Line loss abnormality determination method and system based on power meter data precision compensation

Abstract

The application discloses a line loss abnormality judgment method and system based on electric energy meter data precision compensation, relates to the technical field of data processing, and comprises the following steps: constructing an electric energy meter precision error compensation model, calculating a target electric energy meter, and generating an electric quantity compensation data set; topological analysis is performed on the transformer area, a transformer area topological structure is constructed, line loss calculation is performed according to the transformer area topological structure and the electric quantity compensation data set, and a transformer area line loss residual sequence is generated; time sequence feature decomposition is performed based on the transformer area line loss residual sequence, line loss residual components are obtained for abnormality depth search, abnormality judgment is performed according to the search result, and a line loss abnormality judgment report is obtained. The application solves the technical problem that, in the prior art, line loss calculation is affected by electric energy meter measurement error, leading to low line loss abnormality judgment accuracy, and achieves the technical effects of reducing the interference of measurement error on line loss abnormality judgment results and improving line loss abnormality judgment accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a method and system for determining line loss anomalies based on electricity meter data accuracy compensation. Background Technology

[0002] In distribution area line loss management, it is necessary to statistically analyze the power supply, power sales, and line loss based on the electricity data collected by the electricity meters, and to identify abnormal line loss conditions such as electricity theft, wiring abnormalities, and equipment failures. Because electricity meters are easily affected by factors such as load fluctuations, voltage deviations, current waveform distortion, operating environment, and component aging during actual operation, the collected electricity data may contain measurement errors. These errors are directly transmitted to the line loss calculation process, causing non-abnormal fluctuations caused by meter accuracy deviations to mix into the line loss residuals. This results in inaccurate abnormal characteristics, thus affecting the reliability and accuracy of the line loss anomaly judgment results. Summary of the Invention

[0003] This application provides a method and system for determining line loss anomalies based on electricity meter data accuracy compensation, which is used to address the technical problem in the prior art where the accuracy of line loss anomaly determination is low due to the influence of electricity meter measurement errors on line loss calculation.

[0004] In view of the above problems, this application provides a method and system for determining line loss anomalies based on electricity meter data accuracy compensation.

[0005] The first aspect of this application provides a method for determining line loss anomalies based on electricity meter data accuracy compensation, the method comprising:

[0006] The original electricity time-series data is retrieved to model the metering error of the electricity meter, and an electricity meter accuracy error compensation model is constructed. The target electricity meter is then calculated using the electricity meter accuracy error compensation model to generate an electricity compensation dataset. Topology analysis is performed on the transformer substations to construct the substation topology structure. Line loss is calculated based on the substation topology structure and the electricity compensation dataset to generate a substation line loss residual sequence. Based on the substation line loss residual sequence, time-series feature decomposition is performed to obtain the line loss residual components. Anomaly depth search is then performed, and anomaly determination is made based on the search results to obtain a line loss anomaly determination report.

[0007] A second aspect of this application provides a line loss anomaly determination system based on electricity meter data accuracy compensation, the system comprising: The modeling module retrieves raw electricity time-series data to model electricity meter measurement errors and constructs an electricity meter accuracy error compensation model. The calculation module calculates the target electricity meter using the electricity meter accuracy error compensation model to generate an electricity compensation dataset. The topology analysis module traverses the transformer area to perform topology analysis, constructs the transformer area topology, calculates line loss based on the transformer area topology and the electricity compensation dataset, and generates a transformer area line loss residual sequence. The anomaly detection module performs time-series feature decomposition based on the transformer area line loss residual sequence, obtains line loss residual components, performs anomaly depth search, and determines anomalies based on the search results to obtain a line loss anomaly detection report.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application retrieves original electricity time-series data to model electricity meter measurement errors and constructs an electricity meter accuracy error compensation model. The target electricity meter is then calculated using this model to generate an electricity compensation dataset. A topology analysis is performed across the distribution area to construct its topology. Line loss is calculated based on this topology and the electricity compensation dataset, generating a line loss residual sequence. Time-series feature decomposition is performed on the residual sequence to obtain the line loss components, and an anomaly depth search is conducted. Anomaly determination is then performed based on the search results to obtain a line loss anomaly determination report. This invention addresses the technical problem in existing technologies where electricity meter measurement errors affect line loss calculations, leading to low accuracy in line loss anomaly determination. By constructing an electricity meter accuracy error compensation model to compensate for the target electricity meter data, and performing line loss residual analysis and anomaly search based on the compensated electricity data, the technical effect of reducing the interference of measurement errors on line loss anomaly determination results and improving the accuracy of line loss anomaly determination is achieved. Attached Figure Description

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

[0010] Figure 1 A schematic flowchart of a line loss anomaly determination method based on electricity meter data accuracy compensation provided in an embodiment of this application; Figure 2 This is a schematic diagram of the line loss anomaly determination system based on electricity meter data accuracy compensation provided in an embodiment of this application; Figure 3 The diagram shows the calculation data of the residual sequence of line loss in the transformer substation in the line loss anomaly determination method based on the accuracy compensation of electricity meter data provided in the embodiments of this application.

[0011] Figure labeling: Modeling module 11, Calculation module 12, Topology analysis module 13, Anomaly detection module 14. Detailed Implementation

[0012] This application provides a method and system for determining line loss anomalies based on electricity meter data accuracy compensation. It addresses the technical problem in existing technologies where metering errors affect line loss calculations, leading to low accuracy in determining line loss anomalies. The method constructs an electricity meter accuracy error compensation model to compensate for the target electricity meter data, and then performs line loss residual analysis and anomaly search based on the compensated electricity data. This achieves the technical effect of reducing the interference of metering errors on the line loss anomaly determination results and improving the accuracy of line loss anomaly determination.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a method for determining line loss anomalies based on electricity meter data accuracy compensation, the method comprising: Step S100: Retrieve the original electricity time series data to model the metering error of the electricity meter and construct the electricity meter accuracy error compensation model.

[0016] In this embodiment, the original electricity time-series data is first retrieved, and the original electricity time-series data is segmented using a sliding time window. Active power, voltage deviation, and current waveform distortion rate are calculated separately, and these are combined according to the sliding time window to generate a multi-dimensional operating condition feature parameter set. Next, spatial clustering is performed based on the multi-dimensional operating condition feature parameter set to generate density spatial clustering regions. High-density clustering regions are used as operating condition centers for distribution calculations, defining the feature vector distribution range parameters. The multi-dimensional operating condition feature envelope boundary is determined according to the feature vector distribution range parameters, generating multiple operating condition intervals. Subsequently, error calculation is performed on the target electricity meter based on multiple operating condition intervals using a sliding time window to construct a metering error label. A deep neural network is constructed, and the multi-dimensional operating condition feature parameter set is synchronized to the deep neural network according to multiple operating condition intervals for weighted learning. An initial error compensation model is trained and obtained, and the initial error compensation model is locally updated based on the metering error label to construct an electricity meter accuracy error compensation model.

[0017] Furthermore, the method provided in the application embodiment, which retrieves original electricity time-series data to model the metering error of the electricity meter and constructs an electricity meter accuracy error compensation model, also includes: The original energy time-series data is segmented using a sliding time window to calculate active power; voltage amplitude deviation is calculated based on the original energy time-series data to obtain voltage deviation values; current waveform decomposition is performed based on the original energy time-series data to calculate current waveform distortion rate; the active power, voltage deviation values, and current waveform distortion rate are combined according to the sliding time window to generate a multi-dimensional operating condition feature parameter set; spatial clustering is performed based on the multi-dimensional operating condition feature parameter set to generate density spatial clustering regions, which include high-density clustering regions; the high-density clustering regions are used as the operating condition feature set. The system calculates the distribution of feature vectors based on the operating condition center, defines the distribution range parameters of the feature vectors, determines the multi-dimensional operating condition feature envelope boundary according to the feature vector distribution range parameters, and generates multiple operating condition intervals. Based on the multiple operating condition intervals, it calculates the error of the target energy meter according to a sliding time window and constructs a metering error label. It then constructs a deep neural network, synchronizes the multi-dimensional operating condition feature parameter set to the deep neural network according to the multiple operating condition intervals, performs weighted learning, and trains to obtain an initial error compensation model. Based on the metering error label, it locally updates the initial error compensation model to construct an energy meter accuracy error compensation model.

[0018] In this embodiment, the raw energy time-series data generated by the target energy meter during a continuous sampling period is first retrieved. This raw energy time-series data includes sampling time, cumulative energy value, voltage amplitude, current amplitude, voltage waveform sampling sequence, current waveform sampling sequence, and power factor. The raw energy time-series data is sorted in ascending order according to the sampling time. The window length and sliding step size of the sliding time window are set. The raw energy time-series data within each sliding time window are sequentially extracted according to the window length, and the starting sampling time of adjacent sliding time windows is moved according to the sliding step size. For each sliding time window, the voltage amplitude, current amplitude, and power factor of each sampling point are read. The voltage amplitude, current amplitude, and power factor of the same sampling point are multiplied to obtain the active power of that sampling point. The active power of all sampling points within the sliding time window is added together and then divided by the number of sampling points within the sliding time window to obtain the active power corresponding to that sliding time window.

[0019] Next, voltage amplitude deviation is calculated based on the original power time-series data. The voltage amplitude and preset reference voltage of each sampling point within the same sliding time window are read. The voltage amplitude of each sampling point is subtracted from the preset reference voltage to obtain the voltage amplitude deviation of that sampling point. The absolute value of each voltage amplitude deviation is taken to obtain the voltage deviation amplitude of each sampling point. All voltage deviation amplitudes within the same sliding time window are added together and then divided by the number of sampling points within the sliding time window to obtain the average voltage deviation amplitude. The average voltage deviation amplitude is divided by the preset reference voltage to obtain the voltage deviation value corresponding to that sliding time window.

