A resident multi-period electricity consumption prediction method based on identification resolution

Abstract

The application discloses a kind of based on identification analysis resident multi-period electric quantity prediction method, including the following steps: obtaining the user profile of resident user, metering collection, house type registration, electrical behavior, area attribution and time cycle data and standardization processing, obtain standardization electricity input data;According to identification analysis field, the association link between resident user, metering device, power supply area and electricity account is constructed, and identification analysis association data is generated;Resident user portrait is constructed based on association data, and historical electric quantity sequence is carried out day, week, month cycle division and embedding, and multi-period electricity feature is formed;Multi-period electricity feature is input into improved StemGNN identification analysis multi-period electric quantity prediction model, and day, week, month electric quantity prediction result is output;According to actual meter reading feedback data, it is corrected.The application is based on identification analysis and spectral time sequence diagram modeling, realizes resident electric quantity multi-period prediction.

Description

Technical Field

[0001] This invention relates to the field of power data analysis and identifier resolution technology, specifically to a method for predicting residential multi-cycle electricity consumption based on identifier resolution. Background Technology

[0002] As residential electricity consumption scenarios become increasingly diverse, the number of air conditioners, water heaters, kitchen appliances, smart home devices, and charging equipment in households continues to rise. Changes in residential electricity consumption are influenced by various factors, including apartment layout, number of residents, living habits, holiday schedules, seasonal variations, and the power supply environment of the substation area. When conducting residential electricity consumption analysis, load management, substation operation control, and providing refined power supply services, power supply companies need to predict the electricity consumption of residential users for different periods—daily, weekly, and monthly.

[0003] Existing residential electricity consumption forecasting methods typically use historical electricity consumption sequences as the primary input, employing statistical regression, time series models, or ordinary neural network models for prediction. While these methods can achieve certain predictive results when electricity consumption patterns are stable and data attribution is clear, in actual residential electricity management, user files, metering devices, transformer substation affiliations, and electricity accounts are often scattered across different business systems. This can easily lead to issues such as changes in meter-to-user relationships, unsynchronized meter replacement records, transformer substation adjustments, account changes, and inconsistencies in historical electricity consumption attribution. Directly using this data for prediction can easily result in training sample mismatch, user profile bias, and distorted prediction results.

[0004] Meanwhile, residential electricity consumption exhibits distinct multi-cycle characteristics, with correlations existing between daily fluctuations, weekday and weekend differences, and monthly billing cycle changes. Existing single-cycle forecasting methods struggle to simultaneously account for short-term fluctuations, weekly patterns, and monthly trends, and also find it difficult to continuously revise resident profiles and forecast weights based on actual meter readings.

[0005] Therefore, how to provide a method for predicting residential electricity consumption over multiple cycles based on identifier resolution is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for predicting residential multi-cycle electricity consumption based on identifier resolution. This invention fully utilizes identifier resolution, residential user profile construction, multi-cycle electricity consumption feature extraction, and improved StemGNN spectral time series graph modeling technology. It describes in detail the method of standardizing multi-source residential electricity consumption data, constructing identifier resolution links between residential users, metering devices, power supply areas, and electricity accounts, generating residential user profiles, extracting daily, weekly, and monthly cycle electricity consumption features, and combining actual meter reading feedback data to correct the multi-cycle electricity consumption prediction. This method has the advantages of accurate data attribution, clear identifier association, sufficient fusion of cycle features, and high stability of prediction results.

[0007] A residential multi-cycle electricity consumption prediction method based on identifier resolution according to an embodiment of the present invention includes the following steps:

[0008] S1. Obtain multi-source electricity consumption data of residential users within the predicted time range, and perform identification verification, time alignment, anomaly removal, missing data filling and unified dimension processing on the multi-source electricity consumption data to obtain standardized electricity input data.

[0009] S2. Based on the identifier resolution field in the standardized electricity input data, construct the association link between residential users, metering devices, power supply areas and electricity accounts to obtain residential identifier resolution association data;

[0010] S3. Construct a resident user profile based on standardized electricity input data and resident identifier resolution and correlation data to obtain resident user profile features;

[0011] S4. Based on the characteristics of residential user profiles and historical electricity consumption sequences, daily, weekly and monthly cycles are divided and cycle embedding is performed to obtain multi-cycle electricity consumption characteristics.

[0012] S5. Input the multi-cycle electricity consumption characteristics and the associated data of the residential identifier resolution into the improved StemGNN identifier resolution multi-cycle electricity consumption prediction model. Based on the associated data of the residential identifier resolution, construct the residential identifier resolution electricity consumption map with identifier weight constraints. Then, perform spectral encoding, frequency domain time modeling and cross-cycle fusion prediction on the residential identifier resolution electricity consumption map and the multi-cycle electricity consumption characteristics to obtain the multi-cycle electricity consumption prediction results.

[0013] S6. Based on the multi-cycle electricity prediction results and actual meter reading feedback data, perform deviation calculation, profile update and feedback correction to obtain the residential multi-cycle electricity prediction output data.

[0014] Optionally, step S1 includes the following steps:

[0015] S11. Obtain user profile records, metering collection records, household registration records, electricity consumption behavior records, transformer area affiliation records, and time period records. Extract user number, metering point number, transformer area number, account number, collection time, historical electricity consumption, household area, and electricity consumption behavior fields to obtain raw electricity consumption data.

[0016] S12. Perform field integrity verification, field format verification, and duplicate identification verification on the original electricity consumption data to obtain the electricity consumption data after identification verification.

[0017] S13. Based on the collection time, meter reading time, settlement cycle and predicted time range, the time base of the identified and verified electricity consumption data is unified and the cycle boundary is aligned to obtain time-aligned electricity consumption data.

[0018] S14. Based on the changes in electricity consumption in adjacent cycles, the distribution of electricity consumption of similar residents, and the collection status field, perform anomaly removal, adjacent cycle filling, and unified dimension processing on the time-aligned electricity consumption data to obtain standardized electricity consumption input data.

[0019] Optionally, step S2 includes the following steps:

[0020] S21. Extract user number, metering point number, transformer area number, account number, effective time and business change record from standardized electricity input data, and convert them into residential user nodes, metering device nodes, power supply transformer area nodes and electricity account nodes to obtain a set of identified nodes;

[0021] S22. Based on the user meter binding record, meter area attribution record and user account settlement record, establish binding edge, attribution edge and settlement edge to obtain the initial identifier resolution link;

[0022] S23. Based on the meter change record, transformer area change record, account change record, and binding effective time, perform link merging or link splitting on the initial identifier resolution link to obtain resident identifier resolution associated data;

[0023] S24. Generate an identifier resolution link number based on the resident identifier resolution association data, and write the identifier resolution link number into the corresponding historical electricity consumption sequence and profile basic field to obtain resident electricity consumption data with identifier link number.

[0024] Optionally, step S3 includes the following steps:

[0025] S31. Based on the resident identifier parsing and association data, read the unit area, number of rooms, number of occupants, metering point type, power supply area attribute and account settlement attribute, and perform static encoding on the above fields to obtain static profile features;

[0026] S32. Based on the historical electricity consumption sequence and electricity consumption behavior records in the standardized electricity input data, calculate the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, continuous electricity consumption duration and electricity fluctuation range to obtain dynamic profile features;

[0027] S33. Based on the static profile features, dynamic profile features, and identifier resolution link number, perform feature splicing, noise reduction, completion, and profile category encoding to obtain the resident user profile features;

[0028] S34. Write the resident user profile features into the corresponding resident user node, and write the profile category code into the resident identifier parsing association data to obtain resident identifier parsing association data with profile attributes.

[0029] Optionally, step S4 includes the following steps:

[0030] S41. Based on the collection date, weekday attribute, month attribute, holiday attribute, settlement cycle attribute, and identifier resolution link number in the historical electricity sequence, periodic labels are marked to obtain the periodic labeled electricity sequence;

[0031] S42. Based on the continuous daily electricity consumption, weekly working day electricity consumption, weekly rest day electricity consumption and monthly settlement interval electricity consumption in the periodic marked electricity consumption sequence, segments are divided to obtain daily periodic electricity consumption segments, weekly periodic electricity consumption segments and monthly periodic electricity consumption segments.

[0032] S43. Based on the identifier resolution link number, the resident user profile features are matched with the daily cycle electricity consumption segment, the weekly cycle electricity consumption segment and the monthly cycle electricity consumption segment to obtain three types of cycle profile electricity consumption features.

[0033] S44. Perform periodic location encoding, periodic category encoding, and numerical mapping on the electricity consumption characteristics of the three types of periodic profiles to obtain multi-period electricity consumption characteristics.

[0034] Optionally, step S5 includes the following steps:

[0035] S51. Input the residential identifier resolution association data into the improved StemGNN identifier resolution multi-cycle electricity prediction model, construct the residential identifier resolution electricity consumption map based on the identifier resolution links between residential users, metering devices, power supply areas and electricity accounts, and obtain graph structure data including node codes, edge relationship codes and identifier weight adjacency matrices.

