Electric power measurement data processing method and system based on feature recognition
By analyzing the time series data and Granger causality analysis of electricity metering data, occasional power consumption anomalies can be identified and adjusted, solving the problem of low accuracy in power load forecasting in existing technologies and achieving higher forecasting accuracy and grid management efficiency.
Patent Information
- Application Number
- CN202510946783.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-09
- Publication Date
- 2025-11-18
AI Technical Summary
Existing time-series forecasting models are unable to effectively identify and adapt to occasional changes in electricity consumption, leading to a decrease in the accuracy of power load forecasting and affecting the efficiency of power grid dispatching and energy management.
By performing time-series analysis on electricity metering data, we can identify occasional electricity consumption anomalies using predicted electricity metering data, and distinguish between occasional and regular changes through Granger causality analysis. By adjusting the electricity metering time-series analysis set, we can improve the prediction accuracy.
It effectively avoids the impact of occasional power outages, improves the accuracy of power load forecasting, and ensures the effectiveness of power grid dispatching and energy management.
Smart Images
Figure CN120974358A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data analysis, in particular to a power metering data processing method and system based on feature recognition. BACKGROUND
[0002] In the field of smart grid and energy management, accurate power load forecasting is crucial for the safe and stable operation of the power grid, the optimal allocation of power resources, and the effective trading of the energy market. Currently, time series forecasting methods, such as traditional statistical models (e.g., ARIMA, exponential smoothing) and emerging machine learning models (e.g., LSTM, Transformer), have been widely applied in power load forecasting. These methods usually rely on the time series patterns of historical electricity consumption data, learning the rules of historical data to predict future electricity consumption. However, in actual power systems, electricity consumption behavior is not entirely regular, and occasional electricity consumption changes are common. For example, the sudden start of large industrial equipment and the centralized start of air conditioning or heating equipment under abnormal weather conditions will cause non-periodic and sudden fluctuations in electricity consumption.
[0003] These occasional electricity consumption changes often contradict the regularity of historical time series patterns and are considered anomalies or noise in time series data. When traditional time series forecasting models face these sudden changes, since their training process is mainly based on the stationarity and periodicity assumptions of historical data, they are difficult to effectively identify and adapt to these unexpected fluctuations. Therefore, the existence of occasional electricity consumption changes can significantly interfere with the learning and prediction process of time series forecasting models, leading to a deviation of the prediction results from the actual electricity consumption trend, ultimately reducing the accuracy of electricity consumption time series forecasting, significantly reducing the practical value of the prediction results, and further affecting the scheduling decisions of the power grid and the efficiency of energy management. SUMMARY
[0004] The present application analyzes the time series of user electricity metering data and determines occasional electricity anomalies based on predicted power metering data. During the time series analysis, the predicted power metering data is used to avoid occasional electricity anomalies, avoiding the decrease in prediction accuracy caused by occasional electricity anomalies, and considering regular changes caused by, for example, the addition of electrical appliances or the reduction of electrical appliances, further improving the prediction accuracy.
[0005] The present application provides a power metering data processing method based on feature recognition, comprising:
[0006] acquiring power metering data at a current monitoring time point; forming a power metering time series analysis set by combining the power metering data acquired at the current monitoring time point and power metering data acquired at previous N-1 monitoring time points, and sending the power metering time series analysis set into a power metering time series analysis model for processing to output predicted power metering data at a next monitoring time point;
[0007] Meanwhile, at the current monitoring time point, a deviation value judgment operation is performed, which includes the following contents: performing a deviation value calculation on the acquired power metering data and corresponding predicted power metering data, δ = |U1-U2| / U2, where δ is a deviation value, U1 is the power metering data acquired at the current monitoring time point, U2 is the predicted power metering data corresponding to the current monitoring time point, and it is judged whether "δ>G" is true, where G is a deviation threshold value; if "δ>G" is true, the current monitoring time point is marked as a suspected abnormal time point, and if "δ>G" is not true, the power metering data at the next monitoring time point is acquired; for monitoring time points after the suspected abnormal time point, the power metering data corresponding to the suspected abnormal time point and monitoring time points after the suspected abnormal time point are replaced by corresponding predicted power metering data when the power metering time series analysis set is constructed; the replaced power metering time series analysis set is processed by the power metering time series analysis model to output predicted power metering data corresponding to the monitoring time point; for the next monitoring time point after the suspected abnormal time point, the deviation value judgment operation is performed, if "δ>G" is true, the current monitoring time point is marked as a suspected abnormal time point, the total number M of suspected abnormal time points is counted, and it is judged whether "M>L" is true, where L is an occasional time point threshold value; if "M>L" is true, an electricity abnormality causal judgment operation is performed, and if "M>L" is not true, the total number M of suspected abnormal time points is continuously counted; if "δ>G" is not true, the total number M of suspected abnormal time points is counted, and it is judged whether "M>L" is true; if "M>L" is not true, all suspected abnormal time points are re-marked as occasional electricity abnormal time points, and the power metering data corresponding to the occasional electricity abnormal time points is replaced by corresponding predicted power metering data when the power metering time series analysis set is constructed, and if "M>L" is true, the electricity abnormality causal judgment operation is performed;
[0008] The power consumption anomaly cause-effect judgment operation includes the following contents: all suspected abnormal time points corresponding to the power metering data are sequentially grouped into a dependent variable time series analysis data set, the difference between the suspected abnormal time points corresponding to the power metering data and the predicted power metering data is calculated and recorded as a change value, all suspected abnormal time points corresponding to the change value are sequentially grouped into an independent variable time series analysis data set, Granger cause-effect analysis is performed based on the dependent variable time series analysis data set and the independent variable time series analysis data set, if there is a cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, all suspected abnormal time points are deleted, and when constructing the power metering time series analysis set, the power metering data is not replaced, and if there is no cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, all suspected abnormal time points are re-marked as accidental power consumption abnormal time points.
