Park electric power data privacy protection and recovery device and method based on space-time federation hidden feature learning

Abstract

The invention discloses a space-time federation hidden feature learning-based park power data privacy protection and recovery device and method in the field of power grid incomplete power data recovery. The device comprises a server and a plurality of client devices, each client device particularly comprises a spatial-temporal feature extraction module and a power data module, and each power data module particularly comprises a spatial-temporal feature correlation matrix storage unit and a hidden feature vector storage unit of other clients in the same park. According to the method, a spatial Laplacian matrix and a time difference matrix are constructed, and the hidden features of the power data are explicitly and structurally fused into a model, so that higher power data recovery precision is obtained. The problem that in the prior art, priori knowledge is difficult to use, so that performance in data recovery is limited is solved, it can be ensured that power data recovery is accurately achieved under privacy protection, and recovery precision is effectively improved.

Description

Technical Field

[0001] This application relates to the field of campus power data privacy protection and recovery, and is particularly applicable to a campus power data privacy protection and recovery device and method based on spatiotemporal federated implicit feature learning. Background Technology

[0002] In the era of big data, large-scale sensor networks (SNs) have been widely applied in various scenarios, playing a crucial role, especially in power systems. With the comprehensive deployment of smart meters, energy efficiency monitoring terminals, distributed energy monitoring systems, and integrated energy management platforms in industrial parks, user-side power data has experienced explosive growth. However, to reduce data storage costs and monitoring overhead, intelligent algorithms are often used to dynamically start and stop some sensors, leading to intermittent data collection. Furthermore, different types of parks, such as industrial parks, science parks, and agricultural parks, exhibit significant differences in energy consumption structures, load characteristics, response capabilities, and service requirements, further exacerbating the sparsity of data semantics and the complexity of business scenarios. Therefore, power data for industrial parks inherently possesses the typical characteristic of incompleteness.

[0003] Sensor networks (SNs) typically contain rich and valuable latent knowledge. Because sensor nodes are usually located in fixed positions and sample at predetermined time intervals, their observation data exhibit significant correlations in both spatial and temporal dimensions: readings from adjacent nodes are highly similar spatially, while observations of the same node at adjacent times show smooth temporal changes. For a single monitoring indicator, its data can be naturally organized into a sensor-time matrix, where each row corresponds to a sensor node and each column corresponds to a time point. Benefiting from temporal smoothness and the redundancy of sensor deployment, this matrix typically exhibits low-rank properties and a strong spatiotemporal dependency structure. Latent Feature Analysis (LFA) is a typical representation learning method. Its core idea is to model the SN data as an incomplete sensor-time matrix, explicitly fusing spatiotemporal context information during the modeling process, and using matrix factorization techniques to learn its low-rank approximation, thereby effectively mining hidden latent features and reconstructing missing observations.

[0004] With increasing awareness of data privacy and the implementation of regulations such as the European General Data Protection Regulation (GDPR) and Chinese laws on data security protection, industrial parks are generally unwilling to share raw electricity consumption data. This results in discrete and localized SN data, making traditional LFA methods, which rely on global data, difficult to apply. Against this backdrop, achieving high-precision representation learning while ensuring privacy has become a key challenge. For example, the FedPerGNN proposed in the June 2022 paper "Afederated graph neural network framework for privacy-preserving personalization" published in the British journal *Nature Communications* achieves a good balance between privacy protection and model performance, providing an important solution for securely utilizing graph data for personalized services in data silos. Some solutions are based on federated learning (FL). While learning-based methods offer an effective path for privacy-preserving modeling of distributed data, existing frameworks rely on the low-rank nature of tensor decomposition to indirectly capture correlations in the data. This limits their ability to model the inherent strong spatiotemporal correlations in sensor data, making it difficult to fully utilize prior knowledge of spatial proximity and temporal continuity between nodes. Consequently, performance is limited in tasks such as data recovery, and the recovered data may not be smooth enough in the spatiotemporal dimension, ignoring strong correlations between neighboring sensors or adjacent time points. Furthermore, some solutions use hybrid padding strategies that introduce noise or bias during training, especially when the model is inaccurate in the early stages of iteration. This may interfere with the correct direction of gradients, thus affecting the final model convergence quality and recovery accuracy. On the other hand, some solutions use three-dimensional tensor models, introducing unnecessary complexity, which is "overkill" for the power load data recovery problem.

