed to smooth the time series data, eliminate short-term fluctuations, and extract long-term trends. Smoothing data such as oil temperature, infrared temperature, partial discharge, current, and voltage removes short-term anomalies such as "peaks" and "jitter", and more clearly captures trends (such as temperature rise and partial discharge growth trends), providing more stable input for subsequent time series models such as LSTM and Transformer.

[0025] Step S3 can be used to perform preliminary fault assessment on the edge computing node based on the acquired inspection data. The method for performing preliminary fault assessment can be of various forms known to those skilled in the art. In one example of the present invention, the method for performing preliminary fault assessment may include, for example... Figure 4 The steps are shown. Specifically: In step S31, the two-dimensional temperature matrix of the enhanced image is obtained; In step S32, the temperature of all pixels in the enhanced image is traversed according to the two-dimensional temperature matrix to obtain the highest temperature value; In step S33, the temperature difference between the highest temperature value and the reference temperature is obtained; In step S34, when the temperature difference is detected to be greater than the first threshold, there is a potential for abnormal heating, and an early warning is issued. In step S35, the partial discharge sequence at the current moment is obtained according to formula (6). (6) in, For the entire partial discharge signal sequence, For the first The amplitude of each pulse, This represents the number of discharge pulses detected within the current acquisition period. In step S36, when the number of discharge pulses and the pulse amplitude in the partial discharge sequence are greater than preset values, there is a tendency for insulation breakdown, and an early warning is issued. In step S37, the gas concentration value of the substation at the current moment is obtained according to formula (7). (7) in, In time The gas concentration vector collected by the oil chromatograph at the substation at any given time. It's the hydrogen concentration. methane concentration, Acetylene concentration, and These are the concentrations of ethylene and ethane; In step S38, the gas concentration is diagnosed according to formula (8) to obtain the gas concentration ratio. (8) in, This is the ratio of acetylene concentration to methane concentration. This refers to the acetylene concentration. This refers to the methane concentration. In step S39, when the gas concentration ratio is greater than the second threshold, there is a risk of high-energy discharge, and an early warning is issued. In step S40, when the gas concentration ratio meets the third threshold, a warning is issued indicating a potential thermal fault or common aging problem.

[0026] In Figure 4 In the method shown, step S31 can be used to obtain a two-dimensional temperature matrix of the enhanced image, which can be represented by formula (22). ,(twenty two) in The entire infrared image is in the 1st... Temperature distribution over time; set Each element Indicates the first in the image line, number Temperature values ​​of each pixel (unit: ℃ ); in the formula These represent the image's height (number of rows) and width (number of columns), respectively. Essentially, the above formula defines a... A temperature matrix, where each pixel value represents the actual temperature of a point, is used to analyze hotspot areas and determine temperature rise trends. For example, infrared anomaly detection in high-voltage equipment is based on this matrix.

[0027] After acquiring the infrared image matrix, it is necessary to calculate the number of images in real time. The highest temperature value among all pixels in the entire image at time step S32 can be obtained by iterating through the temperatures of all pixels in the enhanced image based on the two-dimensional temperature matrix. This can be expressed by formula (23). ,(twenty three) in, It is the first line, number Temperature values ​​of column pixels; Iterate through all pixels of the entire image, extract the maximum value, and assign it to... .

[0028] Step S33 can be used to obtain the temperature difference between the highest temperature and the reference temperature, which can be used as a basis for judging abnormal hot spots. It can be calculated according to formula (24). ,(twenty four) in, It is a reference temperature, which can be the ambient temperature, the historical average temperature, or the temperature of an adjacent health device; It is the temperature rise difference. When it exceeds the first threshold, it indicates a potential risk of abnormal heating.

[0029] Step S34 can be used to issue an early warning when the detected temperature difference is greater than the first threshold, indicating a potential for abnormal heating. Step S35 can be used to obtain the partial discharge sequence at the current moment according to formula (6). This sequence reflects the health of the insulation state of electrical equipment. For example, if the number of discharge pulses is large and the amplitude of the pulses is large within a certain period of time, it indicates that there may be a tendency for insulation breakdown. Step S36 can be used to issue an early warning when the number of discharge pulses and the amplitude of the pulses in the partial discharge sequence are greater than the preset value, indicating a tendency for insulation breakdown. Step S37 can be used to obtain the gas concentration value of the substation at the current moment according to formula (7). Step S38 can be used to perform gas concentration diagnosis on the gas in the gas concentration value according to formula (8) and obtain the gas concentration ratio. Step S39 can be used to issue an early warning when the gas concentration ratio is greater than the second threshold, indicating a potential for high-energy discharge. Step S40 can be used to issue an early warning when the gas concentration ratio meets the third threshold, indicating a potential for thermal failure or ordinary aging. In one example of the present invention, when the ratio ＞ 1. This usually indicates the presence of a high-energy discharge (such as an electric arc); if ≈ A value of 0 could indicate a thermal failure or normal aging.