[0020] Subsequently, current waveform decomposition was performed based on the original power time-series data. The current waveform sampling sequence within the same sliding time window was read, and the current sampling values ​​were arranged in the order of sampling time. For the fundamental frequency, cosine reference values ​​and sine reference values ​​were generated according to the fundamental frequency and each sampling time. Each current sampling value was multiplied by the cosine reference value at the same sampling time and summed to obtain the fundamental cosine projection value. Each current sampling value was multiplied by the sine reference value at the same sampling time and summed to obtain the fundamental sine projection value. The square of the fundamental cosine projection value was added to the square of the fundamental sine projection value, the square root of the sum was taken, and then normalized according to the number of sampling points to obtain the amplitude of the fundamental current component. For each harmonic frequency, harmonic cosine reference values ​​and harmonic sine reference values ​​are generated based on the harmonic order, fundamental frequency, and sampling time. Each current sample value is multiplied by the harmonic cosine reference value at the same sampling time and summed to obtain the harmonic cosine projection value. Similarly, each current sample value is multiplied by the harmonic sine reference value at the same sampling time and summed to obtain the harmonic sine projection value. The square of the harmonic cosine projection value is added to the square of the harmonic sine projection value, and the square root of the sum is taken. This result is then normalized according to the number of sampling points to obtain the amplitude of the corresponding harmonic current component. The amplitude of each harmonic current component is squared, and the sum of all squared results is taken to obtain the overall harmonic amplitude. The overall harmonic amplitude is divided by the amplitude of the fundamental current component to obtain the current waveform distortion rate corresponding to the sliding time window.

[0021] Then, according to the time index of the sliding time window, the active power, voltage deviation, and current waveform distortion rate corresponding to the same sliding time window are aligned one by one, and the aligned active power, voltage deviation, and current waveform distortion rate are used as the three feature dimensions of the same operating condition data. All sliding time windows are processed sequentially to generate a multi-dimensional operating condition feature parameter set. Each operating condition data in the multi-dimensional operating condition feature parameter set includes the sliding time window identifier, active power, voltage deviation, and current waveform distortion rate.

[0022] When performing spatial clustering based on a multi-dimensional set of operating condition feature parameters, the mean and standard deviation of active power, voltage deviation, and current waveform distortion rate are calculated over the entire sliding time window. For each operating condition data point, the active power is subtracted from the mean and then divided by the standard deviation to obtain the standardized active power result; the voltage deviation is subtracted from the mean and then divided by the standard deviation to obtain the standardized voltage deviation result; and the current waveform distortion rate is subtracted from the mean and then divided by the standard deviation to obtain the standardized current waveform distortion rate result. Two operating condition data points are randomly selected, and the differences between them in terms of the standardized active power result, the standardized voltage deviation result, and the standardized current waveform distortion rate result are calculated. The squares of these three differences are then summed, and the square root of the sum is taken to obtain the spatial distance between the two operating condition data points.

[0023] Next, based on the preset neighborhood radius, the number of operating condition data points contained within the neighborhood for each operating condition data point is counted. Operating condition data points whose number of operating condition data points within the neighborhood reaches the preset minimum number of neighborhood points are used as the starting point for cluster expansion. Operating condition data points within the neighborhood are then gradually merged according to spatial distance to generate density spatial clustering regions. For each density spatial clustering region, the number of operating condition data points contained within it is counted. The maximum and minimum values ​​of the active power standardization result, voltage deviation standardization result, and current waveform distortion rate standardization result are determined. The maximum value minus the minimum value for the same feature dimension yields the corresponding value span. The three value spans are multiplied to obtain the feature space range occupied by the density spatial clustering region. The number of operating condition data points contained within the density spatial clustering region is divided by the feature space range to obtain the region density. Density spatial clustering regions whose region density reaches the preset high-density condition are identified as high-density clustering regions.

[0024] Then, using the high-density clustered area as the operating condition center, the normalized results of all active power, voltage deviation, and current waveform distortion rates within the high-density clustered area are summed, and then divided by the number of operating condition data points contained in the high-density clustered area to obtain the center values ​​of active power, voltage deviation, and current waveform distortion rates. Using these center values ​​as references, the deviation magnitude of each operating condition data point within the high-density clustered area in the corresponding feature dimension is calculated. All deviation magnitudes in the same feature dimension are summed, and then divided by the number of operating condition data points contained in the high-density clustered area to obtain the distribution radius of active power, voltage deviation, and current waveform distortion rates. Based on this, the feature vector distribution range parameters are defined.

[0025] When determining the feature envelope boundary of the multidimensional operating condition according to the feature vector distribution range parameter, the active power center value is subtracted from and the active power distribution radius is added to obtain the lower limit boundary and the upper limit boundary of active power; the voltage deviation center value is subtracted from and the voltage deviation distribution radius is added to obtain the lower limit boundary and the upper limit boundary of voltage deviation; the current waveform distortion rate center value is subtracted from and the current waveform distortion rate distribution radius is added to obtain the lower limit boundary and the upper limit boundary of current waveform distortion rate; the feature envelope boundary of the multidimensional operating condition is determined by the lower limit boundary and the upper limit boundary of the above three feature dimensions.

[0026] The multi-dimensional operating condition feature parameter set is divided into intervals according to the multi-dimensional operating condition feature envelope boundary. The normalized results of active power, voltage deviation, and current waveform distortion rate for each operating condition data point are matched with the multi-dimensional operating condition feature envelope boundary. When the normalized results of all three feature dimensions are between the corresponding lower and upper bounds, the operating condition data is assigned to the corresponding operating condition interval. Interval division is performed on the multi-dimensional operating condition feature envelope boundary corresponding to each high-density clustering region, generating multiple operating condition intervals.

[0027] When calculating the error of a target energy meter based on multiple operating condition intervals using a sliding time window, the operating condition interval to which the target energy meter belongs is determined according to the operating condition data corresponding to the sliding time window. The cumulative energy value at the start and end of the sampling time of the sliding time window is read. The cumulative energy value at the start of the sampling time is subtracted from the cumulative energy value at the end of the sampling time to obtain the original energy metering value of the target energy meter within that sliding time window. The standard energy cumulative value at the start and end of the sampling time of the same sliding time window is read from the standard metering device. The standard energy cumulative value at the start of the sampling time is subtracted from the standard energy cumulative value at the end of the sampling time to obtain the standard energy value. The original energy metering value is subtracted from the standard energy value to obtain the metering error value. The metering error value is divided by the standard energy value to obtain the metering error rate. A correspondence is established between the sliding time window identifier, operating condition interval identifier, active power, voltage deviation value, current waveform distortion rate, metering error value, and metering error rate to construct a metering error label.

[0028] Subsequently, a deep neural network is constructed, and the multi-dimensional operating condition feature parameter set is synchronized to the deep neural network according to multiple operating condition intervals for weighted learning, training to obtain the initial error compensation model. In this process, a gated recurrent unit layer and a self-attention layer are constructed and cascaded to form a deep neural network for measurement error learning. Then, temporal analysis is performed on the multi-dimensional operating condition feature parameter set to obtain the temporal dependencies between the multi-dimensional operating condition feature parameters. Next, the multi-dimensional operating condition feature parameters are synchronized to the gated recurrent unit layer according to the temporal dependencies for temporal analysis to obtain the hidden state sequence. Then, the self-attention layer assigns weights to the hidden state sequence according to multiple operating condition intervals to obtain multiple influence weight values. These multiple influence weight values ​​are then matched and mapped with the hidden state sequence to generate measurement error prediction values. Finally, the measurement error label is used as a supervision signal, and the deep neural network is backpropagated and iteratively updated based on the measurement error prediction values ​​to train and obtain the initial error compensation model.

[0029] Finally, the initial error compensation model is locally updated based on the metering error label to construct the electricity meter accuracy error compensation model. In this process, multi-dimensional operating condition feature parameters and metering error labels are extracted one by one according to multiple operating condition intervals. The multi-dimensional operating condition feature parameters corresponding to the current operating condition interval are input into the initial error compensation model. Temporal dependencies are extracted and hidden state sequences are generated through a gated recurrent unit layer. The hidden state sequences are weighted according to the current operating condition interval through a self-attention layer, and then the predicted metering error value is obtained through output mapping. The metering error value in the metering error label is read, and the difference between the predicted metering error value and the actual metering error value is calculated. The difference is squared to obtain the error loss value for the current operating condition interval. Based on the error loss value, the parameter gradient is calculated backward from the output mapping parameters, and the gradient is sequentially passed to the self-attention layer and the gated recurrent unit layer. The process begins with a meta-layer that obtains the adjustment directions corresponding to the output mapping parameters, self-attention layer weight parameters, and gated recurrent unit layer parameters. Following a preset learning rate, the gradients corresponding to each parameter are multiplied and subtracted from the current parameters to obtain the updated local model parameters. Within the same operating condition interval, the processes of metering error prediction, error loss value calculation, parameter gradient backpropagation, and local model parameter update are repeatedly executed until the error loss value reaches the preset convergence condition. Then, the process switches to the next operating condition interval and locally updates the model parameters related to that operating condition interval in the initial error compensation model according to the corresponding metering error label. After completing the local updates for multiple operating condition intervals, a power meter accuracy error compensation model capable of outputting metering error prediction values ​​according to different operating condition intervals is obtained.

[0030] Furthermore, in the method provided in the application embodiments, constructing a deep neural network, synchronizing the multi-dimensional working condition feature parameter set to the deep neural network according to the multiple working condition intervals for weighted learning, and training to obtain an initial error compensation model, further includes: A gated recurrent unit layer and a self-attention layer are constructed, and the gated recurrent unit layer and the self-attention layer are cascaded to construct a deep neural network. Temporal analysis is performed on the multi-dimensional operating condition feature parameters to obtain temporal dependencies. The multi-dimensional operating condition feature parameters are synchronized to the gated recurrent unit layer according to the temporal dependencies for further analysis to obtain a hidden state sequence. The self-attention layer assigns weights to the hidden state sequence according to the multiple operating condition intervals to obtain multiple influence weight values. The multiple influence weight values ​​are matched and mapped with the hidden state sequence to generate a measurement error prediction value. The measurement error label is used as a supervision signal, and the deep neural network is backpropagated and iteratively updated according to the measurement error prediction value to construct the initial error compensation model.

