A multi-source energy data fusion system and method based on differential privacy

Abstract

The application discloses a multi-source energy data fusion system and method based on differential privacy, comprising an edge module, a fusion center module and a privacy budget management module. The edge module filters complete batches according to the fusion instruction, cleanses data by using the time sequence correlation in batches, and uploads after adding correlation maintaining noise proportional to the covariance. The fusion center module issues the fusion instruction, receives and verifies the validity of the data packet uploaded by each data source device; maps the timestamp of each data source to a unified time bucket index based on structured data, obtains unified time axis data after cutting, performs boundary missing inference, multi-granularity conversion and time sequence fingerprint verification, obtains boundary inference values, granularity constraint values and relationship confidence matrices, merges to generate multi-dimensional fusion records and adds noise to obtain a fusion result sequence and a privacy audit log; and the privacy budget management module dynamically allocates the privacy budget.

Description

Technical Field

[0001] This invention belongs to the field of data fusion technology, and in particular relates to a multi-source energy data fusion system and method based on differential privacy. Background Technology

[0002] Multi-source energy data fusion is a fundamental technology for applications in the energy sector, such as smart grids, virtual power plants, and carbon emission accounting. Existing fusion methods mainly employ ETL architectures, streaming processing frameworks, or data middleware models, while privacy protection relies on traditional methods such as data anonymization, access control, and encrypted transmission.

[0003] However, existing differential privacy methods generally assume that data arrives in real-time, one record at a time. In real-world power energy scenarios, devices such as smart meters and photovoltaic inverters typically operate with local caching and batch uploads. This disconnect between theoretical models and industrial practice leads to extremely low privacy budget utilization. The independent and identically distributed noise added by traditional differential privacy methods disrupts the temporal correlation of energy data, causing the loss of key features such as the smoothness and periodicity of load curves, affecting the accuracy of downstream forecasting and scheduling applications. Furthermore, the cleaning and alignment processes themselves may leak privacy information, lacking effective protection. In industrial practice, issues such as inconsistent batch windows, different time resolutions, and asynchronous device clocks among different data source devices make accurate alignment of multi-source data difficult. Therefore, existing technologies have shortcomings. Summary of the Invention

[0004] To address the problems of existing technologies, this invention proposes a multi-source energy data fusion system and method based on differential privacy. Through a cloud-edge collaborative framework, the edge collects and cleans data in batches, adds correlations to maintain noise, and then uploads the data. The cloud achieves unified timeline alignment and correlation analysis of multi-source data through time bucket alignment, missing data inference, granularity conversion, and fingerprint verification. After generating multi-dimensional fusion records, the data is published with added noise. This system features adaptability to industrial batch collection modes, high privacy budget utilization, and complete preservation of temporal correlations.

[0005] The technical solution of the present invention is as follows:

[0006] A multi-source energy data fusion system based on differential privacy includes an edge module, a fusion center module, and a privacy budget management module;

[0007] The edge module is used to filter out complete batches according to the fusion instructions, clean each batch of data in the complete batch using time-series correlation, add correlation-preserving noise proportional to the time-series covariance of the batch data, generate noisy batch data, and encapsulate and upload one or more noisy batch data corresponding to the complete batch.

[0008] The fusion center module includes a fusion instruction unit, an alignment unit, a fusion preparation unit, a multi-dimensional fusion unit, and a differential privacy output unit;

[0009] The fusion instruction unit is used to issue unified fusion instructions, receive and verify the validity of data packets uploaded by each data source device, and obtain structured data.

[0010] Alignment unit, used to map timestamps from each data source to a unified time bucket index based on the structured data, to obtain clipped unified timeline data;

[0011] The fusion preparation unit is used to perform boundary missing inference, multi-granularity transformation and temporal fingerprint verification on the clipped unified timeline data to obtain boundary inference values, granularity constraint values ​​and relation confidence matrix.

[0012] The multidimensional fusion unit is used to merge the clipped unified timeline data, boundary inference values, granular constraint values ​​and relation confidence matrix to generate a multidimensional fusion record in which each time bucket contains multi-source information.

[0013] The differential privacy output unit is used to add noise to the multidimensional fusion record to obtain a fusion result sequence and a privacy audit log.

[0014] The privacy budget management module is used to dynamically allocate the privacy budget based on data quality.

[0015] Further, the step of adding correlation-preserving noise proportional to the temporal covariance of the batch data to generate noisy batch data includes:

[0016] Request the allocation of the first privacy budget from the privacy budget management module;

[0017] Based on the first privacy budget and the temporal covariance matrix of the cleaned batch data, correlation-preserving noise is generated to obtain multidimensional Gaussian noise.

[0018] The multidimensional Gaussian noise is added to the cleaned batch data one by one to generate noisy batch data.

[0019] Furthermore, the step of mapping timestamps from various data sources to a unified time bucket index based on the structured data to obtain pruned unified timeline data includes:

[0020] Based on the structured data, the timestamp range and the time series associated with the value of M are standardized to obtain the timestamp corresponding to each noisy data value, where M is the batch data length;

[0021] Map each noisy data value, device identifier, and corresponding timestamp to the corresponding time bucket;

[0022] Within each time bucket, a representative point is selected for each device to generate unified timeline data;

[0023] The unified timeline data is cropped according to the fusion time range to obtain cropped unified timeline data.

[0024] Furthermore, the step of mapping timestamps from various data sources to a unified time bucket index based on the structured data to obtain pruned unified timeline data also includes:

[0025] For each time bucket, perform boundary region processing on the data points, and store the data points corresponding to the timestamps of the boundary regions into the adjacent time buckets.

[0026] Furthermore, the step of performing boundary missing inference, multi-granularity transformation, and temporal fingerprint verification on the cropped unified timeline data to obtain boundary inference values, granularity constraint values, and relationship confidence matrices includes:

[0027] Traverse all devices on the unified timeline data after clipping, and divide the devices into high-frequency devices and low-frequency devices based on the comparison results of the device sampling interval and the target time bucket granularity;

[0028] For high-frequency devices, an asynchronous tolerance window mechanism is used to infer missing data from adjacent time buckets within the preset asynchronous tolerance window duration to obtain boundary inference values;

[0029] For low-frequency devices, the missing data in the middle time bucket is filled in according to the time buckets with valid data points in the adjacent time buckets to obtain the granularity constraint value;

[0030] Extract the temporal fingerprints of each data source device on a unified timeline after cropping, calculate the fingerprint matching degree between different data source devices, and obtain the relationship confidence matrix.

[0031] Furthermore, for low-frequency devices, the process of filling in missing data in intermediate time buckets based on adjacent time buckets containing valid data points to obtain granularity constraint values ​​includes:

[0032] Iterate through each time bucket of the low-frequency device's unified timeline data after clipping to obtain the bucket to be filled;

[0033] For a filling interval consisting of multiple consecutive buckets to be filled, find the time buckets with representative points before and after the filling interval, and record them as the forward reference point and the backward reference point.

[0034] Based on the equipment type, the data characteristics are identified. For status data, linear interpolation is used, and for cumulative data, time ratio allocation is used to calculate the noise-added data value of the bucket to be filled.

[0035] The calculated noisy data values ​​are filled into the corresponding time buckets, and the granularity constraint values ​​are obtained.

[0036] Further, the step of extracting the temporal fingerprints of each data source device on a unified timeline after cropping, calculating the fingerprint matching degree between different data source devices, and obtaining a relationship confidence matrix includes:

[0037] Continuous A time bucket constitutes a time sequence window, and multiple time sequence windows are generated by sliding sequentially along a unified time axis.

[0038] For each timing window, extract the timing fingerprint for each data source device within the timing window;

[0039] All data source devices are paired up, and for each pair of devices, the matching degree of the time fingerprint of the corresponding device pair is calculated within the current time window.

[0040] The matching degree is used as the relationship confidence degree of the corresponding device pair under the current time window, forming a relationship confidence matrix with the time window as the row and the device pair as the column.

[0041] Furthermore, the process of merging the cropped unified timeline data, boundary inference values, granular constraint values, and relationship confidence matrix to generate a multi-dimensional fused record where each time bucket simultaneously contains multi-source information includes:

[0042] For each time bucket, collect the noise-added data values, boundary inference values, and granularity constraint values ​​of all devices within the corresponding time bucket;

[0043] Map the relationship confidence matrix to the current time bucket to generate a time bucket-level relationship confidence matrix;

[0044] The data confidence factor is determined based on the boundary inference value or granularity constraint value corresponding to each device. The association confidence factor is determined based on the average confidence of the relationship between the corresponding device and other devices. The data confidence factor and the association confidence factor are multiplied to obtain the fusion weight of the corresponding device in the current time bucket.

[0045] The data is integrated according to a preset structured format to form a multi-dimensional fused record corresponding to each time bucket.

[0046] Furthermore, the step of adding noise to the multidimensional fusion record to obtain the fusion result sequence and privacy audit log includes:

[0047] Request the allocation of a second total privacy budget from the privacy budget management module;

[0048] Based on the proportion of the fusion weight of each time bucket to the sum of the fusion weights of all pruned time buckets, the total amount of the second privacy budget is dynamically allocated to each time bucket to obtain the second privacy budget for each time bucket.

