Ammeter load real-time monitoring method based on edge calculation

Abstract

The invention discloses an electric meter load real-time monitoring method based on edge calculation. The method comprises the following steps: acquiring load data in real time at an electric meter end and constructing an initial hidden layer network structure; determining an initial reference threshold value of the load data change event; detecting load data change in real time by using an event triggering mechanism and generating a triggering signal; dynamically adjusting a hidden layer node structure based on real-time load data; updating the output weight matrix by adopting a sequential learning method and predicting load data in real time; identifying a key data event according to a prediction result and executing compression packaging; and uploading the packaged key data to a remote monitoring center. According to the invention, local real-time processing of load data is realized, data transmission delay is effectively reduced, and monitoring precision and real-time performance are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of real-time monitoring of power load, and particularly relates to a meter load real-time monitoring method based on edge computing. BACKGROUND

[0002] In recent years, with the continuous expansion of the power system and the rapid development of smart grid technology, real-time monitoring of meter load data has become an important technical means to ensure the safe and stable operation of the power system. At present, the commonly used method is to realize the monitoring of meter load data through a centralized collection mode, that is, the load data collected by the meter is uploaded to a remote data center for unified analysis and processing and decision-making at the center server. It usually includes collecting load data from a large number of meters through a communication network at regular intervals or periodically, and then using conventional data analysis algorithms to analyze and process the load data in the remote data center to determine the load operating state, identify abnormal load conditions and develop appropriate decision-making measures.

[0003] However, the centralized monitoring method has obvious technical defects and deficiencies: Because the load data needs to be transmitted to the data center in large scale and high frequency, the communication network bandwidth pressure increases significantly, which in turn produces high data transmission delay, reducing the real-time performance of the meter load data monitoring; Because data processing and analysis are concentrated in the remote data center, the system response time is significantly delayed, which is not conducive to timely detection and response to abnormal load events at the meter end; Traditional monitoring algorithms generally use fixed model structures and lack adaptive ability, which cannot adapt to the actual scenario of dynamic changes in meter load characteristics over time, thereby reducing the accuracy of load prediction and the reliability of system monitoring.

[0004] The centralized data processing mode is difficult to effectively filter redundant load data, further increasing the data center operation and maintenance cost and data storage cost, and the economy is poor.

[0005] Therefore, how to provide a meter load real-time monitoring method based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide a meter load real-time monitoring method based on edge computing. In view of the problem of high data transmission delay and insufficient real-time performance of the existing centralized processing mode of meter load data, an edge computing node is set up at the meter end to collect load data in real time and dynamically adjust the hidden layer network structure, and through an event triggering mechanism combined with an online sequential extreme learning machine, local real-time processing of meter load data and compression and packaging upload of key data events are realized. The present application has the technical effects of effectively reducing data transmission delay, improving load monitoring real-time performance and accuracy.

[0007] The method for real-time monitoring of electric meter load based on edge computing according to the embodiment of the application comprises: Real-time collection of electric meter load data at the electric meter end, and generation of an initial hidden layer node structure by an improved online sequential extreme learning machine model in the edge computing node at the electric meter end; Off-line analysis of historical load data by using the initial hidden layer node structure to determine an initial reference threshold value of the event triggering threshold value in the event triggering mechanism; Based on the initial reference threshold value, real-time calculation of the variation of the electric meter load data at the current time and the previous time by the event triggering mechanism, and generation of an event triggering signal if the variation exceeds the initial reference threshold value; After receiving the event triggering signal, the improved online sequential extreme learning machine model analyzes the corresponding electric meter load data feature distribution in real time, dynamically adjusts the hidden layer node structure, and obtains a real-time updated hidden layer network structure; Sequential online learning and prediction of the electric meter load data by using the real-time updated hidden layer network structure, real-time calculation of the prediction error value, and execution of local optimization correction of the output weight if the prediction error value exceeds the allowed range; Prediction of real-time electric meter load data by using the optimized output weight, completion of local decision-making at the electric meter end, extraction of key data event results, and data compression and packaging processing; Uploading of the data event results after compression and packaging processing to the remote monitoring center.

[0008] Optionally, the real-time collection of electric meter load data at the electric meter end and the generation of an initial hidden layer node structure by the improved online sequential extreme learning machine model in the edge computing node at the electric meter end are specifically as follows: Real-time collection of voltage, current and power parameters by a 24-bit resolution analog-to-digital converter built-in at the electric meter end with a sampling period of 1 second to obtain clock-synchronized labeled instantaneous electric meter load data; Construction of an initial random weight matrix based on a singular value decomposition orthogonal initialization method with the instantaneous electric meter load data as input data; Determination of the initial number of hidden layer nodes according to the structure of the initial random weight matrix by using the rank of the historical load data sample matrix; Execution of forward propagation operation of the improved online sequential extreme learning machine model by using the initial random weight matrix and the initial number of hidden layer nodes to obtain a hidden layer output matrix; Obtainment of an initial output weight matrix by ridge regression analysis based on the hidden layer output matrix; The initial output weight matrix and the initial random weight matrix are combined for error analysis, and the matching degree of the hidden layer node structure and the instantaneous electric meter load data is evaluated by residual sum of squares to determine the final initial hidden layer node structure.

