An electric power load privacy joint prediction method

Abstract

The application belongs to the technical field of power load prediction and data privacy protection, and discloses a power load privacy joint prediction method, which is applied to a system composed of a server and multiple local subjects. The server generates group prototype statistical information, the local subject calculates the behavior recognizable degree, the portrait uniqueness and the reconfigurability risk, determines the segmented protection strategy, executes the same peak silence replacement, the feature mask, the statistical quantity clipping, the differential privacy disturbance and the encrypted uploading, the server safely aggregates the sufficient statistics and solves the global prediction model parameters, and the local subject fuses the global model and the local private model to output the load prediction result. The privacy joint modeling is carried out based on the multi-subject load data, the load prediction is realized, and the leakage risk of the load data is reduced.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting and data privacy protection technology, specifically to a joint forecasting method for power load privacy. Background Technology

[0002] In power distribution network operation, electricity sales management, demand response, charging load scheduling, and industrial park energy management, it is necessary to predict the future load of multiple electricity users. Existing centralized forecasting methods typically require users to upload hourly load data or high-frequency electricity consumption curves. This type of data may reflect sensitive information such as residents' work and rest schedules, equipment start-up and shutdown, charging behavior, and production rhythms, posing a risk of privacy leakage.

[0003] While existing federated load forecasting methods can avoid the centralized uploading of detailed load curves, isolated spikes can still contribute to the global model through model updates, sufficient statistics, or aggregated results when only a few entities experience charging starts, equipment start-ups / shutdowns, or sudden increases in production load within a certain time window. This allows servers or other participants to infer the entity's electricity consumption behavior. Existing unified differential privacy noise addition or simple homomorphic encryption methods cannot simultaneously address the risks of isolated spike identification and the problem of reduced prediction accuracy.

[0004] Therefore, there is an urgent need for a joint prediction method for electricity load privacy to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a privacy-predicting joint method for power load: based on multi-entity load data, privacy-predicting joint modeling is performed to achieve load forecasting while reducing the risk of load data leakage.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for joint prediction of electricity load privacy, applied to a system including a server and multiple local entities, includes the following steps:

[0008] Local entities calculate corresponding privacy risk indicators based on local load sequences and group prototype statistics of their respective load groups, and generate comprehensive privacy risk values ​​based on these indicators. The local load sequence consists of load samples from local entities.

[0009] Based on the comprehensive privacy risk value, the segmented protection strategy corresponding to the load sample of the local entity is determined, and the load sample of the local entity is subjected to the same peak silent replacement according to the segmented protection strategy to obtain the replacement load sample.

[0010] Based on the alternative load samples, a predicted feature vector is constructed, and the local sufficient statistics required for joint modeling are generated according to the predicted feature vector. The local sufficient statistics are pruned, and the privacy budget of each local subject is determined. Differential privacy perturbation is performed based on the privacy budget to obtain the traditional econometrics to be applied.

[0011] The traditional metrics to be uploaded are encrypted and uploaded to the server. The server aggregates the encrypted statistics uploaded by multiple local entities based on a secure aggregation protocol to obtain a globally aggregated sufficient statistic.

[0012] The server solves the parameters of the global joint prediction model based on the global aggregated sufficient statistics, and then distributes the global joint prediction model parameters to each local entity;

[0013] Each local entity performs local power load forecasting based on the parameters of the global joint forecasting model to obtain the power load forecast value for future time periods.

[0014] Furthermore, the local entity calculates the corresponding privacy risk indicators based on the local load sequence and the group prototype statistics of its load group, including:

[0015] Local subject Calculate the deviation factor of its load relative to the prototype of its load group locally;

[0016] Local entities are statistically analyzed based on deviation coefficients. The corresponding number of peak subjects was determined, and the significance coefficient of repeated behaviors was obtained.

[0017] The behavioral identifiability risk value is calculated based on the deviation coefficient, the number of peak subjects, and the significance coefficient of repeated behaviors. This behavioral identifiability risk value is recorded as a privacy risk indicator.

[0018] Furthermore, the privacy risk indicators calculated by local entities based on local load sequences and group prototype statistics of their respective load groups also include:

[0019] Local subject A normalized load profile is generated based on historical load curves, and local main load data is obtained. The image uniqueness risk value is calculated based on the normalized load image, the image prototype mean, and the image dispersion of the load group. The image uniqueness risk value is recorded as the privacy risk indicator.