[0005] In conclusion, accurately representing the network's serial number (SN) while protecting privacy is a key challenge in current power data recovery for industrial parks. Summary of the Invention

[0006] In view of this, and in response to the problems of the prior art, this invention provides a privacy-preserving recovery device and method for campus power data based on spatiotemporal federated latent feature learning. The aim is to design a latent feature analysis (LFA) model that is globally shared by different distributed nodes under the federated learning (FL) framework, thereby achieving privacy protection of the original data. At the same time, spatiotemporal smoothness constraints are incorporated into the modeling of each distributed campus node to improve the model's representational capability. Ultimately, this achieves the recovery of distributed campus power data under privacy protection.

[0007] To achieve the above objectives, this application provides a campus power data privacy protection and recovery device based on spatiotemporal federated latent feature learning, comprising a server and M client devices, which communicate with each other via a network. The server includes a main data storage module and a main federated learning module. Each client device includes a power data receiving module, a spatiotemporal feature extraction module, a data recovery module, a power data module, and a slave federated learning module. The power data module includes a power load data storage unit, a spatiotemporal feature correlation matrix storage unit, a latent feature vector storage unit for other clients in the same campus, a recovery data storage unit, and a server data storage unit, where M is a positive integer greater than 1. The main data storage module is used to receive and store the federated learning gradient data sent by the server data storage unit in the client device, and to store and send the latent feature matrix data calculated by the main federated learning module to the client device. The main federated learning module updates the latent feature matrix data in the server according to the federated learning gradient data and stores it in the main data storage module. The data receiving module receives incomplete power load data from power data acquisition equipment installed in the power grid within the jurisdiction of the client device, and instructs the power data storage unit in the power data module to store the incomplete power load data. The spatiotemporal feature extraction module receives the physical coordinates of equipment in the same area, calculates the spatiotemporal feature matrix, and stores it in the spatiotemporal feature correlation matrix storage unit. Preferably, the data recovery module uses the incomplete power load data stored in the power data storage unit of the power data module, the latent feature matrix data stored in the server data storage unit, the spatiotemporal feature matrix stored in the spatiotemporal feature correlation matrix storage unit, and the latent feature vectors stored in the latent feature vector storage units of other clients in the same area to perform the incomplete power load data recovery process, and stores the recovered data in the recovery data storage unit of the power data module. The federated learning module uses the incomplete power load data, latent feature matrix data, and recovered data in the power data module to calculate the federated learning gradient data and stores it in the server data storage unit.

[0008] Based on the same inventive concept, this application also provides a method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning, using the aforementioned campus power data privacy protection and recovery device based on spatiotemporal federated implicit feature learning, including the following steps: S1. The main federated learning module in the server initializes the latent feature matrix data and federated learning parameters, and stores them in the main data storage module; S2. The client device spatiotemporal feature extraction module calculates the spatiotemporal feature correlation matrix of the power load data based on the physical coordinates of the monitoring sensors for the power load data in the park. S3. The client device initializes the latent feature vector and obtains the latent feature vectors of other nodes in the same park, and stores them in the recovery data storage unit; S4. The main data storage module sends the hidden feature matrix data to all client devices and stores it in the server data storage unit; S5. The power data receiving module receives incomplete power load data fed back by the power data acquisition equipment installed in the power grid and stores it in the power data storage unit in the power data module. S6. The data recovery module uses the latent feature analysis method to calculate the approximate power load data vector and stores it in the recovery data storage unit; S7. Based on the original power load data of the node, the spatiotemporal feature correlation matrix of the park, and the latent feature vectors of other nodes in the park, the data recovery module updates the latent feature vectors and calculates the federated learning gradient data. S8. The client device sends federated learning gradient data to the server device; S9. The server updates the latent feature matrix data using federated learning gradient data; S10. Repeat steps S4 to S9 until the maximum number of iterations is reached, i.e., the client device has completed the data recovery under privacy protection.

