Energy consumption sub-metering comparison method based on multi-source data fusion

Abstract

This invention relates to the field of energy consumption comparison technology, specifically to a method for comparing electricity consumption based on multi-source data fusion, comprising the following steps: constructing an initial power sequence; preprocessing the initial power sequence; defining normal points and normal sequences, extracting multiple core features from the normal sequences to form high-dimensional feature vectors corresponding to the normal points; constructing a correlation matrix based on the high-dimensional feature vectors, then using hierarchical clustering to divide different normal points into different clusters, matching marker points to the corresponding clusters, analyzing the topological relationships between different clusters, and the final window length adapted within the corresponding clusters; using the final window length to identify and correct abnormal power within the corresponding clusters in real time; then comparing energy consumption based on the total power within the clusters, and based on the energy consumption comparison results, locating the topological branches of abnormal energy consumption.

Description

Technical Field

[0001] This invention relates to the field of energy consumption comparison technology, and more specifically, to a method for energy consumption comparison based on multi-source data fusion and itemized metering of electricity consumption. Background Technology

[0002] In the refined energy-saving management of industrial and building scenarios, the sub-metering of electricity consumption is the core foundation for realizing energy consumption monitoring, anomaly diagnosis and energy efficiency optimization. Specifically, it involves breaking down and statistically analyzing the overall electricity load according to dimensions such as purpose, equipment type, and power distribution lines, accurately locating the energy consumption ratio and operating pattern of each sub-item, and providing data support for subsequent energy consumption analysis, anomaly investigation and energy-saving renovation. By using physical diversion and dedicated metering equipment, direct collection of energy consumption for each item can be achieved. That is, separate collection points are set at the power distribution branches of each power consumption item (such as production lines and air compressors in industrial scenarios, and air conditioning, lighting, and socket circuits in building scenarios). Then, electricity meters, current and voltage sensors and other equipment are installed at the collection points, and the branch current, voltage and power data are collected in real time by relying on technologies such as electromagnetic induction and photoelectric conversion. Since power is the core parameter characterizing the intensity of electrical load and the total energy consumption, and the essence of energy consumption is the integral result of power over time, the comparison of energy consumption must include a multi-dimensional power comparison of the real-time power, time-period average power, and peak power of each sub-branch. However, during the power comparison process, there are significant differences in equipment models, sampling frequencies, and timestamp calibration accuracy at different collection points. Some collection points may experience data drift due to equipment aging or improper installation. At the same time, complex on-site power conditions (such as equipment start-up and shutdown shocks and load fluctuations) can lead to instantaneous power changes. Due to the superposition of the above multiple factors, the collected data deviates significantly from the theoretical power value, thus resulting in abnormal power. At the same time, power comparison is not an isolated comparison of data from a single branch. It usually requires a coordinated comparison of the hierarchical relationship of the distribution topology, historical benchmark data under the same operating conditions, and energy consumption benchmark data of similar branches, so as to clarify the rationality of energy consumption of each branch and its topological affiliation. However, due to implicit topology changes in the power distribution system, such as temporary wiring and switching of backup branches, there are discrepancies between the design drawings and the actual topology after on-site construction and rewiring. Therefore, when comparing power based on the power distribution topology, if the topological relationship between each line cannot be clearly known, the accuracy of the comparison will be affected. In view of this, we propose a method for comparing energy consumption by sub-item metering based on multi-source data fusion. Summary of the Invention

[0003] The purpose of this invention is to solve the problem of energy consumption comparison distortion caused by the above-mentioned abnormal power interference and unclear power distribution topology relationship.

[0004] To achieve the above objectives, this invention provides a method for comparing energy consumption by sub-item metering based on multi-source data fusion, comprising the following steps: S1. Obtain the initial active power monitored in real time at different collection points by dedicated equipment, and construct the initial power sequence; The initial power sequence is preprocessed, including setting the sliding window length, using the sliding window to identify abnormal power in the initial power sequence, correcting abnormal power, and finally obtaining the active power sequence. S2. Define the collection point with abnormal power as the marker point, the active power sequence corresponding to the marker point as the marker sequence, the corrected abnormal power as the corrected power, the time corresponding to the abnormal power as the abnormal time, and all collection points except the marker point as normal points, and the active power sequence corresponding to the normal point as the normal sequence. Extract multiple core features from the normal sequence to form a high-dimensional feature vector corresponding to the normal point; S3. Construct a correlation matrix based on high-dimensional feature vectors, and then use hierarchical clustering to divide different normal points into different clusters, match the marked points to the corresponding clusters, and analyze the topological relationship between different clusters. S4. After topological analysis, retrieve the abnormal power corresponding to the marked point in the cluster, and the normal power corresponding to the abnormal time for all normal points in the cluster except the marked point; calculate the mean of normal power, and adjust the sliding window length in step S1 in turn with the minimum difference between the abnormal power and the mean of normal power as a constraint, and finally determine the final window length for the corresponding cluster. S5. The corresponding abnormal power within the cluster is identified and corrected in real time using the final window length; then, energy consumption is compared based on the total power within the cluster, and the topological branch with abnormal energy consumption is located based on the energy consumption comparison results.

[0005] As a further improvement to this technical solution, the preprocessing in step S1 specifically includes time alignment, identification of abnormal power, and correction of abnormal power; the specific working steps are as follows: S1.1: Obtain the initial active power monitored in real time at different collection points by dedicated equipment, and construct the initial active power sequence corresponding to each collection point by continuously monitoring the initial active power in chronological order. S1.2: Used to time-align the initial power sequences corresponding to different acquisition points. S1.3: Set the sliding window length K, identify abnormal power, and correct abnormal power.