[0009] Optionally, the improved online sequential extreme learning machine model comprises an input layer, a data feature space adaptive hidden layer and an output weight online robust self-calibration output layer. The input layer is configured to receive current time electric meter load data, determine the initial random weight matrix between the input layer and the hidden layer according to the input feature dimension and the number of hidden layer nodes; The data feature space adaptive hidden layer is configured to monitor the feature space distribution change of the electric meter load data in real time, dynamically adjust the number of hidden layer nodes according to the feature space change, and update the initial random weight matrix synchronously; The output weight online robust self-calibration output layer is configured to update the output weight in real time according to the hidden layer output matrix and the target value of the current electric meter load data; After updating the output weight each time, the robustness verification is performed in real time, specifically as follows: The prediction error value after the action of the current output weight is calculated in real time, and an error sliding window sequence is constructed. The least square method is used for linear regression fitting of the sliding window sequence to calculate the prediction error trend coefficient in real time; When the prediction error trend coefficient is continuously positive and exceeds the robustness threshold value, the output weight local optimization process is triggered immediately to obtain the optimized output weight matrix.

[0010] Optionally, the initial hidden layer node structure is used to analyze the historical load data offline to determine the initial baseline threshold value of the event trigger threshold value in the event trigger mechanism, specifically as follows: The hidden layer output matrix of the initial hidden layer node structure is used to construct a historical load data sample matrix, which contains instantaneous sampling values of voltage, current and power; The historical load data sample matrix is subjected to principal component analysis, and the principal components corresponding to the number of principal components when the cumulative variance contribution rate reaches the preset threshold value are intercepted to obtain a load data main feature space; The K-means algorithm is used to cluster the load data main feature space, and the load data is divided by calculating the Euclidean distance between the load data and the cluster center, and the cluster center is iteratively updated until convergence; According to the clustering result, a load change pattern is formed, and the load data change gradient value and the standard deviation of each load change pattern are calculated respectively; Based on the three-sigma rule, the load data change gradient value is identified for abnormality, and the load data gradient distribution is obtained after removing the abnormal data. The preset percentage of the gradient distribution of the calculation load data is taken as an initial reference value of the event trigger mechanism, and the initial reference value is taken as an initial benchmark threshold of real-time event detection.

[0011] Optionally, based on the initial benchmark threshold, the event trigger mechanism is used to calculate the variation of the electric meter load data at the current time and the previous time in real time, and if the variation exceeds the initial benchmark threshold, an event trigger signal is generated, specifically as follows: The electric meter load data collected at the current time is subjected to a difference operation with the electric meter load data collected at the previous time to obtain a load variation, and the absolute value of the load variation is calculated; The absolute value of the load variation is compared with the initial benchmark threshold to determine whether the absolute value of the load variation is greater than the initial benchmark threshold; When the absolute value of the load variation is greater than the initial benchmark threshold, an event trigger signal is generated, and the corresponding time stamp and load variation value are recorded, and the event trigger signal and the load variation value are sent to the improved online sequential extreme learning machine model; When the absolute value of the load variation is less than or equal to the initial benchmark threshold, the electric meter end enters a low-power monitoring state, and no signal is sent to the improved online sequential extreme learning machine model; In the low-power monitoring state, the absolute value of the load variation is continuously calculated at a preset time period, and compared with the initial benchmark threshold, until the absolute value of the load variation is greater than the initial benchmark threshold again, the low-power monitoring state is exited, a new event trigger signal is generated, and the new event trigger signal is sent to the improved online sequential extreme learning machine model.

[0012] Optionally, after receiving the event trigger signal, the improved online sequential extreme learning machine model analyzes the corresponding electric meter load data feature distribution in real time, dynamically adjusts the hidden layer node structure, and obtains a real-time updated hidden layer network structure, specifically as follows: After receiving the event trigger signal, the corresponding electric meter load data is extracted, the electric meter load data is mapped based on an initial random weight matrix to obtain a hidden layer output matrix; Residuals are calculated node by node according to the hidden layer output matrix and the target value of the electric meter load data to form a node residual vector; According to the nodes in the node residual vector whose residual absolute values are greater than a node increase threshold, new hidden layer nodes are added in the initial random weight matrix, and the random weights of the new nodes are calculated by singular value decomposition; The new node random weights are combined with the electric meter load data to perform forward propagation calculation, an updated hidden layer output matrix is obtained, and the node residual vector is recalculated; According to the updated hidden layer output matrix, a node residual absolute value is calculated, and when the node residual absolute value continuously falls below a node deletion threshold, a hidden layer node is deleted, and a corresponding node weight is deleted from the initial random weight matrix; According to the updated hidden layer output matrix, an output weight matrix corresponding to the hidden layer node structure is recalculated and updated to obtain a real-time updated hidden layer network structure.

[0013] Optionally, the real-time updated hidden layer network structure is used for sequential online learning and prediction of the electric meter load data, a prediction error value is calculated in real time, and if the prediction error value exceeds an allowable range, local optimization correction of the output weight is performed, specifically: The real-time updated hidden layer network structure is used for feature mapping of the current time electric meter load data to obtain a hidden layer output matrix; Based on the output weight matrix and the covariance matrix of the previous time, the hidden layer output matrix and the actual target value of the current time electric meter load data are used to update the output weight matrix through a sequential least square algorithm; The updated output weight matrix is used for real-time prediction of the current time electric meter load data to obtain a real-time predicted electric meter load data value; The real-time predicted electric meter load data value and the actual target value are differentiated to obtain a real-time electric meter load prediction error value; The real-time electric meter load prediction error value is added to a prediction error value sliding window sequence, and a prediction error trend coefficient is obtained by linear regression fitting of the sliding window sequence using a least square method; When the prediction error trend coefficient continuously is positive and greater than a robustness threshold, the output weight matrix is recalculated to obtain an optimized output weight matrix.

