Load monitoring method based on weak label streaming learning
By combining weak-label streaming learning with bus active power and abnormal operation status detection, the problem of poor adaptability of load monitoring models to new users is solved, and low-cost, high-precision load monitoring is achieved.
Patent Information
- Application Number
- CN202610163980.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-19
AI Technical Summary
Existing load monitoring models are unable to adapt to the distribution and drift of new data. The cost of labeling the operating status of electrical equipment for new users is high and the information is scarce, resulting in poor model adaptability to new user patterns and low monitoring accuracy.
A weak-label-based streaming learning method is adopted, which combines bus active power and abnormal operation status detection. By querying weak labels and fusing strong labels for inference, the load monitoring model is trained in streaming mode, and a recursive compensation branch is embedded for model update.
It reduces the training cost and annotation burden of the load monitoring model, improves the monitoring performance for new users, and avoids the degradation of model capabilities when data distribution drifts.
Smart Images
Figure CN122068665A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of load monitoring, and in particular to a load monitoring method based on weak label streaming learning. Background Technology
[0002] To fully unlock the value of smart meter data, load monitoring aims to infer device-level electricity consumption information based solely on master meter readings. Fine-grained electricity consumption profiling supports applications such as energy-saving optimization, equipment maintenance, and daily life monitoring, while also helping grid operators identify load flexibility and improve grid economics and low-carbon operation. Therefore, this technology has received widespread attention in recent years.
[0003] Given the highly diverse user behaviors and appliance operating characteristics, load monitoring models need to stream learn the distribution of new data to adapt to new users. Existing methods struggle to stream learn new user data and rely on additional sensors or long-term user records for annotation, failing to address the following three key challenges:
[0004] 1. Catastrophic forgetting caused by the distribution drift of newly added data. Because existing load monitoring models can only consider small batches of newly added electricity consumption data that deviate from the overall distribution during training, they are prone to forgetting the overall monitoring capabilities accumulated in the original load monitoring models, resulting in difficulty in continuously adapting to new user patterns and rapid degradation of old patterns.
[0005] 2. The cost and difficulty in labeling the operating status of monitored electrical equipment among new users are high. Currently, the vast majority of methods still rely on independent sensors directly connected to the monitored electrical equipment to obtain strong labels for operating status, resulting in high installation and maintenance costs. A few studies have attempted to obtain weak labels by having users record their electrical equipment usage over a long period, but this requires storing several months of bus data, and users have difficulty accurately recalling past electricity usage behavior, limiting both implementation cost and reliability.
[0006] 3. Information scarcity of weak labels. Weak labels only indicate whether the monitored electrical equipment is running within the sliding window, lacking key information such as the specific running time and duration of the equipment, as well as key information such as the start / stop time and duration. This makes it difficult for the load monitoring model to accurately characterize the operating characteristics of the equipment, thereby weakening its role in model training. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings and deficiencies of existing technologies and propose a load monitoring method based on weak label streaming learning. This method first combines bus active power, abnormal operation status detection criteria, and load monitoring model to query weak operation status labels for new users. Then, it integrates weak operation status labels, bus active power, and load monitoring model to infer strong operation status labels. Finally, it performs streaming training on the load monitoring model. This method can adapt to the electricity consumption patterns of new users with low labeling cost in streaming data, providing a highly practical solution for the promotion and application of load monitoring models among large-scale users. It is a key technology for load monitoring.
[0008] To achieve the above objectives, the technical solution provided by this invention is: a load monitoring method based on weakly labeled streaming learning, comprising the following steps:
[0009] 1) Collect the bus active power of power users based on smart meters, and collect the active power of the power devices to be monitored based on independent sensors directly connected to the devices to be monitored. Combine the bus active power of power users and the active power of the devices to be monitored to form a training dataset.
[0010] 2) Based on the training dataset, establish an anomaly detection criterion for the operating status of the monitored electrical equipment based on Mahalanobis distance, train and output a load monitoring model that outputs the probabilities of strong and weak operating status of the monitored electrical equipment, and search for a discretization threshold to discretize the probabilities of strong and weak operating status of the monitored electrical equipment into estimated strong and weak labels of the operating status of the monitored electrical equipment.
[0011] 3) Based on the bus active power collected by the smart meter, and combined with the abnormal operation status detection criteria of the monitored electrical equipment in step 2), the load monitoring model and the discretized threshold, query the new user for weak labels of the operation status of the monitored electrical equipment until the preset query budget is reached.
[0012] 4) Based on the weak labels of the operating status of the electrical equipment to be monitored queried in step 3), the bus active power of the new user, and the load monitoring model and discretization threshold in step 2), calculate the inferred strong labels of the electrical equipment to be monitored.
[0013] 5) Based on the inferred strong labels of the electrical equipment to be monitored in step 4) and the bus active power of the new user collected in step 3), the load monitoring model in step 2) is stream-trained. The load monitoring model after stream training outputs the probability of strong operating state and the probability of weak operating state of the electrical equipment to be monitored. Then, based on the discretization threshold in step 2), the estimated strong and weak labels of the operating state of the electrical equipment to be monitored are obtained, so as to realize the load monitoring of the electrical equipment to be monitored by the new user.