[0031] In this embodiment, when constructing the deep neural network, active power, voltage deviation, and current waveform distortion rate from the multi-dimensional operating condition feature parameter set are read and arranged in chronological order according to the sliding time window. Each sliding time window corresponds to one set of multi-dimensional operating condition feature parameters, and each set of multi-dimensional operating condition feature parameters includes three feature dimensions and is associated with the corresponding operating condition interval. In one implementation, 24 consecutive sliding time windows are used as a time-series input, so that the deep neural network receives data for 24 time steps each time, and each time step includes active power, voltage deviation, and current waveform distortion rate. The active power, voltage deviation, and current waveform distortion rate entering the deep neural network are standardized, that is, each feature value is subtracted from the mean of the corresponding feature dimension and then divided by the standard deviation of the corresponding feature dimension to obtain the multi-dimensional operating condition feature parameters entering the gated recurrent unit layer.

[0032] Next, a gated recurrent unit layer and a self-attention layer are constructed, and the output of the gated recurrent unit layer is connected to the input of the self-attention layer to form a cascaded deep neural network. In one implementation, the gated recurrent unit layer is set to 2 layers, each with 64 hidden units and an inter-layer dropout rate of 0.2; the self-attention layer has 4 attention heads, each with a feature dimension of 16; and the output mapping layer has 1 output node for outputting the predicted measurement error value. When performing time-series analysis on the multi-dimensional operating condition feature parameters, the data of each time step is read in the order of the sliding time windows, and the time-series dependency is obtained based on the temporal continuity between adjacent sliding time windows and the change relationship of their respective operating condition intervals; when adjacent sliding time windows belong to the same operating condition interval, a continuous dependency within the same operating condition is established; when adjacent sliding time windows belong to different operating condition intervals, a cross-operating condition transition dependency is established.

[0033] Subsequently, the multi-dimensional operating condition feature parameters are synchronized to the gated loop unit layer for analysis according to their temporal dependencies. For the current time step, the gated loop unit layer reads the active power, voltage deviation, and current waveform distortion rate of the current time step, and reads the hidden state output from the previous time step. The multi-dimensional operating condition feature parameters of the current time step are multiplied by a weight matrix and then superimposed with a bias using the hidden state of the previous time step, respectively. The update gate result and reset gate result are calculated using the sigmoid function. The reset gate result is multiplied by the hidden state of the previous time step, and then multiplied by a weight matrix and superimposed with the multi-dimensional operating condition feature parameters of the current time step, respectively. The candidate hidden state is calculated using the tanh function. The update gate result is used as a weighting ratio to perform a weighted fusion of the hidden state of the previous time step and the candidate hidden state to obtain the hidden state of the current time step. The above calculations are performed sequentially for each of the 24 time steps to obtain a hidden state sequence formed by 24 hidden states arranged in chronological order.

[0034] Next, a self-attention layer is used to assign weights to the hidden state sequence according to multiple operating condition intervals. The self-attention layer first performs a linear transformation on each hidden state in the sequence to obtain query features, key features, and value features. Then, it performs a dot product calculation on the query feature corresponding to any hidden state with the key features corresponding to other hidden states to obtain a relevance score. Based on the operating condition interval to which the sliding time window corresponding to the hidden state belongs, the relevance score is weighted by interval. In one implementation, the relevance score within the same operating condition interval is multiplied by 1.0, and the relevance score between different operating condition intervals is multiplied by 0.6. The interval-weighted relevance score is then normalized to obtain multiple influence weight values, so that each influence weight value corresponds to a hidden state in the hidden state sequence.

[0035] When matching and mapping multiple influence weights to the hidden state sequence, each hidden state is multiplied by its corresponding influence weight to obtain a weighted hidden state. All weighted hidden states within the same time series input are summed to obtain a weighted time series representation result. This weighted time series representation result is input to the output mapping layer, and the predicted measurement error value is calculated through a fully connected layer. The predicted measurement error value is correlated with the input multidimensional operating condition feature parameters, the corresponding sliding time window, and the corresponding operating condition interval.

[0036] Finally, the measurement error label is used as a supervision signal, and the deep neural network is iteratively updated through backpropagation according to the predicted measurement error value. The prediction bias is obtained by subtracting the predicted measurement error value from the measurement error label value; the squared prediction bias is then averaged within the same training batch to obtain the error loss value. In one implementation, the training batch size is set to 32, the initial learning rate is set to 0.001, and the maximum number of iterations is set to 100. Using the error loss value as the basis for backpropagation, the parameter gradient is calculated starting from the output mapping layer and sequentially passed to the self-attention layer and the gated recurrent unit layer; the parameters of the output mapping layer, self-attention layer, and gated recurrent unit layer are updated according to the initial learning rate. The process of inputting multi-dimensional working condition feature parameters, calculating hidden state sequences, allocating influence weights, predicting measurement errors, calculating error loss values, and iteratively updating through backpropagation is repeated. Training stops when the decrease in error loss value over 10 consecutive iterations is less than a preset convergence threshold, or when the number of iterations reaches 100, and the initial error compensation model is obtained.

[0037] Step S200: Calculate the target energy meter using the energy meter accuracy error compensation model to generate an energy compensation dataset.

[0038] In this embodiment, when calculating the target energy meter using the energy meter accuracy error compensation model, the real-time operating condition dataset and the original energy metering value of the target energy meter are collected. The real-time operating condition dataset is synchronized to the energy meter accuracy error compensation model for calculation to obtain the initial metering error prediction value and the confidence evaluation index corresponding to the initial metering error prediction value. Then, the confidence of the initial metering error prediction value is evaluated based on the confidence evaluation index, and the confidence evaluation index is compared with the preset high confidence condition. When the confidence evaluation index meets the high confidence condition, the initial metering error prediction value is used to compensate the original energy metering value to generate the first compensation energy. When the confidence evaluation index does not meet the high confidence condition, the transformer area collaborative inference mechanism is activated to perform inference, calculate the inferred compensation amount, and use the inferred compensation amount to compensate the original energy metering value to generate the second compensation energy. The transformer area energy conservation verification is performed based on the first compensation energy and the second compensation energy to obtain the energy compensation dataset.

[0039] Furthermore, in the method provided in the application embodiments, the calculation of the target energy meter using the energy meter accuracy error compensation model to generate an energy compensation dataset also includes: The system collects real-time operating condition datasets and raw electricity metering values ​​from the target electricity meter. It then synchronizes these datasets to the electricity meter accuracy error compensation model to obtain an initial metering error prediction value. This initial prediction value includes a confidence assessment index, which corresponds to the predicted value. Based on the confidence assessment index, the system evaluates the reliability of the initial prediction value: a high reliability condition is preset, and the confidence assessment index is compared with this condition. If the confidence assessment index meets the high reliability condition, the initial prediction value is used to compensate the raw electricity metering value, generating a first compensation amount. If the confidence assessment index does not meet the high reliability condition, a district collaborative inference mechanism is activated to perform inference, calculate the inferred compensation amount, and compensate the raw electricity metering value, generating a second compensation amount. Finally, based on the first and second compensation amounts, a district energy conservation check is performed to obtain the electricity compensation dataset.

[0040] In this embodiment, the real-time operating condition dataset and raw energy consumption values ​​of the target energy meter are first collected. The real-time operating condition dataset is acquired according to the current acquisition cycle and includes the target energy meter's active power, voltage deviation, current waveform distortion rate, and corresponding acquisition time within that acquisition cycle. The raw energy consumption values ​​are determined based on the energy consumption readings of the target energy meter at the start and end of the current acquisition cycle. The raw energy consumption values ​​for the current acquisition cycle are obtained by subtracting the energy consumption reading at the start of the acquisition cycle from the energy consumption reading at the end of the acquisition cycle.

[0041] Next, the real-time operating condition dataset is synchronized to the electricity meter accuracy error compensation model. First, the dataset is sorted according to the acquisition time, and the active power, voltage deviation, and current waveform distortion rate corresponding to the same acquisition time are aligned. Then, following the feature processing method used in the training phase of the electricity meter accuracy error compensation model, the active power, voltage deviation, and current waveform distortion rate are standardized. The processed real-time operating condition dataset is then input into the electricity meter accuracy error compensation model. The model calculates based on the operating condition interval and time series characteristics corresponding to the real-time operating condition dataset, outputting an initial metering error prediction value and simultaneously outputting a confidence assessment index corresponding to this initial metering error prediction value, ensuring a one-to-one correspondence between the initial metering error prediction value and the confidence assessment index.

[0042] Subsequently, the reliability of the initial metering error prediction value is assessed based on the confidence evaluation index, with a preset high reliability condition. This high reliability condition includes the confidence evaluation index being no less than a preset reliability threshold. The confidence evaluation index output by the electricity meter accuracy error compensation model is numerically compared with the preset reliability threshold. When the confidence evaluation index is greater than or equal to the preset reliability threshold, the confidence evaluation index is determined to meet the high reliability condition, and the original electricity metering value is compensated using the initial metering error prediction value. During compensation, the original electricity metering value is subtracted from the initial metering error prediction value to obtain the first compensated electricity amount. A correspondence is established between the first compensated electricity amount and the target electricity meter identifier, data collection period, initial metering error prediction value, and confidence evaluation index.

[0043] When the confidence assessment index does not meet the high reliability condition, the distribution area collaborative inference mechanism is activated for inference. After activating the distribution area collaborative inference mechanism, the total electricity consumption of the distribution area where the target electricity meter is located within the same collection period, the original electricity consumption value of each electricity meter, and the generated first compensation electricity consumption are read. Electricity meters that have generated first compensation electricity consumption are included in the distribution area electricity consumption summary according to the first compensation electricity consumption, and electricity meters that have not generated first compensation electricity consumption are included in the distribution area electricity consumption summary according to their original electricity consumption value, to obtain the current summary electricity consumption on the user side of the distribution area. The total electricity consumption of the distribution area is subtracted from the current summary electricity consumption on the user side of the distribution area to obtain the electricity consumption difference of the distribution area. Then, according to the proportion of the original electricity consumption value of the target electricity meter to the sum of the original electricity consumption values ​​of all electricity meters that do not meet the high reliability condition, the electricity consumption difference of the distribution area is allocated to obtain the inferred compensation amount corresponding to the target electricity meter.

[0044] When using the inferred compensation amount to compensate the original electricity metering value, the inferred compensation amount is subtracted from the original electricity metering value of the target electricity meter to generate the second compensation amount; the second compensation amount is associated with the target electricity meter identification, collection period, inferred compensation amount and confidence evaluation index.