[0020] Furthermore, the calculation of the corresponding privacy risk index by the local entity based on the local load sequence and the group prototype statistical information of the load group to which it belongs also includes calculating the reconfigurability risk value: the local entity trains a local private model using its own historical load data, collects the private model information corresponding to the local entity, the private model information includes the goodness of fit of the entity's local private model to its own load, the mean absolute error of the entity's local private model and the mean absolute error of the benchmark model; the reconfigurability risk value is calculated based on the private model information, and the reconfigurability risk value is recorded as the privacy risk index.

[0021] Furthermore, according to the segmented protection strategy, the load samples of the local main body are subjected to peak-shifting silent replacement, resulting in replacement load samples including:

[0022] The local entity determines the deviation coefficient corresponding to the current load sample, the number of peak entities under the same time type in the same load group, and the effective co-peak safety threshold corresponding to the current entity based on the segmented protection strategy, and calculates the co-peak participation weight;

[0023] When the participation weight of the same peak is 0, the current load sample is determined to be an isolated spike, and the actual load value of the current load sample does not participate in the joint modeling;

[0024] When the weight of the same peak is greater than 0 and less than 1, the current load sample is determined to participate in joint modeling with reduced weight;

[0025] When the participation weight of the same peak is 1, it is determined that the current load sample is participating in joint modeling normally;

[0026] Based on the participation weights of the same peak, the actual load value of the current load sample, and the local baseline load of the local subject in the corresponding time type, an alternative load sample is constructed.

[0027] Furthermore, constructing a predicted feature vector based on the alternative load sample, and generating the local sufficient statistics required for joint modeling based on the predicted feature vector, includes:

[0028] Construct a single-time-time predicted feature vector; arrange the predicted feature vectors from multiple times vertically in chronological order to form a feature matrix;

[0029] A random mask matrix is ​​generated based on the feature mask ratio in the segmented protection strategy, and the feature matrix is ​​then subjected to feature masking, random occlusion, or dimensional compression using the random mask matrix to obtain the masked feature matrix.

[0030] Based on the statistics in the segmented protection strategy, the attenuation coefficient and the original local sample weight matrix are used to generate the effective sample weight matrix.

[0031] Based on the masked feature matrix, the effective sample weight matrix, and the predicted target vector trained by the local model, calculate the local second-order sufficient statistics matrix, the local first-order sufficient statistics vector, and the local predicted target sum of squares statistics.

[0032] Furthermore, determining the privacy budget for each local entity includes:

[0033] The server obtains the historical modeling contribution and comprehensive privacy risk value of each local entity. When there is no historical modeling contribution in the first round of training, the historical modeling contribution of each participating local entity is initialized to an equal value.

[0034] The server dynamically determines the differential privacy effective budget finally allocated to the local subject based on historical modeling contribution, comprehensive privacy risk value, privacy budget corresponding to privacy risk level, system-wide minimum privacy budget, system-wide maximum privacy budget, numerical smoothing to prevent zero, contribution adjustment index, and privacy risk adjustment index.

[0035] Furthermore, the server solves for the parameters of the global joint prediction model based on the global aggregate sufficient statistics, including:

[0036] The server uses ridge regression regularization to solve for the parameters of the global joint prediction model based on the global aggregate second-order statistic, the global aggregate first-order statistic, and the model regularization penalty coefficient.

[0037] When the matrix corresponding to the global aggregate second-order statistics is singular and cannot be directly inverted, the parameters of the global joint prediction model are solved by the Choleski decomposition method, the conjugate gradient iteration method, or the generalized pseudo-inverse algorithm.

[0038] Furthermore, each local entity performs local power load forecasting based on the parameters of the global joint forecasting model to obtain the power load forecast value for future time periods. This includes: the local entity constructs the time-series forecast features for the future time period to be predicted locally and obtains the global forecast result using the parameters of the global joint forecasting model; the local entity obtains the local forecast result based on its local private model and merges the global forecast result and the local forecast result according to the preset fusion weight to obtain the final power load forecast value for future time periods.

[0039] Compared to existing solutions, the beneficial effects achieved by this invention are:

[0040] This invention identifies privacy risks from multiple dimensions, such as peak load exposure, individual profile differences, and load pattern reconfigurability, through a joint assessment of group prototype statistics, behavioral identifiability risk, profile uniqueness risk, and reconfigurability risk. It also dynamically determines segmented protection strategies based on comprehensive privacy risks, enabling local entities with different risk levels to receive differentiated protection and avoiding insufficient protection or loss of predictive performance caused by using uniform protection strength.