[0049] Based on the second privacy budget of each time bucket, noise is added to the noisy data values ​​of each device using the Laplace mechanism to obtain the fused data values ​​of each device in each time bucket;

[0050] The fused data values ​​of each device in each time bucket are integrated with the unperturbed boundary inference values, granular constraint values, time bucket-level relationship confidence matrix and fusion weights to obtain the fusion result sequence;

[0051] A privacy audit log is generated synchronously, recording the first privacy budget allocation record and the second privacy budget allocation record.

[0052] Furthermore, the dynamic allocation of the privacy budget based on data quality includes:

[0053] Receive the first privacy budget request from the edge device and allocate the first privacy budget according to the preset budget allocation strategy;

[0054] Receive the second privacy budget total request from the cloud and calculate the global remaining privacy budget as the second privacy budget total.

[0055] A method for multi-source energy data fusion based on differential privacy, implemented based on any one of the aforementioned multi-source energy data fusion systems, includes:

[0056] The upload data packets covering the fusion time range of the edge module are obtained, and the structured data of the noise-added data packets are extracted. The upload data packets contain one or more noise-added data packets, and the noise-added data packets contain noise-added batch data, device identifier, batch identifier, timestamp range and value of M, where M is the batch data length.

[0057] Based on the structured data, the timestamps of each data source device are mapped to a unified time bucket index to generate unified timeline data. Combined with the fused time range, the clipped unified timeline data is obtained.

[0058] Boundary missing inference, multi-granularity transformation and temporal fingerprint verification are performed on the cropped unified timeline data to obtain boundary inference values, granularity constraint values ​​and relation confidence matrix;

[0059] The cropped unified timeline data, the boundary inference values, the granularity constraint values, and the relationship confidence matrix are merged to obtain a multidimensional fused record;

[0060] Differential privacy noise is added to the multidimensional fusion record, and the fusion result sequence and privacy audit log are output.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] Through a cloud-edge collaborative architecture, the edge device performs zero-privacy-cost cleaning by leveraging intra-batch temporal correlation and adds correlation-preserving noise proportional to the temporal covariance of the batch data, with each noise addition consuming only one privacy budget. The cloud device achieves unified timeline alignment of multi-source data through time buckets, combining boundary missing data inference, multi-granularity transformation, and temporal fingerprint verification to address issues such as abnormal missing data from high-frequency devices, structural gaps in low-frequency devices, and quantification of reliable inter-device correlations. It also merges data reliability and correlation reliability to generate fusion weights, dynamically allocates a second privacy budget based on these weights, adds noise to obtain the fused result sequence, and simultaneously generates a privacy audit log containing budget consumption data from both the edge and cloud devices. This invention solves the technical problems of point-by-point noise addition leading to rapid depletion of the privacy budget, independent noise destroying temporal characteristics, and inconsistent batch windows for multi-source data making alignment difficult in existing technologies. It achieves synergistic optimization of privacy protection and data utility, asynchronous alignment between batches, and full auditability, significantly improving the usability, security, and traceability of energy data fusion. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the structure of a multi-source energy data fusion system based on differential privacy according to the present invention;

[0064] Figure 2 This is a flowchart illustrating the steps of data acquisition, cleaning, and noise addition in the edge module of this invention.

[0065] Figure 3 This is a flowchart illustrating the steps involved in unifying time bucket alignment in this invention.

[0066] Figure 4 This is a flowchart illustrating the steps of differential privacy output in this invention;

[0067] Figure 5 This is a flowchart illustrating the steps of a multi-source energy data fusion method based on differential privacy according to the present invention. Detailed Implementation

[0068] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0069] Example 1:

[0070] like Figure 1 As shown, this embodiment of the invention provides a multi-source energy data fusion system based on differential privacy, including an edge module, a fusion center module, and a privacy budget management module;

[0071] Among them, the edge module is used to filter out the complete batch according to the fusion instruction, clean each batch of data in the complete batch using time-series correlation, add correlation maintenance noise proportional to the time-series covariance of the batch data, generate noisy batch data, and encapsulate and upload one or more noisy batch data corresponding to the complete batch.

[0072] The fusion center module includes:

[0073] Fusion instruction unit: used to issue unified fusion instructions, receive and verify the validity of data packets uploaded by each data source device, and obtain structured data;

[0074] Alignment Unit: Used to map timestamps from various data sources to a unified time bucket index based on structured data, resulting in clipped unified timeline data;

[0075] Fusion preparation unit: used to perform boundary missing inference, multi-granularity transformation and temporal fingerprint verification on the clipped unified time axis data to obtain boundary inference values, granularity constraint values ​​and relation confidence matrix;

[0076] Multidimensional fusion unit: used to merge the clipped unified timeline data, boundary inference values, granular constraint values ​​and relation confidence matrix to generate a multidimensional fusion record in each time bucket that contains information from multiple sources;

[0077] Differential privacy output unit: used to add noise to multidimensional fusion records to obtain a fusion result sequence and privacy audit log;

[0078] The privacy budget management module is used to dynamically allocate privacy budgets based on data quality.

[0079] The edge module is located on the data source device side, the fusion center module is located on the cloud server, and the privacy budget management module communicates with the edge module and the fusion center module.

[0080] Specifically, edge modules are configured on the data source device side. When there are multiple data source devices, there are corresponding edge modules. Each edge module independently collects raw energy data, filters complete batches according to fusion instructions, completes data cleaning and local noise addition, and then encapsulates and uploads the noisy batch data to the fusion center module. The fusion center module uniformly schedules data uploads and performs integrity verification through the fusion instruction unit. The data then goes through the timeline unification of the alignment unit, the data completion and correlation analysis of the fusion preparation unit, the weight calculation and record generation of the multi-dimensional fusion unit, and the noise addition and audit log output of the differential privacy output unit, finally forming a fusion result sequence and privacy audit log. The privacy budget management module runs through the edge and cloud and is responsible for the dynamic allocation of the first privacy budget and the second privacy budget.

[0081] This embodiment utilizes a cloud-edge collaborative architecture, with batches as the basic processing unit. At the edge, zero-privacy-cost cleaning is achieved by leveraging the temporal correlation within batches, and correlation-preserving noise proportional to the temporal covariance of the batch data is added. Each noise addition consumes only one privacy budget, significantly improving privacy budget utilization. At the cloud, unified timeline alignment of multi-source data is achieved through time bucket alignment, overlapping bucket boundary processing, and representative point selection. Combined with asynchronous tolerance window boundary missing inference, multi-granularity transformation based on data characteristics, and time-series window sliding fingerprint verification, the problems of abnormal missing high-frequency devices, structural gaps in low-frequency devices, and quantification of reliable inter-device associations are addressed respectively. The multi-dimensional fusion unit fuses data credibility and association credibility to generate fusion weights, achieving weighted fusion of multi-source information. The differential privacy output unit dynamically allocates a second privacy budget based on the fusion weights and adds noise before publishing, simultaneously generating privacy audit logs containing budget consumption from both the edge and cloud. This invention solves the technical problems in the prior art, such as the rapid depletion of privacy budget due to point-by-point noise addition, the destruction of temporal characteristics by independent noise, and the difficulty in aligning inconsistent batch windows of multi-source data. It achieves synergistic optimization of privacy protection and data utility, asynchronous alignment between batches, and full auditability, significantly improving the usability, security, and traceability of energy data fusion.

[0082] Furthermore, the edge module is set on the data source device side to filter out the complete batch according to the fusion time range in the fusion instruction, clean each batch of data in the complete batch using time-series correlation, add correlation-preserving noise proportional to the time-series covariance of the batch data to generate noisy batch data, and encapsulate and upload one or more noisy batch data corresponding to the complete batch.

[0083] like Figure 2 As shown, specifically, the edge module collects raw energy data at a preset sampling frequency according to the data acquisition interface. The raw energy data includes at least one of the following: smart meter load data, photovoltaic inverter power generation data, and meteorological sensor environmental data. A preset number of M consecutive timestamps of raw energy data are locally cached to form batch data. Each batch of data contains M consecutive timestamps that constitute a complete time series window, with equal time intervals between adjacent timestamps.

[0084] It should be noted that the preset sampling frequency and batch data length M should be set differently according to the type of data source device, data characteristics and actual application requirements. In actual industrial scenarios, the sampling frequency and M value are not uniform for different data source devices.

[0085] Furthermore, the edge module filters out complete batches based on the fusion time range in the fusion instruction. Specifically, it receives a fusion instruction from the fusion instruction unit in the fusion center module. This instruction includes the current fusion time range. It then filters out complete batches from its local cache whose timestamp range completely covers the current fusion timestamps. These complete batches may contain one or more batches of data. If no complete batch matching the current fusion instruction time range is generated locally, it waits for the batch to be generated naturally before performing subsequent processing.

[0086] Furthermore, the edge module cleans each batch of data in the complete batch by utilizing the temporal correlation within the batch data, generates cleaned batch data, and calculates the data quality score of that batch data.