[0036] S52. Write the multi-cycle electricity consumption features into the residential user nodes in the residential identifier resolution electricity consumption map according to the identifier resolution link number, and map and fuse the residential user profile features with the daily cycle embedding, weekly cycle embedding and monthly cycle embedding to obtain the graph node input vector.

[0037] S53. In the improved StemGNN identifier parsing multi-cycle power prediction model, based on graph structure data and graph node input vectors, a graph Laplace with identifier weight constraints is constructed, a graph Fourier transform is performed, and a spectral domain graph association encoding is performed to obtain the identifier-constrained spectral domain graph association features.

[0038] S54. Perform frequency domain time decomposition and periodic time convolution on the association features of the constrained spectral domain graph to obtain daily periodic time features, weekly periodic time features and monthly periodic time features.

[0039] S55. Based on the daily cycle time series characteristics, weekly cycle time series characteristics, monthly cycle time series characteristics and resident user profile characteristics, cross-cycle fusion and residual correction are performed to obtain cross-cycle fusion prediction characteristics.

[0040] S56. Generate daily, weekly, and monthly predicted electricity consumption based on the cross-cycle fusion prediction characteristics, and bind the predicted electricity consumption with the identifier resolution link number to obtain multi-cycle electricity consumption prediction results.

[0041] Optionally, step S51 includes the following steps:

[0042] S511. In the improved StemGNN identifier resolution multi-cycle electricity prediction model, user number, metering point number, transformer area number, account number, profile category code and business activation time are extracted from the resident identifier resolution associated data, and the extracted content is converted into graph node coded data.

[0043] S512. Based on the user meter binding relationship, meter substation affiliation relationship, user account settlement relationship and the relationship between residents in the same substation, encode the connection category, connection direction and effective time between different identifier nodes to obtain graph edge coding data.

[0044] S513. Establish an initial adjacency matrix based on graph node coding data and graph edge coding data, and perform time segmentation processing on edge connection status according to table change records, transformer area change records and account change records to obtain a time segmentation adjacency matrix.

[0045] S514. Based on the identifier conflict record, binding effective status, station area ownership status and account settlement status, correct the abnormal edge weights in the time segment adjacency matrix to obtain the identifier weight adjacency matrix.

[0046] S515. Combine the graph node encoding data, graph edge encoding data, and label weight adjacency matrix into a residential label resolution electricity consumption graph to obtain the graph structure data for inputting the improved StemGNN label resolution multi-cycle electricity prediction model.

[0047] Optionally, step S53 includes the following steps:

[0048] S531. In the improved StemGNN identifier parsing multi-cycle power prediction model, the identifier weight adjacency matrix is ​​used as the input of the connection structure. Based on the node category, edge relationship category, binding effective status, transformer area affiliation status, account settlement status and resident user profile characteristics, the spectral domain propagation constraint coefficient of each identifier connection relationship is calculated to generate the identifier weight matrix.

[0049] S532. Merge the identifier weight matrix and the identifier weight adjacency matrix. Identify invalid links, conflicting links and expired links based on the binding effective time, binding ineffective time, station area ownership status, account settlement status and identifier conflict records. Suppress the edge weights corresponding to invalid links, conflicting links and expired links, construct a graph Laplacian matrix with identifier resolution constraints, and obtain graph structure data.

[0050] S533. Perform graph Fourier transform on the graph node input vector based on the graph structure data to map the multi-cycle electricity consumption characteristics to the graph domain and obtain the multi-cycle characteristics of the graph domain.

[0051] S534. Perform spectral convolution, spectral filtering, and node association aggregation on the multi-period features of the spectral domain, and adjust the spectral domain association strength according to the label weight matrix to obtain the label-constrained spectral domain graph association features.

[0052] Optionally, S55 includes the following steps:

[0053] S551. In the improved StemGNN identifier resolution multi-cycle power prediction model, the daily cycle time series features, weekly cycle time series features, and monthly cycle time series features are aligned according to the residential user node and the identifier resolution link number to obtain cross-cycle aligned features.

[0054] S552. Calculate the daily-week correlation weight and weekly-month correlation weight based on the cross-cycle alignment features, and perform weighted fusion of the three types of cycle time series features based on the daily-week correlation weight and weekly-month correlation weight to obtain cycle fusion features;

[0055] S553. The periodic fusion features are spliced ​​with the residential user profile features, and profile residual correction features are generated based on the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio and electricity fluctuation range.

[0056] S554. Based on the periodic fusion features and the image residual correction features, a mapping process is performed to obtain the cross-period fusion prediction features.

[0057] Optionally, step S6 includes the following steps:

[0058] S61. Obtain actual meter reading feedback data, and perform identifier matching and time matching on the user number, metering point number, transformer area number, account number and meter reading time in the actual meter reading feedback data to obtain actual feedback electricity data.

[0059] S62. Align the actual feedback power data with the multi-cycle power prediction results and calculate the difference to obtain the daily cycle prediction deviation, weekly cycle prediction deviation and monthly cycle prediction deviation.

[0060] S63. Update the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, continuous electricity consumption duration, and electricity fluctuation range based on the three types of prediction deviations to obtain feedback and update profile features;

[0061] S64. Input the updated profile features and actual feedback power data into the improved StemGNN identifier parsing multi-cycle power prediction model, and update the identifier weight matrix, cycle fusion weight and residual correction coefficient to obtain the updated multi-cycle power prediction results.

[0062] S65. Generate residential multi-cycle electricity prediction output data based on the updated multi-cycle electricity prediction results, identifier resolution link number, residential user profile category and prediction deviation data.

[0063] The beneficial effects of this invention are: This invention constructs a link between residential users, metering devices, power supply areas, and electricity accounts through identifier resolution, enabling data that was originally scattered in user files, metering data collection, area affiliation, and account settlement systems to be linked according to a unified identifier relationship. This reduces data mismatch problems caused by meter replacement, area adjustment, account changes, and inconsistencies in meter-user relationships, and provides an accurate data foundation for residential electricity consumption forecasting.

[0064] This invention constructs a resident user profile based on standardized electricity input data and resident identification resolution data. It transforms factors such as apartment size, number of residents, metering point type, power supply area attribute, peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, and historical electricity fluctuation range into profile features that can participate in model calculations. This enables the prediction model to distinguish the differences in electricity consumption habits and the stability of electricity consumption among different resident users.

[0065] This invention divides residential user profile features and historical electricity consumption sequences into daily, weekly, and monthly cycles and performs cycle embedding processing, so that continuous daily electricity consumption changes, differences between weekdays and rest days, and monthly settlement trends can be simultaneously included in the prediction process, avoiding the problem that single-cycle prediction methods cannot take into account both short-term fluctuations and long-term trends.

[0066] This invention employs an improved StemGNN identifier resolution multi-cycle electricity prediction model, which converts residential identifier resolution associated data into residential identifier resolution electricity consumption maps with identifier weight constraints. In the spectral domain, spectral encoding, frequency domain time modeling, and cross-cycle fusion prediction are performed, enabling the correlation between residential users, between users and metering devices, and between users and transformer substations to participate in the prediction calculation, thereby enhancing the consistency between the multi-cycle electricity prediction results and the actual business identifier relationships.

[0067] The present invention also calculates the prediction deviation based on actual meter reading feedback data, and updates the prediction results by feeding back the characteristics of residential users, the periodic fusion weight and the residual correction coefficient, so that the prediction results can be adjusted according to the changes in residents' recent electricity consumption behavior, thereby improving the stability and adaptability of subsequent daily, weekly and monthly multi-cycle electricity prediction. Attached Figure Description

[0068] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0069] Figure 1 This is an overall flowchart of a residential multi-cycle electricity consumption prediction method based on identifier resolution proposed in this invention;

[0070] Figure 2 This is a schematic diagram illustrating the process of constructing resident identifier resolution and association data, generating resident user profiles, and generating multi-cycle electricity consumption characteristics in this invention.

[0071] Figure 3 This diagram illustrates the process of constructing a residential identifier resolution electricity consumption map, encoding the identifier constraint spectrum, performing cross-cycle fusion prediction, and feedback correction for the improved StemGNN identifier resolution multi-cycle electricity prediction model in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0073] refer to Figures 1-3 A method for predicting residential multi-cycle electricity consumption based on identifier resolution includes the following steps:

[0074] S1. Obtain multi-source electricity consumption data of residential users within the predicted time range, and perform identification verification, time alignment, anomaly removal, missing data filling and unified dimension processing on the multi-source electricity consumption data to obtain standardized electricity input data.