[0045] Subsequently, energy conservation verification of the transformer area is performed based on the first and second compensation amounts. The compensation results of each energy meter are summarized according to the same transformer area and the same data collection period. For energy meters whose confidence assessment indicators meet the high reliability condition, the first compensation amount is used in the summary; for energy meters whose confidence assessment indicators do not meet the high reliability condition, the second compensation amount is used. The compensation amounts of each energy meter are added together to obtain the user-side compensation amount for the transformer area. The total meter consumption of the transformer area is subtracted from the user-side compensation amount to obtain the energy conservation verification difference for the transformer area. The energy conservation verification difference for the transformer area is compared with the preset allowable deviation range.

[0046] When the energy conservation verification difference of the transformer area is within the preset allowable deviation range, the corresponding first compensation amount and second compensation amount are retained. When the energy conservation verification difference of the transformer area is greater than the upper limit of the preset allowable deviation range, the difference between the energy conservation verification difference of the transformer area and the upper limit is calculated to obtain the amount to be increased in correction; the sum of each second compensation amount included in the summary is calculated, and the proportion of each second compensation amount to the sum of all second compensation amounts is calculated, and the amount to be increased in correction is allocated to each second compensation amount according to the proportion; the corresponding allocated correction amount is added to each second compensation amount to obtain the corrected second compensation amount. When the energy conservation verification difference of the transformer area is less than the lower limit of the preset allowable deviation range, the difference between the lower limit and the energy conservation verification difference of the transformer area is calculated to obtain the amount to be deducted in correction; the sum of each second compensation amount included in the summary is calculated, and the proportion of each second compensation amount to the sum of all second compensation amounts is calculated, and the amount to be deducted in correction is allocated to each second compensation amount according to the proportion; the corresponding allocated correction amount is subtracted from each second compensation amount to obtain the corrected second compensation amount.

[0047] After completing the second compensation power correction, the first compensation power and the corrected second compensation power are re-summarized to obtain the updated user-side compensation power for the transformer area. The updated user-side compensation power for the transformer area is then subtracted from the total meter power for the transformer area to obtain the transformer area energy conservation verification difference, which is then compared with the preset allowable deviation range again. If the newly obtained transformer area energy conservation verification difference is within the preset allowable deviation range, the transformer area energy conservation verification is completed; if it still exceeds the preset allowable deviation range, the second compensation power is continued to be corrected according to the transformer area energy conservation verification difference until the transformer area energy conservation verification difference is within the preset allowable deviation range. The first compensation power, the second compensation power, the target energy meter identifier, the collection period, and the confidence evaluation index after the transformer area energy conservation verification are summarized to obtain the power compensation dataset.

[0048] Step S300: Traverse the transformer area to perform topology analysis, construct the transformer area topology structure, calculate line loss according to the transformer area topology structure and power compensation dataset, and generate transformer area line loss residual sequence.

[0049] In this embodiment, the transformer substation is first traversed for topology analysis to construct its topology. This process involves first defining low-load periods and collecting voltage and current waveform data from target energy meters during these periods. Next, disturbance detection is performed based on the voltage waveform data to obtain a first disturbance event set, and disturbance detection is performed based on the current waveform data to obtain a second disturbance event set. Then, correlation analysis is performed between the first and second disturbance event sets to calculate multiple correlation coefficients. Based on these correlation coefficients, the target energy meters are identified to determine multiple electrical connection paths. Finally, topology analysis is performed based on these multiple electrical connection paths to construct the transformer substation topology.

[0050] Next, line loss calculation is performed based on the transformer substation topology and the power compensation dataset. In this process, the power compensation dataset is mapped to the transformer substation topology according to multiple electrical connection paths, and the topology is then back-engineered to calculate the current distribution data of multiple branches. Subsequently, dynamic impedance loss analysis is performed based on the multiple branch current distribution data to calculate the expected line loss time series value for the transformer substation. Then, multi-meter compensation calculations are performed on the electricity meters within the transformer substation based on the power compensation data to obtain the statistical line loss time series value for the transformer substation. Afterwards, multiple time sections are defined, and the expected line loss time series value for the transformer substation is subtracted from the statistical line loss time series value for the transformer substation according to these time sections to calculate the initial residual sequence. Finally, dynamic correction is performed based on the initial residual sequence to obtain the transformer substation line loss residual sequence.

[0051] Furthermore, the method provided in the application embodiments, which involves traversing the transformer areas to perform topology analysis and constructing the transformer area topology structure, also includes: Low-load periods are defined, and voltage and current waveform data of the target energy meter are collected based on these periods. Disturbance detection is performed on the voltage waveform data to obtain a first set of disturbance events. Disturbance detection is also performed on the current waveform data to obtain a second set of disturbance events. Correlation analysis is conducted based on the first and second sets of disturbance events to calculate multiple correlation coefficients for connection identification of the target energy meter, determining multiple electrical connection paths. Topology analysis is then performed based on these multiple electrical connection paths to construct the transformer substation topology.

[0052] In this embodiment, a low-load period is first defined, and voltage and current waveform data of the target energy meters are collected based on the low-load period. During this process, load data of the main distribution meter within a preset statistical period is read, and the total load value of the distribution area is calculated according to the collection time. The total load value of the distribution area corresponding to each collection time is compared with a preset low-load threshold. When the total load value of the distribution area at multiple consecutive collection times is less than the preset low-load threshold, and the continuous duration reaches a preset duration, this continuous time interval is defined as a low-load period. During the low-load period, the target energy meters within the distribution area are synchronously sampled, and voltage and current waveform data of the target energy meters are collected according to a unified sampling clock. The voltage and current waveform data are then correlated with the target energy meter identifier and the sampling time, respectively.

[0053] Next, disturbance detection is performed based on voltage waveform data. The voltage waveform data of the target energy meter is read sequentially according to the sampling time, and the voltage amplitude difference between adjacent sampling points is calculated. The absolute value of the voltage amplitude difference is taken to obtain the voltage amplitude change. The voltage amplitude change is compared with a preset voltage disturbance threshold. When the voltage amplitude change is greater than or equal to the preset voltage disturbance threshold, the corresponding sampling time is determined as the voltage disturbance occurrence time. The voltage amplitude change of subsequent sampling points is continued to be read. When the voltage amplitude change is continuously less than the preset voltage recovery threshold and reaches the preset threshold, the disturbance is detected. When the recovery time is set, the corresponding sampling time is determined as the voltage disturbance termination time; the voltage disturbance termination time is subtracted from the voltage disturbance occurrence time to obtain the disturbance duration; the average voltage amplitude of a preset number of sampling points before the voltage disturbance occurrence time is calculated, and the average voltage amplitude of each sampling point between the voltage disturbance occurrence time and the voltage disturbance termination time is calculated. The absolute value of the difference between the two average values ​​is obtained to obtain the disturbance amplitude; the target energy meter identification, disturbance occurrence time, disturbance termination time, disturbance duration, and disturbance amplitude are recorded to obtain the first disturbance event set.

[0054] Subsequently, disturbance detection is performed based on the current waveform data. The current waveform data of the target energy meter is read sequentially according to the sampling time, and the difference in current amplitude between adjacent sampling points is calculated. The absolute value of the current amplitude difference is taken to obtain the change in current amplitude. The change in current amplitude is compared with a preset current disturbance threshold. When the change in current amplitude is greater than or equal to the preset current disturbance threshold, the corresponding sampling time is determined as the time when the current disturbance occurs. The current amplitude change at subsequent sampling points is continued to be read. When the change in current amplitude is continuously less than a preset current recovery threshold and reaches a preset threshold... When restoring the duration, the corresponding sampling time is determined as the current disturbance termination time; the current disturbance occurrence time is subtracted from the current disturbance termination time to obtain the disturbance duration; the average current amplitude of a preset number of sampling points before the current disturbance occurrence time is calculated, and the average current amplitude of each sampling point between the current disturbance occurrence time and the current disturbance termination time is calculated. The absolute value of the difference between the two average values ​​is obtained to obtain the disturbance amplitude; the target energy meter identification, disturbance occurrence time, disturbance termination time, disturbance duration, and disturbance amplitude are recorded to obtain the second disturbance event set.

[0055] Next, correlation analysis is performed based on the first and second disturbance event sets to calculate multiple correlation coefficients for connection identification of target energy meters and determine multiple electrical connection paths. In this process, the first and second disturbance event sets are matched according to the target energy meter identifier and the disturbance occurrence time, extracting the disturbance occurrence time difference, disturbance duration difference, and disturbance amplitude difference between any two target energy meters. When calculating the disturbance occurrence time difference, the absolute values ​​of the disturbance occurrence times corresponding to the two target energy meters are subtracted to obtain the time difference. This time difference is then divided by a preset time matching window to obtain the normalized time difference. Finally, the normalized time difference is subtracted from 1 to obtain the time correlation coefficient. When calculating the disturbance duration difference, the absolute values ​​of the disturbance durations corresponding to the two target energy meters are subtracted to obtain the duration difference. This duration difference is then divided by the larger of the two disturbance durations to obtain the normalized duration difference. Subtract the normalized duration difference from 1 to obtain the duration correlation coefficient. When calculating the disturbance amplitude difference, subtract the corresponding disturbance amplitudes of the two target energy meters and take the absolute value to obtain the amplitude difference. Divide the amplitude difference by the larger of the two disturbance amplitudes to obtain the normalized amplitude difference. Subtract the normalized amplitude difference from 1 to obtain the amplitude correlation coefficient. Compare the time correlation coefficient, duration correlation coefficient, and amplitude correlation coefficient with the corresponding preset correlation thresholds. When all the above correlation coefficients meet the corresponding preset correlation thresholds, the connection relationship between the corresponding target energy meters is identified. Determine the upstream and downstream relationship between the target energy meters based on the order of the disturbance occurrence time, and determine multiple electrical connection paths based on the connection relationship between the corresponding target energy meters.

[0056] Finally, a topology analysis was performed based on multiple electrical connection paths. The power supply node of the transformer substation was taken as the starting node, and the target energy meter was taken as the ending node. The power supply node, branch node, and target energy meter were connected sequentially according to the upstream and downstream relationships in the multiple electrical connection paths. When multiple electrical connection paths have the same upstream node, the corresponding target energy meter was connected to the same branch node. When there are continuous upstream and downstream relationships between electrical connection paths, the connection level was determined according to the order in which the disturbance occurred. After completing the connection of all electrical connection paths, duplicate connection paths were merged, and the connection relationships between the power supply node, branch node, target energy meter, and multiple electrical connection paths were organized according to the connection level to construct the transformer substation topology.