[0030] Step S4 can be used to send the inspection data after fault assessment into a lightweight prediction model in the edge computing node to perform preliminary defect trend prediction on the inspection data and obtain structured prediction data. The method for performing preliminary defect trend prediction can be of various forms known to those skilled in the art. In one example of the present invention, the method for performing preliminary defect trend prediction may include, for example... Figure 5 The steps are shown. Specifically: In step S51, inspection data is acquired, and multimodal feature fusion is performed on the inspection data; In step S52, the fused inspection data is constructed into a time-series feature sequence and used as the input of the trained lightweight prediction model to obtain the corresponding prediction value. In step S53, the comprehensive risk score is obtained based on the predicted value using formula (9). (9) in, This indicates the overall risk score of the current equipment. This indicates the number of key indicators being monitored. Indicates the first The weight of each indicator Segmented scoring is used to represent the predicted value. The rating; In step S54, a defect evaluation is performed based on the comprehensive risk score.

[0031] Final comprehensive risk score The risk level is divided into the following ranges: Level I (high risk). >80, the predicted trend is rising sharply, the indicator is approaching its limit, and immediate maintenance is required; Level II (Medium Risk), 50< <80, the predicted value is slightly outside the normal range, and early maintenance is recommended; Level III (low risk). If the predicted value is less than 50, it is within an acceptable range, and the regular inspection frequency is maintained. This process does not rely on cloud computing on the provincial side, ensuring real-time prediction and on-site response capabilities. Edge nodes receive and preprocess multimodal data from on-site devices. The aforementioned data types were uniformly constructed into a time-series structured feature vector in the previous step. To ensure low latency and low power consumption operation at the edge, a single-layer LSTM network is used to achieve trend prediction and risk scoring. After completing model loading and initialization, the edge nodes periodically perform inference on the latest collected data, outputting the following structured prediction results: device identifier and timestamp; current risk score (e.g., 0.87); trend judgment (e.g., temperature rise); risk level (e.g., warning level); and model version and inference time. These prediction results will serve as input for subsequent data uploads, alarm triggering, and edge-cloud collaborative decision-making.

[0032] The structured prediction data generated by the edge computing node is first transmitted to the inspection host under the same local area network. The inspection host updates the current preliminary inspection results for the first time, updating the label data such as the operation trend category, hidden danger risk score, and risk level classification trend qualitative judgment of this inspection and collection task.

[0033] In Figure 5In the method shown, step S51 can be used to acquire inspection data and perform multimodal feature fusion on the inspection data. The multimodal feature fusion method can be of various forms known to those skilled in the art. In one example of the present invention, the multimodal feature fusion method may include, for example... Figure 6 The steps are shown. Specifically: In step S61, a feature vector is constructed based on the current maximum temperature, temperature difference, peak value of the local signal, and gas concentration value. In step S62, the feature vector is input into the fully connected layer, and the attention weight of each sub-vector in the feature vector is obtained according to formulas (10) to (13). (10) (11) (12) (13) in, For bias terms, , , and These are the attention weights for the component vectors representing the highest temperature value, temperature difference, peak value of the local signal, and gas concentration value, respectively. , , This is the weight matrix. This is the highest temperature value. This is the temperature difference value. The peak value of the local signal. This represents the gas concentration value. In step S63, the feature layer fusion result is obtained according to formula (14). (14) in, This is the result of feature layer fusion; In step S64, the temporal feature sequence constructed based on the feature layer fusion result is obtained according to formula (15). (15) in, It is a time-series characteristic sequence. For a point in time.