[0014] Furthermore, in step 1), the sampling interval for active power by both the smart meter and the independent sensor is... Bus active power refers to the active power at the power line entry point of the electricity user; the training dataset consists of data with a width of... The sliding step size is The active power data obtained by cutting the sliding window in chronological order from front to back, along with strong and weak labels for the operating status of the monitored electrical equipment, are represented as follows:
[0015] ;
[0016] In the formula, It refers to the first Training dataset for a number of electricity users Indicates the first The electricity user in the first The active power of the bus in each sliding window. Indicates the first The electricity user in the first The active power of the electrical equipment to be monitored in each sliding window. Indicates the first The electricity user in the first A sliding window displays a strong label indicating the operating status of the electrical equipment to be monitored. Indicates the first The electricity user in the first The monitoring system uses a sliding window to identify weak labels for the operating status of the electrical equipment. Strong labels refer to the binary operating status sequence of the equipment at each sampling point within the sliding window. If the equipment is operating at the corresponding sampling point, the strong label is 1; otherwise, it is 0. Weak labels refer to the binary operating status of the equipment within the sliding window. If the equipment is operating within the sliding window, the weak label is 1; otherwise, it is 0. This represents the total number of sliding windows in the training dataset.
[0017] Furthermore, in step 2), establishing the abnormal operation status detection criterion for the monitored electrical equipment based on Mahalanobis distance refers to dividing the strong labels of the operating status of the monitored electrical equipment in each sliding window of the training dataset into a set of continuous operating segments with a value of 1. A collection of operational segments with strong labels representing the operating status of the monitored electrical equipment in a sliding window. , Refers to the first A segment of operation in which the strong tag of the operating status of the electrical equipment to be monitored is continuously 1. Refers to the first The total number of operating segments in which the strong tag of the operating status of the monitored electrical equipment in each sliding window is continuously 1;
[0018] The average length, average interval, and total number of operating segments with consecutive strong labels of 1 for the operating status of the monitored electrical equipment are statistically analyzed to form a set of operating mode feature vectors. , Refers to the first Feature vectors of the operating modes of each sliding window;
[0019] K-Means clustering is performed using Mahalanobis distance as the metric. Feature vector of the operating mode of a sliding window With a single cluster center Mahalanobis distance Represented as:
[0020] ;
[0021] In the formula, Indicates the first The covariance matrix of each cluster;
[0022] The number of clusters is determined by Within the range, the search objective is to minimize the negative log-likelihood cost of assigning feature vectors of each operating mode to a single cluster center. Feature vector of the operating mode of a sliding window Assigned to the Negative log-likelihood cost of each cluster It is expressed as follows:
[0023] ;
[0024] In the formula, Indicates the first The Gaussian probability density function of each cluster. Belongs to the The Gaussian probability density function of each cluster is expressed as follows:
[0025] ;
[0026] After K-Means clustering of the operating mode feature vectors, calculate the Mahalanobis distance set from all operating mode feature vectors to the nearest cluster center for the electrical equipment to be monitored. Mahalanobis distance set quantiles in As an anomaly detection threshold, when the Mahalanobis distance from the feature vector of the running mode to the nearest cluster center is greater than the anomaly detection threshold, the running mode of the sliding window data is judged as abnormal; otherwise, it is judged as normal.
[0027] The load monitoring model that outputs the probabilities of strong and weak operating states of the monitored electrical equipment refers to a function mapping that takes the bus active power of a single sliding window as input and the probabilities of strong and weak operating states of the monitored electrical equipment within the single sliding window as output, expressed as:
[0028] ;
[0029] In the formula, This represents the load monitoring model that outputs the probabilities of strong and weak operating states of the monitored electrical equipment. Indicates the first The electricity user in the first The probability of a strong operating state of the monitored electrical equipment in a sliding window. Indicates the first The electricity user in the first The probability of weak operating status of the monitored electrical equipment in each sliding window;
[0030] The load monitoring model that trains and outputs the probabilities of strong and weak operating states of the monitored electrical equipment refers to a model based on a training dataset. By minimizing the following loss function To optimize and obtain Parameters:
[0031] ;
[0032] In the formula, This represents the total number of electricity users in the training dataset. Represents the binary cross-entropy loss function;
[0033] The search discretization threshold, which estimates the strong and weak operating state probabilities of the monitored electrical equipment by discretizing them, refers to the strong and weak operating state labels of the monitored electrical equipment based on the training dataset. Iterate through the probability thresholds of the strong operating state of the electrical equipment to be monitored. 1. Probability threshold of weak operating state of the electrical equipment to be monitored When the first The electricity user in the first Probability of strong operating state of the monitored electrical equipment in a sliding window At that time, the first The electricity user in the first Strong labeling of the estimate of a sliding window ,otherwise When the first The electricity user in the first Probability of weak operating state of the monitored electrical equipment in a sliding window At that time, the first The electricity user in the first Weak labeling of the estimated power consumption of the monitored equipment in a sliding window. ,otherwise Choice makes Minimize the probability threshold of the monitored electrical equipment in a strong operating state and the probability threshold of weak operating state of the electrical equipment to be monitored As a discretization threshold.
[0034] Furthermore, in step 3), the acquisition refers to the smart meter collecting data at sampling intervals. Real-time acquisition of bus active power of new users, and arranging it into a data structure with a width of [missing value] in chronological order. Non-overlapping sliding window data.
[0035] Furthermore, step 3) includes the following steps performed on the data of each sliding window:
[0036] 3.1) When the probability of the weak operating state of the monitored electrical equipment output by the load monitoring model in step 2) is located at... If the query is within the specified range, continue with step 3.2; otherwise, do not perform a query in the current sliding window and end the execution of step 3).
