Image migration recursive dynamic correction new user load prediction method and system

Abstract

The present application relates to the technical field of power load prediction, and discloses a new user load prediction method and system based on portrait migration recursive dynamic correction, which comprises a data acquisition and quality access module, a portrait generation module, a similar user retrieval module, a basic recursive prediction module, a candidate curve routing module, a dynamic fusion module and a backflow correction module. By constructing a multi-dimensional power user load characteristic portrait and based on the portrait similarity, a mature user cluster is retrieved, capacity normalization, time alignment and similarity weighting processing are performed on the mature user historical curve, and portrait migration load curve data is generated, thereby solving the problem of roughness of the industry average and capacity proportion method curve, enabling the cold start stage to obtain prior load curve reflecting the intra-day shape difference under different industries, shift systems, regions and weather sensitivities, and improving the accuracy of early load prediction and the stability of peak-valley shape.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, specifically to a method and system for forecasting new user load based on profile migration and progressive dynamic correction. Background Technology

[0002] With the rapid and stable development of the economy and society, the orderly progress of new urbanization and the construction of beautiful villages, and especially with the continuous increase in household appliances in rural households, my country's residential electricity load is constantly growing, and the characteristics of residential electricity consumption are also constantly changing. Electricity load forecasting is an important technical foundation for electricity sales companies to carry out electricity forecasting, transaction planning, contract performance and deviation risk control.

[0003] Currently, due to the lack of long-term stable 15-minute historical load curves for newly signed contracts, new market entrants, or newly commissioned users, existing load forecasting methods cannot generate 96-point load priors for the cold start phase. Instead, they rely solely on industry averages or capacity ratios for estimation, resulting in overly coarse load curves that fail to reflect intraday variations across different industries, shift schedules, regions, and weather sensitivities. This leads to low accuracy in early load forecasting and unstable peak-valley patterns, which in turn affects electricity planning and deviation risk control.

[0004] Therefore, a new user load forecasting method and system based on dynamic correction of user profile migration is proposed to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for predicting new user load based on dynamic correction of profile migration, which solves the problems mentioned in the background technology, such as low accuracy of early load prediction and unstable peak-valley patterns, which in turn affect power planning and deviation risk control.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method and system for predicting new user load based on profile migration and progressive dynamic correction, the method comprising the following steps: S1. Collect basic attribute data of target new users and historical load data of mature users. After quality access verification, generate basic attribute data of target new users and historical load data of mature users. S2. Based on the basic attribute data of the target new users, construct a power user load characteristic profile and generate target user profile data; S3. Based on the target user profile data and the mature user historical load data, retrieve mature users similar to the target user and output similar mature user cluster data; S4. Based on the similar mature user cluster data, perform capacity normalization, time alignment and similarity weighting on the historical load curves of mature users to generate profile migration load curve data. S5. Input the target user profile data and the historical load data of the target new users into the basic recursive load prediction model to generate basic predicted load curve data. S6. Based on the historical load data of the target new users, construct a set of candidate curves and combine them with the characteristics of the running status agent to generate candidate curve routing result data; S7. Based on the target user profile data, the similar mature user cluster data, the basic predicted load curve data, and the candidate curve routing result data, calculate the dynamic fusion weight of each prediction channel and generate the final fusion load prediction curve data. S8. Based on the comparison results between the final fusion load prediction curve data and the actual load data, perform backflow correction on the target user profile data, the similar mature user cluster data, the basic recursive load prediction model, and the dynamic fusion weight to generate corrected prediction model parameter data.

[0007] Preferably, the generation of basic attribute data for target new users and historical load data for mature users in step S1 includes the following steps: S11. Obtain the static attributes, capacity attributes, industry attributes, regional attributes, production shift attributes, contract capacity, and application capacity data of the target new users through the data interface, and generate basic attribute data of the target new users; S12. Obtain mature user profiles, historical 96-point load curves, historical prediction errors and load stability data through the data interface, and generate historical load data for mature users. S13. Perform 96-point integrity check, null value check, duplicate time point check and unit caliber check on the 15-minute load data in the target new user basic attribute data and the mature user historical load data, and mark the data that passes the check as the target new user basic attribute data and the mature user historical load data.