[0009] As a preferred aspect of the present application, the power metering time series analysis model includes a power metering data updating layer, a power metering time series analysis layer, and a predicted power metering data output layer, wherein the power metering data updating layer is used to update the power metering data in the power metering time series analysis set based on historical power metering data to construct an updated power metering time series analysis set; the power metering time series analysis layer is established based on an LSTM model and is used to process the updated power metering time series analysis set to construct predicted power metering data; and the predicted power metering data output layer is used to output the predicted power metering data.
[0010] As a preferred aspect of the present application, the power metering data updating layer updates the power metering data in the power metering time series analysis set based on historical power metering data to construct an updated power metering time series analysis set, which specifically includes the following steps:
[0011] The power metering time series analysis set is divided into a plurality of power metering time series analysis sequences through a sliding window with a length of C;
[0012] For each power metering time series analysis sequence, the power metering time series analysis sequence is updated through the following formula: , wherein F2 is the updated power metering time series analysis sequence, F1 is the power metering time series analysis sequence before updating, A 24k-1 is the power metering time series analysis sequence at the 24k-1 monitoring time points before the power metering time series analysis sequence F1, k = 1, 2, 3, …, K, K is the distance upper limit, and a 24k-1 is the power metering time series analysis sequence F1 and the power metering time series analysis sequence A 24k-1The similarity between the updated power metering time series analysis sequence F2 and the power metering time series analysis sequence F1 before updating is calculated, and a self-attention mechanism operation is performed based on the updated power metering time series analysis sequence F2 and the power metering time series analysis sequence F1 before updating, to complete updating of the power metering data in the power metering time series analysis set and construct an updated power metering time series analysis set.
[0013] As a preferred aspect of the present application, a self-attention mechanism operation is performed based on the updated power metering time series analysis sequence F2 and the power metering time series analysis sequence F1 before updating, to complete updating of the power metering data in the power metering time series analysis set and construct an updated power metering time series analysis set, specifically including the following steps:
[0014] All updated power metering time series analysis sequences F2 are spliced from top to bottom in time sequence to construct a target feature matrix, and all power metering time series analysis sequences F1 before updating are spliced from top to bottom in time sequence to construct a query feature matrix; then the target feature matrix is multiplied by the value weight matrix and the key weight matrix respectively to construct the corresponding target value matrix V and target key matrix K, and the query feature matrix is multiplied by the query weight matrix to construct a query matrix Q, and the self-attention mechanism operation is completed through the following formula: H = softmax (QK T / D 0.5 ) V, wherein H is an updated feature matrix, T is a matrix transposition operation, D is the dimension size of the target key matrix K, and the updated feature matrix H is flattened by row to obtain an updated power metering time series analysis set.
[0015] As a preferred aspect of the present application, Granger causality analysis is performed based on the dependent variable time series analysis dataset and the independent variable time series analysis dataset, specifically including the following steps:
[0016] Set a significance level η;
[0017] Construct a baseline model according to the dependent variable time series analysis dataset;
[0018] Increase the independent variable time series analysis dataset based on the baseline model to construct a full model;
[0019] Calculate the residual sum of squares corresponding to the baseline model and the full model respectively, and then perform F test based on the residual sum of squares corresponding to the baseline model and the full model to obtain a P value; if the P value is higher than the significance level η, there is a causal relationship between the dependent variable time series analysis dataset and the independent variable time series analysis dataset; if the P value is not higher than the significance level η, there is no causal relationship between the dependent variable time series analysis dataset and the independent variable time series analysis dataset.