[0057] Furthermore, in the method provided in the application embodiment, the process of calculating line loss according to the transformer area topology and combining it with the power compensation dataset to generate a transformer area line loss residual sequence further includes: The power compensation dataset is mapped to the transformer area topology according to the multiple electrical connection paths for forward and backward iteration to calculate the current distribution data of multiple branches. Dynamic impedance loss analysis is performed based on the multiple branch current distribution data to calculate the expected line loss time series value of the transformer area. Multi-meter compensation calculation is performed on the electricity meters within the transformer area based on the power compensation data to obtain the statistical line loss time series value of the transformer area. Multiple time sections are defined, and the expected line loss time series value of the transformer area is subtracted from the statistical line loss time series value of the transformer area according to the multiple time sections to calculate the initial residual sequence. Dynamic correction is performed based on the initial residual sequence to obtain the line loss residual sequence of the transformer area.

[0058] In this embodiment, the power compensation dataset is first mapped to the transformer substation topology along multiple electrical connection paths for forward and backward substitution to calculate the current distribution data of multiple branches. In this process, the transformer substation power supply nodes, branch nodes, target energy meters, and multiple electrical connection paths are read from the transformer substation topology. Based on the target energy meter identifier, acquisition period, and time segment, the corresponding compensation power is extracted from the power compensation dataset. The compensation power of each target energy meter is mapped to its corresponding end node, and backward substitution is performed along the direction from the target energy meter to the branch node and from the branch node to the transformer substation power supply node. The compensation power of each target energy meter downstream of the same branch node is added to obtain the downstream compensation power of the corresponding branch. The downstream compensation power is converted to Wh, the time segment duration is converted to h, and the downstream compensation power is divided by the corresponding node voltage and... The duration of the time segment is used to obtain the branch current; then, the line length and impedance per unit length of each branch are read according to multiple electrical connection paths. The line length and impedance per unit length are multiplied to obtain the line impedance of the corresponding branch; forward calculation is performed along the direction from the power supply node of the transformer area to the branch node and from the branch node to the target energy meter. The branch current is multiplied by the line impedance to obtain the node voltage drop. The voltage drop of the upstream node is subtracted from the node voltage drop to obtain the voltage of the next level node; the updated node voltage is used to perform back-substitution calculation again, and the forward and back-substitution calculations are repeated until the difference between the branch currents obtained from two adjacent calculations meets the preset iteration stop condition, thus obtaining multiple branch current distribution data.

[0059] Next, dynamic impedance loss analysis is performed based on multiple branch current distribution data to calculate the expected line loss time series value for the transformer area. In this process, branch current distribution data is read according to branch identification within each time segment, along with the corresponding branch's line length, line impedance per unit length, and line operating temperature. The line length is multiplied by the line impedance per unit length to obtain the basic line impedance. The basic line impedance is corrected based on the line operating temperature to obtain the dynamic impedance within the current time segment. The branch current is squared, and the squared result is multiplied by the dynamic impedance to obtain the power loss of that branch within the current time segment. The power loss is multiplied by the duration of the time segment to obtain the electrical energy loss of that branch within the current time segment. The electrical energy losses of all branches within the same time segment are summed to obtain the expected line loss value for the transformer area corresponding to that time segment. These values ​​are then arranged in chronological order according to the multiple time segments to obtain the expected line loss time series value for the transformer area.

[0060] Subsequently, multi-meter compensation calculations were performed on the electricity meters within the distribution area based on the electricity compensation data. The compensation electricity of each target electricity meter in the electricity compensation data was read according to the meter identification, collection period, and time segment. The power supply of the distribution area's total meter within the same time segment was also read. For target electricity meters whose confidence assessment indicators met the high-reliability condition, the corresponding first compensation electricity was read. For target electricity meters whose confidence assessment indicators did not meet the high-reliability condition, the corresponding second compensation electricity was read. The first and second compensation electricity of each target electricity meter within the same time segment were added together to obtain the user-side compensation electricity. The user-side compensation electricity was subtracted from the power supply of the distribution area's total meter to obtain the distribution area's statistical line loss value for that time segment. These values ​​were then arranged in chronological order according to multiple time segments to obtain the distribution area's statistical line loss time-series value.

[0061] Then, multiple time segments are defined, and the expected line loss time series value of the transformer area is subtracted from the statistical line loss time series value of the transformer area according to the multiple time segments to calculate the initial residual sequence. Specifically, the operating time of the transformer area is divided into multiple continuous time segments according to a preset time interval, and the expected line loss time series value and the statistical line loss time series value of the transformer area are aligned one by one according to the time segments; for each time segment, the corresponding expected line loss value and statistical line loss value of the transformer area are read, and the expected line loss value of the transformer area is subtracted from the statistical line loss value to obtain the initial residual value corresponding to that time segment; the initial residual values ​​are arranged in chronological order according to the time segments to obtain the initial residual sequence.

[0062] like Figure 3 As shown, Figure 3 The graph shows the calculated residual data for the line loss sequence of the transformer area. The horizontal axis represents multiple time sections, and the vertical axis represents the electricity consumption per kWh. Figure 3 In the diagram, the expected line loss time series value of the transformer area is represented by a solid broken line, the statistical line loss time series value of the transformer area is represented by a dashed broken line, and the initial residual sequence is represented by a black and white filled bar chart. Within the same time section, the expected line loss time series value of the transformer area is subtracted from the statistical line loss time series value of the transformer area to obtain the initial residual value of the corresponding time section, and these values ​​are arranged in chronological order according to the multiple time sections to form the initial residual sequence.

[0063] Finally, dynamic correction is performed based on the initial residual sequence. In this process, the initial residual sequence is first decomposed into two levels, and then modal decomposition is performed on the initial residual sequence according to multiple time segments to calculate multiple intrinsic mode components. Subsequently, zero-crossing calculation is performed based on the multiple intrinsic mode components to obtain the zero-crossing rate, which is then compared with a preset zero-crossing threshold. When the zero-crossing rate is less than the preset zero-crossing threshold, the corresponding intrinsic mode component is identified as a slow-changing trend component; when the zero-crossing rate is greater than or equal to the preset zero-crossing threshold, the corresponding intrinsic mode component is identified as a fast-changing random component. Then, line loss anomaly determination is performed based on the slow-changing trend component and the fast-changing random component, generating a stray loss estimate. This stray loss estimate is then removed from the initial residual sequence for correction, generating the transformer area line loss residual sequence.

[0064] Furthermore, in the method provided in the application embodiments, dynamic correction is performed based on the initial residual sequence to obtain the transformer area line loss residual sequence, which further includes: The initial residual sequence is decomposed in two layers: modal decomposition is performed on the initial residual sequence according to the multiple time sections to calculate multiple intrinsic mode components; zero-crossing calculation is performed based on the multiple intrinsic mode components to obtain the zero-crossing rate; when the zero-crossing rate is less than a preset zero-crossing threshold, the multiple intrinsic mode components are identified as slow-changing trend components; when the zero-crossing rate is greater than or equal to the preset zero-crossing threshold, the multiple intrinsic mode components are identified as fast-changing random components; line loss anomaly determination is performed based on the slow-changing trend components and the fast-changing random components to generate stray loss estimates; the stray loss estimates are removed and corrected from the initial residual sequence to generate the transformer area line loss residual sequence.

[0065] In this embodiment, when performing a two-level decomposition on the initial residual sequence, modal decomposition is performed on the initial residual sequence according to multiple time segments to calculate multiple intrinsic modal components. First, the initial residual values ​​corresponding to multiple time segments are read in chronological order, and each initial residual value is matched one-to-one with its corresponding time segment. During modal decomposition, the initial residual values ​​of adjacent time segments are compared sequentially to determine local maxima and local minima. An upper envelope sequence is formed based on the local maxima, and a lower envelope sequence is formed based on the local minima. Then, the upper and lower envelope values ​​corresponding to the same time segment are added together and divided by 2 to obtain the envelope mean sequence. The envelope mean sequence is then subtracted from the current residual sequence to be decomposed to obtain the component to be extracted. Intrinsic modal component extraction conditions are then determined for the component to be extracted. These conditions include the difference between the number of local extrema and the number of zero-crossings in the component to be extracted. The absolute value of the value is not greater than 1, and the average amplitude of the envelope mean sequence corresponding to the component to be extracted is not greater than the product of the average amplitude of the component to be extracted and the preset screening threshold. When the component to be extracted meets the intrinsic mode component extraction condition, the component to be extracted is determined as an intrinsic mode component, and the current residual sequence to be decomposed is subtracted from the intrinsic mode component to obtain the remaining residual sequence. When the remaining residual sequence still has no less than 2 local extreme points, and the amplitude change of the remaining residual sequence reaches the preset decomposition threshold, the determination of local extreme points, generation of upper and lower envelope sequences, calculation of envelope mean sequence and judgment of intrinsic mode component extraction condition are continued for the remaining residual sequence until the remaining residual sequence no longer meets the conditions for continued decomposition, and multiple intrinsic mode components are calculated.

[0066] Next, zero-crossing calculations are performed based on multiple intrinsic modal components. The component values ​​of each intrinsic modal component are read sequentially according to the time segments, and the sign of the component values ​​corresponding to two adjacent time segments is determined. A zero-crossing event is recorded when the component value of the previous time segment is positive and the component value of the next time segment is negative; a zero-crossing event is also recorded when the component value of the previous time segment is negative and the component value of the next time segment is positive. When a zero value exists in an adjacent time segment, the non-zero component values ​​before and after the zero value are read, and a zero-crossing event is recorded when the signs of the non-zero component values ​​are opposite. After completing the sign determination of the same intrinsic modal component across all time segments, the number of zero-crossing events for that intrinsic modal component is counted, and the number of zero-crossing events is divided by the number of adjacent time segments to obtain the data zero-crossing rate corresponding to that intrinsic modal component. Data zero-crossing calculations are performed sequentially on multiple intrinsic modal components to obtain multiple data zero-crossing rates.