[0087] Specifically, for each batch of data within a complete batch, the mean and standard deviation of the raw energy data corresponding to the M timestamps within the batch are calculated. For each timestamp, anomaly detection is performed on the raw energy data; data points exceeding three times the standard deviation are identified as outliers and replaced with the mean of adjacent timestamps. If the outlier is the first point of the batch, it is replaced with the next raw energy data point; if the outlier is the last point of the batch, it is replaced with the previous raw energy data point. The percentage of outliers is recorded, and the accuracy score is calculated as the difference between 1 and the percentage of outliers. Next, for missing timestamps within the batch data, linear interpolation is performed based on the raw energy data of the preceding and following timestamps to fill in the missing timestamps, resulting in M ​​complete raw energy data points. The percentage of missing timestamps is recorded, and the completeness score is calculated as the difference between 1 and the percentage of missing timestamps. Then, Kalman filtering is applied to the M raw energy data points for smoothing, removing high-frequency noise while preserving temporal trend characteristics, thus obtaining cleaned batch data. The cleaned batch data includes M cleaned data points corresponding to M consecutive timestamps. Simultaneously, the average absolute deviation of M raw energy data points and the cleaned batch data is calculated. This deviation is used to calculate the consistency score, which is 1 minus the ratio of the average absolute error to the equipment range. The equipment range is preset according to the equipment's factory technical specifications, and this embodiment does not limit it. This yields the data quality score for the batch of data, which is the sum of the accuracy score, the completeness score, and the consistency score.

[0088] For example, the Kalman filter uses a linear discrete system model, with the state transition matrix being an M-dimensional identity matrix and the observation matrix being an M-dimensional identity matrix; the process noise covariance matrix is ​​a diagonal matrix, with diagonal elements being 1%-5% of the standard deviation of the original data, and the measurement noise covariance matrix is ​​a diagonal matrix, with diagonal elements being 0.5%-2% of the equipment range; initialization uses the original energy data at the first moment as the initial state, and the initial error covariance matrix is ​​10 times the identity matrix.

[0089] In this embodiment, the edge module completes data cleaning and calculates data quality scores locally, achieving zero privacy budget consumption during the cleaning process and avoiding the problem of privacy budget consumption in the cleaning stage in traditional solutions. Through the data cleaning mechanism, the integrity and accuracy of the raw energy data are effectively improved, providing high-quality data for the subsequent noise addition stage. Based on this, the data quality score is quantified according to the proportion of outliers, the proportion of missing points, and the average absolute deviation of smoothing filtering recorded during the cleaning process. This provides a basis for the privacy budget management module to dynamically allocate the first privacy budget, so that batches with worse data quality receive more privacy budget and less noise is added. This avoids the superposition of errors in the data itself and noise disturbances, which would lead to a further decrease in data availability, thus achieving differentiated adjustment between privacy protection and data availability.

[0090] Furthermore, the edge module adds correlation-preserving noise proportional to the temporal covariance of the batch data to generate noisy batch data, including:

[0091] Request the allocation of the first privacy budget from the privacy budget management module;

[0092] Based on the first privacy budget and the temporal covariance matrix of the cleaned batch data, correlation-preserving noise is generated to obtain multidimensional Gaussian noise.

[0093] The multidimensional Gaussian noise is added to each of the cleaned batch data to generate noisy batch data.

[0094] Specifically, after batch data cleaning is completed, the edge module requests the allocation of a first privacy budget from the privacy budget management module. This request includes the data source device identifier, batch identifier, and data quality score, which the privacy budget management module uses to differentiate budget allocation requirements for different data source devices and data quality scores. The privacy budget management module then allocates the corresponding first privacy budget to the data batch. and will Feedback is sent to the edge module; among which, A positive number greater than 0. The smaller the value, the stronger the privacy protection, but the greater the corresponding noise intensity. The edge module receives the first privacy budget successfully allocated by the privacy budget management module. .

[0095] Subsequently, the temporal covariance matrix of the cleaned batch data is calculated, and the cleaned batch data is represented in vector form. ,in, For batch data length, For the first The cleaned energy data corresponds to each timestamp, and its time-series covariance matrix is ​​calculated. ,in, The covariance calculation function, for The mean vector, Let be the mathematical expectation.

[0096] Next, based on the first privacy budget and the temporal covariance matrix, multidimensional Gaussian noise proportional to the temporal covariance matrix is ​​generated. Specifically: Performing Cholesky decomposition on the time series covariance matrix yields a lower triangular matrix. The formula is: And generate an M-dimensional standard Gaussian random vector. The vector follows a standard M-dimensional Gaussian distribution, i.e. ,in This is an M-dimensional identity matrix, where the elements on the main diagonal are 1s and the rest are 0s. This generates multidimensional Gaussian noise. The formula is ,in, Sensitivity is used to define the upper bound of the impact of the maximum change in a single data point on the time series covariance matrix. Based on the preset physical value range of energy data, for example, if the data source device collects electricity load data from smart meters, then... The preferred value is the difference between the maximum and minimum electrical loads divided by . Meanwhile, this multidimensional Gaussian noise obey Gaussian distribution, formula is This means that the multidimensional Gaussian noise is proportional to the temporal covariance matrix, strictly preserving the intra-batch temporal correlation structure. Then, the multidimensional Gaussian noise and the cleaned batch data are added element by element, using the formula: Noise-added batch data was obtained. .

[0097] This embodiment calculates the temporal covariance matrix of the cleaned batch data and generates multidimensional Gaussian noise proportional to the temporal covariance matrix. This ensures that the added noise strictly preserves the temporal correlation structure of the data within the batch, avoiding the destruction of temporal characteristics by traditional independent and identically distributed noise. This mechanism adds noise on a batch-by-batch basis, with each noise addition consuming only one privacy budget. Compared to point-by-point noise addition, the privacy budget utilization rate is significantly improved. While ensuring strong privacy protection, it maximizes the preservation of the temporal characteristics of energy data, providing a high-quality noisy data foundation for subsequent multi-source data fusion.

[0098] Furthermore, the noisy batch data, device identifier, batch identifier, timestamp range, and the value of M are encapsulated into a noisy data packet corresponding to that batch. One or more noisy data packets corresponding to the complete batch are combined in chronological order into an upload data packet, which is then uploaded to the fusion center module via the communication network. The device identifier is a unique device code used by the fusion center module to distinguish the data source terminal; the batch identifier is a batch number generated by the edge module according to the acquisition order, used to identify which batch of data, ensuring time sequence traceability, non-repetition, and no out-of-order processing; the timestamp range is from the timestamp of the first data acquisition in this batch to the timestamp of the last data acquisition, used by the fusion center module to align the time sequences of multiple devices and stitch together a complete time series.

[0099] In this embodiment, through the edge module, different data source devices can independently configure sampling frequency and batch size according to their own characteristics, adapting to diverse industrial data acquisition scenarios. Each data source device independently completes anomaly detection, missing data imputation, and smoothing filtering using the temporal correlation within the batch. The cleaning process does not consume privacy budget, and the noise generated during batch noise addition is proportional to the temporal covariance matrix of the original data, ensuring that the noise-added batch data maintains its original temporal correlation structure. Furthermore, by adopting a batch-by-batch noise addition method, each noise addition consumes only one privacy budget, significantly improving privacy budget utilization compared to point-by-point noise addition. The local caching, batch processing, and batch upload working mode is fully compatible with the batch upload mechanisms actually used by various energy devices, such as smart meters, photovoltaic inverters, and weather sensors, without requiring modifications to the transmission logic of existing devices, thus reducing deployment costs. Moreover, all original data and cleaned data are cached and processed locally, with only the noise-added batch data being encapsulated and uploaded, effectively preventing the leakage of original sensitive data during transmission and further enhancing data privacy protection capabilities.

[0100] Furthermore, the fusion center module, located on a cloud server, includes: a fusion instruction unit: used to issue unified fusion instructions to each edge module, receive and verify the validity of data packets uploaded by each data source device, and obtain structured data; an alignment unit: used to map the timestamps of each data source to a unified time bucket index based on the structured data, and obtain pruned unified timeline data; a fusion preparation unit: used to perform boundary missing inference, multi-granularity transformation, and temporal fingerprint verification on the pruned unified timeline data, and obtain boundary inference values, granularity constraint values, and relationship confidence matrices; a multidimensional fusion unit: used to merge the pruned unified timeline data, boundary inference values, granularity constraint values, and relationship confidence matrices, and generate multidimensional fusion records where each time bucket contains information from multiple sources; and a differential privacy output unit: used to add noise to the multidimensional fusion records, and obtain a fusion result sequence and privacy audit logs.

[0101] Specifically, the fusion center module uses the fusion command unit to uniformly schedule the data uploads from each edge module and performs integrity verification on the uploaded data. After the verification is successful, the data is transmitted to the alignment unit for timeline unification. Subsequently, the multidimensional fusion of multi-source heterogeneous energy data is completed through the fusion preparation unit and the multidimensional fusion unit. Finally, the differential privacy output unit performs protection processing to obtain the fusion result sequence and privacy audit log, achieving dual protection of data fusion availability and privacy security.

[0102] Furthermore, the fusion instruction unit is used to issue unified fusion instructions to each edge module, receive and verify the validity of data packets uploaded by each data source device, and obtain structured data. Specifically, based on the preset fusion task cycle or user query request, the fusion time range for this fusion task is determined, a fusion instruction is generated, and sent to the edge modules. The fusion instruction unit receives the uploaded data packets from each edge module, decapsulates the uploaded data packets, traverses each noisy data packet within them, decapsulates a single noisy data packet, extracts the noisy batch data, device identifier, batch identifier, timestamp range, and the value of M, and temporarily stores the extracted structured data. All extracted structured data is verified to ensure that all batch identifiers of each data source device are consecutive and that all timestamp ranges cover the current fusion time range. If the data of a device is incomplete, a fusion instruction is resent to the edge module of that device, requesting a resend of the uploaded data packet. After successful verification, the fusion instruction unit passes all extracted structured data to the alignment unit.