[0020] As a preferred aspect of the present application, the power metering time series analysis model is trained, specifically including the following steps:
[0021] A plurality of power metering time series analysis training samples are obtained, the power metering time series analysis training samples include N+1 power metering data arranged in time, and there is no accidental power consumption anomaly, all the power metering time series analysis training samples form a power metering time series analysis training set, the power metering time series analysis model is trained through the power metering time series analysis training set, during the training, the last power metering data in the power metering time series analysis training sample is taken as a target, whether the training condition is met is judged, if the training condition is met, the trained power metering time series analysis model is output, otherwise, the power metering time series analysis model is continuously trained through the power metering time series analysis training set.
[0022] The application further provides a power metering data processing system based on feature recognition, comprising:
[0023] A power metering data acquisition module is configured to acquire power metering data at a current monitoring time point;
[0024] A power metering time series analysis module is configured to combine the power metering data acquired at the current monitoring time point and the power metering data acquired at the previous N-1 monitoring time points to form a power metering time series analysis set, and then send the power metering time series analysis set to a power metering time series analysis model for processing, so as to output predicted power metering data at a next monitoring time point;
[0025] The monitoring time point judgment module is used for executing a deviation value judgment operation at a current monitoring time point. The deviation value judgment operation includes the following contents: performing deviation value calculation on the acquired power metering data and corresponding predicted power metering data, δ=|U1-U2| / U2, where δ is the deviation value, U1 is the power metering data acquired at the current monitoring time point, U2 is the predicted power metering data corresponding to the current monitoring time point, judging whether "δ>G" is established, where G is a deviation threshold value, if "δ>G" is established, marking the current monitoring time point as a suspected abnormal time point, if "δ>G" is not established, continuing to acquire power metering data at a next monitoring time point; for the monitoring time points after the suspected abnormal time point, replacing the power metering data corresponding to the suspected abnormal time point and the monitoring time points after the suspected abnormal time point with corresponding predicted power metering data when constructing a power metering time sequence analysis set; processing the replaced power metering time sequence analysis set through a power metering time sequence analysis model to output predicted power metering data corresponding to the monitoring time points; for the next monitoring time point after the suspected abnormal time point, executing the deviation value judgment operation, if "δ>G" is established, marking the current monitoring time point as a suspected abnormal time point, counting the total number M of suspected abnormal time points, judging whether "M>L" is established, L being an occasional time point threshold value, if "M>L" is established, executing a power consumption abnormality causal judgment operation, if "M>L" is not established, continuously counting the total number M of suspected abnormal time points; if "δ>G" is not established, counting the total number M of suspected abnormal time points, judging whether "M>L" is established, if "M>L" is not established, re-marking all suspected abnormal time points as occasional power consumption abnormal time points, and replacing the power metering data corresponding to the occasional power consumption abnormal time points with corresponding predicted power metering data when constructing the power metering time sequence analysis set, if "M>L" is established, executing the power consumption abnormality causal judgment operation;
[0026] The power consumption abnormality causal judgment operation module is used for executing a power consumption abnormality causal judgment operation. The power consumption abnormality causal judgment operation includes the following contents: grouping the power metering data corresponding to all suspected abnormal time points into a dependent variable time sequence analysis data set according to time sequence, calculating the difference between the power metering data corresponding to the suspected abnormal time points and the predicted power metering data and recording the difference as a change value, grouping all change values corresponding to the suspected abnormal time points into an independent variable time sequence analysis data set according to time sequence, performing Granger causal analysis based on the dependent variable time sequence analysis data set and the independent variable time sequence analysis data set, if there is a causal relationship between the dependent variable time sequence analysis data set and the independent variable time sequence analysis data set, deleting all suspected abnormal time points, not replacing the power metering data when constructing a power metering time sequence analysis set, if there is no causal relationship between the dependent variable time sequence analysis data set and the independent variable time sequence analysis data set, re-marking all suspected abnormal time points as occasional power consumption abnormal time points.
[0027] The present application has the following advantages:
[0028] 1、The present application judges the occasional power consumption anomaly based on the predicted power consumption data by performing time series analysis on the user's power consumption data, and avoids the occasional power consumption anomaly by the predicted power consumption data in the process of time series analysis, avoids the decline of prediction accuracy caused by occasional power consumption anomaly, and considers the regularity change caused by, for example, adding or reducing electrical appliances, further improves the prediction accuracy.
[0029] 2、The power consumption time series analysis sequence divided by the sliding window can reflect the local time series relationship, and the power consumption data is updated by the power consumption time series analysis sequence, which can pay more attention to the change of power consumption data, and further obtain better updating effect. BRIEF DESCRIPTION OF DRAWINGS
[0030] Figure 1 The structure diagram of the power consumption data processing system based on feature recognition used in the embodiment of the present application. DETAILED DESCRIPTION
[0031] In order to make the person skilled in the art better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application.