[0037] 3.2) Based on the discretization threshold in step 2), Discretize the probability of strong operating conditions of the monitored electrical equipment output by the load monitoring model to obtain the estimated strong operating condition label of the monitored electrical equipment. ;
[0038] Strong labeling of the estimated operating status of the electrical equipment to be monitored Divide the runtime segments into a set of consecutive 1s, the first... A set of operating segments in which the strong label of the operating status of the monitored electrical equipment in each sliding window is continuously 1. ,in, Indicates the first The total number of operating segments in which the strong tag of the operating status of the monitored electrical equipment in each sliding window is continuously 1. Indicates the first The first sliding window A segment of operation where the strong tag of the operating status of the electrical equipment to be monitored is continuously 1, and It consists of the starting sampling point index and the ending sampling point index;
[0039] merge Interval less than The fragments were then filtered. Medium length less than The fragment is multiplied by the rated power of the electrical equipment to be monitored to obtain the estimated active power sequence set of the electrical equipment to be monitored, the first of which is the first. A set of estimated active power sequences for the monitored electrical equipment with sliding windows. ,in, Indicates the first The total number of estimated active power sequences of the monitored electrical equipment in each sliding window. Indicates the first The first sliding window Estimated active power sequence of each monitored electrical device;
[0040] statistics Sequences and relaxation factors The percentage of sampling points whose sum of the products is greater than the bus active power at the corresponding moment of the current sliding window. ,like Greater than the threshold If the query is successful, proceed to step 3.3; otherwise, the current sliding window will not perform a query, and the execution of step 3) will end.
[0041] 3.3) Statistics The average length of the operating segments in which the strong tag of the operating status of the monitored electrical equipment is continuously 1. and average interval ,and Together they form the operating mode feature vector, the first Feature vector of the operating mode of a sliding window ;
[0042] The abnormal operating status detection criteria of the monitored electrical equipment in step 2) are used to determine the situation. If there is an anomaly, query the user for the weak tag of the operating status of the monitored electrical equipment in the current sliding window; otherwise, do not query the current sliding window.
[0043] The query budget refers to the preset maximum number of queries. When the cumulative number of queries achieve At that time, the streaming acquisition of active power data on the new user bus is stopped.
[0044] Furthermore, step 4) includes the following steps:
[0045] 4.1) When the weak label of the operating status of the electrical equipment to be monitored in step 3) is 0, set all the strong labels of the operating status of the electrical equipment to be monitored estimated in step 3) to 0;
[0046] 4.2) For the set of estimated active power sequences of the electrical equipment to be monitored obtained in step 3.2), if the set of estimated active power sequences of the electrical equipment to be monitored is not empty, then filter and relax the set first. The percentage of sampling points whose sum of the products is greater than the active power of the bus corresponding to the sliding window. The sequence is then filtered, and the changes in bus active power corresponding to the start and end sampling point indices are found to be greater than the rated power of the monitored equipment. If the sequence is filtered, the strong label of the estimated operating status of the monitored electrical equipment corresponding to the filtered sequence is set to 0, and proceed to step 4.3); if the set of estimated active power sequences of the monitored electrical equipment is empty, and the weak label of the operating status of the monitored electrical equipment in the single sliding window obtained in step 3) is 1, then it is recorded as the electrical equipment to be activated, and proceed to step 4.4).
[0047] 4.3) For the remaining set of estimated active power sequences of the electrical equipment to be monitored in step 4.2), each sequence is expanded forward and backward by a certain proportion, starting from the initial sampling point index and the end sampling point index, respectively. The sequence length, up to the active power of the bus corresponding to the sliding window. The absolute value of the order difference is greater than the rated power and relaxation factor of the monitored electrical equipment. The product of the two sequences is used to set the strong label of the estimated operating status of the monitored electrical equipment corresponding to the expanded sequence to 1, thus ending the execution of step 4).
[0048] 4.4) Select the rated power and relaxation factor of the electrical equipment to be activated in step 4.2). The product of the products ... The fragments are then filtered to find those with a length less than [a certain value]. After the segment, the estimated operating status label of the monitored electrical equipment corresponding to the segment whose length is closest to the average length is set to 1;
[0049] The estimated strong label of the operating status of the electrical equipment to be monitored after processing in steps 4.1) to 4.4) is used as the inferred strong label of the electrical equipment to be monitored and output to step 5).
[0050] Furthermore, in step 5), the streaming training refers to embedding a compensation branch into the training load monitoring model in step 2). Then, the compensation branch parameters are solved recursively, and the compensation branch... Embedded load monitoring model Represented as:
[0051] ;
[0052] ;
[0053] In the formula, This represents the Sigmoid activation function. Indicates the first The electricity user in the first Feature vectors of a sliding window and The feature extraction part and the fully connected layer part of the load monitoring model in step 2) are respectively represented, and they satisfy the following relationship:
[0054] ;
[0055] Based on the inferred strong tags of the monitored electrical equipment in step 4) and the bus active power acquired by streaming data in step 3), the following formula is recursively solved. parameters :
[0056] ;
[0057] ;
[0058] In the formula, Indicates the current update time step. Indicates the compensation branch at time step Parameters at time, Indicates the compensation branch at time step Parameters at time, Represents the identity matrix. Indicates at time step In step 4), the inference of the strong tag of the electrical equipment to be monitored and The residual, It refers to the first The electricity user in the first The feature vectors of the sliding window at time step The value at time, Indicates time step Intermediate variables at time, Indicates time step Intermediate variables and initial values , This is the regularization coefficient.