[0075] S2. Based on the identifier resolution field in the standardized electricity input data, construct the association link between residential users, metering devices, power supply areas and electricity accounts to obtain residential identifier resolution association data;

[0076] S3. Construct a resident user profile based on standardized electricity input data and resident identifier resolution and correlation data to obtain resident user profile features;

[0077] S4. Based on the characteristics of residential user profiles and historical electricity consumption sequences, daily, weekly and monthly cycles are divided and cycle embedding is performed to obtain multi-cycle electricity consumption characteristics.

[0078] S5. Input the multi-cycle electricity consumption characteristics and the associated data of the residential identifier resolution into the improved StemGNN identifier resolution multi-cycle electricity consumption prediction model. Based on the associated data of the residential identifier resolution, construct the residential identifier resolution electricity consumption map with identifier weight constraints. Then, perform spectral encoding, frequency domain time modeling and cross-cycle fusion prediction on the residential identifier resolution electricity consumption map and the multi-cycle electricity consumption characteristics to obtain the multi-cycle electricity consumption prediction results.

[0079] S6. Based on the multi-cycle electricity prediction results and actual meter reading feedback data, perform deviation calculation, profile update and feedback correction to obtain the residential multi-cycle electricity prediction output data.

[0080] refer to Figure 1 In this embodiment, S1 is the step of obtaining multi-source electricity consumption data of residential users within the predicted time range and processing the multi-source electricity consumption data into standardized electricity input data.

[0081] Obtain user profile records, metering collection records, household registration records, electricity consumption behavior records, transformer area affiliation records, and time period records. Extract user ID, metering point ID, transformer area ID, account ID, collection time, historical electricity consumption, household area, and electricity consumption behavior fields from the above records to obtain raw electricity consumption data.

[0082] When a residential user undergoes meter replacement, account adjustment, or transformer area relocation within the predicted time frame, the same residential user may correspond to multiple metering point numbers, multiple account activation phases, or multiple transformer area affiliation phases. Therefore, the original electricity consumption data retains the service activation time and service change records to facilitate the subsequent identification and resolution link construction.

[0083] The original electricity consumption data is subjected to field integrity verification, field format verification, and duplicate identification verification to obtain the electricity consumption data after identification verification.

[0084] Field integrity verification includes checking whether user ID, metering point ID, transformer area ID, account ID, collection time, and historical electricity consumption are missing; field format verification includes checking whether the ID length, ID character type, collection time format, and electricity value format meet the preset specifications; duplicate identifier verification includes checking whether there are duplicate electricity consumption records at the same collection time, whether the same metering point ID corresponds to multiple user IDs, and whether the same account ID is connected to multiple residential users within the same effective time period.

[0085] For missing fields that can be filled in by business change records, they are filled in by adjacent valid records in the same identifier resolution link; for fields that cannot be filled in, they are marked as records to be verified; for historical electricity records that are collected repeatedly, records with normal collection status, complete meter reading time, and clear business effective relationship are retained first.

[0086] Based on the collection time, meter reading time, settlement cycle, and predicted time range, the electricity consumption data after identification and verification is unified in time reference and aligned in cycle boundary to obtain time-aligned electricity consumption data.

[0087] For daily forecasting scenarios, historical electricity consumption is arranged according to consecutive collection dates; for weekly forecasting scenarios, it is aggregated according to natural weeks or business-defined weekly cycles; for monthly forecasting scenarios, it is summarized according to the settlement cycle and monthly settlement interval. During time alignment, the meter replacement effective time, the transformer area change effective time, and the account change effective time are used as cycle boundaries. Electricity records before and after the boundaries are written into the corresponding identifier resolution stage to avoid incorrect merging of electricity consumption before and after the meter replacement.

[0088] Based on the changes in electricity consumption in adjacent cycles, the distribution of electricity consumption by similar residents, and the data collection status field, the time-aligned electricity consumption data is processed by removing anomalies, filling in adjacent cycles, and unifying the units to obtain standardized electricity consumption input data.

[0089] When removing anomalies, electricity records with significantly larger fluctuations than the range of similar residents in adjacent cycles and with abnormal collection status are marked as abnormal records, and are corrected by combining historical electricity consumption in the same period, electricity consumption in adjacent cycles, and electricity consumption behavior status.

[0090] When filling missing data, for short-term missing data, the electricity consumption of adjacent cycles or similar cycles of the same residential user can be used to fill the missing data; for missing data for a longer period of time, the electricity consumption distribution of residential users of the same profile category, the same transformer area, or similar house type can be used to fill the missing data.

[0091] When processing data in a unified dimension, historical electricity consumption, apartment size, number of occupants, peak-valley electricity consumption ratio, and nighttime electricity consumption ratio are mapped to a unified numerical range, and weekday attributes, month attributes, holiday attributes, and settlement cycle attributes are converted into cycle category codes.

[0092] refer to Figure 2 In this embodiment, S2 is the step of constructing the association link between residential users, metering devices, power supply areas and electricity accounts based on the identifier resolution field in the standardized electricity input data, and obtaining the residential identifier resolution association data.

[0093] Extract user ID, metering point ID, transformer area ID, account ID, effective time, and business change records from standardized electricity input data, and convert them into residential user nodes, metering device nodes, power supply transformer area nodes, and electricity account nodes to obtain a set of identified nodes.

[0094] Based on the user meter binding record, meter area affiliation record, and user account settlement record, establish binding edge, affiliation edge, and settlement edge to obtain the initial identifier resolution link.

[0095] The binding edge connects the residential user node and the metering device node, recording the metering relationship between the residential user and the metering device; the attribution edge connects the metering device node and the power supply area node, recording the power supply area where the metering device is located; the settlement edge connects the residential user node and the electricity account node, recording the settlement relationship between the residential user and the account.

[0096] Based on the meter change record, transformer area change record, account change record, and binding effective time, the initial identifier resolution link is merged or split to obtain resident identifier resolution associated data.

[0097] When the same residential user replaces their meter within the predicted time frame, the metering device node before and after the replacement is merged through the same residential user node and the continuous account settlement edge to form a continuous identifier resolution link; when the metering device is moved from one power supply area to another, the attribution edge is split into two effective phases according to the effective time of the area change; when the residential user's account is changed, the settlement edge is continued or split according to the effective time of the account.

[0098] The identifier resolution link number is generated based on the resident identifier resolution association data, and then written into the corresponding historical electricity consumption sequence and profile basic field to obtain resident electricity consumption data with identifier link number.

[0099] The identifier resolution link number is used throughout the subsequent processes of profile generation, multi-period feature generation, model prediction, and feedback correction. Through the identifier resolution link number, the features of the residential user profile can be accurately written into the corresponding residential user node, the historical electricity consumption sequence can be accurately bound to the corresponding metering device and power supply area, and the multi-period prediction results can be accurately returned to the corresponding electricity account.

[0100] refer to Figure 2 In this embodiment, S3 is the step of constructing a resident user profile based on standardized electricity input data and resident identifier resolution association data, and obtaining the resident user profile features.

[0101] Based on the resident identifier resolution and related data, the unit area, number of rooms, number of occupants, metering point type, power supply area attribute, and account settlement attribute are read, and the above fields are statically encoded to obtain static profile features.

[0102] Based on the historical electricity consumption sequence and electricity consumption behavior records in the standardized electricity input data, the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, continuous electricity consumption duration, and electricity consumption fluctuation range are calculated to obtain dynamic profile features.

[0103] Based on static profile features, dynamic profile features, and identifier resolution link numbers, feature splicing, noise reduction, completion, and profile category encoding are performed to obtain resident user profile features.

[0104] During noise reduction, abnormal profile fields that are inconsistent with the historical behavior of residential users are smoothed or removed. During completion processing, short-term missing profile fields are filled using profile features from adjacent periods or the distribution of users with similar profiles. When encoding profile categories, residential users are classified into stable electricity consumption type, nighttime enhanced electricity consumption type, holiday enhanced electricity consumption type, and high-fluctuation type based on the stability of continuous electricity consumption, nighttime electricity consumption intensity, holiday variation range, and monthly fluctuation range. The profile category encoding serves as input for subsequent graph node attributes and residual correction, participating in the calculation of the improved StemGNN identifier parsing multi-period electricity prediction model.

[0105] The resident user profile features are written into the corresponding resident user nodes, and the profile category code is written into the resident identifier resolution and association data, resulting in resident identifier resolution and association data with profile attributes. The resident identifier resolution and association data with profile attributes not only records the business connection relationships between resident users, metering devices, power supply areas, and electricity accounts, but also records the profile features corresponding to the resident user nodes.

[0106] refer to Figure 2 In this embodiment, S4 is a step of dividing the electricity consumption into daily, weekly and monthly cycles and performing cycle embedding processing based on the characteristics of residential user profiles and historical electricity consumption sequences to obtain multi-cycle electricity consumption characteristics.

[0107] Periodically labeled electricity sequences are obtained by labeling the historical electricity consumption sequence with periodic tags based on the collection date, weekday attribute, month attribute, holiday attribute, settlement cycle attribute, and identifier resolution link number.