[0032] Embodiment 1, a power consumption data processing method based on feature recognition, comprising:
[0033] The power consumption data at the current monitoring time point is obtained, wherein the power consumption data refers to the power consumption, which is used to reflect the user's power consumption; the power consumption data obtained at the current monitoring time point and the power consumption data obtained at the previous N-1 monitoring time points are combined to form a power consumption time series analysis set, and then the power consumption time series analysis set is sent to a power consumption time series analysis model for processing to output the predicted power consumption data at the next monitoring time point; it should be noted that the time interval between the adjacent two monitoring time points is generally 15 minutes, and the value of N is generally 24, and under the assistance of the power consumption time series analysis model, the user's power consumption can be analyzed, and then the user's power consumption can be predicted;
[0034] Meanwhile, at the current monitoring time point, a deviation value judgment operation is performed, and the deviation value judgment operation includes the following contents: performing a deviation value calculation on the obtained power metering data and the corresponding predicted power metering data, δ = |U1-U2| / U2, where δ is the deviation value, U1 is the power metering data obtained at the current monitoring time point, U2 is the predicted power metering data corresponding to the current monitoring time point, and it is judged whether "δ>G" is established, where G is a deviation threshold value, which is used to judge whether the power consumption situation is abnormal, if "δ>G" is established, the current monitoring time point is marked as a suspected abnormal time point, if "δ>G" is not established, the power metering data at the next monitoring time point is continuously obtained; for the monitoring time points after the suspected abnormal time point, when the power metering time series analysis set is constructed, the power metering data corresponding to the suspected abnormal time point and the monitoring time points after the suspected abnormal time point are replaced with the corresponding predicted power metering data, it should be noted that when the power consumption situation is abnormal, it is impossible to judge the source of the power consumption abnormality, which may be caused by the start of an occasional high-power electrical appliance, for example, a kettle, and such occasional power consumption changes will affect the time series prediction of the power consumption, reducing the accuracy of the power consumption prediction, therefore, the predicted power metering data predicted when no abnormality occurs is regarded as the true data for subsequent prediction, avoiding the influence of occasional power consumption abnormalities; the replaced power metering time series analysis set is processed through the power metering time series analysis model to output the predicted power metering data corresponding to the monitoring time point; for the next monitoring time point after the suspected abnormal time point, the deviation value judgment operation is performed, if "δ>G" is established, the current monitoring time point is marked as a suspected abnormal time point, the total number M of suspected abnormal time points is counted, and it is judged whether "M>L" is established, L is an occasional time point threshold value, L is generally 3, if "M>L" is established, a power consumption abnormality causal judgment operation is performed, if "M>L" is not established, the total number M of suspected abnormal time points is continuously counted; if "δ>G" is not established, the total number M of suspected abnormal time points is counted, and it is judged whether "M>L" is established, if "M>L" is not established, all suspected abnormal time points are re-marked as occasional power consumption abnormal time points, and are sent to the user interface, so that the user can check the power consumption situation, and when the power metering time series analysis set is constructed, the power metering data corresponding to the occasional power consumption abnormal time points is replaced with the corresponding predicted power metering data, if "M>L" is established, the power consumption abnormality causal judgment operation is performed;
[0035] The use electricity abnormality cause-effect judgment operation includes the following contents: all suspected abnormal time point corresponding power metering data is composed into a dependent variable time series analysis data set in time sequence, the difference between the suspected abnormal time point corresponding power metering data and the predicted power metering data is calculated and recorded as the change value, all suspected abnormal time point corresponding change value is composed into an independent variable time series analysis data set in time sequence, Granger cause-effect analysis is carried out based on the dependent variable time series analysis data set and the independent variable time series analysis data set, if there is a cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, it shows that the use electricity abnormality is not accidental, but may be due to the regular change caused by the addition or reduction of electrical appliances, all suspected abnormal time points are deleted, and the power metering data is not replaced when constructing the power metering time series analysis set. Since the predicted use electricity abnormality is a regular change, it can be directly integrated into the calculation of the power metering time series analysis model to ensure the prediction accuracy, if there is no cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, all suspected abnormal time points are re-marked as accidental use electricity abnormality time points.
[0036] The application judges accidental use electricity abnormality based on the predicted predicted power metering data by performing time series analysis on the user's use electricity metering data. In the process of time series analysis, the accidental use electricity abnormality is avoided by the predicted predicted power metering data, which avoids the decrease of prediction accuracy caused by accidental use electricity abnormality, and considers the regular change caused by the addition or reduction of electrical appliances, which further improves the prediction accuracy.
[0037] The power metering time series analysis model includes a power metering data updating layer, a power metering time series analysis layer and a predicted power metering data output layer, wherein the power metering data updating layer is used to update the power metering data in the power metering time series analysis set based on the historical power metering data to construct an updated power metering time series analysis set. It should be noted that in the process of daily electricity consumption, regularity between days will be presented, for example, after the white-collar workers go home from work, the electricity consumption will increase. Therefore, updating the power metering data in the power metering time series analysis set by the historical power metering data can improve the confidence of the power metering data and further improve the prediction accuracy; the power metering time series analysis layer is established based on the LSTM model, which is used to process the updated power metering time series analysis set to construct the predicted power metering data; the predicted power metering data output layer is used to output the predicted power metering data.