[0014] Optionally, the optimized output weight is used to predict real-time electric meter load data to complete local decision-making at the electric meter end, and key data event results are extracted and processed for data compression and packaging, specifically: The optimized output weight matrix is used to calculate the current time electric meter load data to obtain a current time predicted electric meter load data result; According to the current time predicted electric meter load data result and the actual electric meter load data of the previous time, a load change gradient value is calculated; The load change gradient value is compared with an initial reference threshold, and if the load change gradient value exceeds the initial reference threshold, the current time predicted electric meter load data result is marked as a significant data event result, otherwise it is marked as a non-significant data event result; The significant data event result is subjected to a data compression method combining differential encoding and run-length encoding for data compression; Adding event trigger signal corresponding time stamp, voltage, current and power parameter value and data check information to data encapsulation of significant data event result; Non-significant data event result does not perform data compression and data encapsulation, and is directly discarded, only the encapsulated significant data event data is reserved.

[0015] The beneficial effects of the present application are:

[0016] (1) The present application realizes real-time local analysis of electric meter load data and efficient identification of key events by setting an edge computing node at the electric meter end, combining an event trigger mechanism and an improved online sequential extreme learning machine model, effectively reducing data transmission delay and communication bandwidth occupation, and enhancing the real-time and response timeliness of load monitoring.

[0017] (2) The present application significantly improves the accuracy and robustness of load prediction through adaptive implicit layer node dynamic adjustment mechanism and real-time robust self-calibration of output weight matrix, and shows better adaptability and stability in complex application scenarios of power load abnormal fluctuation.

[0018] (3) In real-time data processing, the present application realizes online dynamic update of output weight matrix through sequential least squares algorithm, effectively solving the technical problems of fixed model structure and insufficient adaptive ability in the prior art, breaking through the real-time bottleneck under the traditional centralized data processing architecture, realizing significant improvement of load data monitoring accuracy and processing efficiency, and effectively improving the safe and stable operation ability of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 A flowchart of an electric meter load real-time monitoring method based on edge computing is provided. DETAILED DESCRIPTION

[0020] The present application will now be further described in detail in conjunction with the accompanying drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0021] Reference Figure 1 An electric meter load real-time monitoring method based on edge computing, comprising: Real-time acquisition of electric meter load data at the current time at the electric meter end, and setting an edge computing node at the electric meter end, using an improved online sequential extreme learning machine model in the edge computing node to determine the random weight matrix between the initial input layer and the hidden layer, and generate the initial hidden layer node structure; offline analysis of historical load data is performed using the initial hidden layer node structure to determine an initial reference value of the meter load data change event triggering threshold, and the initial reference value is set in the event triggering mechanism to obtain an initial reference threshold for real-time event detection; Based on the initial reference threshold, the event triggering mechanism is used to calculate the change amount between the meter load data collected at the current time and the meter load data at the previous time, to obtain a real-time load change amount. The real-time load change amount is compared with the initial reference threshold to determine whether the real-time load change amount exceeds the initial reference threshold. If it does, an event triggering signal is generated and sent to the improved online sequential extreme learning machine model. Otherwise, the meter end low-power consumption monitoring state is maintained. After receiving the event triggering signal, the improved online sequential extreme learning machine model analyzes the feature distribution of the meter load data corresponding to the event triggering signal in real time, dynamically adds or deletes hidden layer nodes according to the analysis result of the feature distribution, matches the current hidden layer node structure with the meter load data features, and obtains a real-time updated hidden layer network structure. Using the real-time updated hidden layer network structure, the improved online sequential extreme learning machine model performs sequential online learning and prediction on the current meter load data, obtains real-time updated output weights, and uses the output weight online robust self-calibration mechanism to calculate the current meter load prediction error value in real time, assesses whether the prediction error value exceeds the preset allowed range, and if it does, performs real-time local optimization correction of the output weights to obtain an optimized output weight matrix. The optimized output weight matrix is used to predict the real-time collected meter load data, and based on the predicted meter load data result, local decision-making for the current load state is completed at the meter end, the data event results judged as significant or critical in the current real-time prediction process are extracted, and data compression and packaging processing are performed. Using the data event results after compression and packaging processing, data upload is performed to the remote monitoring center to realize real-time, small data volume remote transmission of meter load data.

[0022] In this embodiment, the meter load data at the current time is collected in real time at the meter end, and an edge computing node is set at the meter end. An improved online sequential extreme learning machine model in the edge computing node is used to determine a random weight matrix between the initial input layer and the hidden layer, and generate an initial hidden layer node structure, specifically: In the sampling unit built in the electric meter end, the voltage, current and power parameters of the electric meter at the current time are synchronously collected in real time by a built-in 24-bit resolution analog-to-digital converter with a fixed sampling period of 1 second, and the clock signal synchronization marked electric meter load data is obtained by time. In the edge computing node integrated in the electric meter end, the electric meter load data is taken as initial input data, and an initial random weight matrix is constructed based on a singular value decomposition orthogonal initialization method. Specifically, the initial input data is constructed into an initial input feature matrix, and then singular value decomposition is performed on the initial input feature matrix, and the orthogonal matrix obtained by the decomposition is taken as the initial random weight matrix, so as to ensure the orthogonal independence between the nodes of each hidden layer in the initial random weight matrix. Based on the matrix structure of the initial random weight matrix, the initial number of hidden layer nodes is determined. Specifically, a historical load data sample matrix is constructed by using historical load data, the rank value of the historical load data sample matrix is calculated, and the rank value is taken as the initial number of hidden layer nodes, so as to ensure that the initial number of hidden layer nodes determined strictly corresponds to the feature space dimension of historical load data. The forward propagation operation of the improved online sequential extreme learning machine model is performed by using the determined initial random weight matrix and the initial number of hidden layer nodes. Specifically, the electric meter load data collected at the current time is input into the input layer, and then is transformed by the initial random weight matrix and is nonlinearly mapped by using a Sigmoid function node by node to obtain a hidden layer output matrix corresponding to the initial hidden layer node structure. Ridge regression analysis is performed based on the hidden layer output matrix. Specifically, a ridge regression cost function is constructed by introducing a regularization factor λ, an analytical solution is obtained for the ridge regression cost function, and an initial output weight matrix is obtained by analytical solution calculation. The matching degree between the initial hidden layer node structure and the current electric meter load data is verified by joint error analysis of the initial output weight matrix and the initial random weight matrix. Specifically, the residual sum of squares between the hidden layer output matrix and the actual collected electric meter load data at the current time is taken as an index for matching degree evaluation. When the residual sum of squares does not satisfy a pre-set convergence threshold, the number of hidden layer nodes is gradually increased, and the forward propagation and ridge regression analysis calculation are repeatedly performed until the hidden layer node structure satisfies the pre-set convergence threshold condition, so that a final determined initial hidden layer node structure completely matched with the feature space of the current electric meter load data is obtained.