[0006] As a further improvement to this technical solution, step S1.3, identifying the abnormal power, specifically involves: Each moment is selected sequentially, and the active power of the K moments before and after the selected moment is taken as the center to form a power subset; The normal power range at selected times is constructed using a power subset; Calculate the deviation between the power corresponding to the selected time point and the mean of the corresponding power subset; Compare with normal power range and compare with deviation: If the comparison deviation is within the normal power range, then the active power corresponding to the selected time is judged to be normal. If the comparison deviation is not within the normal power range, the active power corresponding to the selected time is determined to be abnormal power, and the abnormal power is replaced by the mean of the power subset to obtain the corrected active power sequence.

[0007] As a further improvement to this technical solution, step S2 receives the normal sequence of each normal point, and sequentially extracts multiple core features in the time domain that can characterize the running characteristics of the normal point from the basic dataset corresponding to each normal point; and constructs a high-dimensional feature vector of multiple core features corresponding to the normal point in a fixed order. Specifically, the time-domain characteristics include, but are not limited to, average load, load fluctuation rate, peak percentage, and start / stop frequency; the frequency-domain characteristics include, but are not limited to, peak frequency and frequency concentration.

[0008] As a further improvement to this technical solution, step S3 uses the high-dimensional feature vectors of each normal point as a basis to cross-calculate the correlation coefficient between any two different normal points, and uses the correlation coefficients of each pair of normal points as matrix elements to construct a correlation matrix.

[0009] As a further improvement to this technical solution, step S3 uses hierarchical clustering to divide highly correlated normal points into the same cluster, with each cluster corresponding to a distribution topology branch. The specific steps are as follows: Initialization: Define each normal point as an independent cluster; Calculate inter-cluster distances using the correlation matrix. Based on this, the inter-cluster distance is defined as the average of the correlation coefficients of all normal point pairs within a cluster; Cluster merging: The two clusters with the smallest matching distance are merged into a new cluster; and the correlation matrix is ​​updated. Iterative repetition: Repeatedly calculate inter-cluster distances and merge clusters until all clusters are merged into a single root cluster; Generate a clustering tree: Visualize the clustering process. The branch nodes of the clustering tree correspond to the clusters generated during the merging process. Each branch node corresponds to a power distribution topology branch, and the leaf nodes correspond to the initial independent normal points. All normal points contained in each cluster are retrieved, and all clusters are classified into different cluster sets. Each cluster set contains multiple clusters, and the collection points within multiple clusters do not overlap. The topological relationships are then analyzed.

[0010] As a further improvement to this technical solution, step S3 is based on cluster set analysis of topological relationships: Analyze each cluster set in sequence: sequentially retrieve the abnormal times in the marker sequence corresponding to the marker points in step S1, define the sampling time that matches the abnormal time in each cluster set as the matching time, and calculate the intra-cluster mean of the corresponding matching time in each cluster set; Matching the corrected power with the intra-cluster mean: Calculate the difference between the corrected power and the intra-cluster mean in turn, and match the marked points to the corresponding clusters in order of increasing difference; The active power sequence corresponding to the receiving point is received, the average power of each individual point is calculated sequentially, and the average power of multiple points within a single cluster is added together to obtain the total power of the corresponding cluster. Randomly pair all clusters within the cluster set to form multiple upper and lower level combinations, each consisting of two clusters; and set the line power loss corresponding to each upper and lower level combination; if the total power of cluster A in the upper and lower level combination satisfies the total power of cluster B and the line power loss, then cluster A is determined to be the upper-level topology branch of cluster B; otherwise, pair them again.

[0011] As a further improvement to this technical solution, step S3 takes the clusters that have been identified as upper-level clusters as new objects to be verified, and performs matching verification of total power and corresponding line loss with the clusters at higher levels, until all clusters are included in the hierarchical structure with the cluster where the main incoming line collection point is located as the root node, and finally sorts out the membership relationship of each cluster, thereby obtaining the topological relationship between all collection points.

[0012] As a further improvement to this technical solution, if after multiple pairings, all the upper and lower level combinations within the cluster set do not satisfy the topological relationship analysis in step S3, then the next cluster set is selected again until the topological relationship is analyzed.

[0013] As a further improvement to this technical solution, step S4 adjusts the sliding window length K in step S1, with the minimum difference between the average values ​​of abnormal power and normal power as a constraint, and adjusts the sliding window length K sequentially. During the adjustment process, the sliding window length K is not less than 3 and not greater than the length of the corresponding power sequence.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: In this method for comparing energy consumption by sub-metering based on multi-source data fusion, the sliding window length in the preprocessing step S1 is used to first align the initial power sequence in time, and then identify and correct abnormal power through statistical confidence principles. This transforms the scattered, distorted, and asynchronous raw power data into a unified, reliable, and synchronous active power sequence, thereby effectively filtering out interference caused by sensor noise, equipment drift, and instantaneous operating condition fluctuations. This provides a high-quality data foundation for subsequent feature extraction, cluster partitioning, and topology analysis, and avoids power issues affecting the accuracy of the overall method. Step S3 prioritizes dividing different normal points into different clusters using the high-dimensional feature vectors of all normal points. During the division process, relying on the physical characteristic of strong power synchronization among normal points in the same cluster, the marked points identified in the preprocessing are assigned to the corresponding cluster based on the difference between the corrected power of the marked points and the average normal power at the corresponding abnormal time in each cluster, according to the matching logic of the smallest difference. Then, the topological relationship between different clusters is analyzed based on the power balance relationship between the total power of the cluster and the line loss, clarifying the distribution topology affiliation of all collection points. This solves the problem of energy consumption comparison distortion caused by unclear topological relationships in traditional methods, and provides structured topological support for energy consumption anomaly location. Step S4, based on the average power of normal points within the same cluster at abnormal times (which can truly reflect the energy consumption level of the branch at that time), adjusts the sliding window length in the preprocessing process of step S1 with the constraint of minimizing the difference between the abnormal power and the average normal power, to obtain the final window length specific to each cluster. This avoids the problem of misjudgment or missed judgment of anomalies caused by the initial empirical setting of the sliding window length in step S1 not matching the actual fluctuation cycle of the branch. At the same time, it makes the subsequent abnormal power correction more in line with the actual operating characteristics of the branch, providing a solid foundation for the energy consumption comparison in step S5. Ultimately, it realizes the scientificity and operability of energy consumption monitoring, anomaly diagnosis and energy efficiency optimization in refined energy-saving management in industrial, building and other scenarios.