[0041] This invention employs a peak-shifting substitution mechanism to prevent isolated peak loads from entering the joint modeling process in their true form. Peak loads with a certain degree of group participation are weighted and reduced. Combined with feature masking, statistical participation attenuation, sufficient statistical pruning, and differential privacy perturbation, this invention effectively reduces the risk of user electricity consumption behavior, equipment start-up and shutdown habits, and abnormal load patterns being identified or reconstructed, thereby improving the security of power load data in the joint modeling process.

[0042] This invention employs end-to-end encrypted upload and secure aggregation methods, ensuring that the server obtains only sufficient global aggregate statistics without acquiring detailed load curves or individual independent statistics for any local entity. Simultaneously, it solves the global joint prediction model based on the aggregate statistics, and local entities integrate their local private models for personalized prediction. This approach balances cross-entity data collaborative modeling capabilities with local load characteristic adaptation capabilities, which helps improve the accuracy, stability, and practicality of power load forecasting. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0044] Figure 1 This is a flowchart of the first power load privacy joint prediction method according to an embodiment of the present invention;

[0045] Figure 2 This is a flowchart of the second power load privacy joint prediction method according to an embodiment of the present invention;

[0046] Figure 3 This is a flowchart of the third power load privacy joint prediction method according to an embodiment of the present invention. Detailed Implementation

[0047] 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.

[0048] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more exemplary embodiments. Numerous specific details are provided in the following description to give a full understanding of exemplary embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, steps, etc., can be employed. In other instances, well-known structures, methods, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0049] This invention provides a method for privacy-preserving joint prediction of electricity load, applicable to a system comprising a server and multiple local entities, including residential users, industrial and commercial users, industrial parks, charging stations, energy storage nodes, load aggregators, or distribution operation nodes. Before implementing this method, each local entity has been assigned to a corresponding load group based on registration information, data integrity, and irreversible load profile summaries. The server only maintains group-level statistics and does not store detailed load curves for any single entity.

[0050] Figure 1 This is a flowchart of the first power load privacy joint prediction method according to an embodiment of the present invention, as follows: Figure 1 As shown, the method includes the following steps:

[0051] S1: Obtain group prototype statistics

[0052] Based on historical security aggregation results, the server generates and distributes group prototype statistics for each load group under different time types, including: , and ; Indicates a load group; Indicates the time type, which includes at least one of the following: weekday morning rush hour, weekday evening rush hour, weekend, holiday, night, and midday. Indicates group In time type The average load under the current conditions; Indicates group In time type Load dispersion is used to characterize the degree of dispersion of the load of each subject within the same load group relative to the group's prototype load under the same time type. A smaller load dispersion indicates a more concentrated load pattern within the group; a larger load dispersion indicates more significant differences in load among the subjects within the group. Load dispersion is obtained by securely aggregating the local load statistics of multiple subjects; the server cannot obtain detailed load curves for any single subject. The co-peak safety threshold refers to the minimum number of simultaneous peak subjects required for a particular peak load sample to participate in joint modeling in a realistic form within the same load group, the same time type, or the same time window.

[0053] S2: Local entities calculate corresponding privacy risk indicators based on local load sequences and group prototype statistical information of their respective load groups, and generate comprehensive privacy risk values ​​based on these indicators; among which, privacy risk indicators include behavioral identifiability risk value, profile uniqueness risk value, and reconfigurability risk value;

[0054] Figure 2 This is a flowchart of the second power load privacy joint prediction method according to an embodiment of the present invention, as follows: Figure 2 As shown, the local entity calculates the corresponding privacy risk index based on the local load sequence and the group prototype statistics of the load group to which it belongs. The specific process of generating a comprehensive privacy risk value based on the privacy risk index is as follows:

[0055] S21: Risk value for the identifiability of local entity's calculated behavior;

[0056] For the Local entities Its local load sequence is denoted as:

[0057] in, This represents the number of sampling points within the local training cycle.

[0058] Local subject Calculate the deviation factor of its load relative to the prototype of its load group locally:

[0059]

[0060] in: For local entities At any moment The actual local load; For local entities Belongs to the load group; For a moment Corresponding time type; For local entities Belonging to load group At any moment Corresponding time type The average load under the current conditions; For local entities Belonging to load group At any moment Corresponding time type Load dispersion under the following conditions; To prevent extremely small positive numbers with a denominator of zero.

[0061] when At that time, local entities This moment is marked as the peak moment, where, This is the threshold for determining the peak.