[0008] Preferably, generating target user profile data in step S2 includes the following steps: S21. Obtain the basic attribute data of the target new user; S22. Encode the industry category, contract capacity, installed capacity, region, production shift, whether it has its own photovoltaic system, peak-valley power structure, load factor and weather sensitivity in the basic attribute data of the target new user into a multi-dimensional profile vector, construct a power user load characteristic profile, and generate target user profile data.

[0009] Preferably, the output of similar mature user cluster data in S3 includes the following steps: S31. Obtain the target user profile data and the mature user historical load data; S32. Based on the target user profile data, calculate the comprehensive similarity between the target user and each mature user in terms of industry, capacity, region, shift system and weather sensitivity, select the mature users with the highest comprehensive similarity ranking, and output the cluster data of similar mature users.

[0010] Preferably, generating the portrait migration load curve data in step S4 includes the following steps: S41. Obtain the data of the similar mature user clusters; S42. Divide the 96-point historical load curve of each mature user in the similar mature user cluster by its contract capacity, perform capacity normalization processing, and generate capacity normalized curve data. S43. Based on the day type, season, and weather conditions of the predicted day, perform weekday alignment on the capacity normalized curve data to generate time-aligned curve data. S44. Based on the comprehensive similarity, the time alignment curve data is weighted and summed, and then denormalized according to the contract capacity of the target new user to generate profile migration load curve data.

[0011] Preferably, generating the basic forecast load curve data in step S5 includes the following steps: S51. Obtain the target user profile data and the historical load data of the target new user; S52. Using the target user profile data, the historical load data of the target new user, weather data, and calendar data as input features, call the pre-trained time series prediction model to output the basic load curve from the first prediction day to the third prediction day, and generate basic prediction load curve data.

[0012] Preferably, generating candidate curve routing result data in step S6 includes the following steps: S61. Obtain historical load data of the target new user; S62. Based on the historical load data, construct a candidate curve set including yesterday's curve, the curve of the same day last week, and the median curve of the past seven days; S63. Extract the nighttime base load, early ramp speed and recent peak-valley difference of the target prediction day as operational status proxy features, assign weights to each curve in the candidate curve set according to the operational status proxy features, and generate candidate curve routing result data.

[0013] Preferably, generating the final fused load prediction curve data in step S7 includes the following steps: S71. Obtain the target user profile data, the similar mature user cluster data, the basic predicted load curve data, and the candidate curve routing result data; S72. Calculate the quality scores of the profile migration channel, the basic prediction channel, and the candidate curve routing channel based on the number of days of the actual load data of the target new users, the average similarity of the target user profile data, the dispersion of the similar mature user cluster data, and the recent error of the basic model. S73. Based on the quality score, calculate the dynamic fusion weight of each prediction channel using a normalization function, and perform weighted fusion on the portrait migration load curve data, the basic prediction load curve data, and the candidate curve routing result data to generate the final fused load prediction curve data.

[0014] Preferably, generating the corrected prediction model parameter data in step S8 includes the following steps: S81. Obtain the final converged load forecast curve data and the actual load data of the target new users; S82. Calculate the total daily error and the 96-point same-point error between the final fused load prediction curve data and the actual load data; S83. Based on the daily total error and the 96-point same-point error, update the load rate and peak-valley difference attributes in the target user profile data, remove users in the similar mature user cluster data that deviate greatly from the actual curve, and adjust the input features of the basic recursive load prediction model and the dynamic fusion weights to generate corrected prediction model parameter data.

[0015] Preferably, the system includes: The data acquisition and quality access module collects basic attribute data of target new users and historical load data of mature users through the data interface. After 96-point integrity verification, null value verification, duplicate time point verification and unit caliber verification, it outputs basic attribute data of target new users and historical load data of mature users. The profile generation module receives the basic attribute data of the target new user, encodes industry, capacity, region, shift system, and weather sensitivity into profile vectors through the multi-dimensional profile vector encoding unit, and outputs the target user profile data. The similar user retrieval module receives the target user profile data and the mature user historical load data, selects the top-ranked mature users through the comprehensive similarity calculation unit, and outputs similar mature user cluster data. The profile migration curve generation module receives the similar mature user cluster data, processes the historical load curves of mature users through a capacity normalization unit, a time alignment unit, and a similarity weighting unit, and outputs profile migration load curve data. The basic recursive prediction module receives the target user profile data and the target new user historical load data, and outputs the basic load curves for the first to third prediction days through the time series prediction model to generate basic predicted load curve data. The candidate curve routing module receives historical load data of the target new user and outputs candidate curve routing result data through the candidate curve construction unit and the running status agent feature extraction unit. The dynamic fusion module receives the target user profile data, the similar mature user cluster data, the basic predicted load curve data, and the candidate curve routing result data, and outputs the final fused load prediction curve data through the quality score calculation unit and the normalization fusion unit. The backflow correction module receives the final fused load prediction curve data and the actual load data, and outputs the corrected prediction model parameter data through the error calculation unit and the parameter adjustment unit.