[0015] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall steps of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The method for comparing energy consumption by sub-item metering based on multi-source data fusion includes the following steps: S1. Obtain the initial active power monitored in real time at different collection points by dedicated equipment, and construct the initial power sequence; The initial power sequence is preprocessed, including setting the sliding window length K, using the sliding window to identify abnormal power in the initial power sequence, correcting abnormal power, and finally obtaining the active power sequence; In step S1, the initial power sequence and active power sequence corresponding to sampling point i are as follows: Initial power sequence: ; Active power sequence: ; in, For sampling point i at time... active power, Number of samples; in, For the i-th collection point at time... active power, Number of samples; S2. Define the collection point with abnormal power as the marker point, the active power sequence corresponding to the marker point as the marker sequence, the corrected abnormal power as the corrected power, the time corresponding to the abnormal power as the abnormal time, and all collection points except the marker point as normal points, and the active power sequence corresponding to the normal point as the normal sequence. Extract multiple core features from the normal sequence to form a high-dimensional feature vector corresponding to the normal point; S3. Construct a correlation matrix based on high-dimensional feature vectors, and then use hierarchical clustering to divide different normal points into different clusters, match the marked points to the corresponding clusters, and analyze the topological relationship between different clusters. S4. After topological analysis, retrieve the abnormal power corresponding to the marked point in the cluster, and the normal power corresponding to the abnormal time for all normal points in the cluster except the marked point; calculate the mean of normal power, and adjust the sliding window length in step S1 in turn with the minimum difference between the abnormal power and the mean of normal power as a constraint, and finally determine the final window length for the corresponding cluster. S5. The corresponding abnormal power within the cluster is identified and corrected in real time using the final window length; then, energy consumption is compared based on the total power within the cluster, and the topological branch with abnormal energy consumption is located based on the energy consumption comparison results. In the implementation of the above embodiments, the initial power sequence is preprocessed in step S1, thereby providing a normal sequence without abnormal interference and capable of accurately characterizing the actual operating law of the power distribution branch for steps S2 and S3; Step S2 extracts the core features of the normal sequence in the time and frequency domains and constructs the high-dimensional feature vectors corresponding to the normal points. Step S3 prioritizes constructing a correlation matrix using the high-dimensional feature vectors of all normal points. Based on the core criteria of strong synchronization of the operating characteristics of normal points in the same cluster and corresponding to the same distribution topology branch, hierarchical clustering is used to divide different normal points into different clusters. During the division process, relying on the power synchronization of normal points within the cluster, the difference between the power correction of the marker point and the average normal power at the corresponding abnormal time in each cluster can be calculated. Marker points are introduced according to the principle of minimizing the difference. After introducing the marker points, the total power within the cluster is calculated, the clusters are paired up and down and the line power loss is checked. Based on the distribution physics law that the total power of the upper-level cluster = the total power of the lower-level cluster + the line loss, the topological relationship between different clusters is analyzed. After analyzing the topological relationship in step S3, step S4 retrieves the abnormal power before correction, the abnormal time, and the normal power of the corresponding normal point in the same cluster at the same time. With the minimum difference between the average abnormal power and the normal power as a constraint, the sliding window length in the preprocessing process of step S1 is further adjusted so that the subsequent preprocessing of abnormal power in step S5 can more accurately adapt to the actual fluctuation characteristics of the corresponding branch, effectively reducing the false judgment rate and false negative rate of anomaly identification. Furthermore, when step S5 performs multi-dimensional energy consumption comparison based on the reliable total power within the cluster after correction, it can accurately locate the topological branch of energy consumption anomaly, providing a clear optimization target for refined energy-saving management. Furthermore, in step S1, to avoid inherent differences in the initial power sequences from different sampling points due to equipment sampling frequencies and data transmission delays, which could lead to inconsistent clock reference calibration and consequently, inaccurate sampling times and mismatches in acquisition timing, and to prevent data drift caused by differences in equipment models, aging, or improper installation, as well as instantaneous power fluctuations caused by complex operating conditions such as equipment start-up and shutdown shocks and load fluctuations, which could result in deviations between the initial power and normal power, leading to abnormal power and ultimately causing distortion in high-dimensional feature extraction in step S2, cluster partitioning deviations in step S3, and failure of topology relationship analysis, preprocessing of the initial power sequence is necessary to ensure the consistency, accuracy, and reliability of the active power in the preprocessed active power sequence. More detailed steps are as follows: S1.1: Obtain the initial active power monitored in real time by the dedicated equipment at different collection points, and construct the initial active power sequence corresponding to each collection point by continuously monitoring the initial active power in chronological order. Specifically, the initial power sequence of collection point i is as follows: ; in, For sampling point i at time... The initial active power, Number of samples; Considering that the initial active power of each collection point is monitored independently using dedicated equipment, and that during parameter comparison at different initial active power levels, inherent differences in sampling frequency and data transmission delay between different devices prevent the clock references of each dedicated monitoring device from being completely uniformly calibrated, resulting in inconsistent timestamps for the initial active power monitoring at different collection points and inaccurate matching of collection times, S1.2 is used to time-align the initial power sequences corresponding to different collection points to avoid the above situation. The specific working principle is as follows: Set the reference clock and the target sampling interval; Obtain the initial active power and its timestamp collected by each dedicated monitoring device; Using the target sampling time of the reference clock as the anchor point, the initial active power is interpolated and calculated to complete the initial active power at each target time, so as to achieve complete alignment of the timestamps of all dimensions of power consumption parameters. S1.3: Identify Abnormal Power and correct abnormal power The details are as follows: Set the sliding window length K: This is used to accurately identify and correct abnormal power in the initial power sequence, providing reliable data support for subsequent topology analysis. To avoid distortion of local power patterns caused by the center offset of the sliding window length K, and to ensure that the target time is always in the center of the window, with a balanced number of sampling times covered before and after, so as to accurately capture the instantaneous fluctuation characteristics of the distribution branch and avoid sampling deviation in the time dimension, the sliding window length K must be set to an odd number; and the sliding window length K is not less than 3. Each time point is selected sequentially, and a power subset is formed by taking the active power of the K time points before and after the selected time point as the center and using a sliding window of length K. ; Based on the statistical confidence principle, a normal power range for the selected time point is constructed. ,in , (where m is the time index, and its value range is...) ) are the mean and standard deviation of the power subset, respectively. For statistical confidence coefficients (usually taken as 2-3); Calculate the difference between the power at the selected time point and the mean of the corresponding power subset. ; Compared to normal power range and comparison deviation If the comparison deviation Not within the normal power range Within this range, the active power corresponding to the selected time point is determined. For anomalous power, the mean of the corresponding power subset is used. Alternate abnormal power The corrected active power sequence is obtained.