[0062] Each local entity submits spike markers only through secure aggregation, enabling the server to obtain the number of spike entities within the same group and time type. The server cannot determine which specific local entity experienced the spike. Secure aggregation methods include submission via additive homomorphic encryption or secure multi-party computation protocols.

[0063] Before calculating the behavioral identifiability risk value, first... Mapped to the interval using the min-max normalization method. superior.

[0064] Local subject The behavioral identifiability risk value is calculated based on the deviation coefficient, the number of peak subjects, and the significance coefficient of repeated behaviors:

[0065]

[0066] in: Risk value for behavioral identifiability; This is an indicator function; it takes the value 1 if the condition is met and 0 if the condition is not met. The number of main groups that show a peak at the same time; This refers to the peak safety threshold for the corresponding time period; The significance coefficient for recurring behavior can be determined locally by the entity based on the degree of load repetition in similar historical scenarios. A higher degree of repetition indicates that the load behavior is more easily identified and carries a higher risk. The significance coefficient characterizes whether the same or similar load behaviors repeatedly occur in similar scenarios for the local entity. This coefficient is calculated locally by the entity and does not require uploading detailed load data to the server. For example, the local entity can select historical samples that match the current time locally. Matching conditions include one or more of the following: same weekday, same time period, similar weather conditions, and similar electricity price conditions. For instance, if the current sample is a load sample under the conditions of "Friday evening 20:00 to 21:00, low temperature weather, peak electricity price," the local entity can search locally for load segments that meet the same or similar conditions in the past. The local entity then determines whether the current load behavior has repeatedly occurred in similar historical scenarios. Load behaviors can include peak occurrences, charging load startup, concentrated equipment operation, rapid load increase, rapid load decrease, or continuous stable high load operation. The significance coefficient for recurring behavior can be determined using a tiered assignment method. Example values ​​are as follows: The current load behavior has basically not occurred in similar historical scenarios. The value is 0.05; the current load behavior has occurred several times, but the timing or amplitude is unstable. The value is 0.40; the current load behavior occurs frequently and is generally concentrated in the same time period. The value is 0.60; the current load behavior occurs frequently, and the load amplitude and duration are relatively stable. It is 0.80; , , , is a risk weight parameter, set by the server, and . This indicates that the result of the operation will be forcibly truncated and constrained to an interval. Operations within the specified range.

[0067] S22: Calculate the uniqueness risk value of the profile of the local entity;

[0068] Local subject A normalized load profile is generated locally based on historical load curves. The normalized load profile includes at least one of the following: average daily load, peak-to-valley difference, peak rate, ramp rate, load factor, and intraday cyclical characteristics. When constructing a normalized load profile locally, the local entity first processes historical load curves by handling missing values, correcting outliers, and aligning them with time. Then, the load data is divided according to day, weekday, non-weekday, holiday, or peak-valley periods. Subsequently, features such as average daily load, peak-to-valley difference, peak frequency, load ramp-up or ramp-down magnitude, load factor, and intraday load patterns are extracted. These features are then normalized according to preset ranges to ensure that profiles of different capacities and types of entities are on a uniform scale. Finally, the normalized features are combined in a fixed order to form the normalized load profile of the local entity. .

[0069] Local entities calculate the uniqueness risk value of a profile using the following formula:

[0070]

[0071] in: The risk value for the uniqueness of the portrait; For local entities Local load profile; For groups The average of the original portraits; For groups Internal image dispersion can be determined by the average distance, median distance, standard deviation, or quantile distance between each local subject image within the group and the group image prototype. It is a L2 norm. The greater the difference between the local subject profile and the group profile, the easier it is to identify the local subject, and the higher the risk of profile uniqueness.

[0072] S23: Local entity calculates reconfigurability risk value:

[0073]

[0074] in: This represents the reconfigurability risk value. This represents the goodness of fit of the local private model of the local entity to its own load. This refers to a private prediction model trained locally by the entity using its own historical load data. This model's ability to explain or fit the historical load variation patterns of that local entity is crucial. The local private model can be an autoregressive model, sliding window regression model, gradient boosting tree model, random forest model, neural network model, long short-term memory network model, or other local load prediction models capable of outputting validation error and goodness-of-fit. Both training and validation are completed locally by the entity; the server does not obtain model training samples, detailed load curves, or single-time-point prediction errors. The mean absolute error of the local private model is It refers to the average absolute error between the predicted load value and the actual load value after the local private prediction model is used to predict the load of the main body during the verification period. It is used to measure the prediction error level of the local private model for the load of the main body. The mean absolute error of the baseline model refers to the local subject Within the same verification period, after predicting the load using a preset benchmark model, the average absolute error between the predicted load value and the actual load value is taken. The benchmark model can be a model based on the same time yesterday, a model based on the average value of the same period in the same week, a moving average model, a seasonal average model, or other simple prediction models. When the local load is easily fitted with a local model with high accuracy, it indicates that a small amount of statistical information may reconstruct its load change pattern, thus the risk of reconstructibility is high.