[0059] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0060] 1. This invention proposes a recursive load monitoring model streaming training method, which embeds a compensation branch for new users into the load monitoring model. By recursively updating the compensation branch, model training can be achieved without storing historical data and with low computational overhead, while avoiding the degradation of the model's load monitoring capability when the distribution of new data shifts.
[0061] 2. This invention integrates three key information types: bus active power, abnormal operation status detection, and load monitoring model. It only initiates queries to users on weak labels of the operating status of the electrical equipment to be monitored for high-learning-value electricity data, which significantly reduces the streaming training cost of the load monitoring model and the burden of electricity data labeling.
[0062] 3. Based on the weak labels of the operating status of the electrical equipment to be monitored obtained by query, this invention combines the bus active power and load monitoring model to calculate the inferred strong labels of the electrical equipment to be monitored, thereby enhancing the density of supervisory information and improving the monitoring performance of the load monitoring model after the addition of new data through streaming learning. Attached Figure Description
[0063] Figure 1 This is a schematic diagram of the logical flow of the method of the present invention. Detailed Implementation
[0064] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.
[0065] like Figure 1 As shown in the figure, this embodiment discloses a load monitoring method based on weak label streaming learning, the specific details of which are as follows:
[0066] 1) The active power of the power user's bus is collected based on smart meters, and the active power of the monitored electrical equipment is collected based on independent sensors directly connected to the monitored equipment. The collected bus active power of the power user and the active power of the monitored electrical equipment are combined to form a training dataset. The sampling interval for active power of both smart meters and independent sensors is [missing information]. Bus active power refers to the active power at the point where the power line enters the household; training dataset From the width as The sliding step size is The active power data of the sliding window, obtained by cutting it from front to back in chronological order, and the strong and weak labels of the operating status of the monitored electrical equipment are represented as follows:
[0067] ;
[0068] In the formula, It refers to the first Training dataset for a number of electricity users Indicates the first The electricity user in the first The active power of the bus in each sliding window. Indicates the first The electricity user in the first The active power of the electrical equipment to be monitored in each sliding window. Indicates the first The electricity user in the first A sliding window displays a strong label indicating the operating status of the electrical equipment to be monitored. Indicates the first The electricity user in the first The monitoring system uses a sliding window to identify weak labels for the operating status of the electrical equipment. Strong labels refer to the binary operating status sequence of the equipment at each sampling point within the sliding window. If the equipment is operating at the corresponding sampling point, the strong label is 1; otherwise, it is 0. Weak labels refer to the binary operating status of the equipment within the sliding window. If the equipment is operating within the sliding window, the weak label is 1; otherwise, it is 0. This represents the total number of sliding windows in the training dataset, adapted to the sampling conditions of smart meters in practical applications. It can be set to 6 seconds. Set to 540, Set it to 128.
[0069] 2) Based on the training dataset in step 1), establish an anomaly detection criterion for the operating status of the monitored electrical equipment based on Mahalanobis distance, train and output the probability of strong operating status and weak operating status of the monitored electrical equipment, and search for a discretization threshold to discretize the probability of strong operating status and weak operating status of the monitored electrical equipment into estimated strong label and estimated weak label of the operating status of the monitored electrical equipment.
[0070] The establishment of an anomaly detection criterion for the operating status of the monitored electrical equipment based on Mahalanobis distance refers to dividing the strong labels of the operating status of the monitored electrical equipment in each sliding window of the training dataset into a set of continuous operating segments with values of 1. A collection of operational segments with strong labels representing the operating status of the monitored electrical equipment in a sliding window. , Refers to the first A segment of operation in which the strong tag of the operating status of the electrical equipment to be monitored is continuously 1. Refers to the first The total number of operating segments in which the strong tag of the operating status of the monitored electrical equipment in each sliding window is continuously 1;
[0071] The average length, average interval, and total number of operating segments with consecutive strong labels of 1 for the operating status of the monitored electrical equipment are statistically analyzed to form a set of operating mode feature vectors. , Refers to the first Feature vectors of the operating modes of each sliding window;
[0072] K-Means clustering is performed using Mahalanobis distance as the metric. Feature vector of the operating mode of a sliding window With a single cluster center Mahalanobis distance Represented as:
[0073] ;
[0074] In the formula, Indicates the first The covariance matrix of each cluster;
[0075] The number of clusters is determined by Within the range, the search objective is to minimize the negative log-likelihood cost of assigning feature vectors of each operating mode to a single cluster center. Feature vector of the operating mode of a sliding window Assigned to the Negative log-likelihood cost of each cluster It is expressed as follows:
[0076] ;
[0077] In the formula, Indicates the first The Gaussian probability density function of each cluster. Belongs to the The Gaussian probability density function of each cluster is expressed as follows:
[0078] ;
[0079] After K-Means clustering of the operating mode feature vectors, calculate the Mahalanobis distance set from all operating mode feature vectors to the nearest cluster center for the electrical equipment to be monitored. Mahalanobis distance set quantiles in As an anomaly detection threshold, the sliding window data's operating mode is judged as abnormal when the Mahalanobis distance from the feature vector of the operating mode to the nearest cluster center is greater than the anomaly detection threshold; otherwise, it is judged as normal. In practical applications... and They can be set to 3 and 8 respectively. It can be set to 0.9.