[0103] In this embodiment, the fusion command unit issues a unified fusion command to all edge modules, defining the same target fusion time window for each device. This enables devices with different acquisition frequencies and different batch start and end times to respond collaboratively. By decapsulating and extracting the uploaded data packets layer by layer, combined with batch identifier continuity verification and timestamp range coverage verification, it ensures that the batch data of each device entering the fusion process can cover the current fusion time window in terms of time. This solves the problem that multi-source data cannot be aligned within the same time period due to asynchronous device acquisition and inconsistent batch windows, providing a foundation of input data with complete time coverage and standardized format for subsequent timeline unification and fusion analysis.

[0104] Furthermore, this embodiment provides a step for mapping timestamps from various data sources to a unified time bucket index based on structured data to obtain pruned unified timeline data, including:

[0105] Based on structured data, the timestamp range and the time series associated with the value of M are standardized to obtain the timestamp corresponding to each noisy data value;

[0106] Map each noisy data value, device identifier, and corresponding timestamp to the corresponding time bucket;

[0107] Within each time bucket, a representative point is selected for each device to generate unified timeline data;

[0108] The unified timeline data is cropped according to the fusion time range to obtain cropped unified timeline data.

[0109] like Figure 3 As shown, the alignment unit receives structured data transmitted by the fusion instruction unit, including the following for each noisy data packet: noisy batch data, device identifier, batch identifier, timestamp range, and the value of M. Since the timestamp formats of data packets uploaded from different data source devices may differ, it is necessary to first standardize the timestamp range and the time series associated with the value of M. Specifically, the timestamp range is converted into a unified millisecond-level timestamp to eliminate differences in formats. Simultaneously, based on the timestamp range and the value of M, and combined with the uniformity of the sampling frequency at the edge, the noisy batch data is extracted. Each noisy data value The specific value of the corresponding timestamp is denoted as ,in , Represented as the first Noisy data values The corresponding timestamp. At this point, each noisy data value from each data source corresponds to a unique, standardized timestamp, providing a time reference for mapping each timestamp from each data source to a unified time bucket index.

[0110] The basic parameters of the unified time bucket are preset, including: bucket size, which is the length of time covered by a single time bucket. The bucket size is determined based on the maximum sampling period of each data source participating in the fusion, and is preferably set to the least common multiple of the sampling periods of all data sources or 15 minutes; overlap duration. The overlap length between adjacent time buckets is determined based on the time synchronization accuracy of each data source device, with a preferred value of 5% to 15% of the bucket size. The bucket alignment reference point is the reference zero point for dividing the time buckets, with a preferred value of 00:00:00 per day. These parameters together construct a unified time bucket index, which is a time grid composed of a series of continuous, equal-length time buckets arranged in chronological order. Each time bucket has a fixed start time and coverage area, providing a unified time positioning reference for the timestamps of various data sources.

[0111] Based on each noise-added data value Corresponding timestamp Calculate the bucket number of the time bucket to which the timestamp belongs. The formula is ,in, This is the floor function. As the reference point for bucket alignment, The size of the bucket. Each time bucket corresponds to a unique time identifier. The formula is: Timestamp The corresponding noisy data values, device identifier, and timestamp. The original storage bucket number In the corresponding bucket structure.

[0112] Data within each time bucket is organized by device identifier, forming the following data structure: Each time bucket corresponds to a unique time identifier; each time bucket contains a device data set, with the device identifier as the key and the list of data points for that device within the current time bucket as the value; for each device's data point list, each data point contains at least three fields: timestamp, noisy data value, and offset, where the offset is the time difference between the timestamp and the time identifier. When the same device has multiple data points within the same time bucket, the device's data point list contains multiple data point records, sorted in ascending order of timestamp. When the same device has only one data point within the same time bucket, the device's data point list contains only a single data point record.

[0113] This embodiment eliminates the heterogeneity caused by differences in timestamp formats among different data source devices by standardizing the timestamps, thus laying a time benchmark for subsequent unified mapping. Furthermore, a unified time bucket index system is constructed by presetting bucket size, overlap duration, and bucket alignment benchmark point, and each noisy data value is mapped to the corresponding time bucket, realizing the initial alignment of multi-device data in the time dimension. This fully preserves the multi-data point information of each device within the bucket, providing a sufficient data foundation for subsequent selection of representative points.

[0114] Furthermore, this embodiment also provides a step of performing boundary region processing on data points within each time bucket, and simultaneously storing the data points corresponding to the timestamps located in the boundary regions into adjacent time buckets.

[0115] Specifically, to address the boundary effect issue when timestamps are near bucket boundaries, boundary judgment and processing are performed on the timestamps. Specifically, based on a preset overlap duration... The boundary region of each time bucket is determined. This boundary region is the overlapping time within the time limit from the start time of the bucket to the overlapping time limit before the end time of the bucket. and So, calculate the timestamp For the offset of the current time bucket, if the offset is less than the overlap duration, the timestamp is determined to be in the left boundary region; if the offset is greater than the difference between the bucket size and the overlap duration, the timestamp is determined to be in the right boundary region. If the timestamp is in the left boundary region, the data point corresponding to the timestamp is simultaneously stored in the previous adjacent bucket. If the timestamp is located in the right boundary region, then the data point corresponding to that timestamp will also be stored in the next adjacent bucket. If an adjacent bucket does not exist, it is ignored. Additionally, for data points stored in multiple time buckets simultaneously, a boundary marker field is added to the data structure within each bucket to identify that the data point originates from an overlapping region, facilitating identification.

[0116] This embodiment defines the boundary region of each time bucket and calculates the offset of the timestamp within the bucket, thereby enabling the identification of data points near the bucket boundary. For timestamps located in the left or right boundary region, their data points are simultaneously stored in adjacent time buckets, effectively solving the problem of data misalignment caused by slight deviations in device clocks or differences in timestamp accuracy, and eliminating the bucket boundary effect.

[0117] Furthermore, to select a representative point from multiple data points within each time bucket for each data source device, forming a unified timeline data, specifically, all time buckets are traversed, and the following operations are performed for each time bucket:

[0118] If a data source device has only one data point in the current bucket, that data point is directly used as the representative point for that device. If a data source device has multiple data points in the current bucket (i.e., multiple timestamps fall into the same bucket), the data point closest to the bucket's center timestamp is selected as the representative point. Calculate using the following formula: The selection criteria for representative points are as follows: That is, the data point with the smallest absolute time difference from the center of the bucket is selected as the representative point.

[0119] For each selected representative point, record the device identifier, noisy data value, timestamp, and offset. Using the start time of the current time bucket as the time identifier, summarize the representative point information of all devices within that bucket to form a record for that time bucket. Sort all time bucket records according to their start times to form a time series structure, generating unified timeline data. The unified timeline data format is: an ordered list, where each element corresponds to a time bucket record. Each time bucket record contains a bucket start time field and a device data set field. The device data set field uses the device identifier as the key and the representative point data object as the value. The representative point data object at least contains noisy data value, timestamp, and offset information.

[0120] For example, assume the time bucket size is 15 minutes, the current time bucket start time is 08:00:00, and the bucket center time is 08:07:30. Device A has two data points within this bucket: timestamp 08:02:00 with a noisy data value of 2.41, and timestamp 08:12:00 with a noisy data value of 2.56. Calculate the time difference between the two data points and the bucket center time: |08:02:00 - 08:07:30| = 5 minutes and 30 seconds, |08:12:00 - 08:07:30| = 4 minutes and 30 seconds. Select the data point with the smaller time difference, i.e., timestamp 08:12:00 and value 2.56, as the representative point for device A in this bucket.

[0121] For the generated unified timeline data, the time buckets are trimmed according to the fusion time range issued by the fusion instruction unit. All time buckets are traversed, and the time range of each bucket is retained. Time buckets that overlap with the fusion time range are clipped and the unified timeline data only includes time buckets that overlap with the current fusion time window, and are then passed to the fusion preparation unit.

[0122] This embodiment simplifies multiple data points that may exist in a bucket into a single representative point by selecting a representative point for each device within each time bucket. This effectively solves the problem of multiple point attribution caused by the device sampling interval being smaller than the time bucket granularity. At the same time, the time buckets are pruned according to the fusion time range, retaining only the time buckets that overlap with the fusion window, focusing on the effective data range and reducing computational redundancy in subsequent processing.

[0123] Furthermore, this embodiment provides a step for performing boundary missing inference, multi-granularity transformation, and temporal fingerprint verification on the clipped unified timeline data to obtain boundary inference values, granularity constraint values, and relation confidence matrices, including:

[0124] Traverse all devices on the unified timeline data after clipping, and divide the devices into high-frequency devices and low-frequency devices based on the comparison results of the device sampling interval and the target time bucket granularity;

[0125] For high-frequency devices, an asynchronous tolerance window mechanism is used to infer missing data from adjacent time buckets within the preset asynchronous tolerance window duration to obtain boundary inference values;

[0126] For low-frequency devices, the missing data in the middle time bucket is filled in according to the time buckets with valid data points in the adjacent time buckets to obtain the granularity constraint value;

[0127] Extract the temporal fingerprints of each data source device on a unified timeline after cropping, calculate the fingerprint matching degree between different data source devices, and obtain the relationship confidence matrix.