[0023] In this embodiment, the improved online sequential extreme learning machine model includes an input layer, a data feature space adaptive hidden layer and an output weight online robust self-calibration output layer. The input layer is configured to receive current time electric meter load data collected by an electric meter terminal in real time, and determine an initial random weight matrix between the input layer and the hidden layer according to an input feature dimension of the electric meter load data and an initial number of nodes of the hidden layer. The data feature space adaptive hidden layer is configured to monitor a feature space distribution change of the input electric meter load data in real time, and dynamically increase or delete hidden layer nodes according to the feature space distribution change. The specific steps include: calculating a residual sum of squares change rate between the feature space distribution of the current time electric meter load data and the hidden layer node output matrix in real time, when the residual sum of squares change rate exceeds a preset node increase threshold, a new hidden layer node is added in real time, and the initial random weight matrix between the input layer and the hidden layer is updated synchronously; when the residual sum of squares change rate is lower than a preset node deletion threshold, a redundant hidden layer node is deleted in real time, and the initial random weight matrix between the input layer and the hidden layer is updated synchronously, so that the hidden layer structure is always matched with the current electric meter load data feature space in real time. The output weight online robust self-calibration output layer is configured to update the output weight in real time by using a sequential least squares algorithm according to the hidden layer output matrix output by the data feature space adaptive hidden layer and the target value of the current time electric meter load data. The specific steps include: based on the updated output weight and the covariance matrix at the previous time, the Kalman gain matrix is calculated by using the hidden layer output matrix at the current time and the corresponding target value, then the output weight at the current time is updated in real time by using the Kalman gain matrix, and the covariance matrix corresponding to the output weight is updated in real time. The output weight robustness verification is performed in real time after each output weight update is completed. The specific steps include: The prediction error value after the current output weight acts on the current time electric meter load data is calculated in real time, and a sliding window sequence of the prediction error value is constructed, then the sliding window sequence is linearly regressed and fitted by using the least squares method, and the change trend coefficient of the prediction error is calculated in real time. When the prediction error change trend coefficient is continuously positive and exceeds a preset robustness threshold, the output weight local re-optimization process is triggered immediately. The specific method is: the current time electric meter load data feature space distribution and the corresponding hidden layer output matrix are re-analyzed by using the ridge regression analysis, and the optimized output weight matrix is obtained, so as to ensure the stability and robustness of the output weight for the electric meter load data prediction.

[0024] In the embodiment, the initial reference value of the electric meter load data change event trigger threshold is determined by using the initial hidden layer node structure to perform offline analysis on the historical load data, and the initial reference value is set in the event trigger mechanism to obtain an initial reference threshold for real-time event detection. Specifically, An initial hidden layer node structure is determined according to the historical load data, and a hidden layer output matrix of the initial hidden layer node structure is obtained. A principal component analysis is performed on the historical load data sample matrix, a principal component score matrix of the historical load data sample matrix is obtained, and a variance contribution rate of each principal component dimension of the principal component score matrix is calculated. The principal components are cut off according to a principal component number corresponding to an accumulated variance contribution rate of 95%, and a load data main feature space is obtained. The historical load data is clustered by using the cut-off load data main feature space, and the K-means algorithm is used to cluster the load data main feature space. The specific steps include: randomly selecting a clustering center, calculating the Euclidean distance between each load data point and the clustering center and dividing the attribution category, updating the clustering center and repeating the calculation until the position change of the clustering center in two adjacent times is less than a set clustering convergence threshold. According to the clustering result, the load data is divided into different load change modes, and a load data change gradient value is calculated for each load change mode. The load data change gradient value is obtained by difference operation of adjacent time load data, and the standard deviation of the load data change gradient value in each load change mode is calculated. According to the standard deviation of the load data change gradient value corresponding to each load change mode, the historical load data is identified according to the three standard deviation criterion, and the abnormal fluctuation data is removed to obtain the load data gradient distribution after the abnormal fluctuation is removed. The percentile of the gradient value of the load data gradient distribution after the abnormal fluctuation is removed is calculated, and the value corresponding to the 95th percentile of the load data gradient distribution is selected as the initial reference value of the event trigger mechanism. The initial reference value is set as the initial reference threshold used by the real-time event detection. The event trigger mechanism is to calculate the load change between the adjacent time load data in real time, and compare it with the initial reference value. When the real-time load change exceeds the initial reference value, an event trigger signal is generated and sent to the improved online sequential extreme learning machine model to trigger the real-time prediction process.