[0067] When the zero-crossing rate is less than the preset zero-crossing threshold, multiple intrinsic mode components are identified as slow-changing trend components. The zero-crossing rate of each intrinsic mode component is read one by one, and the zero-crossing rate is compared with the preset zero-crossing threshold. When the zero-crossing rate of a certain intrinsic mode component is less than the preset zero-crossing threshold, the component value of the intrinsic mode component in all time segments is retained, and a slow-changing trend component is added to the intrinsic mode component. For the intrinsic mode components identified as slow-changing trend components, their component values, number of zero-crossing events, and zero-crossing rate are recorded in the order of time segments to obtain the slow-changing trend components corresponding to multiple time segments.

[0068] When the zero-crossing rate is greater than or equal to a preset zero-crossing threshold, multiple intrinsic mode components (IMCs) are identified as fast-changing random components. Specifically, IMCs that are not identified as slow-changing trend components are read one by one, and their corresponding zero-crossing rates are compared with the preset zero-crossing threshold. When the zero-crossing rate of an IMC is greater than or equal to the preset zero-crossing threshold, the component values ​​of that IMC across all time segments are retained, and a fast-changing random component identifier is added to that IMC. For IMCs identified as fast-changing random components, their component values, the number of zero-crossing events, and the zero-crossing rate are recorded in time segment order to obtain fast-changing random components corresponding to multiple time segments.

[0069] Finally, based on the slow-changing trend component and the fast-changing random component, line loss anomalies are determined, and stray loss estimates are generated. These stray loss estimates are then removed and corrected from the initial residual sequence to generate the transformer area line loss residual sequence. In this process, the slow-changing trend component, the fast-changing random component, and the initial residual sequence are aligned according to time segments. For each time segment, the component values ​​of all slow-changing trend components within that time segment are summed to obtain a slow-changing trend composite value, and the component values ​​of all fast-changing random components within that time segment are summed to obtain a fast-changing random composite value. The slow-changing trend composite value is compared with a preset trend amplitude range, and the fast-changing random composite value is compared with a preset random amplitude range. When both the slow-changing trend composite value and the fast-changing random composite value are within the preset trend amplitude range, the slow-changing trend composite value of that time segment is compared with... The stray loss estimate for the time segment is obtained by summing the rapidly changing random composite values. When the rapidly changing random composite value exceeds the preset random amplitude range, or the slowly changing trend composite value exceeds the preset trend amplitude range within a continuous time segment, the corresponding time segment is retained as the line loss anomaly determination object, and the slowly changing trend composite value within the preset trend amplitude range within that time segment is used as the stray loss estimate. Subsequently, the initial residual value and the stray loss estimate are read according to the time segment, and the corresponding stray loss estimate is subtracted from the initial residual value to obtain the line loss residual value for that time segment. The line loss residual values ​​are then arranged in the order of multiple time segments to generate the line loss residual sequence for the transformer area.

[0070] Step S400: Perform time-series feature decomposition based on the line loss residual sequence of the transformer area, obtain the line loss residual components, perform anomaly depth search, determine anomalies based on the search results, and obtain a line loss anomaly determination report.

[0071] In this embodiment, when performing time-series feature decomposition and anomaly depth search based on the transformer area line loss residual sequence, the following steps are taken: First, line loss fluctuation analysis is performed on the transformer area line loss residual sequence to calculate the line loss random fluctuation coefficient. Then, time-series feature decomposition is performed on the transformer area line loss residual sequence according to the line loss random fluctuation coefficient to obtain high-frequency residual components. Next, the high-frequency residual components are continuously mapped to multiple residual state data, and multiple residual state data are used as nodes to construct a directed graph of line loss residual state transition. Finally, feature analysis is performed based on the transformer area line loss residual sequence, and the directed graph of line loss residual state transition is searched according to the feature analysis results to obtain search results.

[0072] Finally, anomaly determination is performed based on the search results. The single-source anomaly path and anomaly connected region obtained from the anomaly depth search are read, and the line loss residual value in the transformer area line loss residual sequence is extracted according to the corresponding time section, target energy meter, and electrical connection path. The line loss residual value is compared with a preset anomaly determination threshold. When the line loss residual value corresponding to a single-source anomaly path reaches the preset anomaly determination threshold, the corresponding target energy meter or branch is determined to be a single-source line loss anomaly. When the sum of multiple line loss residual values ​​within an anomaly connected region reaches the preset region anomaly determination threshold, the corresponding region is determined to be a multi-node concurrent line loss anomaly. Subsequently, the anomaly time section, anomaly target energy meter, anomaly electrical connection path, anomaly type, line loss residual value, and search result source are recorded, generating a line loss anomaly determination report.

[0073] Furthermore, in the method provided in the application embodiment, the process of performing time-series feature decomposition based on the line loss residual sequence of the transformer area to obtain the line loss residual components and then performing anomaly depth search further includes: Line loss fluctuation analysis is performed based on the residual sequence of line loss in the transformer area, and the random fluctuation coefficient of line loss is calculated. The time-series feature decomposition of the residual sequence of line loss in the transformer area is performed according to the random fluctuation coefficient of line loss to obtain high-frequency residual components. The high-frequency residual components are continuously mapped to multiple residual state data, and multiple residual state data are used as nodes to construct a directed graph of line loss residual state transition. Feature analysis is performed based on the residual sequence of line loss in the transformer area, and the search results are obtained by searching the directed graph of line loss residual state transition based on the feature analysis results.

[0074] In this embodiment, line loss fluctuation analysis is first performed based on the transformer area line loss residual sequence. The line loss residual values ​​in the transformer area line loss residual sequence are read in chronological order according to multiple time sections. The line loss residual value of the previous time section is subtracted from the line loss residual value of the later time section to obtain the residual change value between adjacent time sections. The absolute value of each residual change value is taken, and all absolute values ​​are added together and divided by the number of residual change values ​​to obtain the average change of line loss residual. The absolute value of each line loss residual value in the transformer area line loss residual sequence is taken, and all absolute values ​​are added together and divided by the number of line loss residual values ​​to obtain the average amplitude of line loss residual. The average change of line loss residual is divided by the average amplitude of line loss residual to obtain the line loss random fluctuation coefficient. When the average amplitude of line loss residual is zero, the line loss random fluctuation coefficient is set to zero.

[0075] Next, the line loss residual sequence of the transformer area is decomposed into time-series features according to the line loss random fluctuation coefficient. The line loss random fluctuation coefficient is compared with the preset fluctuation threshold, and the number of continuous time segments used for smoothing calculation is determined based on the comparison result. When the line loss random fluctuation coefficient is greater than or equal to the preset fluctuation threshold, the current time segment and one time segment before and after it are selected to participate in the smoothing calculation. When the line loss random fluctuation coefficient is less than the preset fluctuation threshold, the current time segment and two time segments before and after it are selected to participate in the smoothing calculation. For cases where there are insufficient time segments before and after the beginning or end, the actual adjacent time segments of the current time segment are used to participate in the smoothing calculation. The line loss residual values ​​participating in the smoothing calculation are added together and divided by the number of time segments participating in the calculation to obtain the low-frequency residual value of the current time segment. The line loss residual value of the current time segment is subtracted from the corresponding low-frequency residual value to obtain the high-frequency residual value of the current time segment. The low-frequency residual value is calculated and subtracted sequentially for all time segments, and the high-frequency residual values ​​are arranged according to the chronological order of the time segments to obtain the high-frequency residual components.

[0076] Next, the high-frequency residual components are continuously mapped to multiple residual state data, and these multiple residual state data are used as nodes to construct a directed graph of line loss residual state transition. In this process, the high-frequency residual values ​​in the high-frequency residual components are read sequentially according to the time segments; the average of all high-frequency residual values ​​is calculated; the average value is subtracted from each high-frequency residual value and squared; the sum of all squared results is divided by the number of high-frequency residual values, and the square root of the result is taken to obtain the standard deviation; the average value is subtracted from the high-frequency residual value of each time segment and divided by the standard deviation to obtain the standardized residual value of that time segment; when the standard deviation is zero, the standardized residual value is set to zero; the standardized residual value is compared with a preset residual interval to determine the residual state corresponding to that time segment. The residual state is obtained by recording the time section, high-frequency residual value, standardized residual value, and residual state. Each residual state data is treated as a node according to the chronological order of the time sections, and the node of the previous time section is pointed to the node of the next time section, forming a time-continuous directed connection edge. When there are residual state data corresponding to multiple target energy meters or multiple electrical connection paths in the same time section, directed connection edges are added between the corresponding nodes according to the connection relationship in the transformer area topology. The directed graph of line loss residual state transition is composed of multiple nodes and directed connection edges.

[0077] Finally, feature analysis is performed based on the transformer area line loss residual sequence, and the directed graph of line loss residual state transition is searched according to the feature analysis results. In this process, feature analysis is performed based on the transformer area line loss residual sequence to generate feature analysis results including step-type abrupt change features and multi-node concurrent fluctuation features. When the feature analysis result is a step-type abrupt change feature, a depth-first search is initiated to traverse the directed graph of line loss residual state transition and extract single-source anomaly paths. When the feature analysis result is a multi-node concurrent fluctuation feature, a breadth-first search is initiated to traverse the directed graph of line loss residual state transition and determine abnormal connected regions. Single-source anomaly paths and abnormal connected regions are added to the search results.

[0078] Furthermore, in the method provided in the application embodiment, feature analysis is performed based on the line loss residual sequence of the transformer area, and the directed graph of the line loss residual state transition is searched according to the feature analysis results to obtain the search results. The method further includes: Feature analysis is performed based on the residual sequence of line loss in the transformer area to generate feature analysis results. The feature analysis results include step-type abrupt change features and multi-node concurrent fluctuation features. When the feature analysis result is a step-type abrupt change feature, a depth-first search is initiated to traverse the directed graph of the residual line loss state transition to extract single-source abnormal paths. When the feature analysis result is a multi-node concurrent fluctuation feature, a breadth-first search is initiated to traverse the directed graph of the residual line loss state transition to determine abnormal connected regions. The single-source abnormal paths and the abnormal connected regions are added to the search results.