[0019] When setting the sliding window length K in step S1.3, the load fluctuation cycle and operating stability of different power distribution scenarios (such as industrial power branches, building lighting branches, air conditioning circuits, etc.) vary significantly. Industry experience has summarized a reasonable range of K values ​​suitable for different scenarios through a large number of practical applications (such as power branches with drastic load fluctuations and frequent equipment start-ups and shutdowns in industrial scenarios, where K is preferably set to 3~5 to focus on instantaneous effective data; and lighting or air conditioning branches with steady operating conditions and gentle fluctuations in building scenarios, where K is preferably set to 7~9 to filter local noise). Specifically, a bootstrap sample set can be generated through repeated sampling with replacement, and the confidence coefficient can be determined based on the statistical characteristics of the sample set. The specific steps are as follows: Step 1: From the current power subset (e.g., 5 power values ​​within a sliding window), sample 1000 times with replacement to generate 1000 bootstrap samples; Step 2: Calculate the mean and standard deviation of each self-help sample to obtain 1000 intervals of mean ± α × standard deviation; Step 3: Calculate the α value covering 95% (or 99.7%) of the self-help sample, which is the confidence coefficient.

[0020] Specifically, in the industrial power branch scenario, the working principle of setting the sliding window length K is as follows: Obtain the mapping table between the indicators and the sliding window length K in the industrial power branch; The load fluctuation coefficient and start-stop frequency of the branch are used as indicators to determine the severity of the branch fluctuation, and then the sliding window length K is matched accordingly. The more severe the fluctuation, the smaller the sliding window length K is (focusing on instantaneous data); the relatively gentle the fluctuation (such as a continuously running motor), the larger the sliding window length K is (balancing noise filtering). The load fluctuation coefficient is: ; in: Let n be the average load of the branch line, and n be the number of samplings. The active power at time k; The corresponding standard deviation; The start / stop frequency is: ; in: Rated power of branch circuit equipment; This is a sign function (takes 1 when the input is >0, -1 when it is <0, and 0 when it is =0). Based on the load fluctuation coefficient Start-stop frequency The matching process corresponds to the sliding window length K in the mapping table. ; Furthermore, in step S2, the collection point with abnormal power is defined as the marker point, the active power sequence corresponding to the marker point is defined as the marker sequence, the corrected abnormal power is defined as the corrected power, and the time corresponding to the abnormal power is defined as the abnormal time. All data collection points other than the marker points are defined as normal points, and the active power sequence corresponding to the normal points is defined as the normal sequence. The system receives the normal sequence of each normal point and, based on the core logic that has strong synchronization of the load operation law of the normal point corresponding to the distribution topology branch, extracts multiple core features in the basic dataset corresponding to each normal point in the time domain (average load, load fluctuation rate, etc.) and frequency domain (peak frequency, frequency concentration, etc.) that can characterize the normal point operation characteristics. Multiple core features corresponding to normal points are constructed into high-dimensional feature vectors in a fixed order, and the feature order used when constructing high-dimensional feature vectors for all normal points is kept consistent to ensure that the dimension and component meaning of the feature vectors of each normal point are consistent, thus providing a comparable quantitative basis for subsequent distribution topology branching. Specifically, taking the extraction of time-domain and frequency-domain features corresponding to normal point i as an example, the working principle of step S2 for extracting core features is as follows: Receive the normal sequence of normal point i in step S1 ; in, For the i-th normal point at time... active power, Number of samples; The temporal feature extraction in step S2 is specifically as follows: Because normal power consumption points within the same distribution topology branch are driven by the same production cycle and control strategy, their active power operation patterns exhibit strong synchronicity. Time-domain characteristics can intuitively and quantitatively capture and characterize this synchronicity, thus enabling the extraction of normal sequences. The time-domain characteristics of the distribution topology are as follows: because the normal points of different distribution topology branches have significant differences in load magnitude, fluctuation mode, full load operation ratio and start-stop coordination logic, and the above significant differences can be accurately distinguished by corresponding time-domain indicators, the specific time-domain characteristics include but are not limited to average load, load fluctuation rate, peak ratio and start-stop frequency, which are used to accurately characterize the operating status of active power at the normal point and the core characteristics of load magnitude, fluctuation mode, full load operation ratio and equipment coordination start-stop rules at the distribution topology branch level. Normal sequence Medium average load Load fluctuation rate Peak percentage Start-stop frequency The specific expression is as follows: Average load: ; Load volatility: ,in Normal sequence Standard deviation; Peak percentage: ,in, Let be the rated power of the device corresponding to the i-th normal point. For indicator functions (specifically if) When the time is right, the value is 1; otherwise, the value is 0. Start-stop frequency: , of which Sign function (if) or When the value is greater than 0, the value is 1; when it is less than 0, the value is -1; when it is equal to 0, the value is 0. The frequency domain feature extraction in step S2 is specifically as follows: Because the normal power consumption points within the same distribution topology branch are driven by fixed production processes or operating cycles, their active power fluctuations exhibit significant periodic characteristics. Frequency domain analysis can accurately uncover and quantify the periodicity hidden in time domain data. Normal power consumption points in different distribution topology branches show significant differences in the dominant load fluctuation period and energy concentration. Therefore, extracting normal power consumption sequences is crucial. The frequency domain characteristics in the data include, but are not limited to, peak frequency and frequency concentration. The specific frequency domain characteristics are used to accurately distinguish the periodic patterns of load fluctuations in different distribution topologies and to verify the operational coordination of the topology branches to which the normal point belongs. Using Fast Fourier Transform to transform normal sequences Convert to frequency domain complex sequence Extracting peak frequency Frequency concentration ; Preprocessing normal sequences Apply Hanning window weighting: in, Weight of Hanning windows (length of Hanning windows) , (The first and last weights are 0). If normal sequence length For normal sequences after windowing Pad with zeros to the NFFT length to ensure FFT computation efficiency; Frequency domain complex sequence for: Normal sequence after windowing and zero padding Convert the time-domain signal into a frequency-domain complex sequence: For rotation factor, Let j be the phase increment of the k-th frequency component, where j is the imaginary unit; Peak frequency for: ; in, For the normal point i, the frequency domain complex sequence In the middle, the fluctuation range The index of the largest frequency component. The sampling frequency; Frequency Concentration for: ; in, For frequency domain complex sequences In the middle, the fluctuation amplitude of the frequency component corresponding to the peak frequency (i.e., the maximum) value); For frequency domain complex sequences Center front The sum of the fluctuation amplitudes of the effective frequency components; The time-domain features (including average load, load fluctuation rate, peak percentage, and start-stop frequency) and frequency-domain features (including peak frequency and frequency concentration) extracted from the normalized normal point i are used to obtain the feature components in the high-dimensional feature vector corresponding to the normal point i. ; in, The normalized feature value (time domain feature or frequency domain feature) of the m-th feature (time domain feature or frequency domain feature) of the normal point i, with a value range of [0,1]; The original value of the m-th feature at normal point i (such as features extracted from average load, peak frequency, etc.). These are the minimum and maximum values ​​of the same feature m for all normal points, respectively. By integrating the normalized feature values ​​in a fixed order of average load, load fluctuation rate, peak percentage, start-stop frequency, peak frequency, and frequency concentration, a high-dimensional feature vector for normal point i is obtained: .