[0025] In the embodiment, based on the initial reference threshold, the change between the current time load data collected by the event trigger mechanism and the previous time load data is calculated to obtain the real-time load change. The real-time load change is compared with the initial reference threshold to determine whether the real-time load change exceeds the initial reference threshold. If it exceeds, an event trigger signal is generated and sent to the improved online sequential extreme learning machine model. Otherwise, the meter end low-power monitoring state is maintained. performing item-by-item difference operation on the meter load data collected in real time at the current moment and the meter load data collected at the previous moment, to obtain a load change quantity between the meter load data at the current moment and the meter load data at the previous moment; performing absolute value conversion on the load change quantity, to obtain an absolute value of the load change quantity; performing numerical comparison between the absolute value of the load change quantity and an initial reference threshold value, to determine in real time whether the absolute value of the load change quantity is greater than the initial reference threshold value; when the absolute value of the load change quantity is greater than the initial reference threshold value, generating an event trigger signal in real time, recording a time stamp corresponding to the event trigger signal and a load change quantity value, and then immediately transmitting the event trigger signal and the corresponding load change quantity value to the improved online sequential extreme learning machine model; when the absolute value of the load change quantity is less than or equal to the initial reference threshold value, the meter end enters a low-power monitoring state in real time, does not generate an event trigger signal, and does not send a signal to the improved online sequential extreme learning machine model; in the low-power monitoring state, the absolute value of the load change quantity of the adjacent time load data is continuously calculated in real time at a period of 1 second, and is continuously compared with the initial reference threshold value, until the absolute value of the load change quantity is greater than the initial reference threshold value again, the low-power monitoring state is exited, a new event trigger signal is generated, and is sent to the improved online sequential extreme learning machine model.

[0026] In the embodiment, after the improved online sequential extreme learning machine model receives the event trigger signal, the feature distribution of the meter load data corresponding to the event trigger signal is analyzed in real time, the hidden layer nodes are dynamically added or deleted according to the analysis result of the feature distribution, the current hidden layer node structure is matched with the feature of the meter load data, the real-time updated hidden layer network structure is obtained, and specifically: after receiving the event trigger signal, the corresponding meter load data is extracted from the event trigger signal, and the feature mapping of the meter load data is performed based on the initial random weight matrix, to obtain a hidden layer output matrix; residual analysis is performed on the hidden layer output matrix and the corresponding meter load data target value, the residual value between the hidden layer output and the actual meter load data is calculated node by node, and a node residual vector of the current hidden layer node structure is formed; according to the node residual vector, the residual absolute value of each hidden layer node is taken as a node adjustment index, for the case that the residual absolute value is greater than a node addition threshold value, a new hidden layer node is added in the initial random weight matrix, and the random weight corresponding to the new node is calculated by using a singular value decomposition orthogonal initialization method; Based on the random weight corresponding to the added node, forward propagation calculation is performed again in combination with the current electric meter load data to obtain an updated hidden layer output matrix in real time, and the node residual vector is recalculated; The residual absolute value of each node is recalculated based on the updated hidden layer output matrix, and when the residual absolute value of any node continuously falls below the node deletion threshold, the hidden layer node is deleted in real time, and the weight corresponding to the node is deleted from the initial random weight matrix; Based on the real-time updated hidden layer output matrix, the output weight matrix corresponding to the hidden layer node structure at the current time is calculated and updated again to obtain a real-time updated hidden layer network structure.

[0027] In this embodiment, the improved online sequential extreme learning machine model uses the real-time updated hidden layer network structure to perform sequential online learning and prediction on the current electric meter load data, obtains real-time updated output weights, and uses an online robust self-calibration mechanism for output weights to calculate the current electric meter load prediction error value in real time, assesses whether the prediction error value exceeds the preset allowed range, and if so, performs local optimization correction of the output weights in real time to obtain an optimized output weight matrix, specifically: The current time electric meter load data is mapped using the real-time updated hidden layer network structure to obtain a current time hidden layer output matrix; Based on the updated output weight matrix and the corresponding covariance matrix at the previous time, the current time output weight matrix is calculated and updated in real time using the current time hidden layer output matrix and the actual target value of the current time electric meter load data through a sequential least squares algorithm, and the sequential least squares algorithm is specifically: based on the output weight matrix and the covariance matrix at the previous time, the Kalman gain matrix corresponding to the current hidden layer output matrix is calculated in real time, and the output weight matrix and the covariance matrix at the current time are adjusted in real time according to the Kalman gain matrix; The current time electric meter load data is predicted in real time using the current time updated output weight matrix to obtain real-time predicted electric meter load data values; The real-time predicted electric meter load data values and the actual target values of the current time electric meter load data are differenced item by item to obtain real-time electric meter load prediction error values; The real-time electric meter load prediction error values are added to a prediction error value sliding window sequence of a preset length, and all error values in the sliding window sequence are linearly regressed and fitted using the least squares method to obtain a prediction error trend coefficient; The prediction error change trend coefficient is compared with a preset robustness threshold, and when the prediction error change trend coefficient is continuously positive and greater than the robustness threshold, local optimization correction of the output weight matrix is started. The specific method is: performing ridge regression analysis on the current time hidden layer output matrix and the corresponding actual target value, recalculating the output weight matrix, and obtaining the optimized output weight matrix.