[0075] S24: Local entities generate a comprehensive privacy risk value based on the behavioral identifiability risk value, the profile uniqueness risk value, and the reconfigurability risk value.

[0076]

[0077] in: A comprehensive privacy risk score; , and These are the weight parameters corresponding to the three types of risk values, and .

[0078] S3: Server Select a combination of segmented protection strategies based on the overall privacy risk score:

[0079]

[0080] in: The statistic participates in the attenuation coefficient, used to represent the local subject. The degree of effective participation of the statistic in joint modeling, group statistics, or parameter aggregation processes; the coefficient can range from 0 to 1. The feature mask ratio represents the proportion of the total number of features in the local subject load profile or local statistical features that are hidden, replaced, disturbed, or not uploaded. Privacy budget value; The effective peak load security threshold for the current local entity refers to the threshold used to determine whether peak load meets the conditions for safe participation, after adjustments based on the basic peak load security threshold, combined with the current group size, profile dispersion, significance of repeated behavior, load peak sensitivity, or privacy risk level.

[0081] Specifically,

[0082] The threshold for determining low risk; This is the threshold for determining high risk. , The statistics for medium and high risk levels are respectively included in the attenuation coefficient, satisfying the following conditions: ; , These represent the feature mask ratios for medium and high risk levels, respectively, satisfying... ; This represents a privacy budget value under low-risk conditions. This represents a privacy budget value under high-risk conditions. This represents the privacy budget value under medium risk. The same peak safety threshold is fixed in increment.

[0083] Therefore, low-risk local entities can participate more fully in joint modeling; medium-risk local entities are downweighted and have some features masked; and high-risk local entities are subject to stronger co-peak constraints, smaller privacy budget caps, and higher proportions of feature masking protection.

[0084] S4: Perform peak-shift silent replacement

[0085] Figure 3 This is a flowchart of the third power load privacy joint prediction method according to an embodiment of the present invention, as follows: Figure 3 As shown, performing a silent replacement of peaks includes the following steps:

[0086] S41: Local entities determine the participation weight for peak periods based on the deviation coefficient and the number of participants in the same peak period;

[0087] For local entities At any moment The formula for calculating the participation weights of co-peak load samples is as follows:

[0088]

[0089] in: Weighting for participation at the same peak; This is the load group deviation coefficient; This refers to the number of peak entities within the same group during the same period; This represents the effective peak security threshold corresponding to the current local entity.

[0090] when When this occurs, it indicates that the peak is an isolated peak, and the actual load sample does not participate in the joint modeling; when This indicates that the peak has a certain degree of group activity, but participation still needs to be reduced in weight; when At that time, the samples were normal and unconstrained in the modeling process.

[0091] S42: Local entities construct alternative load samples based on the participation weights of co-peakers.

[0092]

[0093] This is the equivalent load sample after the silent replacement is completed; For local entities In the corresponding time type The local baseline load can be calculated using the historical load average under similar weather conditions, similar date types, and the same time period as... .

[0094] This step prevents isolated spikes from being included in subsequent statistical calculations in their original form, thereby reducing the risk of individual behavior being identified.

[0095] S5: Constructing the predictive eigenvectors and sufficient statistics

[0096] Local entities based on replaced load samples Construct the predicted feature vector:

[0097]

[0098] in: Predict the feature vector at a single time step; External environment variables The basis function mapping result refers to the mapping of time... Corresponding external influence variables The process of converting external variables into a set of characteristic function values ​​that can be used for model calculations. External variables may include temperature, humidity, weather type, date type, weekday / holiday identifier, time period number, electricity price, solar intensity, regional load status, or other variables that affect the main load changes but are not part of the main private load curve itself; For length is The average replacement load value corresponding to the sliding window refers to the first local entity Based on the privacy risk status of the real payload samples, construct corresponding alternative payload sample sequences locally; then, according to a preset window length... The window is moved time-by-time across the alternative load sample sequence; for each target time... Take the continuous values ​​before that time. Each replacement load value is calculated, and their average value is obtained to obtain the average replacement load value. This represents the time-series change in the replacement load at the previous moment; This refers to the period length, such as the number of sampling points per day.