[0038] The power metering data updating layer updates the power metering data in the power metering time series analysis set based on the historical power metering data to construct the updated power metering time series analysis set, which specifically includes the following steps:
[0039] The power metering time series analysis set is divided into a plurality of power metering time series analysis sequences by a sliding window with a length of C, C being determined by a developer and generally being 4;
[0040] For each power metering time series analysis sequence, the power metering time series analysis sequence is updated by the following formula: , where F2 is the updated power metering time series analysis sequence, F1 is the power metering time series analysis sequence before updating, A 24k-1 is a power metering time series analysis sequence at a monitoring time point 24k-1 away from the power metering time series analysis sequence F1, i.e. 1 day, 2 days, etc., k = 1, 2, 3, …, K, K being a distance upper limit and generally being 4, i.e. at most 4 days, and a 24k-1 is a similarity between the power metering time series analysis sequence F1 and the power metering time series analysis sequence A 24k-1 , generally calculated by a cosine similarity algorithm, and then based on the updated power metering time series analysis sequence F2 and the power metering time series analysis sequence F1 before updating, a self-attention mechanism operation is performed to complete updating of the power metering data in the power metering time series analysis set and build an updated power metering time series analysis set; it should be noted that the power metering time series analysis sequence divided by the sliding window can reflect local time series relationship, and updating the power metering data by the power metering time series analysis sequence can pay more attention to changes in the power metering data, thereby obtaining better updating effect;
[0041] Based on the updated power metering time series analysis sequence F2 and the power metering time series analysis sequence F1 before updating, a self-attention mechanism operation is performed to complete updating of the power metering data in the power metering time series analysis set and build an updated power metering time series analysis set, specifically including the following steps:
[0042] All updated power metering time series analysis sequences F2 are spliced from top to bottom in time sequence to build a target feature matrix, and all power metering time series analysis sequences F1 before updating are spliced from top to bottom in time sequence to build a query feature matrix; then the target feature matrix is respectively multiplied by a value weight matrix and a key weight matrix to build corresponding target value matrix V and target key matrix K, and the query feature matrix is multiplied by a query weight matrix to build a query matrix Q, and a self-attention mechanism operation is completed by the following formula, H = softmax(QK T / D 0.5) V, wherein H is an updated feature matrix, T is a matrix transposition operation, D is the dimension size of the target key matrix K, and the updated feature matrix H is flattened by row to obtain an updated power metering time series analysis set; it should be noted that the value weight matrix, the key weight matrix and the query weight matrix are set based on the multi-head self-attention mechanism in the Transformer model, and the value weight matrix, the key weight matrix and the query weight matrix are adjusted in parameters following the training of the power metering time series analysis model;
[0043] Granger causality analysis is performed based on the dependent variable time series analysis dataset and the independent variable time series analysis dataset, specifically including the following steps:
[0044] A significance level η is set, and η is generally 0.05;
[0045] A baseline model is constructed according to the dependent variable time series analysis dataset, and an autoregressive model AR is adopted, which has the form of:
[0046]
[0047] wherein Y t , Y t-1 , …, Y t-p are power metering data in the dependent variable time series analysis dataset, β0, β1, …, β p are corresponding parameters, is an error term;
[0048] The independent variable time series analysis dataset is added to the baseline model to construct a full model, and a vector autoregressive model VAR is adopted, which has the form of:
[0049]
[0050] wherein X t , X t-1 , …, X t-p are power metering data in the dependent variable time series analysis dataset, γ0, γ1, …, γ q are corresponding parameters;
[0051] The residual sum of squares corresponding to the baseline model and the full model is calculated respectively, and the F test is performed based on the residual sum of squares corresponding to the baseline model and the full model to obtain a P value; if the P value is higher than the significance level η, there is a causal relationship between the dependent variable time series analysis dataset and the independent variable time series analysis dataset; if the P value is not higher than the significance level η, there is no causal relationship between the dependent variable time series analysis dataset and the independent variable time series analysis dataset.
[0052] The power metering time series analysis model is trained, specifically including the following steps:
[0053] A plurality of power metering time series analysis training samples are acquired, the power metering time series analysis training samples include N+1 power metering data arranged according to time, and there is no accidental power consumption anomaly. The power metering time series analysis training samples are acquired according to actual power monitoring. All the power metering time series analysis training samples form a power metering time series analysis training set. The power metering time series analysis model is trained through the power metering time series analysis training set. During the training, the last power metering data in the power metering time series analysis training sample is taken as a target to determine whether a training condition is met. The training condition is generally that the accuracy of the power metering time series analysis model meets the expectation. If the training condition is met, the trained power metering time series analysis model is output. Otherwise, the power metering time series analysis model is continuously trained through the power metering time series analysis training set.