[0128] Specifically, all devices on the unified timeline data after cropping are traversed. Based on the comparison between the device sampling interval and the target time bucket granularity, the devices are divided into high-frequency devices and low-frequency devices. The device sampling interval is the ratio of the timestamp range in the data packet uploaded by the device to M-1, and the unit of the device sampling interval is consistent with the timestamp unit. The target time bucket granularity is the size of the time bucket. Devices with sampling intervals less than or equal to the target time bucket granularity are classified as high-frequency devices, and devices with sampling intervals greater than the target time bucket granularity are classified as low-frequency devices.

[0129] It should be noted that the sampling interval of high-frequency devices is less than or equal to the target time bucket granularity, and under normal circumstances, there should be at least one data point in each time bucket; the sampling interval of low-frequency devices is greater than the target time bucket granularity, and under normal circumstances, only some time buckets have data points, while the remaining time buckets have no data. This is a structural feature rather than an abnormal missing data.

[0130] Furthermore, for high-frequency devices, an asynchronous tolerance window mechanism is used to infer missing data from adjacent time buckets to obtain boundary inference values, specifically including:

[0131] Iterate through each time bucket of the clipped, unified timeline data, checking for missing representative data for that device. If no representative data is found in the current time bucket, mark it as a missing bucket for that device. Multiple consecutive missing buckets constitute a missing interval, and the region adjacent to a time bucket with data is defined as the missing boundary region. Set the asynchronous tolerance window duration based on the clock skew characteristics of the data source device. The tolerance window duration is dynamically adjusted based on the device's historical clock deviation records. The larger the clock deviation, the larger the tolerance window. An exemplary preferred value range is 30 to 120 seconds.

[0132] For a time point marked as a missing bucket, search for a neighboring time bucket with data within the tolerance window. If a neighboring time bucket with data exists within the tolerance window, use linear interpolation to infer the missing value and fill it into the device data of the missing bucket as the noisy data value for that time bucket. If only a forward neighboring bucket or only a backward neighboring bucket exists within the tolerance window, use the neighboring value to directly fill it as the noisy data value for that time bucket. If no neighboring time buckets with data exist within the tolerance window, mark it as uninferable, and the noisy data value for that time bucket is empty.

[0133] Simultaneously, a boundary inference value is output for each time bucket to characterize the reliability of the filling result: 1 for non-missing buckets; 0.8 for two-sided interpolation filling; 0.6 for one-sided neighbor-value filling; and 0.2 for inference failures or null values. The values ​​1, 0.8, 0.6, and 0.2 are example data. Those skilled in the art can adjust the numerical range or grading standards according to actual application scenarios and fusion accuracy requirements, using them as weight constraints for data reliability in subsequent multi-dimensional fusion units.

[0134] This embodiment divides the equipment into high-frequency and low-frequency devices and processes them separately. For high-frequency devices, an asynchronous tolerance window mechanism is introduced to solve the timestamp drift problem caused by the asynchronous clocks of the devices, avoiding the misjudgment of data points that should be matched as missing. Missing data is filled by linear interpolation or neighboring values, and quantified boundary inference values ​​are output according to the filling method. This provides measurable data credibility weights for subsequent multi-dimensional fusion, enabling the fusion process to distinguish between original data and filled data, and between high-credibility filling and low-credibility filling. Thus, while ensuring data integrity, the reliability and accuracy of the fusion results are significantly improved.

[0135] Furthermore, this embodiment provides a step for low-frequency devices to fill in the missing data in the middle time bucket based on adjacent time buckets with valid data points, and obtain granularity constraint values, including:

[0136] Iterate through each time bucket of the low-frequency device's unified timeline data after clipping to obtain the bucket to be filled;

[0137] For a filling interval consisting of multiple consecutive buckets to be filled, find the time buckets with representative points before and after the filling interval, and record them as the forward reference point and the backward reference point.

[0138] Based on the equipment type, the data characteristics are identified. For status data, linear interpolation is used, and for cumulative data, time ratio allocation is used to calculate the noise-added data value of the bucket to be filled.

[0139] The calculated noisy data values ​​are filled into the corresponding time buckets, and the granularity constraint values ​​are obtained.

[0140] Specifically, the process iterates through each time bucket of the clipped, unified timeline data for low-frequency devices, identifying time buckets with representative points and recording their timestamps and noise-added data values ​​as reference points. Simultaneously, time buckets without representative points are marked as buckets to be filled. For a filling interval consisting of multiple consecutive buckets to be filled, the nearest time buckets with representative points before and after this interval are identified and designated as the forward and backward reference points. For each bucket to be filled within the filling interval, noise-added data values ​​are calculated using appropriate methods based on the device's corresponding data characteristics. These calculated noise-added data values ​​are then filled into the representative point position of the corresponding time bucket, serving as the noise-added data value for that time bucket.

[0141] The data characteristics corresponding to each device are identified based on the device type. If the device is a status device, such as a temperature sensor or a barometric pressure sensor, its value is a continuously changing physical quantity, and a linear interpolation method is used. If the device is a cumulative device, such as an electricity meter, its value is a cumulative amount over the collection period, and a proportional allocation method is used. The correspondence between device type and data characteristics can be pre-configured.

[0142] For state-based devices, physical quantities such as temperature exhibit linear changes over a short period, and linear interpolation formulas are used to calculate the noise-added data values ​​for the bucket to be filled. For cumulative devices, the reported noise-added data values ​​represent the cumulative amount during the acquisition period, and a proportional allocation method is needed to divide the incremental acquisition interval into time buckets.

[0143] First, determine the sampling interval to be split and its corresponding total increment. When the filling interval is located between the preceding and following reference points, the total increment is the noisy value of the preceding reference point, because the cumulative quantity reported by the device is the cumulative quantity value for that time period. Then, obtain the cumulative quantity allocation ratio of the time buckets within this sampling interval based on historical data from the same period. Specifically, select historical data from the same type of device in the same period of the past 30 days as a statistical sample, and use a weighted average method to calculate the cumulative quantity ratio of each time bucket. The weight is the time proximity, i.e., the closer the data, the higher the weight. First, remove abnormal historical data that deviates from the mean by 3 times the standard deviation. If there is no historical data, use a uniform allocation ratio, i.e., the allocation ratio of each time bucket is equal. For each bucket to be filled within the filling interval, its cumulative quantity value is the product of the total increment and the cumulative quantity allocation ratio of that time bucket. Fill this cumulative quantity value into the device data of the corresponding time bucket as the noisy data value of that time bucket.

[0144] Simultaneously, a granularity constraint value is output for each time bucket to characterize the uncertainty introduced by the granularity transformation of the filled value: for noisy data that does not require granularity transformation, the granularity constraint value is 1; for noisy data values ​​filled by linear interpolation, the granularity constraint value is 0.8; and for noisy data values ​​obtained by historical proportional allocation, the granularity constraint value is 0.6. The larger the granularity constraint value, the closer the filled value is to the original acquired features, and the higher the fusion weight should be assigned in the subsequent multidimensional fusion process; conversely, a relatively lower fusion weight should be assigned.

[0145] This embodiment addresses the structural characteristic of low-frequency devices having sampling intervals larger than the target time bucket granularity, enabling the completion of low-frequency data on a fine-grained time axis. It distinguishes between state-related and cumulative-related devices based on device type, employing linear interpolation methods conforming to physical characteristics and proportional allocation methods respectively to ensure the physical rationality of the filled values: state-related quantities use linear interpolation to maintain a continuous trend, while cumulative quantities use historical proportional allocation to maintain the total amount conserved over time periods. Simultaneously, based on the filling method, a quantified granularity constraint value is output for each time bucket, providing distinguishable filling reliability weights for subsequent multi-dimensional fusion. This allows the fusion process to dynamically adjust the fusion weights based on the reliability of the data source, effectively improving the usability and reliability of low-frequency device data in the fusion results.

[0146] Furthermore, this embodiment provides a step for extracting the temporal fingerprints of each data source device on a unified timeline after cropping, calculating the fingerprint matching degree between different data source devices, and obtaining a relationship confidence matrix, including:

[0147] Continuous A time window is formed by time buckets, and multiple time windows are generated by sliding sequentially along a unified time axis; L is a preset integer greater than 1, representing the number of time buckets contained in the time window;

[0148] For each timing window, extract the timing fingerprint for each data source device within the timing window;

[0149] All data source devices are paired up, and for each pair of devices, the matching degree of the time fingerprint of the corresponding device pair is calculated within the current time window.

[0150] The matching degree is used as the relationship confidence degree of the corresponding device pair under the current time window, forming a relationship confidence matrix with the time window as the row and the device pair as the column.

[0151] Specifically, after completing boundary missing data inference and multi-granularity transformation, all devices now have complete data in each time bucket of the pruned, unified timeline data. A time bucket constitutes a timing window, and multiple timing windows are generated by sliding sequentially along a unified time axis; the same time bucket can belong to multiple adjacent timing windows at the same time, depending on the sliding step size.

[0152] For each time series window, the time series fingerprint of each data source device within that window is extracted: the change in noisy data values ​​between adjacent time buckets within the time series window is calculated, and the direction of change is determined based on the change in noisy data values: if the change in noisy data values ​​is greater than the change threshold, it is determined to be an increase and recorded as +1; if the change in noisy data values ​​is less than the negative change threshold, it is determined to be a decrease and recorded as -1; otherwise, it is determined to be flat and recorded as 0. The change threshold is a preset value, preferably 5% of the device's range, used to filter out minor fluctuations. Since correlation-preserving noise has been added at the edge, to avoid misjudging trend changes due to minor numerical fluctuations caused by noise, the change threshold is magnified by 1.2 times based on the preset value as the actual judgment threshold. Here, 1.2 is an exemplary value; specifically, multiple sets of values ​​can be selected within the range [1.0, 2.0] for a limited number of experiments to select the optimal value. The directions of change are arranged in chronological order, forming a length of [missing information]. The sequence of changes in direction serves as the timing fingerprint of the device within that timing window.