[0009] Furthermore, the latent feature matrix data mentioned in step S1 is a column latent feature matrix. , is randomly generated during initialization; among which, N This represents the time series number of the power load data. d is the dimension of the latent features, and is a positive number.

[0010] Furthermore, the federated learning parameters include: maximum number of iterations. Learning rate and latent feature regularization coefficients .

[0011] Furthermore, the latent feature vector mentioned in step S3 is in M This represents the sum of the number of sensors across all parks.

[0012] Preferably, step S2 includes the following steps: S201. The spatiotemporal feature extraction module receives an instruction to calculate the spatiotemporal feature correlation matrix; S202. The spatiotemporal feature extraction module reads the physical coordinates of the power sensor from the receiving module; S203. The spatiotemporal feature extraction module calculates spatial features based on the physical coordinates of relevant sensors in the park. First, it calculates the Euclidean distance between sensor coordinates and constructs a distance matrix. OAccording to the KNN algorithm, the distances to the five nearest sensors for each sensor are selected to form a distance matrix. O The values ​​in the remaining positions are set to 0. The weight matrix X is calculated based on the distance matrix O. The weight matrix calculation function is: ,in Representative at the h The park O The first in the matrix i The sensor and the first j The Euclidean distance between the sensors is S204. Calculate the Laplace matrix based on the weight matrix. The Laplace matrix has the following form: ,in, , i ∈{1, 2, ..., C h} represents the weight matrix of the first... i The sum of lines, C h To determine the number of sensors in the h-th zone, S205. Construct the time characteristic relationship, difference matrix. G The format is as follows: , in, N This is the time series number of the power load data.

[0013] Preferably, the data recovery module in step S6 calculates an approximate power load data vector using a latent feature analysis method, including the following steps: S601. The data recovery module received a data recovery command; S602. According to the latent feature analysis method, for the first... h The park is located in the first i A sensor in its known data set Modeling is performed, and its target loss function is constructed by combining spatiotemporal feature relationships. as follows: ,in, Representing the h The park is located in the first i The sensor at the first j The actual value observed at each timestamp and These are the aggregate weights that control the influence of L1 and L2 paradigms, and their specific calculation method is as follows: in, t This represents the number of iteration rounds. σ It is a balancing coefficient used to balance the L1 norm and L2 norm.

[0014] Preferably, step S7 includes the following steps: S701. The data recovery module receives an instruction to update the latent feature vector.

[0015] S702. The objective function is transformed as follows: Calculate its gradient, which is done as follows:

[0016] S703. Traverse all observed data and update the local latent feature vector. .

[0017] S704. Repeat steps S702 and S703 until the local latent feature vector change is within the set threshold range.

[0018] Furthermore, the method for updating the latent feature matrix data using federated learning gradient data by the server in step S9 is as follows: traverse the latent feature dimensions, according to... Update.

[0019] The beneficial effects of this application are as follows: This application provides a privacy-preserving recovery device and method for campus power data based on spatiotemporal federated latent feature learning. First, a federated learning framework for a campus power sensor network (SN) is designed, enabling different distributed nodes to efficiently collaborate and train using a globally shared LFA model while maintaining the privacy of the original data. Then, nodes are divided according to the campus, sharing information from other nodes for accurate power load data recovery while protecting data privacy. Finally, spatiotemporal information is incorporated into the modeling of each distributed node, further improving the model's representational capability. The client in this application only uploads gradient data and does not share the original data. This ensures more accurate power data recovery under privacy protection. This scheme treats power data as a two-dimensional "sensor-time" matrix and learns its low-rank latent feature representation. Through spatiotemporal feature modeling, the Laplacian matrix in the spatial dimension and the difference matrix in the temporal dimension are explicitly introduced as regularization terms to constrain the model to capture the smoothness and correlation of data in spatiotemporal space. The loss function and optimization calculate the gradient piecewise according to the positive and negative errors, resulting in more refined calculations. By explicitly and structurally incorporating the prior knowledge of "similar readings from nearby sensors" and "smooth changes in data at adjacent time points" inherent in power data into the model, higher recovery accuracy and better smoothness are achieved. Moreover, this solution is specifically designed for the "park" scenario, and its spatial feature extraction based on physical coordinate KNN graphs is very suitable for the characteristics of relatively concentrated sensor deployment and clear location relationships within the park. Attached Figure Description