[0108] Based on the continuous daily electricity consumption, weekly working day electricity consumption, weekly rest day electricity consumption, and monthly settlement interval electricity consumption in the periodic marked electricity consumption sequence, segments are obtained to obtain daily periodic electricity consumption segments, weekly periodic electricity consumption segments, and monthly periodic electricity consumption segments.

[0109] The daily cycle electricity consumption segment focuses on continuous daily electricity consumption, retaining short-term fluctuations and changes between adjacent days; the weekly cycle electricity consumption segment focuses on weekdays and rest days within a week, retaining weekday load patterns and rest day electricity consumption changes; the monthly cycle electricity consumption segment focuses on the monthly settlement period and monthly cumulative electricity consumption, retaining monthly trends and settlement cycle changes. All three scales of electricity consumption segments are associated with residential user nodes through identifier resolution link numbers.

[0110] Based on the identifier resolution link number, the resident user profile features are matched with the daily cycle electricity consumption segments, weekly cycle electricity consumption segments, and monthly cycle electricity consumption segments to obtain three types of cycle profile electricity consumption features.

[0111] By performing periodic location encoding, periodic category encoding, and numerical mapping on the electricity consumption characteristics of the three types of periodic profiles, multi-period electricity consumption characteristics are obtained.

[0112] refer to Figure 3 In this embodiment, S5 is the step of inputting the multi-cycle electricity consumption characteristics and the associated data of the resident identification resolution into the improved StemGNN identification resolution multi-cycle electricity prediction model to obtain the multi-cycle electricity prediction results.

[0113] The improved StemGNN identifier parsing multi-cycle electricity prediction model includes a residential identifier parsing electricity map construction unit, a graph node input vector generation unit, an identifier constraint spectrum encoding unit, a frequency domain time modeling unit, a cross-cycle fusion prediction unit, and a multi-cycle electricity output unit.

[0114] The resident identifier resolution and electricity map construction unit receives resident identifier resolution and association data and outputs graph structure data;

[0115] The graph node input vector generation unit receives multi-cycle electricity consumption characteristics and residential user profile characteristics, and outputs graph node input vectors.

[0116] The identifier constraint spectral graph encoding unit receives graph structure data and graph node input vectors, and outputs identifier constraint spectral domain graph association features;

[0117] The frequency domain time modeling unit receives the identifier-constrained spectral domain graph correlation features and outputs daily, weekly, and monthly time series features;

[0118] The cross-cycle fusion prediction unit receives three types of periodic time-series features and resident user profile features, and outputs cross-cycle fusion prediction features.

[0119] The multi-cycle power output unit receives cross-cycle fusion prediction features and outputs daily, weekly, and monthly predicted power consumption.

[0120] The residential identifier resolution associated data is input into the improved StemGNN identifier resolution multi-cycle electricity prediction model. Based on the identifier resolution links between residential users, metering devices, power supply areas and electricity accounts, a residential identifier resolution electricity consumption map is constructed, resulting in graph structure data containing node codes, edge relationship codes and identifier weight adjacency matrices.

[0121] The residential user node in the electricity consumption map records the resident user profile and electricity consumption characteristics, the metering device node records the metering point status and data acquisition status, the power supply area node records the area affiliation and area attributes, and the electricity account node records the account settlement attributes and effective status.

[0122] The edge relationship coding records the binding relationship between residential user nodes and metering device nodes, the affiliation relationship between metering device nodes and power supply area nodes, the settlement relationship between residential user nodes and electricity account nodes, and the association relationship between residential users within the same power supply area.

[0123] The adjacency matrix with identifier weights is generated based on the edge relationship status, the service activation time, and identifier conflict records. The higher the connection credibility, the higher the corresponding edge weight. The edge weights of invalid links, conflicting links, and expired links are reduced or deleted.

[0124] In the improved StemGNN identifier resolution multi-cycle electricity prediction model, user ID, metering point ID, transformer area ID, account ID, profile category code, and service activation time are extracted from the resident identifier resolution associated data, and the extracted content is converted into graph node coded data.

[0125] Based on the user meter binding relationship, meter distribution area affiliation relationship, user account settlement relationship and the relationship between residents in the same distribution area, the connection type, connection direction and effective time between different identifier nodes are encoded to obtain graph edge coding data;

[0126] An initial adjacency matrix is ​​established based on the graph node encoding data and graph edge encoding data. The edge connection status is then segmented over time based on the table change record, transformer area change record, and account change record to obtain a time-segmented adjacency matrix.

[0127] Based on the identifier conflict record, binding effectiveness status, station area ownership status, and account settlement status, the abnormal edge weights in the time segment adjacency matrix are corrected to obtain the identifier weight adjacency matrix.

[0128] The graph node encoding data, graph edge encoding data, and label weight adjacency matrix are combined to form a residential label resolution electricity consumption map, which is then used to obtain the graph structure data for inputting the improved StemGNN label resolution multi-cycle electricity consumption prediction model.

[0129] Multi-cycle electricity consumption characteristics are written into the residential user nodes in the residential identifier resolution electricity consumption map according to the identifier resolution link number, and the residential user profile characteristics are mapped and fused with daily cycle embedding, weekly cycle embedding and monthly cycle embedding to obtain the graph node input vector.

[0130] In the improved StemGNN identifier parsing multi-cycle power prediction model, a graph Laplace with identifier weight constraints is constructed, a graph Fourier transform is performed, and a spectral domain graph association encoding is performed based on graph structure data and graph node input vectors to obtain identifier-constrained spectral domain graph association features.

[0131] The identifier constraint spectrum encoding unit first reads the identifier weight adjacency matrix, node category, edge relationship category and resident user profile features from the resident identifier resolution electricity diagram, and generates the identifier weight matrix.

[0132] The identifier weight matrix is ​​calculated based on the user's meter binding status, transformer area affiliation status, account activation status, and profile category consistency. The correlation strength between different residential users no longer depends entirely on the similarity of electricity consumption sequences, but is also constrained by the credibility of business identifiers.

[0133] Subsequently, the identifier weight matrix and the identifier weight adjacency matrix are merged to suppress the edge weights of invalid links and conflicting links, and a graph Laplacian matrix with identifier parsing constraints is constructed to obtain the graph structure data.

[0134] The graph node input vectors undergo graph Fourier transform under the constraints of graph structure data, and the multi-period electricity consumption characteristics are mapped to the graph domain, forming multi-period features in the graph domain.

[0135] Finally, the multi-period features in the spectral domain are processed by spectral convolution, spectral filtering, and node association aggregation, and the spectral domain association strength is adjusted according to the label weight matrix to obtain the label-constrained spectral domain graph association features.

[0136] In this embodiment, the identifier-constrained spectral encoding unit is one of the core structures of the improved StemGNN identifier parsing multi-cycle power prediction model.

[0137] The original StemGNN model typically learns graph structures based on the relationships between multivariate time series. This implementation introduces resident identifier resolution and association data into the graph structure generation process, so that the graph structure is no longer determined entirely by the similarity of historical electricity consumption, but is jointly determined by the business identifier links between resident users, metering devices, power supply areas, and electricity accounts.

[0138] Since meter replacement, transformer area adjustment, and account changes can cause changes in historical electricity attribution, this implementation method constrains the spectral encoding process by using an adjacency matrix and an identifier weight matrix. This reduces the weight of invalid and conflicting links in the spectral domain association calculation, while maintaining a higher weight for valid service links, thereby reducing the impact of incorrect identifier relationships on the residential electricity prediction results.

[0139] Frequency domain time decomposition and periodic time convolution are performed on the association features of the constrained spectral domain graph to obtain daily, weekly, and monthly periodic time features.

[0140] The frequency domain time modeling unit performs discrete frequency decomposition on the associated features of the constrained spectral domain graph, and splits the features into daily, weekly and monthly periodic frequency features according to the different periodic components in the changes in residents' electricity consumption.

[0141] The daily cycle frequency characteristics mainly correspond to the changes in electricity consumption over consecutive days and the fluctuations between adjacent days; the weekly cycle frequency characteristics mainly correspond to the differences in electricity consumption between working days and rest days; and the monthly cycle frequency characteristics mainly correspond to the changes in cumulative electricity consumption within a month and the monthly settlement trend.

[0142] Subsequently, the daily periodic frequency features are entered into a short-term temporal convolutional structure, the weekly periodic frequency features into a weekly intra-week temporal convolutional structure, and the monthly periodic frequency features into a settlement interval temporal convolutional structure, resulting in daily, weekly, and monthly periodic time-series features, respectively. These three types of periodic time-series features are aligned according to resident user nodes and time sequence to form the input for subsequent cross-period fusion.

[0143] Cross-cycle fusion and residual correction are performed based on daily, weekly, and monthly time-series characteristics and resident user profile characteristics to obtain cross-cycle fusion prediction features.