[0028] In the embodiment, the optimized output weight matrix is used to predict the real-time collected electric meter load data, and based on the predicted electric meter load data result, local decision for the current load state is completed at the electric meter end, data events results judged as significant or key in the current real-time prediction process are extracted, data compression and packaging processing are performed, specifically: The optimized output weight matrix is used to perform forward propagation calculation on the real-time collected electric meter load data at the current time, and real-time predicted electric meter load data result at the current time is obtained; The real-time predicted electric meter load data result is taken as input, the change gradient value between the current time predicted load data and the actual load data at the last time is calculated, and the current time load change gradient value is obtained; The current time load change gradient value is compared with an initial reference threshold, and when the current time load change gradient value exceeds the initial reference threshold, the current time predicted electric meter load data result is marked as a significant data event result, otherwise it is marked as a non-significant data event result; For the current time predicted electric meter load data result marked as a significant data event result, data compression processing is performed through a preset data compression algorithm in the edge computing node, the data compression algorithm is a data compression method based on the combination of difference encoding and run-length encoding, and compressed significant data event data is generated in real time; For the current time predicted electric meter load data result marked as a significant data event result, data packaging processing is performed using a preset data packaging algorithm in the edge computing node, and data packaging specifically includes adding the timestamp corresponding to the event trigger signal, the voltage, current and power parameter values, and data verification information, and packaged significant data event data is generated in real time; The current time predicted electric meter load data result marked as a non-significant data event result is not subjected to data compression and packaging operation, and the non-significant data event result is directly discarded, and only the packaged significant data event data is retained.

[0029] In the embodiment, the data event result after compression and packaging processing is uploaded to the remote monitoring center, realizing real-time, small data volume remote transmission of electric meter load data, specifically: In the edge computing node integrated at the end of the electric meter, the packaged significant data event data is uploaded in real time to the server of the remote monitoring center through the TCP / IP network link based on the MQTT communication protocol and using the communication mode with the QoS level of 1; the server receives the data and analyzes the data packaging content in real time, extracts the corresponding time stamp, voltage, current and power parameters, and stores them into the remote database, thereby completing the remote real-time monitoring of the electric meter load data, significantly reducing the delay in the data transmission process and reducing the network bandwidth occupation.

[0030] Embodiment 1

[0031] In order to verify the feasibility of the application in implementation, the method of the application is applied to the real-time monitoring task of power load in a smart grid demonstration project in a certain area, so as to evaluate the application effect and technical advantage of the method of the application in the actual electric meter load monitoring and data processing scene. In this actual scene, due to the large number of terminal electric meters in the power grid system, the traditional centralized monitoring method relies on periodic uploading of load data to the remote data center for centralized processing and analysis. Due to the delay of data transmission and the limitation of network bandwidth, the overall system response speed is slow, especially when the load abnormal fluctuation occurs, the actual load change trend cannot be accurately reflected in time, causing the delay of power grid operation decision, and potential risks to the safe operation of the power system.

[0032] In the implementation process, a high-performance edge computing node is first built in the electric meter end, and 24-bit high-resolution analog-to-digital converters are used to collect the real-time load data of the electric meter end voltage, current and power at the current time once per second, forming the clock-synchronized instantaneous load data. Based on the above real-time load data, an improved online sequential extreme learning machine model is used to determine the initial random weight matrix between the input layer and the hidden layer through the singular value decomposition orthogonal initialization method, and according to the principal component analysis and K-means clustering analysis of the historical load data, the initial reference threshold of the load change event is determined. Subsequently, the system uses the initial reference threshold to monitor the load change in real time, and when the real-time load data changes exceed the threshold, an event trigger signal is immediately generated and the subsequent prediction is started, otherwise the system enters a low-power state, effectively reducing unnecessary data processing load.

[0033] After receiving the event trigger signal, the model deeply analyzes the current load data characteristics, and dynamically adjusts the hidden layer network structure based on the residual sum of squares change rate, and increases or deletes hidden layer nodes in real time to continuously ensure that the network structure matches the load feature space. The optimized hidden layer structure is used for real-time load data feature mapping, and the output weight matrix is updated by the sequential least squares algorithm. To ensure the robustness of the model output, the model calculates the load prediction error trend in real time, and when the prediction error trend continues to exceed the robustness threshold, the output weight matrix is locally optimized and corrected by the ridge regression method, further improving the prediction accuracy and reliability.

[0034] After real-time prediction of the electric meter load data based on the optimized output weight matrix, the system calculates the load change gradient in real time, identifies and marks significant data events, performs data compression through a combination algorithm of difference encoding and run-length encoding, and encapsulates the data with the time stamp, voltage, current, power, and data verification information of the load data. The encapsulated significant data events are uploaded to the remote monitoring center in real time through the MQTT protocol, and the remote center quickly completes data unpacking and analysis, effectively reduces transmission delay, and significantly improves the real-time performance and reliability of power grid load monitoring.

[0035] In order to evaluate the technical performance and application effect of the present application, five electric meter load monitoring points in actual application scenarios were selected for long-term tracking test, and the actual load data of each electric meter and the real-time predicted load data of the present application were recorded and compared in detail. The data is shown in Table 1 as follows.