[0153] All data source devices are paired to form multiple device pairs. For each device pair, the matching degree of their temporal fingerprints is calculated within the current time window. The matching degree is the ratio of the number of devices with the same direction of change at the same location to the total length of the fingerprints. "Same direction of change at the same location" means that the change direction values ​​of the two devices are equal within the same time interval.

[0154] The matching degree is used as the relationship confidence of the device pair within the current time window. The relationship confidence value ranges from [0,1]. A higher value indicates a stronger consistency in the changing trends of the two devices within this time period and a higher degree of association credibility. When there is no association verification requirement or the time fingerprint is invalid, the relationship confidence value takes the default value of 0.5. Therefore, a relationship confidence matrix is ​​formed with the time window as the row and the device pair as the column. The same device can simultaneously form independent device pair confidence with multiple other devices, and the same time bucket can participate in the confidence calculation within multiple time windows, thus comprehensively representing the association credibility of multi-source devices in different time periods. This relationship confidence matrix is ​​not attached to a single time bucket or a single data point, but serves as the global weight basis at the window level and device pair level, and is input into the subsequent multi-dimensional fusion unit to dynamically adjust the weight allocation of each data source during the fusion process.

[0155] This embodiment uses continuous time buckets to form a sliding window. By calculating the change in the noise-added data value between adjacent time buckets and determining the three directions of change (rising, falling, and flat), a sequence of change directions of the device within the time window is constructed as a time fingerprint. By matching the time fingerprints of device pairs within the same time window bit by bit, the fingerprint matching degree is calculated as the relationship confidence. The final relationship confidence matrix, with the time window as the row and the device pair as the column, fully characterizes the degree of association credibility of multi-source devices in different time periods, providing a window-level and device-pair-level global weight basis for subsequent multi-dimensional fusion units.

[0156] Furthermore, this embodiment provides a step for merging the cropped unified timeline data, boundary inference values, granular constraint values, and relationship confidence matrix to generate a multi-dimensional fused record where each time bucket simultaneously contains multi-source information, including:

[0157] For each time bucket, collect the noise-added data values, boundary inference values, and granularity constraint values ​​of all devices within the corresponding time bucket;

[0158] Map the relationship confidence matrix to the current time bucket to generate a time bucket-level relationship confidence matrix;

[0159] The data confidence factor is determined based on the boundary inference value or granularity constraint value corresponding to each device. The association confidence factor is determined based on the average confidence of the relationship between the corresponding device and other devices. The data confidence factor and the association confidence factor are multiplied to obtain the fusion weight of the corresponding device in the current time bucket. The sum of the fusion weights of each device in the current time bucket is determined as the time bucket fusion weight of the current time bucket.

[0160] The data is integrated according to a preset structured format to form a multi-dimensional fused record corresponding to each time bucket.

[0161] Specifically, based on the fusion preparation unit, the clipped unified time bucket data of each device is supplemented. Combining the boundary inference value and granularity constraint value that characterize the reliability of single-device data, and the relationship confidence matrix that characterizes the correlation of change trends among multiple devices, multi-dimensional information aggregation and weight calculation are performed with the time bucket as the smallest fusion unit, ultimately forming a structured multi-dimensional fusion record. The specific steps are as follows:

[0162] First, for each time bucket in the pruned unified time bucket data, the noisy data values ​​of all devices within that time bucket, as well as the boundary inference values ​​and granular constraint values ​​corresponding to each device, are collected. Simultaneously, the relationship confidence matrix at the time-series window level is mapped to the current time bucket. Specifically, the mapping method is as follows: traverse all time-series windows containing the time bucket; for each pair of devices, extract the relationship confidence of that pair within each time-series window; calculate the arithmetic mean of all extracted relationship confidence values; and use this arithmetic mean as the relationship confidence of that pair of devices within the current time bucket. Similarly, calculate the relationship confidence of all pair of devices within the time bucket in the same way, generating the time bucket-level relationship confidence matrix corresponding to the current time bucket. The time bucket-level relationship confidence matrix contains the relationship confidence of each pair of devices within the current time bucket.

[0163] Secondly, for each data source device within each time bucket, its fusion weight is calculated. The boundary inference value or granularity constraint value corresponding to that device is used as the data reliability factor. The arithmetic mean of the confidence scores of the relationships between this device and all other devices in the current time bucket is used as the association confidence factor. Multiplying the two factors yields the fusion weight of the device in the current time bucket. The calculation formula is: ,in, The value range is [0,1]. The larger the value, the higher the reliability of the data from this device and its correlation with other devices, and the higher its weight will be given in the subsequent fusion process.

[0164] Finally, the core information of the current time bucket is integrated according to a preset structured format to form a multi-dimensional fusion record corresponding to that time bucket. This multi-dimensional fusion record specifically includes: bucket number, bucket start time, noisy data values ​​of each device, boundary inference values ​​or granularity constraint values, time bucket-level relationship confidence matrix, and fusion weights of each device. The above steps are repeated for all time buckets on the cropped unified timeline data to generate a complete multi-dimensional fusion record sequence, providing a unified, reliable, and traceable multi-source fusion data foundation for subsequent differential privacy processing.

[0165] This embodiment maps the relationship confidence matrix at the time window level to the time bucket level, realizing the transfer of window-level association information to the time bucket level fusion granularity. On this basis, the reliability of the device's own data and the consistency of the change trend between devices are organically integrated, so that the fusion weight can simultaneously reflect the data quality of a single device and the association strength of multiple devices. The resulting multi-dimensional fusion record provides a unified, reliable, and traceable multi-source fusion data foundation for the subsequent differential privacy output unit, enabling differential privacy processing to dynamically allocate the privacy budget based on the fusion weight, realizing a complete closed loop from data completion and association analysis to fusion weight calculation and differential privacy output.

[0166] Furthermore, this embodiment provides a step for adding noise to multidimensional fusion records to obtain a fusion result sequence and a privacy audit log, including:

[0167] Request the allocation of a second total privacy budget from the privacy budget management module;

[0168] Based on the proportion of the time bucket fusion weight of each time bucket to the sum of the fusion weights of all pruned time buckets, the total amount of the second privacy budget is dynamically allocated to each time bucket to obtain the second privacy budget for each time bucket.

[0169] Based on the second privacy budget for each time bucket, noise is added to the noisy data values ​​of each device using the Laplace mechanism to obtain the fused data values ​​of each device in each time bucket.

[0170] The fused data values ​​of each device in each time bucket are integrated with the unperturbed boundary inference values, granular constraint values, time bucket-level relationship confidence matrix and fusion weights to obtain the fusion result sequence;

[0171] A privacy audit log is generated synchronously, recording the first privacy budget allocation record and the second privacy budget allocation record.

[0172] Specifically, such as Figure 4 As shown, after generating the multidimensional fused record, the differential privacy output unit requests the privacy budget management module to allocate a second privacy budget. Based on the fusion weight within each time bucket, the second privacy budget is dynamically allocated: time buckets with higher fusion weights receive more privacy budget to reduce noise disturbances and ensure data availability; time buckets with lower overall fusion weights receive less privacy budget to enhance privacy protection. This dynamic allocation is based on the proportion of each time bucket's fusion weight to the sum of the fusion weights of all clipped time buckets. The second privacy budget is distributed to each time bucket according to this proportion, resulting in the second privacy budget for each time bucket. Here, all clipped time buckets specifically refer to the valid time buckets participating in this fusion process after being clipped from the fusion time range, excluding invalid time buckets that do not cover the fusion time range, ensuring that the privacy budget is allocated only within the valid fusion range.

[0173] Then, according to the second privacy budget of each time bucket, the noisy data values ​​of each device are perturbed using a Laplace mechanism. Specifically, the scale parameter of the corresponding Laplace noise is determined based on the second privacy budget of each time bucket. This scale parameter is the reciprocal of the second privacy budget of each time bucket. A noise variable following a Laplace distribution is generated based on this scale parameter, and the noise is superimposed on the noisy data values ​​of each device to obtain the fused data value of each device in that time bucket. This step is used to perform differential privacy perturbation on the noisy data values ​​of each device in each time bucket. After the perturbation is completed, the remaining unperturbed information in each time bucket, including boundary inference values ​​or granularity constraint values, time bucket-level relationship confidence matrix, and fusion weights, are retained and integrated together with the fused data values ​​of each device in that time bucket to constitute the fusion result of that time bucket.

[0174] The fusion results of all time buckets are integrated in chronological order to form a complete sequence of fusion results.

[0175] It should be noted that differential privacy perturbation is only performed on the noisy data values ​​of each device in the multidimensional fusion record. This is because the noisy data values ​​come from the original collection information of each data source device, which is sensitive business data and poses a risk of privacy leakage. The boundary inference values, granular constraint values, time bucket-level relationship confidence matrix and fusion weights of each device are all data credibility, correlation degree and fusion control indicators, which do not contain the original sensitive information of the devices, so there is no need to apply noise perturbation.