[0020] To illustrate the objectives and technical solutions of this invention, the following figures are provided: Figure 1 The flowchart shows the campus power data privacy protection and recovery method based on spatiotemporal federated implicit feature learning according to the present invention. Figure 2 This is a schematic diagram of the multi-campus power system in Embodiment 1 of the present invention; Figure 3 This is a framework diagram of federated learning in Embodiment 1 of the present invention; Figure 4 This is a framework diagram of the campus power data privacy protection and recovery device based on spatiotemporal federated implicit feature learning in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram illustrating the working principle of data recovery for client devices in the same park in Embodiment 2 of the present invention; Among them, 1-server, 11-main data storage module, 12-main federated learning module, 2-client device, 2a-first client, 2m-mth client, 2M-Mth client, 21-power data receiving module, 22-spatiotemporal feature extraction module, 23-data recovery module, 24-power data module, 241-power load data storage unit, 242-spatiotemporal feature correlation matrix storage unit, 243-latent feature vector storage unit for other clients in the same park, 244-recovery data storage unit, 245-server data storage unit, 25-from federated learning module, A-first park, B-second park, C-third park. Detailed Implementation

[0021] To make the objectives and technical solutions of this application clearer, the application will be described in detail below with reference to the accompanying drawings and embodiments.

[0022] Example 1: Combination Figure 2 There are multiple power parks whose power systems are monitored as a whole by a server. These parks are divided into three sub-parks: Park A, Park B, and Park C. Each sub-park has a number of power load sensors. Assume that the number of power load sensors in each sub-park is different, and then data is exchanged with the cloud server using power line carrier communication.

[0023] Due to environmental factors, equipment short circuits, and other reasons, power load data collected by sensors on various clients is missing within specific time slots. However, considering user privacy and security, data from different parks and servers is no longer shared between clients. To recover the missing data of the park while protecting privacy, this application provides a park power data privacy protection recovery device based on spatiotemporal federated implicit feature learning.

[0024] Combination Figure 3 The device specifically consists of one server 1 and M client devices 2. Server 1 corresponds to a cloud server, and client devices 2 correspond to clients in the park: for example, the first client 2a, the m-th client 2m, and the M-th client 2M. They communicate with each other via a wireless network connection. Note that logically, one sensor is one client device 2, and there are a total of M devices.

[0025] Combination Figure 4 The server 1 includes: a main data storage unit 11 and a main federated learning module 12; the client device 2 includes: a power data receiving module 21, a spatiotemporal feature extraction module 22, a data recovery module 23, a power data module 24, and a slave federated learning module 25; the power data module 24 includes a power load data storage unit 241, a spatiotemporal feature correlation matrix storage unit 242, a latent feature vector storage unit 243 for other clients in the same park, a recovery data storage unit 244, and a server data storage unit 245.

[0026] Preferably, the main data storage module 11 is used to receive and store the federated learning gradient data sent by the server data storage unit 245 in the client device 2, and to store and send the latent feature matrix data calculated by the main federated learning module 12 to the client device 2.

[0027] The main federated learning module 12 updates the latent feature matrix data in the server according to the federated learning gradient data and stores it in the main data storage module 11.

[0028] The power data receiving module 21 is used to receive incomplete power load data fed back by the power data acquisition equipment installed in the power grid under the jurisdiction of the client device, and to instruct the power data storage unit 241 in the power data module to store the incomplete power load data.

[0029] The spatiotemporal feature extraction module 22 is used to receive the physical coordinates of devices in the same park, calculate the spatiotemporal feature matrix, and store it in the spatiotemporal feature correlation matrix storage unit 242.