[0021] To clarify the topological relationships of all collection points, step S3 constructs a correlation matrix based on the correlation coefficients between different normal points; and uses hierarchical clustering to divide different collection points into different clusters, and analyzes the topological relationships between different clusters. Furthermore, when analyzing the topological relationships between different clusters using hierarchical clustering, normal points are prioritized for classification. Then, based on the normal sequence within the classified clusters, the clusters to which the marker points belong are further assigned. During the assignment of marker points, the physical principles of strong power synchronization between sampling points on the same branch of the power distribution system and power balance between upper and lower level branches are strictly followed. The core criterion is minimizing the difference between the corrected power of the marker point and the average normal power of normal points within the cluster during the same period. This ensures that the cluster to which the marker point belongs is completely matched with its actual power supply branch, thereby achieving the goal of ensuring that the operating characteristics of all sampling points within the cluster are compatible and the topological hierarchy between clusters conforms to the physical structure of the power distribution system. To further ensure the accuracy of cluster division and the reliability of topology relationships, this effectively avoids the situation where, after the abnormal power is corrected by the sliding window length K in step S1, traces of abnormal interference may still remain in the marker sequence of the marker point. If the marker point is included in the cluster first, it will destroy the strong synchronization of normal points within the cluster. Consequently, when extracting the core features of the corresponding marker sequence (average load and load fluctuation rate in the time domain, peak frequency and frequency concentration in the frequency domain, etc.) in step S2, the feature values ​​deviate from the true branch characteristics and become distorted, ultimately causing the cluster division and distribution topology branches to mismatch and the topology relationship determination to fail. The working principle of step S3 in more detail is as follows: Constructing a correlation matrix: Based on the high-dimensional feature vectors of each normal point, the correlation coefficient between any two different normal points is calculated cross-referenced to quantify the similarity between normal points. Using the pairwise correlation coefficients of normal points as matrix elements, a correlation matrix reflecting the synchronous correlation among normal points is constructed. ; Wherein, the high-dimensional feature vectors corresponding to normal point i and normal point j are respectively , (d is the dimension of the feature vector, i.e., the total number of extracted core features); the correlation coefficients between normal point i and normal point j are as follows: ; in, , These are the mean values ​​of the high-dimensional feature vectors corresponding to normal point i and normal point j, respectively. In the scenario of power distribution topology level determination, since multiple collection points on the same power distribution line share the same power supply circuit, and are affected by the unified load scheduling strategy, voltage fluctuation law and line impedance characteristics, their power, current and other operating characteristics will show a high degree of synchronization. Thus, there may be a situation where multiple collection points are on the same line. Therefore, in order to eliminate the redundancy of monitoring data of a single collection point and avoid the interference of local instantaneous fluctuations on the topology determination results. Meanwhile, step S1 has completed the time alignment of the initial power sequence, the identification and correction of abnormal power, and identified the collection points (marked points) with abnormal power and the normal points without abnormalities. Although the marked points have been corrected for abnormalities, the power sequence may still have residual traces of abnormal interference. If they are directly used in clustering with normal points, it will destroy the consistency of features within the cluster. However, the power sequence of normal points can accurately represent the actual operating rules of the distribution branch. The high-dimensional feature vector composed of the core features extracted from the time domain (average load, load fluctuation rate, etc.) and frequency domain (peak frequency, frequency concentration, etc.) can quantify the synchronization of operation through the correlation coefficient, providing a reliable basis for branch division. Therefore, step S3 uses hierarchical clustering to divide the normal points with high correlation into the same cluster. Each cluster corresponds to a distribution topology branch. The specific steps and corresponding working principles are as follows: Initialization: Since the topological relationships of each normal point are not clearly defined in the initial state, each normal point is defined as an independent cluster to ensure that each potential topological branch can be initially identified; Calculate inter-cluster distances using the correlation matrix. Based on this, the inter-cluster distance is defined as the average of the correlation coefficients between all normal points within a cluster. The smaller the inter-cluster distance, the stronger the synchronization of the operational characteristics of normal points within the cluster, and therefore the higher the cluster similarity. The specific inter-cluster distance is obtained by summing the correlation coefficients of all normal points and then dividing by the number of normal points within the cluster. Cluster merging: Since the two clusters with the smallest distance have the strongest synchronization and conform to the physical characteristics of high coordination of the same power distribution branch, the two clusters with the smallest inter-cluster distance are merged into a new cluster; and the correlation matrix is ​​updated. Iterative repetition: Repeatedly calculate the distance between clusters and merge clusters until all clusters are merged into a root cluster. The topology of the specific power transmission and distribution system is hierarchical. The merging process from a single normal point to the root cluster can simulate the hierarchical aggregation logic of the distribution branch from the end branch to the main incoming line. Through iteration, the topological association path of all normal points can be completely restored. Furthermore, because the correlation matrix is ​​constructed by cross-calculating the correlation coefficient between any two different normal points, it can comprehensively cover the operational synchronization association of all normal point pairs without missing any point pair relationships. Before clustering, it initializes (defining each normal point as an independent cluster) and then combines bottom-up agglomerative hierarchical clustering logic to calculate the distance between clusters (taking the average of the correlation coefficients of all normal point pairs in two clusters) and iteratively merges the clusters with the highest similarity. At the same time, normal points on the same topology branch in the power distribution scenario are affected by the same power supply circuit and load strategy, and their operating characteristics are highly synchronized and their correlation coefficients are significantly higher than those of normal points on different branches, forming corresponding cluster division boundaries. There will be no misjudgment across branches, so there will be no omissions in the subsequent clustering of multiple normal points.