[0099] The predicted features from multiple time points are arranged vertically to form a feature matrix: ;in, Represents the transpose of a vector;

[0100] The local entity is protected according to the feature mask ratio in the segmentation protection strategy. Generate a random mask matrix:

[0101]

[0102] in, This is a feature masking operator used to perform zeroing, random occlusion, or dimensionality compression on a specified proportion of predicted features. The masking proportion is determined by the subject's risk level.

[0103] Local entities further participate in the attenuation coefficient based on statistical measures. Solve for the effective sample weight matrix:

[0104]

[0105] in, This is the original local sample weight matrix, for each local subject. It corresponds to a pre-set original local sample weight matrix.

[0106] The local entity solves the necessary statistics for joint modeling of the three types offline:

[0107]

[0108]

[0109]

[0110] in: It is a local second-order sufficient statistics matrix; The vector of local first-order sufficient statistics; For local forecast target sum of squares statistic; The predicted target vector trained for the local model represents the future. The predicted load value corresponding to each step.

[0111] This step eliminates the need for the entity to upload hourly raw load data, instead generating lightweight statistics locally that can be used for iterative solutions to the global model.

[0112] S6: Fully Statistical Pruning

[0113] To limit the interference weight of anomalous data from a single subject on the global model, the local subject applies amplitude pruning constraints to the local sufficient statistics:

[0114]

[0115]

[0116] in: This is a second-order sufficient statistic after the clipping is completed; This is a first-order sufficient statistic after clipping; The threshold for pruning the maximum amplitude of the second-order statistic; The threshold for pruning the maximum amplitude of the first-order statistic; It is the Frobenius norm; It is a 2-norm.

[0117] S7: Allocate the privacy budget based on contribution and overall privacy risk.

[0118] The server dynamically allocates differential privacy available budgets to each participating local entity based on the local entity's historical modeling contribution and overall privacy risks.

[0119] Local subject Historical modeling contribution is recorded as , This can be obtained from the previous round of training; in the first round of training without historical data, the contribution of all participating local entities is uniformly initialized to an equal value.

[0120] Furthermore, allocate the available budget for differential privacy:

[0121]

[0122] in: For local entities The final allocated differential privacy effective budget; Privacy budgets are determined by the privacy risk levels of local entities. Set a minimum privacy budget for the entire system; Set a maximum privacy budget for the entire system; The comprehensive privacy risk score for local entities; For numerical smoothing and zero-prevention term; , These are the contribution adjustment index and the privacy risk adjustment index, respectively. In one implementation, the contribution adjustment index... The value ranges from 0.3 to 0.8, and the privacy risk adjustment index is... The value range is from 1.0 to 1.8. For example, 0.6 is acceptable. Option 1.2 is acceptable. The higher the contribution and the lower the privacy risk, the larger the privacy budget can be allocated to the local entity, and the amplitude of subsequent added disturbance noise will be smaller; the higher the privacy risk, the entity is strictly constrained by the budget cap, and the privacy noise protection efforts are forced to be increased.

[0123] S8: Apply differential privacy perturbation to sufficient statistics

[0124] After integrating and trimming the local subjects, all statistics are combined into a statistical set: ;

[0125] Based on the allocated privacy budget Add standard Gaussian privacy noise to the set of statistics :

[0126]

[0127] The standard deviation of Gaussian noise Complete solution formula:

[0128]

[0129] in: This is to be applied to traditional metrology after differential privacy perturbation; This serves as the upper bound for global sensitivity to local sufficient statistics. These are the parameters for differential privacy relaxation; It is an identity matrix of the same dimension.

[0130] S9: Encrypt uploads and perform secure aggregations

[0131] The local entity performs end-to-end encryption on the statistics after privacy perturbation:

[0132]

[0133] in, For encryption functions, this can include Paillier encryption functions;

[0134] Encrypt ciphertext Uploaded to the server.

[0135] The server relies on an additive homomorphic encryption protocol and a secure multi-party computation framework to perform secure cross-subject ciphertext aggregation operations:

[0136]

[0137] in: It is a global aggregate second-order statistic; This is a global aggregation of first-order statistics; This is the global aggregate target sum of squares statistic; This represents the total number of valid electricity users participating in the current round of joint training. This is the decryption function;

[0138] The server can only obtain the aggregated statistics throughout the entire process, and cannot reverse-analyze the original detailed load or individual sufficient statistics of any participating entity, thus reducing the risk of leakage of original data during transmission and aggregation.