[0080] The load monitoring model that outputs the probabilities of strong and weak operating states of the monitored electrical equipment refers to a function mapping that takes the bus active power of a single sliding window as input and outputs the probabilities of strong and weak operating states of the monitored electrical equipment within the single sliding window, expressed as:
[0081] ;
[0082] In the formula, This represents the load monitoring model that outputs the probabilities of the monitored electrical equipment in a strong operating state and the probabilities of the monitored electrical equipment in a weak operating state. Indicates the first The electricity user in the first The probability of a strong operating state of the monitored electrical equipment in a sliding window. Indicates the first The electricity user in the first The probability of weak operating status of the monitored electrical equipment in each sliding window;
[0083] The load monitoring model that trains and outputs the probabilities of strong and weak operating states of the monitored electrical equipment refers to a model based on the training dataset. The optimization is achieved by minimizing the following loss function. Parameters:
[0084] ;
[0085] In the formula, This represents the total number of electricity users in the training dataset. Represents the binary cross-entropy loss function;
[0086] The search discretization threshold, which discretizes the probabilities of strong and weak operating states of the monitored electrical equipment into estimated strong and weak operating state labels, refers to the threshold based on the training dataset. Iterate through the probability thresholds of the strong operating state of the electrical equipment to be monitored. 1. Probability threshold of weak operating state of the electrical equipment to be monitored When the first The electricity user in the first Probability of strong operating state of the monitored electrical equipment in a sliding window At that time, the first The electricity user in the first Strong labeling of the estimate of a sliding window ,otherwise When the first The electricity user in the first Probability of weak operating state of the monitored electrical equipment in a sliding window At that time, the first The electricity user in the first Weak labeling of the estimated power consumption of the monitored equipment in a sliding window. ,otherwise Choice makes Minimize the probability threshold of the monitored electrical equipment in a strong operating state and the probability threshold of weak operating state of the electrical equipment to be monitored As a discretization threshold.
[0087] 3) Based on the bus active power collected by the smart meter for new users, and combined with the abnormal operation status detection criteria, load monitoring model, and discretized threshold from step 2), query the new user for weak labels of the operating status of the electrical equipment to be monitored, until a preset query budget is reached. The collection refers to the smart meter collecting data at sampling intervals. Real-time acquisition of bus active power of new users, and arranging it into a data structure with a width of [missing value] in chronological order. Non-overlapping sliding window data, in practical applications It can be set to 6 seconds. Set to 540; this includes the following steps performed on the data for each sliding window:
[0088] 3.1) When the probability of the weak operating state of the monitored electrical equipment output by the load monitoring model in step 2) is located at... If the query is within the specified range, continue with step 3.2; otherwise, do not perform a query in the current sliding window and end the execution of step 3).
[0089] 3.2) Based on the discretization threshold in step 2), Discretize the probability of strong operating conditions of the monitored electrical equipment output by the load monitoring model to obtain the estimated strong operating condition label of the monitored electrical equipment. ;
[0090] Strong labeling of the estimated operating status of the electrical equipment to be monitored Divide the runtime segments into a set of consecutive 1s, the first... A set of operating segments in which the strong label of the operating status of the monitored electrical equipment in each sliding window is continuously 1. ,in, Indicates the first The total number of operating segments in which the strong tag of the operating status of the monitored electrical equipment in each sliding window is continuously 1. Indicates the first The first sliding window A segment of operation where the strong tag of the operating status of the electrical equipment to be monitored is continuously 1, and It consists of the starting sampling point index and the ending sampling point index;
[0091] merge Interval less than The fragments were then filtered. Medium length less than The fragment is multiplied by the rated power of the electrical equipment to be monitored to obtain the estimated active power sequence set of the electrical equipment to be monitored, the first of which is the first. A set of estimated active power sequences for the monitored electrical equipment with sliding windows. ,in, Indicates the first The total number of estimated active power sequences of the monitored electrical equipment in each sliding window. Indicates the first The first sliding window Estimated active power sequence of each monitored electrical device;
[0092] statistics Sequences and relaxation factors The percentage of sampling points whose sum of the products is greater than the bus active power at the corresponding moment of the current sliding window. ,like Greater than the threshold If the query is successful, proceed to step 3.3; otherwise, the current sliding window will not perform a query, and the execution of step 3) will end. In practical applications, It can be set to 0.3;
[0093] 3.3) Statistics The average length of the operating segments in which the strong tag of the operating status of the monitored electrical equipment is continuously 1. and average interval ,and Together they form the operating mode feature vector, the first Feature vector of the operating mode of a sliding window ;
[0094] The abnormal operating status detection criteria of the monitored electrical equipment in step 2) are used to determine the situation. If there is an anomaly, query the user for the weak tag of the operating status of the monitored electrical equipment in the current sliding window; otherwise, do not query the current sliding window.
[0095] The query budget refers to the preset maximum number of queries. When the cumulative number of queries achieve At that time, the streaming acquisition of active power data on the new user bus is stopped.