[0036] Table 1 Comparison of real-time prediction and actual measurement performance of electric meter load monitoring Electricity meter number Measured load (kW) Predicted load (kW) Prediction error (kW) Data transmission delay (ms) Bandwidth occupation (kbps) DB-001 325.4 323.7 1.7 85 12.3 DB-002 410.2 412.9 2.7 78 11.9 DB-003 278.8 280.2 1.4 91 13.1 DB-004 356.5 354.8 1.7 83 12.6 DB-005 299.1 297.3 1.8 89 12.8 From the comparison results of the real-time prediction and the actual measurement performance of the ammeter load monitoring in Table 1, it can be observed that the method has stable advantages in key performance indicators such as load prediction accuracy, data transmission delay, and network bandwidth occupation. Specifically, the prediction error between the predicted load value and the actual measured load value of the five monitoring points is less than 3 kW. For example, the actual measured load value of the ammeter numbered DB-003 is 278.8 kW, and the predicted value is 280.2 kW, with an error of only 1.4 kW, which reflects good prediction accuracy. In terms of data transmission delay, the delay time of all monitoring points does not exceed 100 milliseconds, and the delay of the ammeter numbered DB-002 is as low as 78 milliseconds. Compared with the traditional centralized data processing method with an average delay of more than 300 milliseconds, the delay is significantly reduced, effectively ensuring the timeliness of load real-time monitoring. In addition, the network bandwidth occupation of the method is significantly reduced, and the bandwidth occupation of all monitoring points does not exceed 13 kbps, of which the bandwidth occupation of the ammeter numbered DB-002 is 11.9 kbps. Compared with the traditional method with a bandwidth generally higher than 20 kbps, the bandwidth usage efficiency is significantly improved, greatly reducing the network load and data processing cost. In summary, the application of edge computing, event triggering mechanism, and online sequential learning prediction algorithm in the method has achieved obvious effects in improving load data prediction accuracy, reducing data transmission delay, and reducing network resource occupation, and has effectively improved the overall performance of real-time power load monitoring.

[0037] The results of the embodiment verify the significant advantages of the method in real-time power load monitoring, data transmission efficiency, and prediction accuracy, effectively solving the key technical bottlenecks of high data delay and low prediction accuracy in the traditional centralized load monitoring mode, and effectively improving the performance of real-time power load monitoring, with high industrial promotion value.

[0038] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto. Any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for real-time monitoring of electricity meter load based on edge computing, characterized in that, include: Real-time data collection of electricity meter load at the meter terminal; initial hidden layer node structure is generated using an improved online sequential extreme learning machine model in the edge computing node at the meter terminal. Offline analysis of historical load data is performed using the initial hidden layer node structure to determine the initial baseline threshold for the event triggering threshold in the event triggering mechanism; Based on an initial baseline threshold, the change in meter load data between the current time and the previous time is calculated in real time through an event triggering mechanism. If the change exceeds the initial baseline threshold, an event triggering signal is generated. After receiving the event trigger signal, the improved online sequential extreme learning machine model analyzes the characteristic distribution of the corresponding electricity meter load data in real time, dynamically adjusts the hidden layer node structure, and obtains the real-time updated hidden layer network structure. The hidden layer network structure updated in real time is used to perform sequential online learning and prediction of the electricity meter load data. The prediction error value is calculated in real time. If the prediction error value exceeds the allowable range, the local optimization correction of the output weight is performed. The optimized output weights are used to predict real-time electricity meter load data, complete local decision-making at the electricity meter, extract key data event results, and perform data compression and encapsulation processing. The compressed and encapsulated data event results are uploaded to the remote monitoring center.

2. The method for real-time monitoring of electricity meter load based on edge computing according to claim 1, characterized in that, The process of collecting electricity meter load data in real time at the meter and generating an initial hidden layer node structure using an improved online sequential extreme learning machine model in the edge computing node at the meter is as follows: With a sampling period of 1 second, voltage, current and power parameters are collected in real time through the built-in 24-bit resolution analog-to-digital converter at the meter terminal to obtain instantaneous meter load data with clock synchronization mark; Using instantaneous electricity meter load data as input data, an initial random weight matrix is ​​constructed based on the singular value decomposition orthogonal initialization method; Based on the structure of the initial random weight matrix, the initial number of hidden layer nodes is determined using the rank of the historical load data sample matrix; Using the initial random weight matrix and the initial number of hidden layer nodes, the forward propagation operation of the improved online sequential extreme learning machine model is performed to obtain the hidden layer output matrix; Based on the hidden layer output matrix, the initial output weight matrix is ​​calculated through ridge regression analysis. By using joint error analysis of the initial output weight matrix and the initial random weight matrix, the matching degree between the hidden layer node structure and the instantaneous electricity meter load data is evaluated through the residual sum of squares, and the final initial hidden layer node structure is determined.

3. The method for real-time monitoring of electricity meter load based on edge computing according to claim 1, characterized in that, The improved online sequential extreme learning machine model includes an input layer, an adaptive hidden layer for the data feature space, and an online robust self-calibrating output layer for output weights. The input layer is used to receive the current meter load data and determine the initial random weight matrix between the input layer and the hidden layer based on the input feature dimension and the number of hidden layer nodes. The adaptive hidden layer of the data feature space is used to monitor the changes in the distribution of the load data feature space of the electricity meter in real time, dynamically adjust the number of hidden layer nodes according to the changes in the feature space, and synchronously update the initial random weight matrix. The output weight online robust self-calibrating output layer is used to update the output weight in real time based on the hidden layer output matrix and the target value of the current electricity meter load data. After each update of the output weights, a robustness check is performed in real time, specifically as follows: The prediction error value after the current output weight is applied is calculated in real time and an error sliding window sequence is constructed. The least squares method is used to perform linear regression fitting on the sliding window sequence, and the prediction error change trend coefficient is calculated in real time. When the trend coefficient of the prediction error change remains positive and exceeds the robustness threshold, the local optimization process of the output weight is triggered immediately to obtain the optimized output weight matrix.