[0176] Simultaneously, a privacy audit log is generated during the disturbance process. This privacy audit log includes a generation timestamp, an edge-side first privacy budget allocation record, and a cloud-side second privacy budget allocation record. The edge-side first privacy budget allocation record includes the first privacy budget allocated to each device and each batch, as well as a multidimensional Gaussian noise mechanism. This multidimensional Gaussian noise mechanism is generated based on the temporal covariance matrix of the cleaned batch data, and is a correlation-preserving multidimensional Gaussian noise proportional to the temporal covariance matrix. Its generation logic is consistent with the mechanism used for local noise addition at the edge. Only the mechanism type is recorded, without involving specific covariance matrices, sensitivity, or other sensitive parameters. The cloud-side second privacy budget allocation record includes the second privacy budget allocated to each time bucket and the noise scale parameter corresponding to the Laplace mechanism.

[0177] Ultimately, the fusion result sequence and privacy audit logs are obtained, which preserve complete credibility and correlation information while protecting data privacy, supporting the credible analysis of downstream tasks.

[0178] This embodiment dynamically allocates privacy budgets based on the proportion of each time bucket's fusion weight to the sum of all fusion weights of the pruned time buckets. This enables on-demand allocation of privacy budgets, with time buckets having higher fusion weights and stronger data credibility receiving more budgets and less noise, thus maximizing the availability of high-value data while protecting privacy. Simultaneously, it fully preserves unperturbed information such as boundary inference values, granular constraint values, relationship confidence matrices, and fusion weights, ensuring that the fusion results, while meeting differential privacy protection requirements, still contain complete data quality assessment and correlation analysis metadata. This provides an interpretable and traceable data foundation for downstream tasks. The synchronously generated privacy audit logs fully record the first privacy budget allocation records for each batch on each edge device and the second privacy budget allocation records for each time bucket in the cloud, enabling traceability of privacy consumption throughout the entire process from data collection to result publication, effectively supporting compliance verification and auditing requirements.

[0179] Furthermore, this embodiment provides a step for dynamically allocating a privacy budget based on data quality, including:

[0180] Receive the first privacy budget request from the edge device and allocate the first privacy budget according to the preset budget allocation strategy;

[0181] Receive the second privacy budget total request from the cloud and calculate the global remaining privacy budget as the second privacy budget total.

[0182] Specifically, the privacy budget management module receives the first privacy budget request from the edge device. This request includes the data source device identifier, batch identifier, and data quality score. Each batch independently requests and allocates its first privacy budget. According to a preset budget allocation strategy, the first privacy budget is allocated to each device and each batch. The preset budget allocation strategy sets a base allocation amount for the first privacy budget. The preferred value is 0.1. This basic allocation is the lower limit for the privacy budget allocated to each device per batch, ensuring that each batch of data receives at least a minimum level of privacy protection. This is based on the data quality score reported by the device. Calculate the quality adjustment factor , ,in, This is a preset quality sensitivity coefficient. Therefore, the smaller the data quality score, the larger the quality adjustment coefficient, and the more privacy budget is allocated to provide stronger privacy protection to mask the errors in the data itself. Thus, the initial privacy budget for this batch of devices is... And return it to the edge module.

[0183] For example, the preset quality sensitivity coefficient is related to the device type. For devices with large data fluctuations, such as photovoltaic inverters, Take the larger value, preferably 0.8; for devices with relatively stable data, such as smart meters, Take a moderate value, with 0.5 being the preferred value; for equipment with relatively stable data changes, such as weather stations, Choose the smaller value, preferably 0.3.

[0184] The privacy budget management module receives the second privacy budget application from the fusion center module. The privacy budget management module calculates the global remaining privacy budget, which is the total global privacy budget minus the sum of the first privacy budgets of all devices and all batches in this fusion task. The total global privacy budget is a preset value, preferably 1.0, but this embodiment does not limit it.

[0185] This embodiment receives the first privacy budget request from the edge device and allocates independent and differentiated budgets for each device and each batch. The preset budget allocation strategy uses a basic allocation amount as the lower limit and maps the data quality score to the privacy budget through a quality adjustment coefficient. The worse the data quality, the more budget is allocated to it, which strongerly masks data errors with noise. At the same time, the quality sensitivity coefficient is linked to the device type, so that devices with large fluctuations receive more budget and stable devices receive less budget, thus achieving differentiated allocation of privacy protection strength. The total amount of the second privacy budget in the cloud is calculated by subtracting the sum of all allocated first privacy budgets from the global total privacy budget, ensuring that the privacy consumption of the entire fusion task is always controlled within the preset upper limit. This provides budget resource guarantee for differential privacy protection throughout the entire process from data collection to result release, and realizes closed-loop management and controllable consumption of the privacy budget.

[0186] Example 2:

[0187] like Figure 5 As shown, this embodiment of the invention provides a multi-source energy data fusion method based on differential privacy, including:

[0188] The upload data packets covering the fusion time range of the edge module are obtained, and the structured data of the noise-added data packets are extracted. The upload data packets contain one or more noise-added data packets, and the noise-added data packets contain noise-added batch data, device identifier, batch identifier, timestamp range and the value of M, where M is the batch data length.

[0189] Based on structured data, timestamps from various data source devices are mapped to a unified time bucket index to generate unified timeline data. Combined with the fused time range, the clipped unified timeline data is obtained.

[0190] Boundary missing data inference, multi-granularity transformation and temporal fingerprint verification are performed on the clipped unified timeline data to obtain boundary inference values, granularity constraint values ​​and relation confidence matrix;

[0191] The cropped unified timeline data, boundary inference values, granularity constraint values, and relationship confidence matrix are merged to obtain a multidimensional fused record.

[0192] Add differential privacy noise to the multidimensional fusion record, and output the fusion result sequence and privacy audit log.

[0193] Specifically, taking a single fusion task as an example, the fusion command unit of the fusion center module sends a fusion command containing the time range of this fusion to the edge modules. After receiving the fusion command, each edge module selects a complete batch whose timestamp range is completely covered from its local cache. For each batch of data, data cleaning is performed sequentially, a data quality score is calculated, and a first privacy budget is requested from the privacy budget management module. Based on the allocated first privacy budget, correlation-preserving noise proportional to the temporal covariance of the batch data is generated, and noisy batch data is formed after adding noise. One or more noisy batch data corresponding to the complete batch are encapsulated into an upload data packet and uploaded to the fusion center module.

[0194] The fusion command unit of the fusion center module receives data packets uploaded from each edge device, decapsulates them, and extracts the noisy batch data, device identifier, batch identifier, timestamp range, and M-value. It verifies the continuity of each device batch identifier and the completeness of the timestamp range's coverage of the fusion time range. After successful verification, the structured data is transmitted to the alignment unit.

[0195] After receiving the structured data, the alignment unit standardizes the timestamps and maps them to a preset unified time bucket index. Timestamps located in the bucket boundary area are simultaneously stored in adjacent time buckets and boundary markers are added. Within each bucket, a representative point closest to the bucket center time is selected for each device to generate unified timeline data. Based on the fusion time range, time buckets that intersect with the fusion time range are retained, resulting in truncated unified timeline data, which is then passed to the fusion preparation unit.

[0196] The fusion preparation unit divides devices into high-frequency and low-frequency devices based on their sampling intervals. For high-frequency devices, an asynchronous tolerance window mechanism is used: it iterates through each time bucket, identifies missing buckets, searches for neighboring time buckets with data within the tolerance window, fills the missing data using linear interpolation or neighboring values, and outputs boundary inference values. For low-frequency devices, it identifies buckets to be filled, fills in the missing data using linear interpolation or proportional allocation based on the device type, and outputs granular constraint values. A time series window is formed by multiple time buckets, and multiple time series windows are generated by sliding. The sequence of changes in the direction of change of each device within the time series window is extracted as a time series fingerprint. The fingerprint matching degree between device pairs is calculated to form a relationship confidence matrix with windows as rows and device pairs as columns.

[0197] The multidimensional fusion unit maps the temporal window-level relationship confidence matrix to the time bucket level, obtaining the relationship confidence of each device pair within each time bucket. For each device within each time bucket, a boundary inference value or granular constraint value is used as the data confidence factor. High-frequency devices use the boundary inference value, while low-frequency devices use the granular constraint value. The arithmetic mean of the relationship confidence of this device with other devices is used as the association confidence factor, and the two are multiplied to obtain the fusion weight. The bucket start time, noisy data values ​​of each device, boundary inference value / granular constraint value, inter-device relationship confidence, and fusion weight are integrated to form the multidimensional fusion record for that time bucket. All truncated time buckets are traversed to generate a complete multidimensional fusion record sequence.

[0198] The differential privacy output unit requests a second privacy budget from the privacy budget management module. The total second privacy budget is the global total privacy budget minus the sum of all allocated first privacy budgets. Based on the proportion of each time bucket's fusion weight to the sum of all pruned time bucket fusion weights, the total second privacy budget is dynamically allocated to each time bucket. For each time bucket, a Laplace mechanism is used to add noise to its noisy data values, retaining the remaining undisturbed information to form a fusion result. The fusion results of all pruned time buckets are integrated into a fusion result sequence in chronological order. Simultaneously, a privacy audit log is generated, recording the first and second privacy budget allocation records. The fusion result sequence and privacy audit log are then output together.