[0016] Compared with existing technologies, this invention provides a method and system for predicting new user load based on profile migration and progressive dynamic correction, which has the following advantages: 1. In this invention, by constructing a multi-dimensional load characteristic profile of power users and retrieving mature user clusters based on profile similarity, capacity normalization, time alignment, and similarity weighting are performed on the historical curves of mature users to generate profile migration load curve data. This solves the problem of coarse curves from industry averages and capacity ratio methods, enabling the acquisition of prior load curves reflecting intraday morphological differences under different industries, shift systems, regions, and meteorological sensitivities during the cold start phase, thereby improving the accuracy of early load forecasting and the stability of peak-valley morphology.

[0017] 2. In this invention, by inputting target user profile data and historical load data into the basic recursive load prediction model, a candidate curve set is constructed and candidate curve routing results are generated by combining the proxy features of the operating status. Then, the dynamic fusion weight of each prediction channel is calculated based on profile similarity, cluster dispersion, and model error factors. This achieves adaptive weighted fusion of the three channels of profile migration, basic recursion, and candidate curve routing, avoiding the limitations of a single prediction method and enabling the prediction system to maintain prediction accuracy during the cold start to stable transition period.

[0018] 3. In this invention, by comparing the final fused load forecast curve data with the actual load data, the daily total error and the 96-point similarity error are calculated. Based on the error results, the load rate and peak-valley difference attributes in the target user profile data are updated, users with large deviations in similar mature user clusters are removed, and the input features and dynamic fusion weights of the basic recursive load forecast model are adjusted. This achieves closed-loop backflow correction of all elements, solves the problem of traditional methods lacking adaptive correction capabilities, and improves the long-term accuracy and reliability of new user load forecasting. Attached Figure Description

[0019] Figure 1This is a flowchart of the new user load prediction method based on profile migration and progressive dynamic correction according to the present invention; Figure 2 This is a diagram illustrating the architecture of the new user load prediction system based on dynamic correction of image migration according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] For specific implementation examples, please refer to: Figures 1-2 A method and system for predicting new user load based on dynamic correction driven by user profile migration, the method includes the following steps: S1. Collect basic attribute data of target new users and historical load data of mature users. After quality access verification, generate basic attribute data of target new users and historical load data of mature users. S2. Based on the basic attribute data of the target new users, construct a power user load characteristic profile and generate target user profile data; S3. Based on the target user profile data and the historical load data of mature users, retrieve mature users similar to the target users and output similar mature user cluster data. S4. Based on similar mature user cluster data, perform capacity normalization, time alignment and similarity weighting on the historical load curves of mature users to generate profile migration load curve data. S5. Input the target user profile data and the historical load data of the target new users into the basic recursive load forecasting model to generate basic forecast load curve data. S6. Based on the historical load data of the target new users, construct a set of candidate curves and combine them with the characteristics of the running status agent to generate candidate curve routing result data; S7. Based on the target user profile data, similar mature user cluster data, basic predicted load curve data, and candidate curve routing result data, calculate the dynamic fusion weight of each prediction channel and generate the final fusion load prediction curve data. S8. Based on the comparison results between the final fused load forecast curve data and the actual load data, perform backflow correction on the target user profile data, similar mature user cluster data, basic recursive load forecast model and dynamic fused weights, and generate the corrected forecast model parameter data.