4. The method for real-time monitoring of electricity meter load based on edge computing according to claim 1, characterized in that, The process of using the initial hidden layer node structure to perform offline analysis of historical load data to determine the initial baseline threshold for the event triggering mechanism is as follows: A historical load data sample matrix is ​​constructed using the hidden layer output matrix of the initial hidden layer node structure. The sample matrix contains instantaneous sampled values ​​of voltage, current, and power. Principal component analysis is performed on the historical load data sample matrix, and the number of principal components corresponding to the cumulative variance contribution rate reaching a preset threshold is extracted to obtain the main feature space of the load data. The K-means algorithm is used to cluster the main feature space of the load data. The load data is divided by calculating the Euclidean distance between the load data and the cluster centers, and the cluster centers are iteratively updated until convergence. Based on the clustering results, load change patterns are formed, and the load data change gradient values ​​and standard deviations for each load change pattern are calculated. Anomalies in the gradient values ​​of load data changes are identified based on the three-standard-deviation criterion, and the gradient distribution of load data is obtained after removing outlier data. The preset percentile of the gradient distribution of the load data is used as the initial reference value for the event triggering mechanism, and the initial reference value is used as the initial baseline threshold for real-time event detection.

5. The method for real-time monitoring of electricity meter load based on edge computing according to claim 1, characterized in that, Based on an initial benchmark threshold, an event-triggered mechanism is used to calculate the change in meter load data between the current and previous moments in real time. If the change exceeds the initial benchmark threshold, an event-triggered signal is generated. Specifically: The load change is obtained by performing a differential calculation between the current load data collected from the meter and the load data collected from the previous time, and the absolute value of the load change is calculated. The absolute value of the load change is compared with the initial baseline threshold to determine whether the absolute value of the load change is greater than the initial baseline threshold. When the absolute value of the load change exceeds the initial baseline threshold, an event trigger signal is generated, and the corresponding timestamp and load change value are recorded. The event trigger signal and load change value are then sent to the improved online sequential extreme learning machine model. When the absolute value of the load change is less than or equal to the initial baseline threshold, the meter enters a low-power monitoring state and does not send signals to the improved online sequential extreme learning machine model. In the low-power monitoring state, the absolute value of the load change continues to be calculated at a preset time period and compared with the initial benchmark threshold until the absolute value of the load change is greater than the initial benchmark threshold again. Then, the low-power monitoring state is exited, a new event trigger signal is generated and sent to the improved online sequential extreme learning machine model.

6. The method for real-time monitoring of electricity meter load based on edge computing according to claim 1, characterized in that, The improved online sequential extreme learning machine model receives the event trigger signal, analyzes the corresponding meter load data feature distribution in real time, dynamically adjusts the hidden layer node structure, and obtains the real-time updated hidden layer network structure, specifically: Upon receiving an event trigger signal, the corresponding meter load data is extracted, and the meter load data is mapped based on the initial random weight matrix to obtain the hidden layer output matrix; Based on the hidden layer output matrix and the target value of the meter load data, the residual is calculated node by node to form the node residual vector; For nodes whose absolute residual value in the node residual vector is greater than the node addition threshold, new hidden layer nodes are added to the initial random weight matrix, and the random weights of the newly added nodes are calculated using singular value decomposition. The forward propagation calculation is performed by combining the random weights of the newly added nodes with the meter load data to obtain the updated hidden layer output matrix and recalculate the node residual vector. The absolute value of the node residual is calculated based on the updated hidden layer output matrix. When the absolute value of the node residual is continuously lower than the node deletion threshold, the hidden layer node is deleted and the corresponding node weight is deleted from the initial random weight matrix. The output weight matrix corresponding to the hidden layer node structure is recalculated and updated based on the updated hidden layer output matrix to obtain the real-time updated hidden layer network structure.

7. The method for real-time monitoring of electricity meter load based on edge computing according to claim 1, characterized in that, The method involves sequential online learning and prediction of meter load data using a real-time updated hidden layer network structure, calculating the prediction error in real time, and performing local optimization correction of the output weights if the prediction error exceeds the allowable range. Specifically: The hidden layer network structure, updated in real time, is used to perform feature mapping on the current meter load data to obtain the hidden layer output matrix; Based on the output weight matrix and covariance matrix of the previous time step, the output weight matrix is ​​updated by using the hidden layer output matrix and the actual target value of the current meter load data through the sequential least squares algorithm. The updated output weight matrix is ​​used to predict the current meter load data in real time, and the real-time predicted meter load data value is obtained. The difference between the real-time predicted electricity meter load data and the actual target value is calculated to obtain the real-time electricity meter load prediction error value; The real-time electricity meter load prediction error value is added to the prediction error value sliding window sequence, and the least squares method is used to perform linear regression fitting on the sliding window sequence to obtain the prediction error change trend coefficient. When the trend coefficient of the prediction error change is consistently positive and greater than the robustness threshold, the output weight matrix is ​​recalculated to obtain the optimized output weight matrix.

8. The method for real-time monitoring of electricity meter load based on edge computing according to claim 1, characterized in that, The process involves using optimized output weights to predict real-time meter load data, completing local decision-making at the meter, extracting key data event results, and performing data compression and encapsulation. Specifically: The optimized output weight matrix is ​​used to calculate the current meter load data to obtain the predicted meter load data for the current time. The load change gradient value is calculated based on the predicted meter load data at the current moment and the actual meter load data at the previous moment. The load change gradient value is compared with the initial baseline threshold. If it exceeds the initial baseline threshold, the predicted meter load data result at the current time is marked as a significant data event result; otherwise, it is marked as a non-significant data event result. For significant data event results, a data compression method combining differential coding and run-length coding is used for data compression; For significant data event results, add the timestamp, voltage, current, and power parameter values ​​corresponding to the event trigger signal, as well as data verification information for data encapsulation; Non-significant data event results are not compressed or encapsulated and are discarded directly, with only the encapsulated significant data event data retained.

Images