[0034] In Figure 6In the method shown, step S61 can be used to construct a feature vector based on the current maximum temperature, temperature difference, peak value of the local signal, and gas concentration value. Step S62 can be used to input the feature vector into the fully connected layer and obtain the attention weight of each sub-vector in the feature vector according to formulas (10) to (13). Step S63 can be used to obtain the feature layer fusion result according to formula (14). Obtaining the feature layer fusion result also includes... Figure 7 The steps shown are as follows: In step S631, the weight matrix is ​​trained and optimized according to the loss function of formula (16), and the gradient of each weight matrix is ​​obtained. (16) in, It is the output of feature layer fusion. This is the number of samples input this time. The loss function; In step S632, the gradient is updated according to formula (17). (17) in, It's the learning rate, which controls the step size for each update. For gradient; In step S633, each weight value is determined when the loss function converges to its minimum value.

[0035] Step S64 can be used to obtain the temporal feature sequence constructed based on the feature layer fusion result according to formula (15).

[0036] Step S52 can be used to construct a time-series feature sequence from the fused inspection data and use it as input to the trained lightweight prediction model to obtain the corresponding predicted value. The method for obtaining the corresponding predicted value can be of various forms known to those skilled in the art. In one example of the present invention, the method for obtaining the corresponding predicted value may include, for example... Figure 8 The steps are shown. Specifically: In step S71, the temporal feature sequence is used as the prediction input gate and input into the network of the LSTM model; In step S72, the temporal feature sequence of the current time and the proportion of the hidden state of the previous time that needs to be retained after entering the forget gate are obtained according to formula (18). (18) in, For the forgetting rate, It is the hidden state from the previous moment. It is the input feature from the previous time step. The weight matrix of the forget gate, Bias term, yes The activation function outputs a value in the range [0,1], which controls the proportion of information forgotten. In step S73, the predicted state time series feature sequence for the next time step is output based on the current time series feature sequence and the forgetting ratio.

[0037] In Figure 8 In the method shown, each time step of the temporal feature sequence is input into the network, and the flow and updating of information are controlled through the input gate, forget gate, and output gate. Step S71 can be used to input the temporal feature sequence as the prediction input gate into the LSTM model network. The function of the forget gate is to determine which information in the memory unit of the previous time step needs to be forgotten. Step S72 can be used to obtain the forgetting ratio that needs to be retained after the current temporal feature sequence and the hidden state of the previous time step enter the forget gate according to formula (18). When the value is close to 0, it indicates that more historical information has been forgotten; when... A value close to 1 indicates that more historical information is retained. The output gate determines how much information to extract from the forget gate and generates the output for the current time step. Step S73 can be used to output the predicted state time-series feature sequence for the next time step based on the current time-series feature sequence and the forgetting ratio.

[0038] In one example of this invention, the lightweight prediction model also needs to be trained. The specific training method may include, for example: Figure 9 The steps are shown. Specifically: In step S81, training data is acquired and fed into the LSTM model to obtain lightweight prediction data; In step S82, the root mean square error value over a time period is obtained according to formula (19) as an evaluation of the LSTM model prediction results. (19) in, Predict feature sequence values ​​for the LSTM model. The actual feature sequence value, For a period of time, This is the root mean square error; In step S83, the parameter model of the LSTM model is optimized based on the evaluation results to complete the training of the LSTM model.

[0039] After the initial inspection results are transmitted to the inspection host, the inspection host at the substation side establishes a communication connection with the provincial cloud computing center, transferring the results data of this task to the provincial cloud platform data center, thus enabling data sharing between the side and the cloud. The inspection hosts at each substation side are network-isolated and independent, belonging to a peer-to-peer architecture. During the provincial remote measurement process of data transmission from multiple substation side devices within the province, high concurrency may occur. The system follows the service request transmission order, with the substation side device that sent the transmission request first transmitting first, and then proceeding to transmit the data from the next substation side device after the first request is completed. Step S5 can be used to upload the inspection data and structured prediction data as status data to the provincial cloud platform computing center.