[0022] The steps involved in generating basic attribute data for target new users and historical workload data for mature users in S1 are as follows: S11. Obtain the static attributes, capacity attributes, industry attributes, regional attributes, production shift attributes, contract capacity, and application capacity data of the target new users through the data interface, and generate basic attribute data of the target new users; S12. Obtain mature user profiles, historical 96-point load curves, historical prediction errors and load stability data through the data interface, and generate historical load data for mature users. S13. Perform 96-point integrity checks, null value checks, duplicate time point checks, and unit caliber checks on the 15-minute load data in the target new user basic attribute data and mature user historical load data. Mark the data that passes the checks as target new user basic attribute data and mature user historical load data, including the following steps: S131. Check if each 15-minute load data entry contains 96 time points. If it does not, mark it as incomplete. S132. Check whether the load value at each time point is a valid value. If it is a null value or a non-numeric value, mark it as a null value abnormality. S133. Check if the timestamps of adjacent time points are incrementing. If there are identical timestamps, mark them as duplicate time points. S134. Check whether the unit of the load data is consistent with the preset unit. If they are inconsistent, perform unit conversion. S135. Mark all data that passes the verification as qualified data.

[0023] Generating target user profile data in S2 includes the following steps: S21. Obtain basic attribute data of the target new user; S22. Encode the industry category, contracted capacity, applied capacity, region, production shift system, whether it has its own photovoltaic system, peak-valley power structure, load factor, and weather sensitivity in the basic attribute data of the target new users into multi-dimensional profile vectors, construct a power user load characteristic profile, and generate target user profile data, including the following steps: S221. Use one-hot encoding to convert industry category, region and production shift into binary vector; S222. Contracted capacity, installed capacity, and load factor are mapped to an interval using maximum and minimum value normalization. The expression for maximum and minimum value normalization is: ; in, The original attribute value. The minimum value of this attribute across all users. The maximum value, The value is the normalized value, and its range is [0,1]. S223. Boolean value coding is used for whether the photovoltaic system is owned; a three-dimensional vector representation of the peak, flat and valley power structure is used for the power ratio of each time period; and a level numerical coding is used for meteorological sensitivity. S224. Concatenate all encoded vectors into a single multidimensional portrait vector of uniform dimension.

[0024] Outputting similar mature user cluster data in S3 includes the following steps: S31. Obtain target user profile data and historical load data of mature users; S32. Based on the target user profile data, calculate the comprehensive similarity between the target user and each mature user in terms of industry, capacity, region, shift system, and weather sensitivity. Select the mature users with the highest comprehensive similarity ranking and output the cluster data of similar mature users, including the following steps: S321. Obtain the category codes of each attribute in the target user profile data; S322. Calculate the industry similarity between the target user and each mature user in terms of industry attribute, capacity similarity in terms of capacity attribute, region similarity in terms of region attribute, shift similarity in terms of production shift attribute, and weather sensitivity similarity in terms of weather sensitivity attribute. S323. Assign a preset weight coefficient to each attribute, and then sum the weighted similarities of industry similarity, capacity similarity, region similarity, shift system similarity, and weather sensitivity similarity to obtain the comprehensive similarity. The expression for the weighted summation is: ; in, Indicates target user With mature users The overall similarity For industry similarity, For capacity similarity, For regional similarity, For class system similarity, For meteorological sensitivity similarity, to These are the preset weight coefficients for each attribute, and the sum of all weight coefficients is 1.

[0025] Generating portrait migration load curve data in S4 includes the following steps: S41. Obtain data on similar mature user clusters; S42. Divide the 96-point historical load curve of each mature user in the similar mature user cluster by its contract capacity, perform capacity normalization processing, and generate capacity normalization curve data. S43. Based on the day type, season, and weather conditions of the forecast day, align the capacity normalized curve data to weekdays to generate time-aligned curve data, including the following steps: S431. Obtain the day type, season, and weather conditions for the forecast day, and determine the forecast day as a working day; S432. Select the capacity normalization curves corresponding to the days with the same prediction day type from similar mature user clusters; S433. Further match the selected curves with curves that are similar to the predicted season and weather conditions, and use them as time-aligned curve data. S44. Based on the comprehensive similarity, perform weighted summation on the time alignment curve data and perform inverse normalization based on the contract capacity of the target new user to generate profile migration load curve data.