[0030] The data recovery module 23 utilizes the power data storage unit 241 in the power data module 24 to store incomplete power load data, the hidden feature matrix data stored in the server data storage unit 245, the spatiotemporal feature matrix stored in the spatiotemporal feature correlation matrix storage unit 242, and the hidden feature vectors stored in the hidden feature vector storage unit 243 of other clients in the same park to perform the incomplete power load data recovery process of the power grid, and stores the recovered data in the recovery data storage unit 244 in the power data module 24.

[0031] The federated learning module 25 uses the incomplete power load data, latent feature matrix data and recovery data in the power data module 24 to calculate the federated learning gradient data and store it in the server data storage unit 245.

[0032] Example 2: Based on the device of Example 1, this application also provides a campus power data privacy protection recovery method based on spatiotemporal federated implicit feature learning, so as to realize the recovery of missing data in the transformer area under privacy protection. Campus A has multiple client devices, of which one is client 1 and the other is client C1.

[0033] Combination Figure 1 Specifically, it includes the following steps: Step S1. The main federated learning module 12 in server 1 initializes the latent feature matrix data and federated learning parameters, and stores them in the main data storage module 11.

[0034] The latent feature matrix data is a column latent feature matrix. , is randomly generated during initialization; among which, N This represents the time series number of the power load data. d is the dimension of the latent features, and is a positive number.

[0035] The federated learning parameters include: maximum number of iterations. Learning rate and latent feature regularization coefficients .

[0036] Step S2. The spatiotemporal feature extraction module 22 in the client device 2 calculates the spatiotemporal feature correlation matrix of the power load data based on the physical coordinates of the monitoring sensor for the power load data in the park, and stores it in the spatiotemporal feature correlation matrix storage unit 242.

[0037] Upon receiving the instruction to calculate the spatiotemporal feature correlation matrix, the spatiotemporal feature extraction module 22 reads the physical coordinates of the power load sensors from the receiving module. First, it calculates the Euclidean distance between the coordinates of the power load sensors and constructs a distance matrix. O Based on the KNN algorithm, a distance matrix is ​​constructed by selecting the five nearest power load sensors for each power load sensor. O The values ​​in the remaining positions are set to 0. The weight matrix X is calculated based on the distance matrix O. The function for calculating the elements of the weight matrix is: in Representative at the h The park O The first in the matrix i The sensor and the first j Euclidean distance between the sensors; Then, the Laplace matrix is ​​calculated based on the weight matrix. The Laplace matrix has the following form: ,in, , indicating the weight matrix of the first i The sum of lines, i ∈{1, 2, ..., C h}, C h Let h be the number of sensors in the h-th zone; In addition, construct the time characteristic relationship, difference matrix G The format is as follows: , in, N This represents the time series number of the power load data. Subsequently, the Laplace matrix is... S sum difference matrix G Stored in the spatiotemporal feature correlation matrix storage unit 242.

[0038] Step S3. Client device 2 initializes the hidden feature vector and obtains the hidden feature vectors of other nodes in the same park, and stores them in the recovery data storage unit 244.

[0039] Step S4. The main data storage module 11 sends the hidden feature matrix data to all client devices 2 and stores it in the server data storage unit 245.

[0040] Step S5. The power data receiving module 21 receives incomplete power load data fed back by the power data acquisition equipment installed in the power grid and stores it in the power load data storage unit 221 in the power data module 24.

[0041] Step S6. The data recovery module 23 calculates the approximate power load data vector using the hidden feature analysis method and stores it in the recovery data storage unit 244.

[0042] The data recovery module 23 uses latent feature analysis to calculate an approximate power load data vector, including: According to the latent feature analysis method, for the first h The park is located in the first i A power load sensor in its known data set Modeling is performed, incorporating spatiotemporal features. Furthermore, to enhance the model's robustness to noise and anomalies, the L1 and L2 norms are combined and embedded into the training process with adaptively adjusted weights based on error, constructing its target loss function. as follows: in, Representing the h The park is located in the firsti The sensor at the first j The actual value observed at each timestamp and These are the aggregate weights that control the influence of L1 and L2 paradigms, and their specific calculation method is as follows: in, t This represents the number of iteration rounds. σ It is a balancing coefficient used to balance the L1 norm and L2 norm.