[0040] The lightweight prediction model deployed on the edge computing nodes of the substation is mainly located inside the substation and focuses on processing real-time data from a single station. This data comes from sensors within the substation, such as temperature, voltage, airflow, and partial discharge. The goal is to monitor and predict defects in equipment within a single substation in real time. The lightweight prediction model uses sensor data to determine the health status of equipment in real time and makes short-term predictions. However, the model has limitations; it is only applicable to the current station in the short term and cannot combine its own data characteristics with those of other stations for long-term defect identification and trend prediction. The cloud platform processes comprehensive data from multiple substations. Its main task is to perform trend analysis and long-term prediction of substation data across the entire region. The model has regional multi-site universality, improving its generalization ability and robustness. Step S6 can be used to build a deep prediction model on the provincial cloud platform computing center based on historical data. The provincial cloud platform GNN model has multiple task head outputs: Task head 1 is risk level classification (multi-classification), outputting the equipment health level (normal / suspicious / warning / high risk); Task head 2 is maintenance strategy recommendation, outputting suggested handling measures (such as review, manual inspection, dispatching, etc.). By deploying a dual-model approach of Transformer + GNN, Transformer is used for long-term trend prediction, while GNN is used for modeling device structure dependence, thereby improving model efficiency and generalization ability, adapting to the output requirements of multiple scenarios, and forming the intelligent modeling core of the provincial cloud platform.

[0041] Step S7 can be used to predict the status data uploaded by each edge computing node, output the defect prediction result, and provide a corresponding operation and maintenance strategy recommendation based on the defect prediction result. The method for providing the corresponding operation and maintenance strategy recommendation can be of various forms known to those skilled in the art. In one example of the present invention, the method for providing the corresponding operation and maintenance strategy recommendation may include, for example... Figure 10 The steps are shown. Specifically: In step S91, the current device status data is obtained according to formula (20). (20) in, It is equipment At any moment state, It is equipment Other dimensions of data, It is equipment At any moment Input features, It is an aggregate function used for synthesizing devices. Status information in various dimensions; In step S92, the state data is classified using a fully connected layer and a softmax activation function, and the prediction result of the device is obtained according to formula (21). ,(twenty one) in, and These are the weights and biases of the fully connected layer. For the predicted results; In step S93, an operation and maintenance strategy recommendation is output based on the prediction results.

[0042] In Task Head 1 detection, GNN processes data using a graph structure. Each device is treated as a node in the graph, and nodes are connected by edges, representing dependencies between devices. Each device (graph node) status It will be updated based on the status information of other dimensions of its device. Step S91 can be used to obtain the status data of the device at the current time according to formula (20). GNN passes the status information of each device to its neighboring devices through the message passing mechanism, and updates the health representation of the current device according to the status of the neighboring devices. This process is carried out continuously in multiple graph convolutional layers until the model converges. Once the device status is updated, the GNN model will classify the risk level according to the updated device status. The output of Task 1 is the risk level of the device, which is classified through a fully connected layer and a softmax activation function. Step S92 can be used to classify the status data through a fully connected layer and a softmax activation function, and obtain the prediction result of the device according to formula (21). Step S93 can be used to output the operation and maintenance strategy recommendation according to the prediction result. In the detection of Task Head 2, the risk level result of Task 1 is output and the device is compared with the risk level of Task Head 2. Based on the equipment type and current real-time data across various dimensions, combined with substation operation and maintenance procedures, the union and deduplication of the corresponding substation operation and maintenance procedure steps under each dimension are calculated, and operation and maintenance strategy recommendations are output.

[0043] Step S8 can be used to distribute the defect prediction results and corresponding operation and maintenance strategies to the corresponding edge computing nodes. Step S9 can be used to detect the equipment based on the defect prediction results and corresponding operation and maintenance strategies, and feed the monitoring results back to the edge computing nodes and the provincial cloud platform computing center, so as to optimize the lightweight prediction model deployed on the edge computing nodes and the deep prediction model of the provincial cloud platform computing center.

[0044] The purpose of this invention, based on the above technical solutions, is to provide a method for predicting substation hidden dangers and defects based on cloud-edge collaboration. By combining edge computing with deep modeling in the cloud, it achieves hierarchical processing and distributed intelligent prediction of multi-source equipment data. Deploying a lightweight prediction model on the substation-side edge computing nodes enables rapid assessment and real-time early warning of critical equipment status, improving system response speed and operational efficiency. Constructing a deep prediction model integrating Transformer and graph neural networks at the provincial cloud platform computing center enhances the medium- and long-term prediction capabilities of equipment operating trends, improving the accuracy and stability of prediction results. This invention effectively improves the prediction level of substation equipment defects and intelligent operation and maintenance capabilities, and is applicable to power system status monitoring and risk early warning in multiple scenarios, improving operation and maintenance efficiency.