[0026] Generating basic forecast load curve data in S5 includes the following steps: S51. Obtain target user profile data and historical load data of target new users; S52. Using target user profile data, historical load data of target new users, weather data, and calendar data as input features, call the pre-trained time-series forecasting model to output the basic load curve from the first forecast day to the third forecast day, generating basic forecast load curve data; wherein, the pre-trained time-series forecasting model is trained through the following steps: S521. Collect historical load data, corresponding date weather data, and calendar data from multiple mature users as training samples, and use the actual load curve as the label; S522. Input the training samples into the time series prediction model to be trained, and output the predicted load curve; S523. Calculate the loss function value between the predicted load curve and the actual load curve. The expression for the loss function is: ; in, The total number of training samples, For the first The model prediction value for each sample. For the first The actual load value of each sample This represents the mean squared error loss value; S524. Update the model parameters by backpropagation based on the loss function value, iterate until convergence, and obtain the pre-trained time series prediction model.

[0027] The steps involved in generating candidate curve routing result data in S6 are as follows: S61. Obtain historical load data of the target new user; S62. Based on historical load data, construct a candidate curve set including yesterday's curve, the curve of the same day last week, and the median curve of the past seven days; S63. Extract the nighttime base load, early ramp speed, and recent peak-valley difference for the target prediction day as operational status proxy features. Assign weights to each curve in the candidate curve set based on the operational status proxy features to generate candidate curve routing result data, including the following steps: S631. Obtain 96 points of historical load data for the target new user within the most recent preset number of days; S632. The average load value at each time point from 00:00 to 05:00 every day is taken as the nighttime base load, calculated as follows: ; in, For the preset number of days, For the first Heavenly At that point in time, The load values ​​correspond to 96 data points from 00:00 to 04:45 each day. For nighttime base load; S633. Take the average load change rate at each time point from 6:00 AM to 9:00 AM every day as the early morning climbing speed. The calculation formula is: ; in, This index corresponds to 12 time points from 6:00 AM to 8:45 AM. The number of time points, The hour is a 15-minute interval. For time point The load value, To increase the speed of climbing; S634. The difference between the daily peak load and the minimum load is taken as the average of the recent preset number of days, and is used as the recent peak-valley difference. The calculation formula is as follows: ; in, For the preset number of days, For the first Tian Shike The load value, and Take the maximum and minimum values ​​at 96 time points on the same day. This represents the recent peak-to-valley difference.

[0028] Generating the final merged load forecast curve data in S7 includes the following steps: S71. Obtain target user profile data, similar mature user cluster data, basic predicted load curve data, and candidate curve routing result data; S72. Based on the number of days of actual load data for target new users, the average similarity of target user profile data, the dispersion of similar mature user cluster data, and the recent error of the basic model, calculate the quality scores of the profile migration channel, the basic prediction channel, and the candidate curve routing channel, including the following steps: S721. Obtain the number of days of actual load data for the target new user, and divide the number of days by the preset stable days threshold to obtain the days factor; S722. Obtain the average similarity between the target user profile data and each user in a similar mature user cluster, and use it as the profile similarity factor; S723. Calculate the standard deviation of each user profile vector in a cluster of similar mature users, and take its reciprocal as the cluster dispersion factor. S724. Obtain the average prediction error of the basic recursive load forecasting model over the most recent preset number of days, and take its reciprocal as the model error factor. The calculation formula is as follows: ; in, For the preset number of days, For the first Daily actual load value, For the first Daily basic recursive load forecasting model output value, This represents the average prediction error; S725. Multiply the number of days factor, image similarity factor, cluster dispersion factor, and model error factor by their respective preset weights and then sum them to obtain the quality score for each channel. The calculation formula is as follows: ; in, For the first Quality score of each predicted channel The number of days is the factor. For portrait similarity factor, Cluster dispersion factor For model error factor, , , , These are the preset weights for the corresponding factors, and the sum of each weight is 1; S726. After calculating the quality score of each channel, it needs to be converted into dynamic fusion weights using a normalization function. To ensure that the sum of the weights is 1, the normalization formula is: ; in, For the first Quality score of each predicted channel It is a natural exponential function, and the denominator is... For all predicted channel quality score indices, For the first The final dynamic fusion weights of each prediction channel; S73. Based on the quality score, calculate the dynamic fusion weight of each prediction channel through a normalization function, and perform weighted fusion on the profile migration load curve data, basic prediction load curve data and candidate curve routing result data to generate the final fused load prediction curve data.