[0054] In embodiment 2, a power metering data processing system based on feature recognition is provided, as shown in Figure 1 , comprising:
[0055] A power metering data acquisition module is configured to acquire power metering data at a current monitoring time point. The power metering data refers to power consumption, which is used to reflect the power consumption of a user.
[0056] A power metering time series analysis module is configured to form a power metering time series analysis set by combining the power metering data acquired at the current monitoring time point and the power metering data acquired at the previous N-1 monitoring time points, and then send the power metering time series analysis set to a power metering time series analysis model for processing to output predicted power metering data at a next monitoring time point. It should be noted that the time interval between two adjacent monitoring time points is generally 15 minutes, and the value of N is generally 24. With the assistance of the power metering time series analysis model, the power consumption of the user can be analyzed, and the power consumption of the user can be predicted.
[0057] The monitoring time point judgment module is used for executing a deviation value judgment operation at a current monitoring time point. The deviation value judgment operation includes the following contents: performing deviation value calculation on the acquired power metering data and corresponding predicted power metering data, δ = |U1-U2| / U2, wherein δ is the deviation value, U1 is the power metering data acquired at the current monitoring time point, U2 is the predicted power metering data corresponding to the current monitoring time point, and it is judged whether "δ>G" is established, wherein G is a deviation threshold value, which is used for judging whether the power consumption situation is abnormal. If "δ>G" is established, the current monitoring time point is marked as a suspected abnormal time point. If "δ>G" is not established, the power metering data is continuously acquired at the next monitoring time point. For the monitoring time points after the suspected abnormal time point, the power metering data corresponding to the suspected abnormal time point and the monitoring time points after the suspected abnormal time point is replaced by the corresponding predicted power metering data when the power metering time series analysis set is constructed. It should be noted that when the power consumption situation is abnormal, it is impossible to judge the source of the power consumption abnormality. It may be caused by the start of an occasional high-power electrical appliance, such as a kettle. Such occasional power consumption changes will affect the time series prediction of the power consumption, reducing the accuracy of the power consumption prediction. Therefore, the predicted power metering data predicted when no abnormality occurs is regarded as real data for subsequent prediction, avoiding the influence of occasional power consumption abnormalities. The replaced power metering time series analysis set is processed through the power metering time series analysis model, and the predicted power metering data corresponding to the monitoring time point is output. For the next monitoring time point after the suspected abnormal time point, the deviation value judgment operation is executed. If "δ>G" is established, the current monitoring time point is marked as a suspected abnormal time point. The total number M of suspected abnormal time points is counted. It is judged whether "M>L" is established. L is an occasional time point threshold value. L is generally 3. If "M>L" is established, the power consumption abnormality causality judgment operation is executed. If "M>L" is not established, the total number M of suspected abnormal time points is continuously counted. If "δ>G" is not established, the total number M of suspected abnormal time points is counted. It is judged whether "M>L" is established. If "M>L" is not established, all suspected abnormal time points are re-marked as occasional power consumption abnormal time points and sent to the user interface, so that the user can check the power consumption situation. When the power metering time series analysis set is constructed, the power metering data corresponding to the occasional power consumption abnormal time points is replaced by the corresponding predicted power metering data. If "M>L" is established, the power consumption abnormality causality judgment operation is executed.
[0058] The power consumption anomaly cause-effect judgment operation module is configured to perform power consumption anomaly cause-effect judgment operation, which includes the following contents: all suspected anomaly time point corresponding power consumption metering data is arranged in time sequence to form a dependent variable time series analysis data set, the difference between the suspected anomaly time point corresponding power consumption metering data and the predicted power consumption metering data is calculated and recorded as a change value, all suspected anomaly time point corresponding change values are arranged in time sequence to form an independent variable time series analysis data set, Granger cause-effect analysis is performed based on the dependent variable time series analysis data set and the independent variable time series analysis data set, if there is a cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, it is indicated that the power consumption anomaly is not accidental, but may be a regular change caused by adding or reducing electrical appliances, all suspected anomaly time points are deleted, and when the power consumption time series analysis set is constructed, the power consumption metering data is not replaced, since the predicted power consumption anomaly is a regular change, it can be directly integrated into the calculation of the power consumption time series analysis model, ensuring the accuracy of the prediction, and if there is no cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, all suspected anomaly time points are re-labeled as accidental power consumption anomaly time points.
[0059] It should be understood that, for those skilled in the art, improvements or changes can be made according to the above description, and all these improvements and changes shall belong to the protection scope of the appended claims of the present application. The parts not described in detail in the specification belong to the prior art known to those skilled in the art.