[0022] Generate a clustering tree: Visualize the clustering process. The branch nodes of the clustering tree correspond to the clusters generated during the merging process. Each branch node corresponds to a power distribution topology branch, and the leaf nodes correspond to the initial independent normal points, thus providing a clear basis for the topology relationship and avoiding confusion of branch affiliation caused by abstract data. Retrieve all normal points contained in each cluster. Since the core attribute of a cluster is that the sampling point belongs to a unique entity and each cluster must independently correspond to a power distribution topology branch, all clusters are classified into different cluster sets. Each cluster set contains multiple clusters, and the sampling points within multiple clusters do not overlap. This further filters for those that conform to the circuit topology relationship. Specifically: Analyze each cluster set in sequence: sequentially retrieve the abnormal times in the marker sequence corresponding to the marker points in step S1, define the sampling time that matches the abnormal time in each cluster set as the matching time, and calculate the intra-cluster mean of the corresponding matching time in each cluster set; Matching the corrected power with the intra-cluster mean: Calculate the difference between the corrected power and the intra-cluster mean in turn, and match the marked points to the corresponding clusters in order of increasing difference; The active power sequence corresponding to the receiving points (including normal points and marked points) is received, and the average power of each individual receiving point is calculated in turn to smooth out instantaneous fluctuations and offset random measurement errors, so as to obtain a quantitative value (i.e. average power) that can stably reflect the long-term load level of the receiving points. Then, the average power of multiple receiving points in a single cluster is added together to obtain the total power of the corresponding cluster. Randomly pair all clusters within the cluster set to form multiple hierarchical combinations of two clusters; and set the line power loss corresponding to each hierarchical combination; if the total power of cluster A in the hierarchical combination satisfies the total power of cluster B and the line power loss, then cluster A is determined to be the upper-level topology branch of cluster B; otherwise, pair them again. Clusters that have been identified as superiors are treated as new objects to be verified. The total power and corresponding line loss of the clusters are matched and verified with the higher-level clusters until all clusters are included in the hierarchical structure with the cluster where the main incoming line collection point is located as the root node. Finally, the membership relationship of each cluster is sorted out, thereby obtaining the topological relationship between all collection points. If, after multiple pairings (including classifying labeled clusters into different clusters and forming different hierarchical combinations), all hierarchical combinations within a cluster set do not satisfy the topological relationship analysis, then the next cluster set is selected again until the topological relationship is analyzed. The specific topological relationships are as follows: Upper and lower topology branches: When power balance verification is performed using line power loss, if the physical law of the power distribution system that the total power of the upper cluster = the total power of the lower cluster + line power loss can be satisfied, then the two clusters are determined to be upper and lower topology branches. The upper cluster corresponds to the main power distribution line, and the lower cluster corresponds to the branch circuits branched from the main line, which conforms to the hierarchical architecture of the power distribution system. Same-level topology branch: When using line power loss for power balance verification, if the total power of two clusters is in the same power order and the difference between each cluster and the total power of the common upper-level cluster is within a reasonable line loss range, but the two clusters cannot form a subordinate relationship where the total power of one cluster = the total power of the other cluster + line power loss, then the two clusters are determined to be same-level topology branches, corresponding to parallel same-level branches in the power distribution system, powered by the same upper-level branch and without direct subordinate relationship.