[0144] The cross-cycle fusion prediction unit first aligns the daily, weekly, and monthly time-series features according to the resident user nodes and identifier resolution link numbers to obtain cross-cycle aligned features;

[0145] Then, the daily-week correlation weight and weekly-month correlation weight are calculated based on the cross-cycle alignment features, and the three types of cycle time series features are weighted and fused based on the daily-week correlation weight and weekly-month correlation weight to obtain the cycle fusion features;

[0146] Subsequently, the periodic fusion features are spliced ​​with the residential user profile features, and profile residual correction features are generated based on the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, and electricity fluctuation range.

[0147] Finally, mapping is performed based on the periodic fusion features and the image residual correction features to obtain cross-period fusion prediction features.

[0148] In this embodiment, the cross-cycle fusion prediction unit compensates for cycle prediction biases based on residential user profile characteristics. For residential users with stable electricity consumption, the daily cycle characteristics and monthly cycle trends are usually relatively smooth, and the residual correction coefficient can be kept low.

[0149] For residential users with enhanced nighttime electricity consumption, the proportion of nighttime electricity consumption is relatively high, and the nighttime changes in the daily cycle time series characteristics have a significant impact on the daily predicted electricity consumption.

[0150] For residential users with enhanced holiday consumption, the proportion of electricity consumption during holidays is relatively high, and the changes in rest days in the weekly time series characteristics need to be given higher weight.

[0151] For residential users with high volatility, the fluctuation range of electricity consumption is large, and the residual correction features of the user profile need to compensate for sudden increases and decreases in electricity consumption.

[0152] Through the aforementioned cross-cycle fusion and residual correction, the model can simultaneously retain daily short-term changes, weekly behavioral changes, and monthly settlement trends, while reducing prediction bias caused by differences in individual residents' behaviors.

[0153] Based on the cross-cycle fusion prediction characteristics, daily, weekly, and monthly predicted electricity volumes are generated, and the predicted electricity volumes are bound to the identifier resolution link number to obtain multi-cycle electricity volume prediction results.

[0154] The multi-cycle electricity forecast results include the residential user identifier resolution link number, residential user profile category, forecast time range, daily forecast electricity, weekly forecast electricity, monthly forecast electricity, and forecast generation status.

[0155] In this embodiment, the improved StemGNN identifier parsing multi-cycle power prediction model adopts a hierarchical transmission relationship in data transmission, from business identifier relationship to graph structure, from multi-cycle power sequence to node input, from graph domain association to frequency domain time series, and from cross-cycle fusion to multi-cycle output.

[0156] The resident identification resolution and associated data is first converted into a resident identification resolution electricity consumption map, which forms the structural input of the model;

[0157] Multi-cycle electricity consumption characteristics and residential user profile characteristics are written into the residential user node to form the numerical input of the model;

[0158] The adjacency matrix and the weight matrix together constrain the graph domain encoding, making the spectral domain association calculation affected by the credibility of the business identifier.

[0159] The frequency domain time modeling unit decomposes the spectral domain graph correlation features into different periodic components; the cross-period fusion prediction unit fuses different periodic features and corrects residuals based on resident user profile features;

[0160] The multi-cycle power output unit maps the fused prediction features into power output results for daily, weekly, and monthly cycles. The above data transfer relationships and references... Figure 3 The arrows in the diagram correspond to each other.

[0161] In this embodiment, the improved StemGNN identifier parsing multi-cycle power prediction model has several improvements over the original StemGNN model.

[0162] First, the original StemGNN model is usually geared towards general multivariate time series, and the graph structure between variables is mainly learned through data correlation. This implementation explicitly transforms the identifier resolution link between residential users, metering devices, power supply areas and electricity accounts into a residential identifier resolution electricity consumption graph, so that the graph structure has clear business meaning.

[0163] Second, the graph association relationships in the original StemGNN model usually do not distinguish between table changes, station area changes, account changes, and identifier conflicts. This implementation assigns different weights to different service links through the identifier weight adjacency matrix and the identifier weight matrix, so that valid links participate in the spectral domain association calculation, while invalid links and conflicting links are suppressed in the spectral domain association calculation.

[0164] Third, the original StemGNN model is usually geared towards a uniform time scale sequence. This implementation embeds the daily, weekly, and monthly electricity consumption features into the residential user nodes respectively, and extracts features at different time scales through frequency domain time decomposition and periodic time convolution, so that the model can simultaneously obtain short-term fluctuations, intra-week patterns, and monthly trends.

[0165] Fourth, this implementation introduces residual correction of resident user profiles in the cross-cycle fusion stage, so that the characteristics of the household type, electricity consumption behavior and electricity fluctuation status can correct the prediction deviation and improve the prediction adaptability in scenarios with obvious individual differences among residents.

[0166] In this embodiment, the training data for the improved StemGNN identifier parsing multi-cycle electricity prediction model comes from historical residential electricity consumption data, user profile data, transformer area attribution data, account settlement data, electricity consumption behavior statistics data, and actual meter reading feedback data.

[0167] The training samples are residential users. Each sample contains standardized electricity input data within a historical time range, resident identifier resolution and association data, resident user profile features and multi-cycle electricity consumption features. The training labels are the subsequent actual daily electricity consumption, actual weekly electricity consumption and actual monthly electricity consumption.

[0168] The training samples are organized according to resident users and identifier resolution link numbers. The same resident user forms a continuous sample sequence at different business change stages. Change of form, change of substation area and change of account are recorded as edge state changes in the graph structure.

[0169] The training set, validation set, and test set are divided according to the residential user dimension or the time dimension to ensure that the actual electricity consumption in subsequent cycles does not enter the preceding prediction input in advance.

[0170] In this embodiment, during model training, the node input vector consists of resident user profile features, daily periodic embedding, weekly periodic embedding, monthly periodic embedding, and node category encoding.

[0171] The node input vector takes residential user nodes as the main calculation object, while metering device nodes, power supply area nodes, and electricity account nodes participate in the graph domain association calculation as identification relationship nodes.

[0172] After the node input vector enters the identifier-constrained spectral encoding unit, the graph Fourier transform maps the node features to the graph spectral domain. The spectral convolution structure propagates graph relationships based on the graph Laplacian matrix with identifier parse constraints. The spectral domain filtering structure suppresses invalid links and noise associations. The node association aggregation structure aggregates valid association information from the same identifier link or the same transformer area to the resident user node.

[0173] Subsequently, the frequency domain temporal modeling unit decomposes the spectral domain association results into periodic components, and the periodic temporal convolutional structure processes the daily, weekly, and monthly periodic features respectively. The cross-period fusion prediction unit generates the final prediction features based on the daily-weekly association weights, weekly-monthly association weights, and profile residual correction features.

[0174] In this embodiment, the model training loss may include daily cycle prediction error, weekly cycle prediction error, monthly cycle prediction error, label weight constraint error, and residual correction error.

[0175] Daily cycle forecast error constraint limits the deviation between daily forecasted electricity consumption and actual daily electricity consumption; weekly cycle forecast error constraint limits the deviation between weekly forecasted electricity consumption and actual weekly electricity consumption.

[0176] Monthly cycle forecast error constraint on the deviation between monthly forecasted electricity consumption and actual monthly electricity consumption;

[0177] The model identifies the spectral domain correlation strength of the weighted constraint error constraint model on invalid, conflicting, and expired links; the residual correction error constraint profiles the difference between the residual correction results and the actual meter reading deviation.

[0178] Daily, weekly, and monthly prediction errors can be calculated using absolute error or squared error. Identification weight constraint error can be calculated based on the edge state and edge weight difference. Residual correction error can be calculated based on the change in prediction deviation before and after feedback.

[0179] During training, the number of training epochs can be set to 80 to 150, the batch sample size to 64 to 256, and the initial learning rate to 0.001 to 0.0001. The learning rate is decayed based on the decrease in validation set error. When the combined daily, weekly, and monthly prediction error on the validation set decreases below a preset threshold within several consecutive epochs, or when the number of training epochs reaches a preset upper limit, training is stopped and the model parameters are saved.

[0180] In this implementation, sample balancing can be applied to different types of residential users during model training. Users with stable electricity consumption are typically more numerous, while users with high fluctuations, those with increased electricity consumption during holidays, and those with increased electricity consumption at night may be relatively fewer.

[0181] To avoid bias in the model towards users with stable electricity consumption, the sampling ratio of users with high volatility and those experiencing enhanced consumption during holidays can be increased during the batch sample construction stage. Alternatively, the weight of high volatility samples can be increased in the residual error correction.

[0182] For samples involving meter replacement, transformer area adjustment, and account changes, the training weights of edge state change samples can be increased in the identifier weight constraint error, allowing the model to more fully learn the impact of identifier relationship changes on the prediction results. Through the above training method, the improved StemGNN identifier parsing multi-cycle electricity prediction model can simultaneously adapt to both ordinary electricity consumption scenarios for residential users and scenarios involving changes in business relationships.