[0045] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0049] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0050] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0051] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0052] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0053] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A substation hidden danger defect trend prediction method based on cloud edge collaboration, characterized in that, The prediction method comprises: obtaining inspection data of a transformer substation by a patrol host, and deploying an edge computing node on the transformer substation side; preprocessing the inspection data; performing preliminary fault judgment on the edge computing node according to the obtained inspection data; sending the inspection data after the fault judgment into a lightweight prediction model in the edge computing node, performing preliminary defect trend prediction on the inspection data, and obtaining structured prediction data; uploading the inspection data and the structured prediction data as state data to a provincial cloud platform computing center; constructing a deep prediction model in the provincial cloud platform computing center according to historical data; predicting the state data uploaded by each edge computing node, outputting a defect prediction result, and giving a corresponding operation and maintenance strategy recommendation according to the defect prediction result; downloading the defect prediction result and the corresponding operation and maintenance strategy recommendation to the corresponding edge computing node; detecting the equipment according to the defect prediction result and the corresponding operation and maintenance strategy, and feeding back the monitoring result to the edge computing node and the provincial cloud platform computing center, so as to optimize the lightweight prediction model deployed by the edge computing node and the deep prediction model of the provincial cloud platform computing center.

2. The prediction method of claim 1, wherein, The preprocessing of the inspection data comprises: normalizing image data of the infrared image in the inspection data to obtain a first infrared image with uniform size; determining whether the first infrared image is a single-channel image; in a case where the first infrared image is determined to be not a single-channel image, converting each pixel of the first infrared image into an integer gray value according to formula (1), ，（1） wherein, is the image is the gray value of the pixel in the image is the temperature value of the pixel in the image is the minimum temperature value of the image is the maximum temperature value of the image, the target range of the gray value is [0, ], 8-bit image corresponds to 0 to 255, therefore = 256; in a case where the first infrared image is determined to be a single-channel image, traversing each pixel of the first infrared image according to formula (2), counting the occurrence frequency of the gray value in a target range, and generating a histogram, ，（2） wherein, is a histogram of gray values in the target range, is an image width, is an image height, is a target range; based on the histogram, obtaining the total number of pixels with a gray value less than or equal to the target range according to formula (3), ，（3） wherein, is a cumulative distribution function, indicating the total number of pixels having a gray value less than or equal to the target range, , is the total number of pixels, is a histogram of gray values less than or equal to the target range; normalizing the total number of pixels according to formula (4) to generate a mapping pixel, ，（4） wherein, represents normalizing the cumulative distribution function to the target gray value range, generating a mapping function, represents the minimum non-zero value in the cumulative distribution function; traversing each pixel of the first infrared image, mapping the gray value of each pixel to the mapping pixel, and generating an enhanced image; determining whether the gray value of each pixel of the enhanced image is in the target range; in a case where the gray value of each pixel of the enhanced image is determined to be in the target range, outputting the enhanced image.

3. The prediction method of claim 1, wherein, The preprocessing of the inspection data comprises: obtaining total time series data of the inspection data; obtaining a moving average value of the time series data at the current moment according to formula (5), ，（5） wherein, denotes a moving average at time , is a moving window size, is the th data point in the time series data.

4. The prediction method of claim 2, wherein, the preliminary fault judgment on the edge computing node according to the obtained inspection data comprises: obtaining a two-dimensional temperature matrix of the enhanced image; traversing the temperature of all pixels in the enhanced image according to the two-dimensional temperature matrix to obtain a highest temperature value; obtaining a temperature difference value between the highest temperature value and a reference temperature; when the temperature difference value is detected to be greater than a first threshold value, an abnormal heating hidden danger exists, and a warning is issued; obtaining a local discharge sequence at the current moment according to formula (6), ，（6） wherein, is the entire sequence of partial discharge signals, is the amplitude of the th pulse, is the number of detected discharge pulses within the current acquisition period; When the number of discharge pulses and the pulse amplitude in the partial discharge sequence are greater than preset values, there is a tendency of insulation breakdown, and a warning is issued; According to formula (7), the gas concentration value of the substation at the current time is obtained, ，（7） wherein, is the time is the gas concentration vector collected by the substation oil chromatographic detector at time is the hydrogen concentration, is the methane concentration, is the acetylene concentration, and is the ethylene, ethane concentration; According to formula (8), the gas concentration of the gas in the gas concentration value is diagnosed, and a gas concentration ratio is obtained, ，（8） wherein, is the ratio of acetylene concentration to methane concentration, is the acetylene concentration, is the methane concentration; When the gas concentration ratio is greater than a second threshold value, there is a high-energy discharge hidden danger, and a warning is issued; When the gas concentration ratio meets a third threshold value, there is a thermal fault or ordinary aging hidden danger, and a warning is issued.