[0023] Since step S1 does not obtain the distribution topology association information corresponding to each collection point when identifying and correcting abnormal power (the topology relationship needs to be determined by the subsequent step S3), it is impossible to know which collection points belong to the same distribution line and have synchronous operation. Therefore, the sliding window length K can only be set based on industry general experience, which is difficult to adapt to the actual power fluctuation characteristics of different branches. When identifying and correcting abnormal power using the sliding window length K, the local continuous power data of a single acquisition point in the corresponding active power sequence is used (which can only reflect the time dimension fluctuation of a single point and cannot rely on the spatial synchronicity of other points in the same branch). However, after analyzing the topology in step S3, if multiple acquisition points are in the same branch and are affected by the same power supply circuit, load scheduling strategy and line impedance characteristics, their power operation characteristics are strongly synchronized. Therefore, they will be divided into the same cluster in step S3. The normal power at the time corresponding to multiple collection points in the same cluster as the abnormal power is the actual operating power of all normal points in the cluster at that time (no abnormal interference, high data reliability, and can jointly characterize the actual energy consumption level of the branch at that time). Therefore, step S4 is used to retrieve the abnormal power and abnormal time before correction after step S3 has completed the analysis of the topology relationship, as well as the normal power at the abnormal time corresponding to all normal points in the same cluster as the marked point. Adjust the sliding window length in step S1, using the minimum difference between the average abnormal power and normal power as a constraint. Adjust the sliding window length K sequentially (K is strictly limited to no less than 3 during adjustment (ensuring the window contains enough data points to avoid insufficient data to capture the inherent operating patterns of branches, leading to deviations in the normal power range definition), and the maximum length should not exceed the length of the corresponding power sequence (to avoid the window exceeding the data range, resulting in empty windows without effective data points or data redundancy)). Finally, determine the final window length adapted within the corresponding cluster. Therefore, when carrying out anomaly identification and data preprocessing in the subsequent work, the normal power range in step S1 can be further set according to the cluster-specific final window length. The core working principle is: the final window length has been optimized by the synchronization of normal points in the cluster to accurately adapt to the actual power fluctuation cycle of the corresponding topology branch. Continuing the statistical confidence principle, the normal power range in step S1 is further set according to the final window length, thereby avoiding the range definition deviation caused by the initial general K in step S1 and significantly improving the accuracy of anomaly power identification. Furthermore, after the final window length is determined, steps S1 to S3 can be iterated again. The working principle is as follows: the optimized final window length can significantly improve the quality of anomaly identification and data correction in S1, and make the high-dimensional feature vector extracted in S2 more accurately represent the operating characteristics of normal points. Based on this feature vector, the correlation matrix construction, hierarchical clustering and topology analysis in S3 can be carried out again to verify whether the original topology is distorted due to improper initial K-fitting. If the topology still satisfies the power distribution physical law of strong synchronization of the sampling points within the cluster and power balance between upper and lower levels between clusters (total power of the upper cluster = total power of the lower cluster + line loss), then the original topology is reliable. If there is an anomaly, the cluster where the marked point is located needs to be rematched until a stable topology that perfectly matches the actual power distribution structure is found. In summary: This example first constructs an initial power sequence through step S1, and then uses a pre-set sliding window length K to identify anomalous power in the initial power sequence. and correct abnormal power This allows for the transformation of scattered and distorted raw power data into a unified, synchronized, and reliable active power sequence. It effectively filters out transient anomalies caused by sensor noise, equipment drift, and fluctuations in operating conditions, preventing data issues from interfering with subsequent topology analysis and energy consumption comparison. This provides a high-quality data foundation for subsequent feature extraction and cluster partitioning. Then, the topological relationship is further analyzed through steps S2 and S3. During the analysis, multiple clusters are constructed based on normal points. Then, the correction power after the sliding window length K is corrected, and the difference between the mean values ​​of each cluster is used to match the calibration points to different clusters. Since the sliding window length K is an odd number and has been set based on preliminary experience, the corrected abnormal power (corrected power) has eliminated obvious distortion components and is comparable to the power of normal points in the same branch. Therefore, by calculating the difference between the corrected power and the average normal power at the corresponding abnormal time in each cluster, the marker point (including the collection point of abnormal power) is matched to different clusters according to the principle of minimum difference. Furthermore, based on the power distribution physical law that the power of the collection points in the same branch is highly consistent at the same time, it is ensured that the cluster to which the marker point belongs is completely consistent with its actual power supply branch. Furthermore, based on this, the cluster to which the final collection point belongs is determined, and after analyzing the cluster, step S4 again determines the corresponding final window length within each cluster based on the abnormal power that has not yet been corrected in step S1. This allows the sliding window to accurately adapt to the actual power fluctuation characteristics of the corresponding branch, avoiding the inability of small K to capture the steady-state branch pattern due to insufficient data, and avoiding the introduction of irrelevant data by large K to mask the instantaneous anomalies of branches with violent fluctuations. At the same time, it provides an optimization basis for subsequent iterations to verify the reliability of the topology. Step S5, after determining the topological relationship, can identify and correct newly emerging abnormal power within the cluster in real time using the final window length, ensuring the authenticity and consistency of power data within the cluster. Then, based on the corrected power of each collection point within the cluster, the instantaneous total power of the cluster and the total energy consumption of the time period are calculated. A multi-dimensional energy consumption comparison system is constructed through power balance comparison between upper and lower level clusters, benchmark comparison between clusters at the same level, and historical trend comparison. This not only solves the problem of energy consumption comparison distortion caused by the ambiguity of topological relationship and the deviation of anomaly identification in traditional methods, but also realizes a closed loop of the entire process from power preprocessing, topology analysis, anomaly optimization, and energy consumption monitoring. It provides clear optimization targets for refined energy-saving management in industrial, building and other scenarios, helping to reduce energy waste and improve the scientificity and operability of energy efficiency optimization decisions.