[0183] In this embodiment, the model convergence criteria may include four types of indicators: daily prediction error of the validation set, weekly prediction error of the validation set, monthly prediction error of the validation set, and prediction consistency of the identifier link.

[0184] Verifying daily prediction errors reflects short-term prediction capability; verifying weekly prediction errors reflects intra-week pattern learning capability; verifying monthly prediction errors reflects settlement cycle trend learning capability; and identifying link prediction consistency reflects whether the prediction results are correctly bound to the identifier resolution link.

[0185] When the above indicators meet the preset conditions, it indicates that the model can reach a stable state between prediction accuracy and business identifier consistency. When saving the model, the spectral encoding parameters, frequency domain time modeling parameters, cross-cycle fusion weights, residual correction coefficients, and identifier weight generation rules are also saved to facilitate local updates in the subsequent feedback correction stage.

[0186] refer to Figure 3 In this embodiment, S6 is the step of performing deviation calculation, profile update and feedback correction based on the multi-cycle electricity prediction results and actual meter reading feedback data to obtain the residential multi-cycle electricity prediction output data.

[0187] Obtain actual meter reading feedback data, and perform identifier matching and time matching on the user number, metering point number, transformer area number, account number and meter reading time in the actual meter reading feedback data to obtain the actual feedback electricity data.

[0188] Before entering the feedback correction process, the actual meter reading feedback data is first matched with the residential user node, metering device node, power supply area node, and electricity account node by identifying and resolving the link number. Then, it is matched according to the predicted time range to ensure that the actual feedback electricity and the corresponding prediction result belong to the same residential user, the same metering relationship, and the same prediction cycle.

[0189] The actual feedback power data is aligned with the multi-cycle power prediction results and the difference is calculated to obtain the daily cycle prediction deviation, weekly cycle prediction deviation and monthly cycle prediction deviation.

[0190] The daily cycle forecast deviation is obtained from the difference between the actual daily electricity consumption and the daily forecast electricity consumption; the weekly cycle forecast deviation is obtained from the difference between the actual weekly electricity consumption and the weekly forecast electricity consumption; and the monthly cycle forecast deviation is obtained from the difference between the actual monthly electricity consumption and the monthly forecast electricity consumption.

[0191] During periodic alignment, if the actual meter reading feedback data spans the time points of meter replacement, transformer area change, or account change, the feedback data will be split into the corresponding business stages according to the identifier resolution link number and the business effective time to avoid the prediction deviations of different link stages being mixed and calculated.

[0192] Based on the three types of prediction bias, the peak-valley electricity consumption ratio, the nighttime electricity consumption ratio, the holiday electricity consumption ratio, the continuous electricity consumption duration, and the electricity fluctuation range are updated to obtain feedback and updated profile features.

[0193] When the daily cycle prediction deviation remains high, priority is given to updating the continuous electricity consumption duration and electricity fluctuation range; when the weekly cycle prediction deviation is high near rest days or holidays, priority is given to updating the proportion of electricity consumption during holidays and the characteristics of electricity consumption changes within the week; when the monthly cycle prediction deviation is high, priority is given to updating the cumulative electricity consumption changes within the month and the characteristics related to account settlement attributes. Feedback updates the profile features and write them into the residential user node, which then serve as part of the graph node input vector for the next round of prediction.

[0194] The improved StemGNN identifier parsing multi-cycle power prediction model is input with the updated profile features and actual feedback power data. The identifier weight matrix, cycle fusion weights and residual correction coefficients are updated with feedback to obtain the updated multi-cycle power prediction results.

[0195] When updating the identification weight matrix, if the actual feedback data confirms that a certain table replacement link, station area affiliation link, or account settlement link is valid, the weight of the corresponding edge is increased; if the actual feedback data confirms that a certain link has a conflict or has expired, the weight of the corresponding edge is decreased.

[0196] When updating the periodic fusion weight feedback, the daily-week correlation weight and weekly-month correlation weight are adjusted according to the magnitude of the prediction deviation for the daily, weekly, and monthly periods. When updating the residual correction coefficient feedback, the intensity of the residual correction is adjusted according to the relationship between the characteristics of the residential user profile and the prediction deviation, so that subsequent predictions can better adapt to recent changes in residents' electricity consumption.

[0197] The residential multi-cycle electricity prediction output data is generated based on the updated multi-cycle electricity prediction results, the identifier resolution link number, the residential user profile category, and the prediction deviation data.

[0198] The output data for multi-period electricity consumption forecasting for residents includes the resident user identifier resolution link number, resident user profile category, forecast time range, daily forecast electricity consumption, weekly forecast electricity consumption, monthly forecast electricity consumption, daily forecast deviation, weekly forecast deviation, monthly forecast deviation, and feedback correction status.

[0199] The output data can be grouped and saved according to residential users, power supply areas, profile categories, prediction cycles, and prediction deviation levels.

[0200] Power supply companies can conduct short-term electricity consumption analysis based on daily forecasts, manage weekly load based on weekly forecasts, analyze settlement cycle trends based on monthly forecasts, and screen for abnormal electricity consumption based on forecast deviations and profile categories.

[0201] Example 1: To verify the feasibility of this invention in practice, it was applied to the residential electricity consumption analysis business of a power supply company. In this business, the user files, metering records, household registration records, electricity consumption behavior records, transformer substation affiliation records, and account settlement records of residential users come from different systems, resulting in problems such as meter replacement, transformer substation adjustment, account changes, delayed synchronization of meter-user relationships, and inconsistencies in historical electricity consumption affiliation. This invention focuses on the technical direction of collecting and processing user identification resolution information, household characteristics, electricity consumption behavior, historical electricity consumption, and time period data, and establishing a residential user profile and a periodic electricity consumption correlation model.

[0202] This embodiment selects 10,000 residential users in a certain residential area as the sample. Each household has a corresponding user ID, metering point ID, transformer area ID, account ID, apartment size, number of residents, daily electricity consumption, peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, and settlement cycle attributes. To verify the model training effect and prediction stability, the 10,000 residential user samples are divided according to training, validation, and testing purposes. Among them, 7,000 residential users are used as training samples, 1,500 residential users as validation samples, and 1,500 residential users as testing samples. The number of test residential users in Table 1 corresponds to the above 1,500 test samples. The traditional historical sequence prediction method, the initial prediction results of this invention, and the results after feedback correction of this invention are all compared based on the same test sample. In the collected raw data, approximately 3.1% of the records have inconsistent metering point IDs before and after meter replacement, approximately 2.0% of the records have transformer area adjustments, approximately 1.3% of the records have account number changes, and approximately 4.5% of the historical electricity records have short-term missing or duplicate collections.

[0203] The system first performs field integrity checks, duplicate identification checks, and time boundary alignment on the above data, unifying the collection time, meter reading time, and settlement cycle within the predicted time range. For data that is missing for no more than two consecutive collection cycles, the electricity consumption of the adjacent cycle is used to fill the gaps. For data that suddenly increases by more than 2.5 times the average fluctuation range of similar residents, abnormal data is removed or corrected in conjunction with the collection status field to obtain standardized electricity input data.

[0204] Subsequently, the system constructs an identifier resolution link based on the user ID, metering point ID, transformer area ID, and account ID from the standardized electricity input data. Residential users are converted into residential user nodes, metering points into metering device nodes, transformer areas into power supply transformer area nodes, and accounts into electricity account nodes. The system establishes binding edges, attribution edges, and settlement edges based on user meter binding records, meter transformer area affiliation records, and user account settlement records, and merges or splits the links based on meter replacement records, transformer area change records, and account change records. After processing, each residential user generates a corresponding identifier resolution link number, and historical electricity consumption sequences and profile basic fields are bound according to the identifier resolution link number.

[0205] During the residential user profile construction phase, the system reads unit area, number of rooms, number of occupants, metering point type, power supply area attribute, and account settlement attribute from the resident identifier parsing and association data to form static profile features. It calculates peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, continuous electricity consumption duration, and electricity fluctuation range from historical electricity consumption sequences and electricity consumption behavior records to form dynamic profile features. The system then stitches, denoises, and encodes the static and dynamic profile features, classifying residential users into four categories: stable electricity consumption type, enhanced nighttime consumption type, enhanced holiday consumption type, and high fluctuation type. Stable electricity consumption type accounts for approximately 46%, enhanced nighttime consumption type approximately 17%, enhanced holiday consumption type approximately 22%, and high fluctuation type approximately 15%.

[0206] In the multi-cycle electricity consumption feature generation stage, the system labels historical electricity consumption sequences with periodic tags based on the collection date, weekday attribute, month attribute, holiday attribute, and settlement cycle attribute, and divides the periodically labeled electricity consumption sequences into daily, weekly, and monthly periodic electricity consumption segments. The daily periodic electricity consumption segment reflects continuous daily electricity consumption changes, the weekly periodic electricity consumption segment reflects the difference in electricity consumption between weekdays and rest days, and the monthly periodic electricity consumption segment reflects the cumulative electricity consumption changes within a month and the settlement cycle trend. Based on the identifier resolution link number, the system matches and concatenates the residential user profile features with the three types of periodic electricity consumption segments to obtain multi-cycle electricity consumption features.