Claims
1. A feature recognition based power metering data processing method, characterized by, Comprise: acquiring power metering data at a current monitoring time point; composing a power metering time series analysis set with the power metering data acquired at the current monitoring time point and the power metering data acquired at the previous N-1 monitoring time points, and sending the power metering time series analysis set into a power metering time series analysis model for processing to output predicted power metering data at a next monitoring time point; Meanwhile, at the current monitoring time point, performing a deviation value judgment operation, which comprises the following contents: performing deviation value calculation on the acquired power metering data and the corresponding predicted power metering data, δ = |U1-U2| / U2, where δ is the deviation value, U1 is the power metering data acquired at the current monitoring time point, U2 is the predicted power metering data corresponding to the current monitoring time point, judging whether "δ>G" is true, where G is a deviation threshold value, if "δ>G" is true, marking the current monitoring time point as a suspected abnormal time point, if "δ>G" is not true, continuing to acquire power metering data at the next monitoring time point; for the monitoring time points after the suspected abnormal time point, replacing the power metering data corresponding to the suspected abnormal time point and the monitoring time points after the suspected abnormal time point with the corresponding predicted power metering data when constructing the power metering time series analysis set; processing the replaced power metering time series analysis set through the power metering time series analysis model to output the predicted power metering data corresponding to the monitoring time point; for the next monitoring time point after the suspected abnormal time point, performing the deviation value judgment operation, if "δ>G" is true, marking the current monitoring time point as a suspected abnormal time point, counting the total number M of suspected abnormal time points, judging whether "M>L" is true, L is an occasional time point threshold value, if "M>L" is true, performing power consumption abnormality causal judgment operation, if "M>L" is not true, continuously counting the total number M of suspected abnormal time points; if "δ>G" is not true, counting the total number M of suspected abnormal time points, judging whether "M>L" is true, if "M>L" is not true, marking all suspected abnormal time points as occasional power consumption abnormal time points again, and replacing the power metering data corresponding to the occasional power consumption abnormal time points with the corresponding predicted power metering data when constructing the power metering time series analysis set, if "M>L" is true, performing power consumption abnormality causal judgment operation; The power consumption anomaly cause-effect judgment operation includes the following contents: all suspected abnormal time points corresponding to the power metering data are sequentially arranged to form a dependent variable time series analysis data set, the difference between the suspected abnormal time points corresponding to the power metering data and the predicted power metering data is calculated and recorded as a change value, all suspected abnormal time points corresponding to the change value are sequentially arranged to form an independent variable time series analysis data set, Granger cause-effect analysis is performed based on the dependent variable time series analysis data set and the independent variable time series analysis data set, if there is a cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, all suspected abnormal time points are deleted, and the power metering data is not replaced when constructing the power metering time series analysis set, if there is no cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, all suspected abnormal time points are re-marked as accidental power consumption abnormal time points.
2. The feature recognition based power metering data processing method of claim 1, wherein, The power metering time series analysis model includes a power metering data updating layer, a power metering time series analysis layer, and a predicted power metering data output layer, wherein the power metering data updating layer is used to update the power metering data in the power metering time series analysis set based on historical power metering data to construct an updated power metering time series analysis set; The power metering time series analysis layer is established based on an LSTM model and is used to process the updated power metering time series analysis set to construct predicted power metering data; The predicted power metering data output layer is used to output the predicted power metering data.
3. The feature recognition based power metering data processing method of claim 2, wherein, The power metering data updating layer updates the power metering data in the power metering time series analysis set based on historical power metering data to construct an updated power metering time series analysis set, specifically including the following steps: The power metering time series analysis set is divided into a plurality of power metering time series analysis sequences through a sliding window with a length of C; For each power metering time series analysis sequence, the power metering time series analysis sequence is updated by the following formula: wherein F2 is the updated power metering time series analysis sequence, F1 is the power metering time series analysis sequence before updating, A 24k-1 is the power metering time series analysis sequence at the 24k-1 monitoring time points before the power metering time series analysis sequence F1, k = 1, 2, 3, …, K, K is the distance upper limit, and 24k-1 is the similarity between the power metering time series analysis sequence F1 and the power metering time series analysis sequence A 24k-1 Based on the updated power metering time series analysis sequence F2 and the power metering time series analysis sequence F1 before updating, the self-attention mechanism operation is performed to complete the updating of the power metering data in the power metering time series analysis set and construct the updated power metering time series analysis set.
4. The feature recognition based power metering data processing method of claim 3, wherein, The updated power metering time series analysis sequence F2 and the power metering time series analysis sequence F1 before updating are used to perform a self-attention mechanism operation to complete the updating of the power metering data in the power metering time series analysis set and construct an updated power metering time series analysis set, specifically including the following steps: All updated power metering time series analysis sequences F2 are spliced from top to bottom in time sequence to construct a target feature matrix, and all pre-updated power metering time series analysis sequences F1 are spliced from top to bottom in time sequence to construct a query feature matrix; then the target feature matrix is multiplied by the value weight matrix and the key weight matrix respectively to construct corresponding target value matrix V and target key matrix K, and the query feature matrix is multiplied by the query weight matrix to construct a query matrix Q, and the self-attention mechanism operation is completed through the following formula: H = softmax (QK T / D 0.5 ) V, where H is an updated feature matrix, T is a matrix transposition operation, D is the dimension size of the target key matrix K, and then the updated feature matrix H is flattened by row to obtain an updated power metering time series analysis set.