[0043] Step S7. Based on the power load data storage unit 221, the spatiotemporal feature correlation matrix storage unit 242, the server data storage unit 245, and the latent feature vector storage unit 243 of other clients in the same park in the power data module 24, the data recovery module 23 updates the latent feature vector and calculates the federated learning gradient data.

[0044] Specifically, data recovery module 23 receives the instruction to update the latent feature vector. Considering the calculation of absolute value in the L1 paradigm, the objective function becomes:

[0045] Therefore, its gradient is calculated as follows:

[0046] Iterate through all observed data and update the local latent feature vector. Repeat the process of calculating gradients, traversing data, and updating local latent feature vectors multiple times until the changes in local latent feature vectors fall within the set threshold range.

[0047] Step S8. Client device 2 sends federated learning gradient data to server 1; Step S9. Server 1 updates the latent feature matrix data using federated learning gradient data; Step S10. Repeat S4~S9 until the maximum number of iterations. At this point, all client devices 2 have completed data recovery under privacy protection.

[0048] To facilitate the demonstration of the method of this application, the pseudocode of the operation procedures of server 1 and client device 2 is provided in Tables 1 and 2 below.

[0049] Table 1 Server-related operations

[0050] Table 2. Client Device Corresponding Operations

[0051] Furthermore, based on the daily power load data of various industrial parks in Yuhuatai District, Nanjing City, Jiangsu Province for the whole year of 2023 (January-December), some data was artificially missing, and a data recovery experiment was carried out according to this embodiment. The root mean square error (RMSE) and mean absolute error (MAE) accuracy were compared with the FedPerGNN method mentioned in the aforementioned background paper. The results are shown in Table 3. It can be seen that the method of the present invention has high data recovery accuracy under privacy protection.

[0052] Table 3 Comparison of experimental results

[0053] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A campus power data privacy protection and recovery device based on spatiotemporal federated implicit feature learning, comprising a server (1) and M client devices (2), which communicate with each other via a network, wherein M is a positive integer greater than 1; the server (1) includes a main data storage module (11) and a main federated learning module (12); each client device (2) includes a power data receiving module (21), a data recovery module (23), a power data module (24), and a slave federated learning module (25); the power data module (24) includes a power load data storage unit (241), a recovery data storage unit (244), and a server data storage unit (245); the power data receiving module (21) is used to receive incomplete power load data fed back by the power data acquisition equipment installed in the power grid under the jurisdiction of the client device (2), and instruct the power load data storage unit (241) in the power data module (24) to store the incomplete power load data; characterized in that: Each of the client devices (2) also includes a spatiotemporal feature extraction module (22), and the power data module (24) also includes a spatiotemporal feature correlation matrix storage unit (242) and a latent feature vector storage unit (243) for other clients in the same park. The main data storage module (11) is used to receive and store the federated learning gradient data sent by the server data storage unit (245) in the client device (2), and to store and send the latent feature matrix data calculated by the main federated learning module (12) to the client device (2). The main federated learning module (12) updates the latent feature matrix data in the server (1) according to the federated learning gradient data and stores it in the main data storage module (11); the spatiotemporal feature extraction module (22) is used to receive the physical coordinates of the equipment in the same park, calculate the spatiotemporal feature matrix and store it in the spatiotemporal feature correlation matrix storage unit (242); the data recovery module (23) is used to store incomplete power load data in the power load data storage unit (241) in the power data module (24), the latent feature matrix data stored in the server data storage unit (245), and the spatiotemporal feature matrix. The spatiotemporal feature matrix stored in the feature correlation matrix storage unit (242) and the latent feature vectors stored in the latent feature vector storage unit (243) of other clients in the same park are used to perform the incomplete power load data recovery process of the power grid and store the recovered data in the recovery data storage unit (244) in the power data module (24). The federated learning module (25) is used to use the incomplete power load data, latent feature matrix data and recovery data in the power data module (24) to realize the calculation of federated learning gradient data and store it in the server data storage unit (245).