[0207] In the prediction phase, the system inputs multi-cycle electricity consumption characteristics and residential identifier resolution-related data into an improved StemGNN identifier resolution multi-cycle electricity consumption prediction model. The model first constructs a residential identifier resolution electricity consumption graph based on the identifier resolution links, and generates an identifier weight adjacency matrix based on identifier conflict records, binding effectiveness status, transformer area affiliation status, and account settlement status, suppressing the edge weights corresponding to invalid, conflicting, and expired links. Subsequently, the model writes residential user profile features, daily, weekly, and monthly periodic embeddings into the corresponding residential user nodes, forming graph node input vectors. It then performs graph Fourier transform, spectral domain graph association coding, frequency domain time modeling, and cross-cycle fusion prediction on the graph structure with identifier weight constraints, outputting daily, weekly, and monthly predicted electricity consumption.

[0208] Test results show that the average absolute error of the traditional historical sequence prediction method is 1.38 kWh for daily prediction, 8.21 kWh for weekly prediction, and 32.6 kWh for monthly prediction. After adopting this invention, the average absolute error of daily prediction is reduced to 0.82 kWh, the average absolute error of weekly prediction is reduced to 5.13 kWh, and the average absolute error of monthly prediction is reduced to 20.7 kWh. For residential users who have undergone meter replacement, transformer area adjustment, or account changes, the average absolute error of monthly prediction is reduced from 38.4 kWh to 23.5 kWh. After the prediction time range ends, the system obtains actual meter reading feedback data, calculates the daily, weekly, and monthly prediction deviations, and updates the profile features, cycle fusion weights, and residual correction coefficients to generate multi-cycle residential electricity prediction output data.

[0209] Table 1. Comparison of Residential Multi-Period Electricity Prediction Data Based on Identifier Resolution

[0210] Number of residential users / households tested 1500 1500 1500 Abnormal record recognition rate / % 63.5 94.2 95.6 Mismatch rate of customer relationship / % 3.8 1.2 0.9 Channel attribution anomaly correction rate / % 58.7 91.5 93.1 Account change association accuracy rate / % 71.4 95.3 96.2 Historical electricity valid sample retention rate / % 88.6 95.8 96.4 Daily forecast mean absolute error / kWh 1.38 0.82 0.74 Weekly forecast mean absolute error / kWh 8.21 5.13 4.76 Monthly forecast mean absolute error / kWh 32.6 20.7 18.9 Daily forecast error for stable power consumption type / kWh 0.91 0.63 0.58 Nighttime Enhanced Daily Forecast Error / kWh 1.24 0.78 0.71 Enhanced Weekly Forecast Error During Holidays / kWh 8.96 5.62 5.08 High volatility daily forecast error / kWh 1.87 1.12 1.05 There is a monthly forecast error of / kWh for users who have changed their meters. 38.4 23.5 21.8 There is a monthly forecast error for users in the transformer area adjustment / kWh 36.9 22.8 20.6 Multi-period prediction result completeness rate / % 86.2 97.1 98.0 Prediction results binding identifier link ratio / % 74.8 98.3 98.7 Feedback profile update completion rate / % 62.5 91.6 95.2 Number of abnormal samples manually reviewed / items 276 96 73

[0211] As shown in Table 1, traditional historical sequence prediction methods mainly rely on extrapolation based on historical electricity consumption changes. They lack sufficient processing of the identification relationships between residential users, metering devices, power supply areas, and electricity accounts. Therefore, they are weak in identifying abnormal identification records, correcting meter-to-user mismatches, correcting abnormal distribution area attribution, and associating with account changes. The traditional method achieves an abnormal identification record identification rate of 63.5%, a meter-to-user mismatch rate of 3.8%, and a prediction result binding identification link ratio of only 74.8%. This indicates that in residential electricity consumption data involving meter replacement, account changes, or distribution area adjustments, the traditional method is prone to inaccurate historical electricity consumption attribution.

[0212] This invention unifies the association between residential users, metering devices, power supply areas, and electricity accounts through identifier resolution links, increasing the identification rate of abnormal identifier records to 94.2%, and further to 95.6% after feedback correction; the meter-to-user mismatch rate is reduced from 3.8% in the traditional method to 0.9% after feedback correction; and the accuracy of account change association is improved from 71.4% to 96.2%. These results demonstrate that identifier resolution processing can effectively reduce mismatch problems between multi-source residential electricity consumption data, providing a more reliable data foundation for subsequent profile construction and electricity consumption forecasting.

[0213] Regarding prediction errors, the traditional method has an average absolute error of 1.38 kWh for daily predictions, 8.21 kWh for weekly predictions, and 32.6 kWh for monthly predictions. The initial prediction results of this invention are reduced to 0.82 kWh, 5.13 kWh, and 20.7 kWh, respectively, and further reduced to 0.74 kWh, 4.76 kWh, and 18.9 kWh after feedback correction. This change demonstrates that combining residential user profiles, multi-cycle electricity consumption characteristics, and the improved StemGNN identifier parsing multi-cycle electricity prediction model can simultaneously utilize short-term daily fluctuations, weekly electricity consumption patterns, and monthly settlement trends, thereby improving the accuracy of multi-cycle predictions.

[0214] Looking at different user types, the prediction errors for users with stable electricity consumption, those with enhanced nighttime consumption, those with enhanced holiday consumption, and those with high volatility all decreased. Specifically, the daily prediction error for high volatility users decreased from 1.87 kWh to 1.05 kWh after feedback correction, and the weekly prediction error for those with enhanced holiday consumption decreased from 8.96 kWh to 5.08 kWh, indicating that this invention can handle residential users with significant changes in electricity consumption behavior through dynamic profile features and residual correction mechanisms. For users who have undergone meter replacement and transformer area adjustments, the monthly prediction errors decreased from 38.4 kWh and 36.9 kWh to 21.8 kWh and 20.6 kWh, respectively, further demonstrating that the identifier weight constraint can improve the prediction bias caused by changes in business relationships.

[0215] Furthermore, the completeness rate of multi-cycle prediction results in this invention has increased from 86.2% to 98.0%, the completion rate of feedback profile updates has increased from 62.5% to 95.2%, and the number of abnormal samples requiring manual review has decreased from 276 to 73. Overall, this invention not only improves the accuracy of daily, weekly, and monthly electricity consumption predictions for residents, but also enhances the correlation between prediction results and the identifier resolution link, resident profiles, and actual meter reading feedback, thus better supporting residential electricity consumption analysis, load management, and refined power supply services.

[0216] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting residential multi-period electricity consumption based on identifier resolution, characterized in that, Includes the following steps: S1. Obtain multi-source electricity consumption data of residential users within the predicted time range, and perform identification verification, time alignment, anomaly removal, missing data filling and unified dimension processing on the multi-source electricity consumption data to obtain standardized electricity input data. S2. Based on the identifier resolution field in the standardized electricity input data, construct the association link between residential users, metering devices, power supply areas and electricity accounts to obtain residential identifier resolution association data; S3. Construct a resident user profile based on standardized electricity input data and resident identifier resolution and correlation data to obtain resident user profile features; S4. Based on the characteristics of residential user profiles and historical electricity consumption sequences, daily, weekly and monthly cycles are divided and cycle embedding is performed to obtain multi-cycle electricity consumption characteristics. S5. Input the multi-cycle electricity consumption characteristics and the associated data of the residential identifier resolution into the improved StemGNN identifier resolution multi-cycle electricity consumption prediction model. Based on the associated data of the residential identifier resolution, construct the residential identifier resolution electricity consumption map with identifier weight constraints. Then, perform spectral encoding, frequency domain time modeling and cross-cycle fusion prediction on the residential identifier resolution electricity consumption map and the multi-cycle electricity consumption characteristics to obtain the multi-cycle electricity consumption prediction results. S6. Based on the multi-cycle electricity prediction results and actual meter reading feedback data, perform deviation calculation, profile update and feedback correction to obtain the residential multi-cycle electricity prediction output data.

2. The residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 1, characterized in that, S1 includes the following steps: S11. Obtain user profile records, metering collection records, household registration records, electricity consumption behavior records, transformer area affiliation records, and time period records. Extract user number, metering point number, transformer area number, account number, collection time, historical electricity consumption, household area, and electricity consumption behavior fields to obtain raw electricity consumption data. S12. Perform field integrity verification, field format verification, and duplicate identification verification on the original electricity consumption data to obtain the electricity consumption data after identification verification. S13. Based on the collection time, meter reading time, settlement cycle and predicted time range, the time base of the identified and verified electricity consumption data is unified and the cycle boundary is aligned to obtain time-aligned electricity consumption data. S14. Based on the changes in electricity consumption in adjacent cycles, the distribution of electricity consumption of similar residents, and the collection status field, perform anomaly removal, adjacent cycle filling, and unified dimension processing on the time-aligned electricity consumption data to obtain standardized electricity consumption input data.