[0139] S10: Solving the joint prediction model

[0140] The server relies on globally aggregated sufficient statistics and employs ridge regression regularization to solve for the optimal global joint prediction model parameters:

[0141]

[0142] in: To train the parameters of the globally optimal joint prediction model after convergence; This is the model regularization penalty coefficient; It is an identity matrix of the same dimension.

[0143] When matrix When there are singularities that cannot be directly inverted, the Cholesky decomposition method, conjugate gradient iteration method, or generalized pseudo-inverse algorithm can be used to solve for the parameters.

[0144] After the model training is complete, the server will assign globally optimal model parameters. Distribute to all relevant entities.

[0145] S11: Local entities perform personalized load forecasting

[0146] Each local entity constructs time-series prediction features for future time periods offline. And use the global model to obtain the global prediction results:

[0147]

[0148] The local entity can also retain its local private model to obtain local prediction results: ;

[0149] By integrating global general trends with local personalized features, the final load forecast value is obtained:

[0150]

[0151] in: The system will ultimately output the predicted load values ​​for future time periods for the local entity. To predict the step size; The weighting coefficients are used to fuse the global model results. The value ranges from 0.20 to 0.85. For example, when the local entity has limited local historical data or the local private model has a high prediction error, set... The value is set to 0.75 to 0.85 to increase the proportion of predictions from the global model; when the local subject is in a normal risk state and the errors of the local model and the global model are close, the value is set to... The value is set to 0.55; when the uniqueness risk of the local subject profile is high, the load behavior has obvious individual characteristics, and the prediction error of the local private model is lower than that of the global model, the value is set to 0.55. The value is set to 0.25 to 0.35 to increase the proportion of predictions from the local private model.

[0152] In one implementation, before issuing group prototype statistics for any load group, the server first determines whether the number of valid subjects in the corresponding time type of the load group is not less than a preset minimum group size Nmin. If the number of valid subjects is less than Nmin, the server merges the load group with the nearest neighboring group in the load profile, or does not issue group prototype statistics for that time type. The group prototype statistics are obtained by securely aggregating the statistical values ​​calculated locally by each subject. The server only obtains the aggregated mean, median, dispersion, or quantile statistics and cannot obtain the individual statistical values ​​of any subject.

[0153] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0154] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0155] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0156] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0157] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0158] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for joint prediction of electricity load privacy, characterized in that, When applied to a system that includes a server and multiple local entities, the method includes the following steps: Local entities calculate corresponding privacy risk indicators based on local load sequences and group prototype statistics of their respective load groups, and generate comprehensive privacy risk values ​​based on these indicators. The local load sequence consists of load samples from local entities. Based on the comprehensive privacy risk value, the segmented protection strategy corresponding to the local entity is determined, and the load sample of the local entity is subjected to the same peak silent replacement according to the segmented protection strategy to obtain the replacement load sample; Based on the alternative load samples, a predicted feature vector is constructed, and the local sufficient statistics required for joint modeling are generated according to the predicted feature vector. The local sufficient statistics are pruned, and the privacy budget of each local subject is determined. Differential privacy perturbation is performed based on the privacy budget to obtain the traditional econometrics to be applied. The traditional metrics to be uploaded are encrypted and uploaded to the server. The server aggregates the encrypted statistics uploaded by multiple local entities based on a secure aggregation protocol to obtain a globally aggregated sufficient statistic. The server solves the parameters of the global joint prediction model based on the global aggregated sufficient statistics, and then distributes the global joint prediction model parameters to each local entity; Each local entity performs local power load forecasting based on the parameters of the global joint forecasting model to obtain the power load forecast value for future time periods.

2. The electricity load privacy joint prediction method according to claim 1, characterized in that, The privacy risk indicators calculated by local entities based on local load sequences and group prototype statistics of their respective load groups include: Local subject Calculate the deviation factor of its load relative to the group prototype of its load group locally; Local entities are statistically analyzed based on deviation coefficients. The corresponding number of peak subjects was determined, and the significance coefficient of repeated behaviors was obtained. The behavioral identifiability risk value is calculated based on the deviation coefficient, the number of peak subjects, and the significance coefficient of repeated behaviors. This behavioral identifiability risk value is recorded as a privacy risk indicator.

3. The electricity load privacy joint prediction method according to claim 1, characterized in that, The privacy risk indicators calculated by local entities based on local load sequences and group prototype statistics of their respective load groups also include: Local subject A normalized load profile is generated based on historical load curves, and local main load data is obtained. The image uniqueness risk value is calculated based on the normalized load image, the image prototype mean, and the image dispersion of the load group. The image uniqueness risk value is recorded as the privacy risk indicator.