[0096] 4) Based on the weak label of the operating status of the electrical equipment to be monitored queried in step 3), the bus active power, the load monitoring model and discretization threshold in step 2), calculate the inferred strong label of the electrical equipment to be monitored, including the following steps:
[0097] 4.1) When the weak label of the operating status of the electrical equipment to be monitored in step 3) is 0, set all the strong labels of the operating status of the electrical equipment to be monitored estimated in step 3) to 0;
[0098] 4.2) For the set of estimated active power sequences of the electrical equipment to be monitored obtained in step 3.2), if the set of estimated active power sequences of the electrical equipment to be monitored is not empty, then filter and relax the set first. The percentage of sampling points whose sum of the products is greater than the active power of the bus corresponding to the sliding window. The sequence is then filtered, and the changes in bus active power corresponding to the start and end sampling point indices are found to be greater than the rated power of the monitored equipment. If the sequence is multiples of the filtered sequence, the strong label of the estimated operating status of the monitored electrical equipment is set to 0, and proceed to step 4.3); if the set of estimated active power sequences of the monitored electrical equipment is empty, and the weak label of the operating status of the monitored electrical equipment in the single sliding window obtained in step 3) is 1, then it is recorded as the electrical equipment to be activated, and proceed to step 4.4). In practical applications, It can be set to 3;
[0099] 4.3) For the remaining set of estimated active power sequences of the electrical equipment to be monitored in step 4.2), each sequence is expanded forward and backward by a certain proportion, starting from the initial sampling point index and the end sampling point index, respectively. The sequence length, up to the active power of the bus corresponding to the sliding window. The absolute value of the order difference is greater than the rated power and relaxation factor of the monitored electrical equipment. The product of these terms sets the estimated operating status strong label of the monitored electrical equipment corresponding to the expanded sequence to 1, thus ending the execution of step 4). In practical applications, It can be set to 0.6. It can be set to 2. It can be set to 0.6;
[0100] 4.4) Select the rated power and relaxation factor of the electrical equipment to be activated in step 4.2). The product of the products ... The fragments are then filtered to find those with a length less than [a certain value]. After the segment, the estimated operating status label of the monitored electrical equipment corresponding to the segment whose length is closest to the average length is set to 1;
[0101] The estimated strong label of the operating status of the electrical equipment to be monitored after processing in steps 4.1) to 4.4) is used as the inferred strong label of the electrical equipment to be monitored and output to step 5).
[0102] 5) Based on the inferred strong labels of the electrical equipment to be monitored in step 4) and the bus active power collected in step 3), the load monitoring model in step 2) is stream-trained. The load monitoring model after stream training outputs the strong operating state probability and weak operating state probability of the electrical equipment to be monitored. Based on the discretization threshold in step 2), the estimated strong and weak labels of the operating state of the electrical equipment to be monitored are obtained, thereby realizing the load monitoring of the electrical equipment to be monitored for new users.
[0103] Streaming training refers to embedding a compensation branch into the training load monitoring model in step 2), and then recursively solving for the parameters of the compensation branch. Embedded load monitoring model Represented as:
[0104] ;
[0105] ;
[0106] In the formula, This represents the Sigmoid activation function. Indicates the first The electricity user in the first Feature vectors of a sliding window and The feature extraction part and the fully connected layer part of the load monitoring model in step 2) are respectively represented, and they satisfy the following relationship:
[0107] ;
[0108] Based on the inferred strong tags of the monitored electrical equipment in step 4) and the bus active power collected in step 3), the following formula is recursively solved. parameters :
[0109] ;
[0110] ;
[0111] In the formula, Indicates the current update time step. Indicates the compensation branch at time step Parameters at time, Indicates the compensation branch at time step Parameters at time, Represents the identity matrix. Indicates at time step In step 4), the inference of the strong tag of the electrical equipment to be monitored and The residual, It refers to the first The electricity user in the first The feature vectors of the sliding window at time step The value at time, Indicates time step Intermediate variables at time, Indicates time step Intermediate variables and initial values , This is the regularization coefficient, which is used in practical applications. It can be set to 100.
[0112] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A load monitoring method based on weakly labeled streaming learning, characterized in that, Includes the following steps: 1) Collect the bus active power of power users based on smart meters, and collect the active power of the power devices to be monitored based on independent sensors directly connected to the devices to be monitored. Combine the bus active power of power users and the active power of the devices to be monitored to form a training dataset. 2) Based on the training dataset, establish an anomaly detection criterion for the operating status of the monitored electrical equipment based on Mahalanobis distance, train and output a load monitoring model that outputs the probabilities of strong and weak operating status of the monitored electrical equipment, and search for a discretization threshold to discretize the probabilities of strong and weak operating status of the monitored electrical equipment into estimated strong and weak labels of the operating status of the monitored electrical equipment. 3) Based on the bus active power collected by the smart meter, and combined with the abnormal operation status detection criteria of the monitored electrical equipment in step 2), the load monitoring model and the discretized threshold, query the new user for weak labels of the operation status of the monitored electrical equipment until the preset query budget is reached. 4) Based on the weak labels of the operating status of the electrical equipment to be monitored queried in step 3), the bus active power of the new user, and the load monitoring model and discretization threshold in step 2), calculate the inferred strong labels of the electrical equipment to be monitored. 5) Based on the inferred strong labels of the electrical equipment to be monitored in step 4) and the bus active power of the new user collected in step 3), the load monitoring model in step 2) is stream-trained. The load monitoring model after stream training outputs the probability of strong operating state and the probability of weak operating state of the electrical equipment to be monitored. Then, based on the discretization threshold in step 2), the estimated strong and weak labels of the operating state of the electrical equipment to be monitored are obtained, so as to realize the load monitoring of the electrical equipment to be monitored by the new user.