[0079] In this embodiment, when performing feature analysis based on the transformer area line loss residual sequence, the line loss residual values ​​in the transformer area line loss residual sequence are read in chronological order of time segments, and analyzed in conjunction with the residual state data corresponding to each node in the directed graph of line loss residual state transition; for adjacent time segments, the difference between the line loss residual value of the later time segment and the line loss residual value of the previous time segment is calculated, and the absolute value of the difference is compared with a preset step judgment threshold. When the absolute value of the difference reaches the preset step judgment threshold, and the line loss residual value after the sudden change remains stable within a preset number of consecutive time segments... When the residual value is within the preset abnormal range, the corresponding time segment and the corresponding node are marked as step-type mutation features. For multiple nodes within the same time segment, the line loss residual value and residual status data of each node are read, and the number of nodes that reach the preset abnormal residual range is counted. When the number of nodes reaches the preset concurrent number threshold, and the corresponding nodes have an electrical connection relationship in the transformer area topology, the corresponding time segment and the corresponding node are marked as multi-node concurrent fluctuation features. After completing the above feature analysis, feature analysis results containing step-type mutation features and multi-node concurrent fluctuation features are generated.

[0080] When the feature analysis result is a step-type abrupt change feature, a depth-first search is initiated to traverse the directed graph of line loss residual state transition. The node corresponding to the step-type abrupt change feature is taken as the starting search node, and the time section, line loss residual value, and electrical connection path corresponding to the starting search node are read. Starting from the starting search node, adjacent nodes are visited level by level along the directed connection edges in the directed graph of line loss residual state transition. For each node visited, it is determined whether the line loss residual value corresponding to the node is within the preset abnormal residual range, and whether the node and the previous node maintain a continuous time section relationship or the same electrical connection path relationship. When the above judgments are satisfied, the node is added to the current search path, and the search continues downward along the next directed connection edge of the node. When the line loss residual value corresponding to the node is not within the preset abnormal residual range, or the node and the previous node do not satisfy the continuous time section relationship and the same electrical connection path relationship, the branch search is stopped and the previous node is returned to continue searching for other directed connection edges. After the depth-first search is completed, the sequence of abnormal nodes continuously extended from the starting search node is determined as the single-source abnormal path.

[0081] When the feature analysis result indicates a multi-node concurrent fluctuation characteristic, a breadth-first search is initiated to traverse the directed graph of line loss residual state transition. Multiple nodes corresponding to the multi-node concurrent fluctuation characteristic are used as starting search nodes, and search queues are established according to the same or adjacent time segments. Nodes are sequentially retrieved from the search queue, and their corresponding line loss residual values, residual state data, and electrical connection paths are read. Nodes adjacent to the node are then accessed along the directed connection edges in the directed graph of line loss residual state transition. When the line loss residual value of an adjacent node is within a preset abnormal residual range, and the adjacent node has an electrical connection with the current node in the transformer area topology, the adjacent node is added to the search queue, and its connection relationship with the current node is recorded. The node dequeueing, adjacent node judgment, and enqueueing operations are repeated until the search queue is empty. All abnormal nodes and their connection relationships recorded during the search process are aggregated to determine the abnormal connectivity region.

[0082] Finally, single-source abnormal paths and abnormal connected regions are added to the search results. Single-source abnormal paths extracted by depth-first search are read, and the starting node, ending node, time section, target energy meter, electrical connection path, and line loss residual value corresponding to each path are recorded. Abnormal connected regions determined by breadth-first search are read, and the time section, number of target energy meters, number of abnormal nodes, electrical connection path, and line loss residual value corresponding to each region are recorded. These recorded results are then organized according to feature analysis results, search methods, and node connection relationships, so that step-type abrupt change features correspond to single-source abnormal paths, and multi-node concurrent fluctuation features correspond to abnormal connected regions, thus forming the search results.

[0083] In summary, the embodiments of this application have at least the following technical effects: This application retrieves original electricity time-series data to model electricity meter measurement errors and constructs an electricity meter accuracy error compensation model. The target electricity meter is then calculated using this model to generate an electricity compensation dataset. A topology analysis is performed across the distribution area to construct its topology. Line loss is calculated based on this topology and the electricity compensation dataset, generating a line loss residual sequence. Time-series feature decomposition is performed on the residual sequence to obtain the line loss components, and an anomaly depth search is conducted. Anomaly determination is then performed based on the search results to obtain a line loss anomaly determination report. This invention addresses the technical problem in existing technologies where electricity meter measurement errors affect line loss calculations, leading to low accuracy in line loss anomaly determination. By constructing an electricity meter accuracy error compensation model to compensate for the target electricity meter data, and performing line loss residual analysis and anomaly search based on the compensated electricity data, the technical effect of reducing the interference of measurement errors on line loss anomaly determination results and improving the accuracy of line loss anomaly determination is achieved.

[0084] Example 2 is based on the same inventive concept as the line loss anomaly determination method based on electricity meter data accuracy compensation in the previous examples, such as... Figure 2 As shown, this application provides a line loss anomaly determination system based on electricity meter data accuracy compensation. The system and method embodiments in this application are based on the same inventive concept. The system includes: Modeling module 11 is used to retrieve original electricity time-series data to model the metering error of the electricity meter and construct an electricity meter accuracy error compensation model; calculation module 12 is used to calculate the target electricity meter through the electricity meter accuracy error compensation model and generate an electricity compensation dataset; topology analysis module 13 is used to traverse the transformer area to perform topology analysis, construct the transformer area topology structure, and calculate line loss according to the transformer area topology structure combined with the electricity compensation dataset to generate a transformer area line loss residual sequence; anomaly judgment module 14 is used to perform time-series feature decomposition based on the transformer area line loss residual sequence, obtain the line loss residual components for anomaly depth search, perform anomaly judgment based on the search results, and obtain a line loss anomaly judgment report.

[0085] Furthermore, the system is also used to implement the following functions: The original energy time-series data is segmented using a sliding time window to calculate active power; voltage amplitude deviation is calculated based on the original energy time-series data to obtain voltage deviation values; current waveform decomposition is performed based on the original energy time-series data to calculate current waveform distortion rate; the active power, voltage deviation values, and current waveform distortion rate are combined according to the sliding time window to generate a multi-dimensional operating condition feature parameter set; spatial clustering is performed based on the multi-dimensional operating condition feature parameter set to generate density spatial clustering regions, which include high-density clustering regions; the high-density clustering regions are used as the operating condition feature set. The system calculates the distribution of feature vectors based on the operating condition center, defines the distribution range parameters of the feature vectors, determines the multi-dimensional operating condition feature envelope boundary according to the feature vector distribution range parameters, and generates multiple operating condition intervals. Based on the multiple operating condition intervals, it calculates the error of the target energy meter according to a sliding time window and constructs a metering error label. It then constructs a deep neural network, synchronizes the multi-dimensional operating condition feature parameter set to the deep neural network according to the multiple operating condition intervals, performs weighted learning, and trains to obtain an initial error compensation model. Based on the metering error label, it locally updates the initial error compensation model to construct an energy meter accuracy error compensation model.

[0086] Furthermore, the system is also used to implement the following functions: A gated recurrent unit layer and a self-attention layer are constructed, and the gated recurrent unit layer and the self-attention layer are cascaded to construct a deep neural network. Temporal analysis is performed on the multi-dimensional operating condition feature parameters to obtain temporal dependencies. The multi-dimensional operating condition feature parameters are synchronized to the gated recurrent unit layer according to the temporal dependencies for further analysis to obtain a hidden state sequence. The self-attention layer assigns weights to the hidden state sequence according to the multiple operating condition intervals to obtain multiple influence weight values. The multiple influence weight values ​​are matched and mapped with the hidden state sequence to generate a measurement error prediction value. The measurement error label is used as a supervision signal, and the deep neural network is backpropagated and iteratively updated according to the measurement error prediction value to construct the initial error compensation model.

[0087] Furthermore, the system is also used to implement the following functions: The system collects real-time operating condition datasets and raw electricity metering values ​​from the target electricity meter. It then synchronizes these datasets to the electricity meter accuracy error compensation model to obtain an initial metering error prediction value. This initial prediction value includes a confidence assessment index, which corresponds to the predicted value. Based on the confidence assessment index, the system evaluates the reliability of the initial prediction value: a high reliability condition is preset, and the confidence assessment index is compared with this condition. If the confidence assessment index meets the high reliability condition, the initial prediction value is used to compensate the raw electricity metering value, generating a first compensation amount. If the confidence assessment index does not meet the high reliability condition, a district collaborative inference mechanism is activated to perform inference, calculate the inferred compensation amount, and compensate the raw electricity metering value, generating a second compensation amount. Finally, based on the first and second compensation amounts, a district energy conservation check is performed to obtain the electricity compensation dataset.

[0088] Furthermore, the system is also used to implement the following functions: Low-load periods are defined, and voltage and current waveform data of the target energy meter are collected based on these periods. Disturbance detection is performed on the voltage waveform data to obtain a first set of disturbance events. Disturbance detection is also performed on the current waveform data to obtain a second set of disturbance events. Correlation analysis is conducted based on the first and second sets of disturbance events to calculate multiple correlation coefficients for connection identification of the target energy meter, determining multiple electrical connection paths. Topology analysis is then performed based on these multiple electrical connection paths to construct the transformer substation topology.

[0089] Furthermore, the system is also used to implement the following functions: The power compensation dataset is mapped to the transformer area topology according to the multiple electrical connection paths for forward and backward iteration to calculate the current distribution data of multiple branches. Dynamic impedance loss analysis is performed based on the multiple branch current distribution data to calculate the expected line loss time series value of the transformer area. Multi-meter compensation calculation is performed on the electricity meters within the transformer area based on the power compensation data to obtain the statistical line loss time series value of the transformer area. Multiple time sections are defined, and the expected line loss time series value of the transformer area is subtracted from the statistical line loss time series value of the transformer area according to the multiple time sections to calculate the initial residual sequence. Dynamic correction is performed based on the initial residual sequence to obtain the line loss residual sequence of the transformer area.

[0090] Furthermore, the system is also used to implement the following functions: The initial residual sequence is decomposed in two layers: modal decomposition is performed on the initial residual sequence according to the multiple time sections to calculate multiple intrinsic mode components; zero-crossing calculation is performed based on the multiple intrinsic mode components to obtain the zero-crossing rate; when the zero-crossing rate is less than a preset zero-crossing threshold, the multiple intrinsic mode components are identified as slow-changing trend components; when the zero-crossing rate is greater than or equal to the preset zero-crossing threshold, the multiple intrinsic mode components are identified as fast-changing random components; line loss anomaly determination is performed based on the slow-changing trend components and the fast-changing random components to generate stray loss estimates; the stray loss estimates are removed and corrected from the initial residual sequence to generate the transformer area line loss residual sequence.