3. The residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 1, characterized in that, S2 includes the following steps: S21. Extract user number, metering point number, transformer area number, account number, effective time and business change record from standardized electricity input data, and convert them into residential user nodes, metering device nodes, power supply transformer area nodes and electricity account nodes to obtain a set of identified nodes; S22. Based on the user meter binding record, meter area attribution record and user account settlement record, establish binding edge, attribution edge and settlement edge to obtain the initial identifier resolution link; S23. Based on the meter change record, transformer area change record, account change record, and binding effective time, perform link merging or link splitting on the initial identifier resolution link to obtain resident identifier resolution associated data; S24. Generate an identifier resolution link number based on the resident identifier resolution association data, and write the identifier resolution link number into the corresponding historical electricity consumption sequence and profile basic field to obtain resident electricity consumption data with identifier link number.

4. The residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 1, characterized in that, S3 includes the following steps: S31. Based on the resident identifier parsing and association data, read the unit area, number of rooms, number of occupants, metering point type, power supply area attribute and account settlement attribute, and perform static encoding on the above fields to obtain static profile features; S32. Based on the historical electricity consumption sequence and electricity consumption behavior records in the standardized electricity input data, calculate the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, continuous electricity consumption duration and electricity fluctuation range to obtain dynamic profile features; S33. Based on the static profile features, dynamic profile features, and identifier resolution link number, perform feature splicing, noise reduction, completion, and profile category encoding to obtain the resident user profile features; S34. Write the resident user profile features into the corresponding resident user node, and write the profile category code into the resident identifier parsing association data to obtain resident identifier parsing association data with profile attributes.

5. The residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 1, characterized in that, S4 includes the following steps: S41. Based on the collection date, weekday attribute, month attribute, holiday attribute, settlement cycle attribute, and identifier resolution link number in the historical electricity sequence, periodic labels are marked to obtain the periodic labeled electricity sequence; S42. Based on the continuous daily electricity consumption, weekly working day electricity consumption, weekly rest day electricity consumption and monthly settlement interval electricity consumption in the periodic marked electricity consumption sequence, segments are divided to obtain daily periodic electricity consumption segments, weekly periodic electricity consumption segments and monthly periodic electricity consumption segments. S43. Based on the identifier resolution link number, the resident user profile features are matched with the daily cycle electricity consumption segment, the weekly cycle electricity consumption segment and the monthly cycle electricity consumption segment to obtain three types of cycle profile electricity consumption features. S44. Perform periodic location encoding, periodic category encoding, and numerical mapping on the electricity consumption characteristics of the three types of periodic profiles to obtain multi-period electricity consumption characteristics.

6. The residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 1, characterized in that, S5 includes the following steps: S51. Input the residential identifier resolution association data into the improved StemGNN identifier resolution multi-cycle electricity prediction model, construct the residential identifier resolution electricity consumption map based on the identifier resolution links between residential users, metering devices, power supply areas and electricity accounts, and obtain graph structure data including node codes, edge relationship codes and identifier weight adjacency matrices. S52. Write the multi-cycle electricity consumption features into the residential user nodes in the residential identifier resolution electricity consumption map according to the identifier resolution link number, and map and fuse the residential user profile features with the daily cycle embedding, weekly cycle embedding and monthly cycle embedding to obtain the graph node input vector. S53. In the improved StemGNN identifier parsing multi-cycle power prediction model, based on graph structure data and graph node input vectors, a graph Laplace with identifier weight constraints is constructed, a graph Fourier transform is performed, and a spectral domain graph association encoding is performed to obtain the identifier-constrained spectral domain graph association features. S54. Perform frequency domain time decomposition and periodic time convolution on the association features of the constrained spectral domain graph to obtain daily periodic time features, weekly periodic time features and monthly periodic time features. S55. Based on the daily cycle time series characteristics, weekly cycle time series characteristics, monthly cycle time series characteristics and resident user profile characteristics, cross-cycle fusion and residual correction are performed to obtain cross-cycle fusion prediction characteristics. S56. Generate daily, weekly, and monthly predicted electricity consumption based on the cross-cycle fusion prediction characteristics, and bind the predicted electricity consumption with the identifier resolution link number to obtain multi-cycle electricity consumption prediction results.

7. A residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 6, characterized in that, S51 includes the following steps: S511. In the improved StemGNN identifier resolution multi-cycle electricity prediction model, user number, metering point number, transformer area number, account number, profile category code and business activation time are extracted from the resident identifier resolution associated data, and the extracted content is converted into graph node coded data. S512. Based on the user meter binding relationship, meter substation affiliation relationship, user account settlement relationship and the relationship between residents in the same substation, encode the connection category, connection direction and effective time between different identifier nodes to obtain graph edge coding data. S513. Establish an initial adjacency matrix based on graph node coding data and graph edge coding data, and perform time segmentation processing on edge connection status according to table change records, transformer area change records and account change records to obtain a time segmentation adjacency matrix. S514. Based on the identifier conflict record, binding effective status, station area ownership status and account settlement status, correct the abnormal edge weights in the time segment adjacency matrix to obtain the identifier weight adjacency matrix. S515. Combine the graph node encoding data, graph edge encoding data, and label weight adjacency matrix into a residential label resolution electricity consumption graph to obtain the graph structure data for inputting the improved StemGNN label resolution multi-cycle electricity prediction model.

8. A residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 6, characterized in that, S53 includes the following steps: S531. In the improved StemGNN identifier parsing multi-cycle power prediction model, the identifier weight adjacency matrix is ​​used as the input of the connection structure. Based on the node category, edge relationship category, binding effective status, transformer area affiliation status, account settlement status and resident user profile characteristics, the spectral domain propagation constraint coefficient of each identifier connection relationship is calculated to generate the identifier weight matrix. S532. Merge the identifier weight matrix and the identifier weight adjacency matrix. Identify invalid links, conflicting links and expired links based on the binding effective time, binding ineffective time, station area ownership status, account settlement status and identifier conflict records. Suppress the edge weights corresponding to invalid links, conflicting links and expired links, construct a graph Laplacian matrix with identifier resolution constraints, and obtain graph structure data. S533. Perform graph Fourier transform on the graph node input vector based on the graph structure data to map the multi-cycle electricity consumption characteristics to the graph domain and obtain the multi-cycle characteristics of the graph domain. S534. Perform spectral convolution, spectral filtering, and node association aggregation on the multi-period features of the spectral domain, and adjust the spectral domain association strength according to the label weight matrix to obtain the label-constrained spectral domain graph association features.

9. A residential multi-cycle electricity consumption prediction method based on identifier resolution according to claim 6, characterized in that, S55 includes the following steps: S551. In the improved StemGNN identifier resolution multi-cycle power prediction model, the daily cycle time series features, weekly cycle time series features, and monthly cycle time series features are aligned according to the residential user node and the identifier resolution link number to obtain cross-cycle aligned features. S552. Calculate the daily-week correlation weight and weekly-month correlation weight based on the cross-cycle alignment features, and perform weighted fusion of the three types of cycle time series features based on the daily-week correlation weight and weekly-month correlation weight to obtain cycle fusion features; S553. The periodic fusion features are spliced ​​with the residential user profile features, and profile residual correction features are generated based on the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio and electricity fluctuation range. S554. Based on the periodic fusion features and the image residual correction features, a mapping process is performed to obtain the cross-period fusion prediction features.

10. A method for predicting residential multi-cycle electricity consumption based on identifier resolution according to claim 1, characterized in that, S6 includes the following steps: S61. Obtain actual meter reading feedback data, and perform identifier matching and time matching on the user number, metering point number, transformer area number, account number and meter reading time in the actual meter reading feedback data to obtain actual feedback electricity data. S62. Align the actual feedback power data with the multi-cycle power prediction results and calculate the difference to obtain the daily cycle prediction deviation, weekly cycle prediction deviation and monthly cycle prediction deviation. S63. Update the peak-valley electricity consumption ratio, nighttime electricity consumption ratio, holiday electricity consumption ratio, continuous electricity consumption duration, and electricity fluctuation range based on the three types of prediction deviations to obtain feedback and update profile features; S64. Input the updated profile features and actual feedback power data into the improved StemGNN identifier parsing multi-cycle power prediction model, and update the identifier weight matrix, cycle fusion weight and residual correction coefficient to obtain the updated multi-cycle power prediction results. S65. Generate residential multi-cycle electricity prediction output data based on the updated multi-cycle electricity prediction results, identifier resolution link number, residential user profile category and prediction deviation data.

Images