[0199] Through the above steps, this embodiment completes the entire process from edge data acquisition and noise-added uploading, to cloud verification, alignment, completion, and fusion, and finally to differential privacy output and auditing, outputting a fusion result sequence that meets differential privacy protection requirements and a complete privacy audit log. The specific implementation steps are as described in Embodiment 1, and will not be repeated in this embodiment.

[0200] This embodiment collects and cleans data in batches at the edge, strictly aligning the noise covariance with the temporal covariance of the batch data to ensure that the noisy data retains the original temporal structure. In the cloud, multi-source data alignment is achieved based on a unified time bucket index. Boundary overlap processing eliminates attribution misalignment caused by clock skew. Asynchronous tolerance windows are used to fill high-frequency gaps based on device sampling frequency differences, and granular transformation based on data characteristics is used to fill low-frequency gaps. A sliding window extracts the change direction sequence as a temporal fingerprint to quantify the reliability of inter-device associations. Finally, data reliability and association reliability are merged into weights, dynamically allocating differential privacy budgets and adding noise to sensitive data before release, simultaneously generating traceable privacy audit logs. This fundamentally solves the technical problems of low privacy budget utilization, destruction of temporal characteristics, and difficulty in aligning multi-source data in existing technologies, achieving synergistic optimization of privacy protection and data utility.

[0201] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0202] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0203] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0204] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-source energy data fusion system based on differential privacy, characterized in that, This includes an edge module, a convergence center module, and a privacy budget management module; The edge module is used to filter out complete batches according to the fusion instructions, clean each batch of data in the complete batch using temporal correlation, add correlation maintenance noise proportional to the temporal covariance of the batch data, generate noisy batch data, and encapsulate and upload one or more noisy batch data corresponding to the complete batch. The step of adding correlation-preserving noise proportional to the temporal covariance of the batch data to generate noisy batch data includes: Request the allocation of the first privacy budget from the privacy budget management module; Based on the first privacy budget and the temporal covariance matrix of the cleaned batch data, correlation-preserving noise is generated to obtain multidimensional Gaussian noise. The multidimensional Gaussian noise is added to the cleaned batch data one by one to generate noisy batch data; The fusion center module includes a fusion instruction unit, an alignment unit, a fusion preparation unit, a multi-dimensional fusion unit, and a differential privacy output unit; The fusion instruction unit is used to issue unified fusion instructions, receive and verify the validity of data packets uploaded by each data source device, and obtain structured data. Alignment unit, used to map timestamps from each data source to a unified time bucket index based on the structured data, to obtain clipped unified timeline data; The fusion preparation unit is used to perform boundary missing inference, multi-granularity transformation and temporal fingerprint verification on the clipped unified timeline data to obtain boundary inference values, granularity constraint values ​​and relation confidence matrix. The process of performing boundary missing inference, multi-granularity transformation, and temporal fingerprint verification on the cropped unified timeline data yields boundary inference values, granularity constraint values, and a relation confidence matrix, including: Traverse all devices on the unified timeline data after clipping, and divide the devices into high-frequency devices and low-frequency devices based on the comparison results of the device sampling interval and the target time bucket granularity; For high-frequency devices, an asynchronous tolerance window mechanism is used to infer missing data from adjacent time buckets within the preset asynchronous tolerance window duration to obtain boundary inference values; For low-frequency devices, the missing data in the middle time bucket is filled in according to the time buckets with valid data points in the adjacent time buckets to obtain the granularity constraint value; Extract the temporal fingerprints of each data source device on the unified timeline after clipping, calculate the fingerprint matching degree between different data source devices, and obtain the relationship confidence matrix; The multidimensional fusion unit is used to merge the clipped unified timeline data, boundary inference values, granular constraint values ​​and relation confidence matrix to generate a multidimensional fusion record in which each time bucket contains multi-source information. The differential privacy output unit is used to add noise to the multidimensional fusion record to obtain a fusion result sequence and a privacy audit log. The privacy budget management module is used to dynamically allocate the privacy budget based on data quality.

2. The multi-source energy data fusion system based on differential privacy according to claim 1, characterized in that, The process of mapping timestamps from various data sources to a unified time bucket index based on the structured data to obtain pruned unified timeline data includes: Based on the structured data, the timestamp range and the time series associated with the value of M are standardized to obtain the timestamp corresponding to each noisy data value, where M is the batch data length; Map each noisy data value, device identifier, and corresponding timestamp to the corresponding time bucket; Within each time bucket, a representative point is selected for each device to generate unified timeline data; The unified timeline data is cropped according to the fusion time range to obtain cropped unified timeline data.

3. The multi-source energy data fusion system based on differential privacy according to claim 1, characterized in that, The step of mapping timestamps from various data sources to a unified time bucket index based on the structured data to obtain pruned unified timeline data also includes: For each time bucket, perform boundary region processing on the data points, and store the data points corresponding to the timestamps of the boundary regions into the adjacent time buckets.

4. The multi-source energy data fusion system based on differential privacy according to claim 3, characterized in that, For low-frequency devices, the missing data in the middle time bucket is filled according to the adjacent time buckets with valid data points to obtain the granularity constraint value, including: Iterate through each time bucket of the low-frequency device's unified timeline data after clipping to obtain the bucket to be filled; For a filling interval consisting of multiple consecutive buckets to be filled, find the time buckets with representative points before and after the filling interval, and record them as the forward reference point and the backward reference point. Based on the equipment type, the data characteristics are identified. For status data, linear interpolation is used, and for cumulative data, time ratio allocation is used to calculate the noise-added data value of the bucket to be filled. The calculated noisy data values ​​are filled into the corresponding time buckets, and the granularity constraint values ​​are obtained.

5. A multi-source energy data fusion system based on differential privacy according to claim 4, characterized in that, The step involves extracting the temporal fingerprints of each data source device on a unified timeline after cropping, calculating the fingerprint matching degree between different data source devices, and obtaining a relationship confidence matrix, including: Continuous A time bucket constitutes a time sequence window, and multiple time sequence windows are generated by sliding sequentially along a unified time axis. For each timing window, extract the timing fingerprint for each data source device within the timing window; All data source devices are paired up, and for each pair of devices, the matching degree of the time fingerprint of the corresponding device pair is calculated within the current time window. The matching degree is used as the relationship confidence degree of the corresponding device pair under the current time window, forming a relationship confidence matrix with the time window as the row and the device pair as the column.

6. A multi-source energy data fusion system based on differential privacy according to claim 5, characterized in that, The process of merging the cropped unified timeline data, boundary inference values, granularity constraint values, and relationship confidence matrix to generate a multi-dimensional fused record where each time bucket simultaneously contains information from multiple sources includes: For each time bucket, collect the noise-added data values, boundary inference values, and granularity constraint values ​​of all devices within the corresponding time bucket; Map the relationship confidence matrix to the current time bucket to generate a time bucket-level relationship confidence matrix; The data confidence factor is determined based on the boundary inference value or granularity constraint value corresponding to each device. The association confidence factor is determined based on the average confidence of the relationship between the corresponding device and other devices. The data confidence factor and the association confidence factor are multiplied to obtain the fusion weight of the corresponding device in the current time bucket. The data is integrated according to a preset structured format to form a multi-dimensional fused record corresponding to each time bucket.

7. A multi-source energy data fusion system based on differential privacy according to claim 6, characterized in that, The step of adding noise to the multidimensional fusion record to obtain the fusion result sequence and privacy audit log includes: Request the allocation of a second total privacy budget from the privacy budget management module; Based on the proportion of the fusion weight of each time bucket to the sum of the fusion weights of all pruned time buckets, the total amount of the second privacy budget is dynamically allocated to each time bucket to obtain the second privacy budget for each time bucket. Based on the second privacy budget of each time bucket, noise is added to the noisy data values ​​of each device using the Laplace mechanism to obtain the fused data values ​​of each device in each time bucket; The fused data values ​​of each device in each time bucket are integrated with the unperturbed boundary inference values, granular constraint values, time bucket-level relationship confidence matrix and fusion weights to obtain the fusion result sequence; A privacy audit log is generated synchronously, recording the first privacy budget allocation record and the second privacy budget allocation record.

8. A multi-source energy data fusion system based on differential privacy according to claim 7, characterized in that, The dynamic allocation of privacy budget based on data quality includes: Receive the first privacy budget request from the edge device and allocate the first privacy budget according to the preset budget allocation strategy; Receive the second privacy budget total request from the cloud and calculate the global remaining privacy budget as the second privacy budget total.

9. A method for multi-source energy data fusion based on differential privacy, implemented based on the multi-source energy data fusion system based on differential privacy as described in any one of claims 1-8, characterized in that, include: The upload data packets covering the fusion time range of the edge module are obtained, and the structured data of the noise-added data packets are extracted. The upload data packets contain one or more noise-added data packets, and the noise-added data packets contain noise-added batch data, device identifier, batch identifier, timestamp range and value of M, where M is the batch data length. Based on the structured data, the timestamps of each data source device are mapped to a unified time bucket index to generate unified timeline data. Combined with the fused time range, the clipped unified timeline data is obtained. Boundary missing inference, multi-granularity transformation and temporal fingerprint verification are performed on the cropped unified timeline data to obtain boundary inference values, granularity constraint values ​​and relation confidence matrix; The cropped unified timeline data, the boundary inference values, the granularity constraint values, and the relationship confidence matrix are merged to obtain a multidimensional fused record; Differential privacy noise is added to the multidimensional fusion record, and the fusion result sequence and privacy audit log are output.

Images