2. The load monitoring method based on weakly labeled streaming learning according to claim 1, characterized in that, In step 1), the sampling interval for active power by both the smart meter and the independent sensor is... Bus active power refers to the active power at the point where the power line enters the household of the power user. The training dataset consists of datasets with a width of The sliding step size is The active power data obtained by cutting the sliding window in chronological order from front to back, along with strong and weak labels for the operating status of the monitored electrical equipment, are represented as follows: ； In the formula, It refers to the first Training dataset for a number of electricity users Indicates the first The electricity user in the first The active power of the bus in each sliding window. Indicates the first The electricity user in the first The active power of the electrical equipment to be monitored in each sliding window. Indicates the first The electricity user in the first A sliding window displays a strong label indicating the operating status of the electrical equipment to be monitored. Indicates the first The electricity user in the first The monitoring system uses a sliding window to identify weak labels for the operating status of the electrical equipment. Strong labels refer to the binary operating status sequence of the equipment at each sampling point within the sliding window. If the equipment is operating at the corresponding sampling point, the strong label is 1; otherwise, it is 0. Weak labels refer to the binary operating status of the equipment within the sliding window. If the equipment is operating within the sliding window, the weak label is 1; otherwise, it is 0. This represents the total number of sliding windows in the training dataset.
3. The load monitoring method based on weakly labeled streaming learning according to claim 2, characterized in that, In step 2), establishing the abnormal operation status detection criterion for the monitored electrical equipment based on Mahalanobis distance refers to dividing the strong labels of the operating status of the monitored electrical equipment in each sliding window of the training dataset into a set of continuous operation segments with a value of 1. A collection of operational segments with strong labels representing the operating status of the monitored electrical equipment in a sliding window. , Refers to the first A segment of operation in which the strong tag of the operating status of the electrical equipment to be monitored is continuously 1. Refers to the first The total number of operating segments in which the strong tag of the operating status of the monitored electrical equipment in each sliding window is continuously 1; The average length, average interval, and total number of operating segments with consecutive strong labels of 1 for the operating status of the monitored electrical equipment are statistically analyzed to form a set of operating mode feature vectors. , Refers to the first Feature vectors of the operating modes of each sliding window; K-Means clustering is performed using Mahalanobis distance as the metric. Feature vector of the operating mode of a sliding window With a single cluster center Mahalanobis distance Represented as: ； In the formula, Indicates the first The covariance matrix of each cluster; The number of clusters is determined by Within the range, the search objective is to minimize the negative log-likelihood cost of assigning feature vectors of each operating mode to a single cluster center. Feature vector of the operating mode of a sliding window Assigned to the Negative log-likelihood cost of each cluster It is expressed as follows: ； In the formula, Indicates the first The Gaussian probability density function of each cluster. Belongs to the The Gaussian probability density function of each cluster is expressed as follows: ； After K-Means clustering of the operating mode feature vectors, calculate the Mahalanobis distance set from all operating mode feature vectors to the nearest cluster center for the electrical equipment to be monitored. Mahalanobis distance set quantiles in As an anomaly detection threshold, when the Mahalanobis distance from the feature vector of the running mode to the nearest cluster center is greater than the anomaly detection threshold, the running mode of the sliding window data is judged as abnormal; otherwise, it is judged as normal. The load monitoring model that outputs the probabilities of strong and weak operating states of the monitored electrical equipment refers to a function mapping that takes the bus active power of a single sliding window as input and the probabilities of strong and weak operating states of the monitored electrical equipment within the single sliding window as output, expressed as: ； In the formula, This represents the load monitoring model that outputs the probabilities of strong and weak operating states of the monitored electrical equipment. Indicates the first The electricity user in the first The probability of a strong operating state of the monitored electrical equipment in a sliding window. Indicates the first The electricity user in the first The probability of weak operating status of the monitored electrical equipment in each sliding window; The load monitoring model that trains and outputs the probabilities of strong and weak operating states of the monitored electrical equipment refers to a model based on a training dataset. By minimizing the following loss function To optimize and obtain Parameters: ； In the formula, This represents the total number of electricity users in the training dataset. Represents the binary cross-entropy loss function; The search discretization threshold, which estimates the strong and weak operating state probabilities of the monitored electrical equipment by discretizing them, refers to the strong and weak operating state labels of the monitored electrical equipment based on the training dataset. Iterate through the probability thresholds of the strong operating state of the electrical equipment to be monitored.
1. Probability threshold of weak operating state of the electrical equipment to be monitored When the first The electricity user in the first Probability of strong operating state of the monitored electrical equipment in a sliding window At that time, the first The electricity user in the first Strong labeling of the estimate of a sliding window ,otherwise When the first The electricity user in the first Probability of weak operating state of the monitored electrical equipment in a sliding window At that time, the first The electricity user in the first Weak labeling of the estimated power consumption of the monitored equipment in a sliding window. ,otherwise Choice makes Minimize the probability threshold of the monitored electrical equipment in a strong operating state and the probability threshold of weak operating state of the electrical equipment to be monitored As a discretization threshold.
4. The load monitoring method based on weakly labeled streaming learning according to claim 3, characterized in that, In step 3), the acquisition refers to the smart meter collecting data at sampling intervals. Real-time acquisition of bus active power of new users, and arranging it into a data structure with a width of [missing information - likely a value] in chronological order. Non-overlapping sliding window data.