[0029] Generating corrected prediction model parameter data in S8 includes the following steps: S81. Obtain the final converged load forecast curve data and the actual load data of the target new users; S82. Calculate the total daily error and the 96-point same-point error between the final fused load forecast curve data and the actual load data. S83. Based on the daily total error and the 96-point same-point error, update the load rate and peak-valley difference attributes in the target user profile data, remove users in the similar mature user cluster data that deviate significantly from the actual curve, and adjust the input features and dynamic fusion weights of the basic recursive load prediction model to generate the corrected prediction model parameter data, including the following steps: S831. Obtain the actual load data of 96 points for the target new user in the most recent preset number of days; S832. Calculate the ratio of the daily average load to the maximum load, and take the average value as the new load rate attribute value. The calculation formula is: ; in, For the preset number of days, For the first Average load per day For the first The maximum load of the day For the new load factor attribute value, For point-in-time index; S833. Calculate the difference between the daily maximum load and the minimum load, and take the average value as the new peak-valley difference attribute value. The calculation formula is: ; in, For the new peak-valley difference attribute value; S834. Replace the original load rate and peak-valley difference attributes in the target user profile data with new load rate and peak-valley difference attribute values.

[0030] The system includes: The data acquisition and quality access module collects basic attribute data of target new users and historical load data of mature users through the data interface. After 96-point integrity verification, null value verification, duplicate time point verification and unit caliber verification, it outputs basic attribute data of target new users and historical load data of mature users. The profile generation module receives basic attribute data of the target new user, encodes industry, capacity, region, shift system, and weather sensitivity into profile vectors through the multi-dimensional profile vector encoding unit, and outputs the target user profile data. The similar user retrieval module receives target user profile data and mature user historical load data, selects the top-ranked mature users through the comprehensive similarity calculation unit, and outputs similar mature user cluster data. The profile migration curve generation module receives similar mature user cluster data, processes the historical load curves of mature users through capacity normalization unit, time alignment unit and similarity weighting unit, and outputs profile migration load curve data. The basic recursive forecasting module receives target user profile data and target new user historical load data, and outputs the basic load curves for the first to third forecast days through the time series forecasting model, generating basic forecast load curve data. The candidate curve routing module receives historical load data of the target new user and outputs candidate curve routing result data through the candidate curve construction unit and the running status agent feature extraction unit. The dynamic fusion module receives target user profile data, similar mature user cluster data, basic predicted load curve data, and candidate curve routing result data. Through the quality score calculation unit and the normalization fusion unit, it outputs the final fused load prediction curve data. The backflow correction module receives the final fused load forecast curve data and the actual load data, and outputs the corrected forecast model parameter data through the error calculation unit and parameter adjustment unit.

[0031] The operation steps of the new user load forecasting method and system based on profile migration and progressive dynamic correction are as follows: Step 1: Data Acquisition and Quality Access Basic attribute data of target new users is collected, including static attributes, capacity attributes, industry attributes, regional attributes, production shift attributes, contracted capacity, and applied capacity. Historical load data of established users is also collected, including user profiles, historical 96-point load curves, historical forecast errors, and load stability data. The collected 15-minute load data undergoes 96-point integrity checks, null value checks, duplicate time point checks, and unit caliber checks. Data that passes the checks is marked as basic attribute data for target new users and historical load data for established users.

[0032] Step 2: User Profile Construction and Similar User Retrieval The industry category, contracted capacity, applied capacity, region, production shift system, whether it has its own photovoltaic system, peak-valley power structure, load factor, and weather sensitivity of the target new user's basic attribute data are encoded into multi-dimensional profile vectors to generate target user profile data. Based on the target user profile data, the comprehensive similarity between the target user and each mature user in terms of industry, capacity, region, shift system, and weather sensitivity is calculated. The mature users with the highest comprehensive similarity ranking are selected, and the cluster data of similar mature users is output.

[0033] Step 3: Image migration curve generation: The historical load curves of 96 points for each mature user in the similar mature user cluster are divided by their contracted capacity to perform capacity normalization, generating capacity normalized curve data. Based on the day type, season, and weather conditions of the forecast date, the capacity normalized curve data is aligned to weekdays to generate time-aligned curve data. Based on comprehensive similarity, the time-aligned curve data is weighted and summed, and then de-normalized according to the contracted capacity of the target new user to generate profile migration load curve data.