4. The electricity load privacy joint prediction method according to claim 1, characterized in that, The local entity calculates the corresponding privacy risk index based on the local load sequence and the group prototype statistics of the load group to which it belongs. This also includes calculating the reconfigurability risk value. Local entities train their own private models using their own historical load data, and collect information about the private models corresponding to the local entities. The private model information includes the goodness of fit of the entity's local private model to its own load, the mean absolute error of the entity's local private model and the mean absolute error of the benchmark model. Based on the private model information, a reconfigurability risk value is calculated and recorded as a privacy risk indicator. Substituting the private model information corresponding to the local entity into the first correlation formula, the reconfigurability risk value is calculated. The private model information corresponding to the local entity includes the goodness of fit of the entity's local private model to its own load, the mean absolute error of the entity's local private model, and the mean absolute error of the baseline model.

5. The electricity load privacy joint prediction method according to claim 1, characterized in that, According to the segmented protection strategy, the load samples of the local main body are subjected to same-peak silent replacement, and the replacement load samples include: The local entity determines the deviation coefficient corresponding to the current load sample, the number of peak entities under the same time type in the same load group, and the effective co-peak safety threshold corresponding to the current entity based on the segmented protection strategy, and calculates the co-peak participation weight; When the participation weight of the same peak is 0, the current load sample is determined to be an isolated spike, and the actual load value of the current load sample does not participate in the joint modeling; When the weight of the same peak is greater than 0 and less than 1, the current load sample is determined to participate in joint modeling with reduced weight; When the participation weight of the same peak is 1, it is determined that the current load sample is participating in joint modeling normally; Based on the participation weights of the same peak, the actual load value of the current load sample, and the local baseline load of the local subject in the corresponding time type, an alternative load sample is constructed.

6. The method for joint prediction of electricity load privacy according to claim 1, characterized in that, Constructing a predicted feature vector based on the alternative load sample, and generating the local sufficient statistics required for joint modeling based on the predicted feature vector, includes: Construct a single-time-time predicted feature vector; arrange the predicted feature vectors from multiple times vertically in chronological order to form a feature matrix; A random mask matrix is ​​generated based on the feature mask ratio in the segmented protection strategy, and the feature matrix is ​​then subjected to feature masking, random occlusion, or dimensional compression using the random mask matrix to obtain the masked feature matrix. Based on the statistics in the segmented protection strategy, the attenuation coefficient and the original local sample weight matrix are used to generate the effective sample weight matrix. Based on the masked feature matrix, the effective sample weight matrix, and the predicted target vector trained by the local model, calculate the local second-order sufficient statistics matrix, the local first-order sufficient statistics vector, and the local predicted target sum of squares statistics.

7. The method for joint prediction of electricity load privacy according to claim 1, characterized in that, Determining the privacy budget for each local entity includes: The server obtains the historical modeling contribution and comprehensive privacy risk value of each local entity. When there is no historical modeling contribution in the first round of training, the historical modeling contribution of each participating local entity is initialized to an equal value. The server dynamically determines the differential privacy effective budget finally allocated to the local subject based on historical modeling contribution, comprehensive privacy risk value, privacy budget corresponding to privacy risk level, system-wide minimum privacy budget, system-wide maximum privacy budget, numerical smoothing to prevent zero, contribution adjustment index, and privacy risk adjustment index.

8. The method for joint prediction of electricity load privacy according to claim 1, characterized in that, The server solves for the parameters of the global joint prediction model based on global aggregate sufficient statistics, including: The server uses ridge regression regularization to solve for the parameters of the global joint prediction model based on the global aggregate second-order statistic, the global aggregate first-order statistic, and the model regularization penalty coefficient. When the matrix corresponding to the global aggregate second-order statistics is singular and cannot be directly inverted, the parameters of the global joint prediction model are solved by the Choleski decomposition method, the conjugate gradient iteration method, or the generalized pseudo-inverse algorithm.

9. The method for joint prediction of electricity load privacy according to claim 1, characterized in that, Each local entity performs local power load forecasting based on the parameters of the global joint forecasting model, and obtains the power load forecast value for future time periods. This includes: the local entity constructs the time-series forecast features for the future time period to be predicted locally, and obtains the global forecast result using the parameters of the global joint forecasting model; the local entity obtains the local forecast result based on its local private model, and merges the global forecast result and the local forecast result according to the preset fusion weight to obtain the final power load forecast value for future time periods.

Images