[0091] Furthermore, the system is also used to implement the following functions: Line loss fluctuation analysis is performed based on the residual sequence of line loss in the transformer area, and the random fluctuation coefficient of line loss is calculated. The time-series feature decomposition of the residual sequence of line loss in the transformer area is performed according to the random fluctuation coefficient of line loss to obtain high-frequency residual components. The high-frequency residual components are continuously mapped to multiple residual state data, and multiple residual state data are used as nodes to construct a directed graph of line loss residual state transition. Feature analysis is performed based on the residual sequence of line loss in the transformer area, and the search results are obtained by searching the directed graph of line loss residual state transition based on the feature analysis results.

[0092] Furthermore, the system is also used to implement the following functions: Feature analysis is performed based on the residual sequence of line loss in the transformer area to generate feature analysis results. The feature analysis results include step-type abrupt change features and multi-node concurrent fluctuation features. When the feature analysis result is a step-type abrupt change feature, a depth-first search is initiated to traverse the directed graph of the residual line loss state transition to extract single-source abnormal paths. When the feature analysis result is a multi-node concurrent fluctuation feature, a breadth-first search is initiated to traverse the directed graph of the residual line loss state transition to determine abnormal connected regions. The single-source abnormal paths and the abnormal connected regions are added to the search results.

[0093] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for determining abnormal line losses based on electricity meter data accuracy compensation, characterized in that, The method includes: Original electricity time-series data is retrieved to model the metering error of the electricity meter and a model for compensating for the accuracy error of the electricity meter is constructed. The target energy meter is calculated using the energy meter accuracy error compensation model to generate an energy compensation dataset. The topology analysis is performed by traversing the transformer area to construct the transformer area topology structure. Line loss is calculated according to the transformer area topology structure and the power compensation dataset to generate the transformer area line loss residual sequence. Based on the residual sequence of line loss in the transformer area, time-series feature decomposition is performed to obtain the residual components of line loss. Anomaly depth search is then performed, and anomaly determination is made based on the search results to obtain a line loss anomaly determination report. Collect real-time operating condition data and raw electricity metering values ​​of the target electricity meter; The real-time operating condition dataset is synchronized to the electricity meter accuracy error compensation model to obtain an initial metering error prediction value. The initial metering error prediction value includes a confidence evaluation index, and the confidence evaluation index has a corresponding relationship with the initial metering error prediction value. The confidence level assessment index is used to evaluate the reliability of the initial measurement error prediction value: A high-confidence condition is preset, and the confidence evaluation index is compared with the high-confidence condition. When the confidence evaluation index meets the high-confidence condition, the original power measurement value is compensated by the initial measurement error prediction value to generate the first compensated power. When the confidence assessment index does not meet the high reliability condition, the transformer area collaborative inference mechanism is activated to make inference, calculate the inferred compensation amount to compensate the original power metering value, and generate the second compensation power. Based on the first and second compensation amounts, the energy conservation of the transformer area is verified to obtain the power compensation dataset. Based on the residual sequence of line loss in the transformer area, line loss fluctuation analysis is performed, and the random fluctuation coefficient of line loss is calculated. The line loss residual sequence of the transformer area is decomposed into time-series features according to the line loss random fluctuation coefficient to obtain high-frequency residual components; The high-frequency residual components are continuously mapped to multiple residual state data, and the multiple residual state data are used as nodes to construct a directed graph of line loss residual state transition. Based on the residual sequence of line loss in the transformer area, feature analysis is performed, and the directed graph of the state transition of the residual line loss is searched according to the feature analysis results to obtain the search results; Based on the residual sequence of line loss in the transformer area, feature analysis is performed to generate feature analysis results, which include step-type abrupt change features and multi-node concurrent fluctuation features. When the feature analysis result is a step-type mutation feature, a depth-first search is initiated to traverse the directed graph of the line loss residual state transition to extract single-source abnormal paths; when the feature analysis result is a multi-node concurrent fluctuation feature, a breadth-first search is initiated to traverse the directed graph of the line loss residual state transition to determine abnormal connected regions. Add the single-source abnormal path and the abnormal connected region to the search results.

2. The method for determining line loss anomalies based on electricity meter data accuracy compensation as described in claim 1, characterized in that, The method involves retrieving raw electricity time-series data to model the metering error of an energy meter and constructing a compensation model for the accuracy error of the energy meter. The original power time-series data is segmented using a sliding time window to calculate the active power. Based on the original power time-series data, voltage amplitude deviation is calculated to obtain the voltage deviation value; Based on the original power time-series data, the current waveform is decomposed and the current waveform distortion rate is calculated. The active power, voltage deviation, and current waveform distortion rate are combined according to a sliding time window to generate a multi-dimensional operating condition characteristic parameter set. Spatial clustering is performed based on the multidimensional working condition feature parameter set to generate density spatial clustering regions, which include high-density clustering regions. The high-density clustered area is taken as the working condition center, and the distribution calculation is performed according to the working condition center to define the feature vector distribution range parameters. The multidimensional working condition feature envelope boundary is determined according to the feature vector distribution range parameter, and multiple working condition intervals are generated. Based on the multiple operating condition intervals, the error of the target energy meter is calculated according to the sliding time window, and a metering error label is constructed. A deep neural network is constructed, and the multi-dimensional working condition feature parameter set is synchronized to the deep neural network according to the multiple working condition intervals for weighted learning, and an initial error compensation model is obtained through training. Based on the metering error label, the initial error compensation model is locally updated to construct an electricity meter accuracy error compensation model.

3. The method for determining line loss anomalies based on electricity meter data accuracy compensation as described in claim 2, characterized in that, Constructing a deep neural network, synchronizing the multi-dimensional working condition feature parameter set according to the multiple working condition intervals to the deep neural network for weighted learning, and training to obtain an initial error compensation model, the method includes: Construct a gated recurrent unit layer and a self-attention layer, and cascade the gated recurrent unit layer and the self-attention layer to build a deep neural network; Time series analysis is performed on the multidimensional operating condition characteristic parameters to obtain time series dependencies; The multidimensional operating condition feature parameters are synchronized to the gated loop unit layer according to the time-series dependency relationship for analysis to obtain the hidden state sequence. The hidden state sequence is weighted according to the multiple working condition intervals through the self-attention layer to obtain multiple influence weight values; The multiple influence weight values ​​are matched and mapped with the hidden state sequence to generate the measurement error prediction value; Using the measurement error label as a supervision signal, the deep neural network is backpropagated and iteratively updated according to the predicted measurement error value to construct the initial error compensation model.

4. The method for determining line loss anomalies based on electricity meter data accuracy compensation as described in claim 1, characterized in that, The method involves traversing the transformer areas to perform topology analysis and constructing the transformer area topology structure, including: Define low-load periods and collect voltage and current waveform data of the target energy meter based on the low-load periods; Based on the voltage waveform data, disturbance detection is performed to obtain a first disturbance event set; Based on the current waveform data, disturbance detection is performed to obtain a second disturbance event set; Correlation analysis is performed based on the first set of disturbance events and the second set of disturbance events. Multiple correlation coefficients are calculated to identify the target energy meter and determine multiple electrical connection paths. Based on the multiple electrical connection paths, a topology analysis is performed to construct the transformer substation topology.

5. The method for determining line loss anomalies based on electricity meter data accuracy compensation as described in claim 4, characterized in that, Based on the aforementioned transformer area topology and the aforementioned power compensation dataset, line loss is calculated to generate a transformer area line loss residual sequence. The method includes: The power compensation dataset is mapped to the transformer area topology according to the multiple electrical connection paths to perform forward and backward iterations, and the current distribution data of multiple branches are calculated. Dynamic impedance loss analysis is performed based on the current distribution data of the multiple branches to calculate the expected line loss timing value of the transformer area. Based on the power compensation data, multi-meter compensation calculations are performed on the power meters in the transformer area to obtain the time series value of statistical line loss in the transformer area. Multiple time sections are defined, and the expected line loss time series value of the transformer area is subtracted from the statistical line loss time series value of the transformer area according to the multiple time sections to calculate the initial residual sequence; Dynamic correction is performed based on the initial residual sequence to obtain the line loss residual sequence of the transformer area.

6. The method for determining line loss anomalies based on electricity meter data accuracy compensation as described in claim 5, characterized in that, Dynamic correction is performed based on the initial residual sequence to obtain the line loss residual sequence for the transformer area. The method includes: The initial residual sequence is then decomposed into two levels: The initial residual sequence is decomposed according to the multiple time segments to calculate multiple intrinsic mode components. Data zero-crossing rate is obtained by calculating the zero-crossing rate based on the multiple intrinsic mode components. When the zero-crossing rate of the data is less than a preset zero-crossing threshold, the multiple intrinsic mode components are identified as slow-changing trend components. When the data zero-crossing rate is greater than or equal to a preset zero-crossing threshold, the multiple intrinsic mode components are identified as fast-changing random components. Line loss anomaly determination is performed based on the slow-changing trend component and the fast-changing random component, and stray loss estimate is generated. The stray loss estimate is then removed and corrected from the initial residual sequence to generate the transformer area line loss residual sequence.

7. A line loss anomaly determination system based on electricity meter data accuracy compensation, characterized in that, The system is used to execute the line loss anomaly determination method based on electricity meter data accuracy compensation as described in any one of claims 1-6, and the system includes: The modeling module is used to retrieve raw electricity time-series data to model the metering error of the electricity meter and build a compensation model for the accuracy error of the electricity meter. The calculation module is used to calculate the target energy meter using the energy meter accuracy error compensation model and generate an energy compensation dataset. The topology analysis module is used to traverse the transformer area to perform topology analysis, construct the transformer area topology structure, calculate line loss according to the transformer area topology structure and the power compensation dataset, and generate the transformer area line loss residual sequence. The anomaly determination module is used to perform time-series feature decomposition based on the line loss residual sequence of the transformer area, obtain the line loss residual components, perform anomaly depth search, determine anomalies based on the search results, and obtain a line loss anomaly determination report.

Images