5. The prediction method of claim 4, wherein, The inspection data after fault judgment is sent to the light prediction model in the edge computing node, the inspection data is preliminarily predicted for defect trend, and structured prediction data is obtained, including: Obtain the inspection data and perform multi-modal feature fusion on the inspection data; The fused inspection data is constructed as a time sequence feature sequence and is input into the trained light prediction model as an input to obtain a corresponding prediction value; According to the prediction value, a comprehensive risk score is obtained by formula (9), ，（9） wherein, represents the overall risk score of the current device, represents the number of key indicators being monitored, represents the weight of the th indicator, represents the score for the predicted value using piecewise scoring. According to the comprehensive risk score, the defect is evaluated.

6. The prediction method of claim 5, wherein, Obtain the inspection data and perform multi-modal feature fusion on the inspection data, including: According to the temperature maximum value, temperature difference value, local signal peak value and gas concentration value at the current time, a feature vector is constructed; The feature vector is input into a fully connected layer, and the attention weight of each sub-vector in the feature vector is obtained according to formula (10) to formula (13), ，（10） ，（11） ，（12） ，（13） wherein, is a bias term, , , and are attention weights of the temperature maximum value, the temperature difference value, the peak value of the local signal, and the concentration value of the gas, respectively, , , is a weight matrix, is a temperature maximum value, is a temperature difference value, is a peak value of a local signal, is a concentration value of a gas; According to formula (14), a feature layer fusion result is obtained, ，（14） wherein, is the feature layer fusion result; According to formula (15), a time sequence feature sequence constructed according to the feature layer fusion result is obtained, ，（15） wherein is a sequence of timing features, is a time point.

7. The prediction method of claim 6, wherein, Obtaining a feature layer fusion result, including: According to the loss function of formula (16), the weight matrix is trained and optimized to obtain the gradient of each weight matrix, ，（16） wherein, is the output of the feature layer fusion, is the number of samples of the current input, is the loss function; According to formula (17), the gradient is updated, ，（17） wherein, is the learning rate, controlling the step size of each update, is the gradient; When the loss function converges to a minimum value, each weight value is determined.

8. The prediction method of claim 5, wherein, The fused inspection data is constructed as a time sequence feature sequence and is input into the trained light prediction model as an input to obtain a corresponding prediction value, including: The time sequence feature sequence as a prediction input gate is input into the network of the LSTM model; According to formula (18), the forgetting proportion that needs to be retained after the time sequence feature sequence at the current time and the hidden state at the previous time enter the forgetting gate is obtained, ，（18） wherein, is a forgetting ratio, is a hidden state of the previous time, is an input feature of the previous time, is a weight matrix of the forgetting gate, is a bias term, is The activation function outputs a value in the range [0, 1] to control the forgetting ratio of the information. According to the current time sequence feature sequence and the forgetting proportion, the prediction state time sequence feature sequence at the next time is output.

9. The prediction method of claim 8, wherein, Training the light prediction model, including: Obtain the training data and input the training data into the LSTM model to obtain the light prediction data; According to formula (19), the root mean square error value in a time period is obtained as an evaluation of the prediction result of the LSTM model, ，（19） wherein, is the predicted feature sequence value for the LSTM model, is the actual feature sequence value, is a time period, is the root mean square error; According to the evaluation result, the parameter model of the LSTM model is optimized to complete the training of the LSTM model.

10. The prediction method of claim 1, wherein, Predicting the state data uploaded by each edge computing node, outputting a defect prediction result, and giving a corresponding operation and maintenance strategy recommendation according to the defect prediction result, including: According to formula (20), the state data of the device at the current time is obtained, ，（20） wherein, is a device at a time point in time, is a device other dimension data, is a device input feature at a time point in time, is an aggregation function for synthesizing state information of a device in each dimension; Classify the state data through a full connection layer and a softmax activation function, obtain a prediction result of the device according to formula (21), ，（21） wherein, and are weights and biases of the fully connected layer, is the predicted result; Output an operation and maintenance strategy recommendation according to the prediction result.

Images