[0034] Step 4: Basic Recursive Prediction and Candidate Curve Routing Using target user profile data, historical load data of target new users, weather data, and calendar data as input features, a pre-trained time-series forecasting model is invoked to output the base load curve from the first forecast day to the third forecast day, generating base forecast load curve data. Simultaneously, based on the historical load data of target new users, a candidate curve set is constructed, including yesterday's curve, the curve from the same day last week, and the median curve of the past seven days. Nighttime base load, early ramp-up speed, and recent peak-to-valley difference are extracted as proxy features of operational status to generate candidate curve routing result data.

[0035] Step 5: Dynamic Fusion and Closed-Loop Correction Based on the number of days of actual load data for target new users, the average similarity of target user profile data, the dispersion of similar mature user cluster data, and the recent error of the basic model, the quality scores of the profile migration channel, the basic prediction channel, and the candidate curve routing channel are calculated. The dynamic fusion weights of each prediction channel are calculated using a normalization function. The profile migration load curve data, the basic prediction load curve data, and the candidate curve routing results are then weighted and fused to generate the final fused load prediction curve data. The final fused load prediction curve data is compared with the actual load data, and the daily total error and the 96-point similarity error are calculated. Based on the error results, the load rate and peak-valley difference attributes in the target user profile data are updated. Users in the similar mature user cluster data that deviate significantly from the actual curve are removed. The input features and dynamic fusion weights of the basic recursive load prediction model are adjusted to generate corrected prediction model parameter data.

[0036] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A new user load forecasting method based on profile migration and progressive dynamic correction, characterized in that: The method includes the following steps: S1. Collect basic attribute data of target new users and historical load data of mature users. After quality access verification, generate basic attribute data of target new users and historical load data of mature users. S2. Based on the basic attribute data of the target new users, construct a power user load characteristic profile and generate target user profile data; S3. Based on the target user profile data and the mature user historical load data, retrieve mature users similar to the target user and output similar mature user cluster data; S4. Based on the similar mature user cluster data, perform capacity normalization, time alignment and similarity weighting on the historical load curves of mature users to generate profile migration load curve data. S5. Input the target user profile data and the historical load data of the target new users into the basic recursive load prediction model to generate basic predicted load curve data; S6. Based on the historical load data of the target new users, construct a set of candidate curves and combine them with the characteristics of the running status agent to generate candidate curve routing result data; S7. Based on the target user profile data, the similar mature user cluster data, the basic predicted load curve data, and the candidate curve routing result data, calculate the dynamic fusion weight of each prediction channel and generate the final fusion load prediction curve data. S8. Based on the comparison results between the final fusion load prediction curve data and the actual load data, perform backflow correction on the target user profile data, the similar mature user cluster data, the basic recursive load prediction model, and the dynamic fusion weight to generate corrected prediction model parameter data.

2. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 1, characterized in that: The steps involved in generating basic attribute data for new target users and historical load data for mature users in step S1 are as follows: S11. Obtain the static attributes, capacity attributes, industry attributes, regional attributes, production shift attributes, contract capacity, and application capacity data of the target new users through the data interface, and generate basic attribute data of the target new users; S12. Obtain mature user profiles, historical 96-point load curves, historical prediction errors and load stability data through the data interface, and generate historical load data for mature users. S13. Perform 96-point integrity check, null value check, duplicate time point check and unit caliber check on the 15-minute load data in the target new user basic attribute data and the mature user historical load data, and mark the data that passes the check as the target new user basic attribute data and the mature user historical load data.

3. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 2, characterized in that: The process of generating target user profile data in S2 includes the following steps: S21. Obtain the basic attribute data of the target new user; S22. Encode the industry category, contract capacity, installed capacity, region, production shift, whether it has its own photovoltaic system, peak-valley power structure, load factor and weather sensitivity in the basic attribute data of the target new user into a multi-dimensional profile vector, construct a power user load characteristic profile, and generate target user profile data.

4. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 3, characterized in that: The output of similar mature user cluster data in S3 includes the following steps: S31. Obtain the target user profile data and the historical load data of the mature user; S32. Based on the target user profile data, calculate the comprehensive similarity between the target user and each mature user in terms of industry, capacity, region, shift system and weather sensitivity, select the mature users with the highest comprehensive similarity ranking, and output the cluster data of similar mature users.

5. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 4, characterized in that: The steps involved in generating the portrait migration load curve data in S4 are as follows: S41. Obtain the data of the similar mature user clusters; S42. Divide the 96-point historical load curve of each mature user in the similar mature user cluster by its contract capacity, perform capacity normalization processing, and generate capacity normalized curve data. S43. Based on the day type, season, and weather conditions of the predicted day, perform weekday alignment on the capacity normalized curve data to generate time-aligned curve data. S44. Based on the comprehensive similarity, the time alignment curve data is weighted and summed, and then denormalized according to the contract capacity of the target new user to generate profile migration load curve data.

6. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 5, characterized in that: The process of generating the basic forecast load curve data in S5 includes the following steps: S51. Obtain the target user profile data and the historical load data of the target new user; S52. Using the target user profile data, the historical load data of the target new user, weather data, and calendar data as input features, call the pre-trained time series prediction model to output the basic load curve from the first prediction day to the third prediction day, and generate basic prediction load curve data.

7. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 6, characterized in that: The process of generating candidate curve routing result data in S6 includes the following steps: S61. Obtain historical load data of the target new user; S62. Based on the historical load data, construct a candidate curve set including yesterday's curve, the curve of the same day last week, and the median curve of the past seven days; S63. Extract the nighttime base load, early ramp speed and recent peak-valley difference of the target prediction day as operational status proxy features, assign weights to each curve in the candidate curve set according to the operational status proxy features, and generate candidate curve routing result data.

8. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 7, characterized in that: The process of generating the final fused load prediction curve data in S7 includes the following steps: S71. Obtain the target user profile data, the similar mature user cluster data, the basic predicted load curve data, and the candidate curve routing result data; S72. Calculate the quality scores of the profile migration channel, the basic prediction channel, and the candidate curve routing channel based on the number of days of the actual load data of the target new users, the average similarity of the target user profile data, the dispersion of the similar mature user cluster data, and the recent error of the basic model. S73. Based on the quality score, calculate the dynamic fusion weight of each prediction channel using a normalization function, and perform weighted fusion on the portrait migration load curve data, the basic prediction load curve data, and the candidate curve routing result data to generate the final fused load prediction curve data.

9. The method for predicting new user load based on profile migration and progressive dynamic correction according to claim 8, characterized in that: The step of generating the corrected prediction model parameter data in S8 includes the following steps: S81. Obtain the final converged load forecast curve data and the actual load data of the target new users; S82. Calculate the total daily error and the 96-point same-point error between the final fused load prediction curve data and the actual load data; S83. Based on the daily total error and the 96-point same-point error, update the load rate and peak-valley difference attributes in the target user profile data, remove users in the similar mature user cluster data that deviate greatly from the actual curve, and adjust the input features of the basic recursive load prediction model and the dynamic fusion weights to generate corrected prediction model parameter data.

10. A new user load forecasting system based on profile migration and progressive dynamic correction, used to implement the new user load forecasting method based on profile migration and progressive dynamic correction as described in any one of claims 1-9, characterized in that: The system includes: The data acquisition and quality access module collects basic attribute data of target new users and historical load data of mature users through the data interface. After 96-point integrity verification, null value verification, duplicate time point verification and unit caliber verification, it outputs basic attribute data of target new users and historical load data of mature users. The profile generation module receives the basic attribute data of the target new user, encodes industry, capacity, region, shift system, and weather sensitivity into profile vectors through the multi-dimensional profile vector encoding unit, and outputs the target user profile data. The similar user retrieval module receives the target user profile data and the mature user historical load data, selects the top-ranked mature users through the comprehensive similarity calculation unit, and outputs similar mature user cluster data. The profile migration curve generation module receives the similar mature user cluster data, processes the historical load curves of mature users through a capacity normalization unit, a time alignment unit, and a similarity weighting unit, and outputs profile migration load curve data. The basic recursive prediction module receives the target user profile data and the target new user historical load data, and outputs the basic load curves for the first to third prediction days through the time series prediction model to generate basic predicted load curve data. The candidate curve routing module receives historical load data of the target new user and outputs candidate curve routing result data through the candidate curve construction unit and the running status agent feature extraction unit. The dynamic fusion module receives the target user profile data, the similar mature user cluster data, the basic predicted load curve data, and the candidate curve routing result data, and outputs the final fused load prediction curve data through the quality score calculation unit and the normalization fusion unit. The backflow correction module receives the final fused load prediction curve data and the actual load data, and outputs the corrected prediction model parameter data through the error calculation unit and the parameter adjustment unit.

Images