5. The load monitoring method based on weakly labeled streaming learning according to claim 4, characterized in that, Step 3) includes the following steps performed on the data of each sliding window: 3.1) When the probability of the weak operating state of the monitored electrical equipment output by the load monitoring model in step 2) is located at... If the query is within the specified range, continue with step 3.2; otherwise, do not perform a query in the current sliding window and end the execution of step 3). 3.2) Based on the discretization threshold in step 2), Discretize the probability of strong operating conditions of the monitored electrical equipment output by the load monitoring model to obtain the estimated strong operating condition label of the monitored electrical equipment. ; Strong labeling of the estimated operating status of the electrical equipment to be monitored Divide the execution into a set of consecutive segments of 1, the first... A set of operating segments in which the strong label of the operating status of the monitored electrical equipment in each sliding window is continuously 1. ,in, Indicates the first The total number of operating segments in which the strong tag of the operating status of the monitored electrical equipment in each sliding window is continuously 1. Indicates the first The first sliding window A segment of operation where the strong tag of the operating status of the electrical equipment to be monitored is continuously 1, and It consists of the starting sampling point index and the ending sampling point index; merge Interval less than The fragments were then filtered. Medium length less than The fragment is multiplied by the rated power of the electrical equipment to be monitored to obtain the estimated active power sequence set of the electrical equipment to be monitored, the first of which is the first. A set of estimated active power sequences for the monitored electrical equipment with sliding windows. ,in, Indicates the first The total number of estimated active power sequences of the monitored electrical equipment in each sliding window. Indicates the first The first sliding window Estimated active power sequence of each monitored electrical device; statistics Sequences and relaxation factors The percentage of sampling points whose sum of the products is greater than the bus active power at the corresponding moment of the current sliding window. ,like Greater than the threshold If the query is successful, proceed to step 3.3; otherwise, the current sliding window will not perform a query, and the execution of step 3) will end. 3.3) Statistics The average length of the operating segments in which the strong tag of the operating status of the monitored electrical equipment is continuously 1. and average interval ,and Together they form the operating mode feature vector, the first Feature vector of the operating mode of a sliding window ; The abnormal operating status detection criteria of the monitored electrical equipment in step 2) are used to determine the situation. If there is an anomaly, query the user for the weak tag of the operating status of the monitored electrical equipment in the current sliding window; otherwise, do not query the current sliding window. The query budget refers to the preset maximum number of queries. When the cumulative number of queries achieve At that time, the streaming acquisition of active power data on the new user bus is stopped.
6. The load monitoring method based on weakly labeled streaming learning according to claim 5, characterized in that, Step 4) includes the following steps: 4.1) When the weak label of the operating status of the electrical equipment to be monitored in step 3) is 0, set all the strong labels of the operating status of the electrical equipment to be monitored estimated in step 3) to 0; 4.2) For the set of estimated active power sequences of the electrical equipment to be monitored obtained in step 3.2), if the set of estimated active power sequences of the electrical equipment to be monitored is not empty, then filter and relax the set first. The percentage of sampling points whose sum of the products is greater than the active power of the bus corresponding to the sliding window. The sequence is then filtered, and the changes in bus active power corresponding to the start and end sampling point indices are found to be greater than the rated power of the monitored equipment. If the sequence is filtered, the strong label of the estimated operating status of the monitored electrical equipment corresponding to the filtered sequence is set to 0, and proceed to step 4.3); if the set of estimated active power sequences of the monitored electrical equipment is empty, and the weak label of the operating status of the monitored electrical equipment in the single sliding window obtained in step 3) is 1, then it is recorded as the electrical equipment to be activated, and proceed to step 4.4). 4.3) For the remaining set of estimated active power sequences of the electrical equipment to be monitored in step 4.2), each sequence is expanded forward and backward by a certain proportion, starting from the initial sampling point index and the end sampling point index, respectively. The sequence length, up to the active power of the bus corresponding to the sliding window. The absolute value of the order difference is greater than the rated power and relaxation factor of the monitored electrical equipment. The product of the two sequences is used to set the strong label of the estimated operating status of the monitored electrical equipment corresponding to the expanded sequence to 1, thus ending the execution of step 4). 4.4) Select the rated power and relaxation factor of the electrical equipment to be activated in step 4.2). The product of the products ... The fragments are then filtered to find those with a length less than [a certain value]. After the segment, the estimated operating status label of the monitored electrical equipment corresponding to the segment whose length is closest to the average length is set to 1; The estimated strong label of the operating status of the electrical equipment to be monitored after processing in steps 4.1) to 4.4) is used as the inferred strong label of the electrical equipment to be monitored and output to step 5).
7. The load monitoring method based on weakly labeled streaming learning according to claim 6, characterized in that, In step 5), the streaming training refers to embedding a compensation branch into the training load monitoring model in step 2). Then, the compensation branch parameters are solved recursively, and the compensation branch... Embedded load monitoring model Represented as: ； ； In the formula, This represents the Sigmoid activation function. Indicates the first The electricity user in the first Feature vectors of a sliding window and These represent the load monitoring models in step 2). The feature extraction part and the fully connected layer part satisfy the following relationship: ； Based on the inferred strong tags of the monitored electrical equipment in step 4) and the bus active power acquired by streaming data in step 3), the following formula is recursively solved. parameters : ； ； In the formula, Indicates the current update time step. Indicates the compensation branch at time step Parameters at time, Indicates the compensation branch at time step Parameters at time, Represents the identity matrix. Indicates at time step In step 4), the inference of the strong tag of the electrical equipment to be monitored and The residual, It refers to the first The electricity user in the first The feature vectors of the sliding window at time step The value at time, Indicates time step Intermediate variables at time, Indicates time step Intermediate variables and initial values , This is the regularization coefficient.