5. The feature recognition based power metering data processing method of claim 4, wherein, Granger cause-effect analysis is performed based on the dependent variable time series analysis data set and the independent variable time series analysis data set, specifically including the following steps: A significance level η is set; A baseline model is constructed according to the dependent variable time series analysis data set; An overall model is constructed by adding the independent variable time series analysis data set to the baseline model; The residual sum of squares corresponding to the baseline model and the overall model is calculated, and F test is performed based on the residual sum of squares corresponding to the baseline model and the overall model to obtain a P value, if the P value is higher than the significance level η, there is a cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, if the P value is not higher than the significance level η, there is no cause-effect relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set.
6. The feature recognition based power metering data processing method of claim 5, wherein, The power metering time series analysis model is trained, specifically including the following steps: A plurality of power metering time series analysis training samples are acquired, the power metering time series analysis training samples include N+1 power metering data arranged according to time, and there is no accidental power consumption anomaly, all the power metering time series analysis training samples are combined to form a power metering time series analysis training set, the power metering time series analysis model is trained through the power metering time series analysis training set, during the training, the last power metering data in the power metering time series analysis training sample is taken as a target, whether the training condition is met is judged, if the training condition is met, the trained power metering time series analysis model is output, otherwise, the power metering time series analysis model is continuously trained through the power metering time series analysis training set.
7. A feature recognition based power metering data processing system characterized in that, The system applies the power metering data processing method based on feature recognition in any one of claims 1-6, comprising: a power metering data acquisition module, configured to acquire power metering data at a current monitoring time point; a power metering time series analysis module, configured to combine the power metering data acquired at the current monitoring time point and the power metering data acquired at the previous N-1 monitoring time points to form a power metering time series analysis set, and then send the power metering time series analysis set to a power metering time series analysis model for processing, so as to output predicted power metering data at a next monitoring time point; The monitoring time point judgment module is configured to perform a deviation value judgment operation at a current monitoring time point. The deviation value judgment operation includes the following contents: performing a deviation value calculation on the obtained power metering data and corresponding predicted power metering data, δ=|U1-U2| / U2, where δ is the deviation value, U1 is the power metering data obtained at the current monitoring time point, U2 is the predicted power metering data corresponding to the current monitoring time point, judging whether "δ>G" is true, where G is a deviation threshold value, if "δ>G" is true, marking the current monitoring time point as a suspected abnormal time point, if "δ>G" is not true, continuing to obtain power metering data at a next monitoring time point; for the monitoring time points after the suspected abnormal time point, replacing the power metering data corresponding to the suspected abnormal time point and the monitoring time points after the suspected abnormal time point with corresponding predicted power metering data when constructing a power metering time series analysis set; processing the power metering time series analysis set after the replacement through a power metering time series analysis model to output predicted power metering data corresponding to the monitoring time points; for the next monitoring time point after the suspected abnormal time point, performing the deviation value judgment operation, if "δ>G" is true, marking the current monitoring time point as a suspected abnormal time point, counting a total number M of suspected abnormal time points, judging whether "M>L" is true, L being an occasional time point threshold value, if "M>L" is true, performing a power consumption abnormality causal judgment operation, if "M>L" is not true, continuously counting the total number M of suspected abnormal time points; if "δ>G" is not true, counting the total number M of suspected abnormal time points, judging whether "M>L" is true, if "M>L" is not true, re-marking all suspected abnormal time points as occasional power consumption abnormal time points, and replacing the power metering data corresponding to the occasional power consumption abnormal time points with corresponding predicted power metering data when constructing the power metering time series analysis set, if "M>L" is true, performing the power consumption abnormality causal judgment operation; The power consumption abnormality causal judgment operation module is configured to perform a power consumption abnormality causal judgment operation. The power consumption abnormality causal judgment operation includes the following contents: grouping the power metering data corresponding to all suspected abnormal time points into a dependent variable time series analysis data set in chronological order, calculating the difference between the power metering data corresponding to the suspected abnormal time points and the predicted power metering data and recording the difference as a change value, grouping the change values corresponding to all suspected abnormal time points into an independent variable time series analysis data set in chronological order, performing Granger causal analysis based on the dependent variable time series analysis data set and the independent variable time series analysis data set, if there is a causal relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, deleting all suspected abnormal time points, and not replacing the power metering data when constructing a power metering time series analysis set, if there is no causal relationship between the dependent variable time series analysis data set and the independent variable time series analysis data set, re-marking all suspected abnormal time points as occasional power consumption abnormal time points.