[0024] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for comparing energy consumption by sub-item metering based on multi-source data fusion, characterized in that, Includes the following steps: S1. Obtain the initial active power monitored in real time at different collection points by dedicated equipment, and construct the initial power sequence; The initial power sequence is preprocessed, including setting the sliding window length, using the sliding window to identify abnormal power in the initial power sequence, correcting abnormal power, and finally obtaining the active power sequence. S2. Define the collection point with abnormal power as the marker point, the active power sequence corresponding to the marker point as the marker sequence, the corrected abnormal power as the corrected power, the time corresponding to the abnormal power as the abnormal time, and all collection points except the marker point as normal points, and the active power sequence corresponding to the normal point as the normal sequence. Extract multiple core features from the normal sequence to form a high-dimensional feature vector corresponding to the normal point; S3. Construct a correlation matrix based on high-dimensional feature vectors, and then use hierarchical clustering to divide different normal points into different clusters, match the marked points to the corresponding clusters, and analyze the topological relationship between different clusters. S4. After topological analysis, retrieve the abnormal power corresponding to the marked point in the cluster, and the normal power corresponding to the abnormal time for all normal points in the cluster except the marked point; calculate the mean of normal power, and adjust the sliding window length in step S1 in turn with the minimum difference between the abnormal power and the mean of normal power as a constraint, and finally determine the final window length for the corresponding cluster. S5. The corresponding abnormal power within the cluster is identified and corrected in real time using the final window length; then, energy consumption is compared based on the total power within the cluster, and the topological branch with abnormal energy consumption is located based on the energy consumption comparison results.

2. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 1, characterized in that: The preprocessing in step S1 specifically includes time alignment, identification of abnormal power, and correction of abnormal power; the specific working steps are as follows: S1.1: Obtain the initial active power monitored in real time at different collection points by dedicated equipment, and construct the initial active power sequence corresponding to each collection point by continuously monitoring the initial active power in chronological order. S1.2: Used to time-align the initial power sequences corresponding to different acquisition points. S1.3: Set the sliding window length K, identify abnormal power, and correct abnormal power.

3. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 3, characterized in that: Step S1.3, identifying the abnormal power, specifically involves: Each moment is selected sequentially, and the active power of the K moments before and after the selected moment is taken as the center to form a power subset; The normal power range at selected times is constructed using power subsets; Calculate the deviation between the power corresponding to the selected time point and the mean of the corresponding power subset; Compare with normal power range and compare with deviation: If the comparison deviation is within the normal power range, then the active power corresponding to the selected time is judged to be normal. If the comparison deviation is not within the normal power range, the active power corresponding to the selected time is determined to be abnormal power, and the abnormal power is replaced by the mean of the power subset to obtain the corrected active power sequence.

4. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 3, characterized in that: Step S2 receives the normal sequence of each normal point and sequentially extracts multiple core features in the time domain that can characterize the operating characteristics of the normal point from the basic dataset corresponding to each normal point. Construct high-dimensional feature vectors from multiple core features corresponding to normal points in a fixed order; Specifically, the time-domain characteristics include, but are not limited to, average load, load fluctuation rate, peak percentage, and start / stop frequency; the frequency-domain characteristics include, but are not limited to, peak frequency and frequency concentration.

5. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 4, characterized in that: Step S3 is based on the high-dimensional feature vectors of each normal point, cross-calculates the correlation coefficient between any two different normal points, and constructs a correlation matrix using the correlation coefficients of each pair of normal points as matrix elements.

6. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 5, characterized in that: Step S3 uses hierarchical clustering to divide highly correlated normal points into the same cluster, with each cluster corresponding to a distribution topology branch. The specific steps are as follows: Initialization: Define each normal point as an independent cluster; Calculate inter-cluster distances using the correlation matrix. Based on this, the inter-cluster distance is defined as the average of the correlation coefficients of all normal point pairs within a cluster; Cluster Merging: The two clusters with the smallest matching distance are merged into a new cluster; And update the correlation matrix; Iterative repetition: Repeatedly calculate inter-cluster distances and merge clusters until all clusters are merged into a single root cluster; Generate a clustering tree: Visualize the clustering process. The branch nodes of the clustering tree correspond to the clusters generated during the merging process. Each branch node corresponds to a power distribution topology branch, and the leaf nodes correspond to the initial independent normal points. All normal points contained in each cluster are retrieved, and all clusters are classified into different cluster sets. Each cluster set contains multiple clusters, and the collection points within multiple clusters do not overlap. The topological relationships are then analyzed.

7. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 6, characterized in that: Step S3 is based on cluster set analysis of topological relationships: Analyze each cluster set in sequence: sequentially retrieve the abnormal times in the marker sequence corresponding to the marker points in step S1, define the sampling time that matches the abnormal time in each cluster set as the matching time, and calculate the intra-cluster mean of the corresponding matching time in each cluster set; Matching the corrected power with the cluster mean: Calculate the difference between the corrected power and the mean of each cluster in turn, and match the marked points to the corresponding clusters in order of increasing difference; The active power sequence corresponding to the receiving point is received, the average power of each individual point is calculated sequentially, and the average power of multiple points within a single cluster is added together to obtain the total power of the corresponding cluster. Randomly pair all clusters within the cluster set to form multiple upper and lower level combinations, each consisting of two clusters; and set the line power loss corresponding to each upper and lower level combination; if the total power of cluster A in the upper and lower level combination satisfies the total power of cluster B and the line power loss, then cluster A is determined to be the upper-level topology branch of cluster B; otherwise, pair them again.

8. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 7, characterized in that: In step S3, the clusters that have been identified as superiors are treated as new objects to be verified. The total power and corresponding line loss of the clusters are matched and verified with the higher-level clusters until all clusters are included in the hierarchical structure with the cluster where the main incoming line collection point is located as the root node. Finally, the membership relationship of each cluster is sorted out, thereby obtaining the topological relationship between all collection points.

9. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 8, characterized in that: If, after multiple pairings, none of the upper and lower level combinations within the cluster set satisfy the topological relationship analysis in step S3, then the next cluster set is selected again until the topological relationship is analyzed.

10. The method for comparing energy consumption by sub-item metering based on multi-source data fusion according to claim 9, characterized in that: In step S4, the sliding window length K in step S1 is adjusted to minimize the difference between the average value of abnormal power and normal power, and the sliding window length K is adjusted sequentially. During the adjustment process, the sliding window length K is not less than 3 and not greater than the length of the corresponding power sequence.

Images