2. A method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning, characterized in that, The campus power data privacy protection and recovery device based on spatiotemporal federated implicit feature learning as described in claim 1 includes the following steps: S1. The main federated learning module (12) in the server (1) initializes the latent feature matrix data and federated learning parameters and stores them in the main data storage module (11). S2. The spatiotemporal feature extraction module (22) of the client device (2) calculates the spatiotemporal feature correlation matrix of the power load data based on the physical coordinates of the monitoring sensor of the power load data in the park. S3. The client device (2) initializes the hidden feature vector and obtains the hidden feature vectors of other nodes in the same park, and stores them in the recovery data storage unit (244). S4. The main data storage module (11) sends the hidden feature matrix data to all client devices (2) and stores it in the server data storage unit (245). S5. The power data receiving module (21) receives incomplete power load data fed back by the power data acquisition equipment installed in the power grid and stores it in the power load data storage unit (241) in the power data module. S6. The data recovery module (23) uses the hidden feature analysis method to calculate the approximate power load data vector and stores it in the recovery data storage unit (244). S7. Based on the original power load data of the node, the spatiotemporal feature correlation matrix of the park and the latent feature vectors of other nodes in the park, the data recovery module (23) updates the latent feature vectors and calculates the federated learning gradient data. S8. The client device (2) sends federated learning gradient data to the server (1). S9. The server (1) updates the latent feature matrix data using federated learning gradient data; S10. Repeat steps S4 to S9 until the maximum number of iterations, i.e., the client device (2), has completed the data recovery under privacy protection.

3. The method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning according to claim 2, characterized in that: The latent feature matrix data mentioned in step S1 is a column latent feature matrix. , is randomly generated during initialization; among which, N This represents the time series number of the power load data. d is the dimension of the latent features, and is a positive number.

4. The method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning according to claim 2, characterized in that: The federated learning parameters include: maximum number of iterations. Learning rate and latent feature regularization coefficients .

5. The method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning according to claim 2, characterized in that: The hidden feature vector mentioned in step S3 is in M This represents the sum of the number of sensors across all parks.

6. The method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning according to claim 2, characterized in that, Step S2 includes the following steps: S201. The spatiotemporal feature extraction module (22) receives the instruction to calculate the spatiotemporal feature correlation matrix; S202. The spatiotemporal feature extraction module (22) reads the physical coordinates of the power sensor from the receiving module; S203. Spatiotemporal Feature Extraction Module (22) calculates spatial features based on the physical coordinates of relevant sensors in the park. First, it calculates the Euclidean distance between sensor coordinates and constructs a distance matrix. O According to the KNN algorithm, the distances to the five nearest sensors for each sensor are selected to form a distance matrix. O The values ​​at the remaining positions are set to 0, according to the distance matrix. O Calculate the weight matrix X The function for calculating the elements of the weight matrix is: ,in Representative at the h The park O The first in the matrix i The sensor and the first j The Euclidean distance between the sensors is S204. Calculate the Laplace matrix based on the weight matrix. The Laplace matrix has the following form: ,in, The weight matrix represents the first... i The sum of lines, i ∈{1, 2, ..., C h }, C h For the number of sensors in the h-th park, S205. Construct the time characteristic relationship, difference matrix G The format is as follows: ,in, N This is the time series number of the power load data.

7. The method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning according to claim 2, characterized in that, The data recovery module (23) in step S6 calculates an approximate power load data vector using the latent feature analysis method, including the following steps: S601. The data recovery module (23) receives the data recovery instruction; S602. According to the latent feature analysis method, for the first... h The park is located in the first i A sensor in its known data set Modeling is performed, and its target loss function is constructed by combining spatiotemporal feature relationships. as follows: ， in, Representing the h The park is located in the first i The sensor at the first j The actual value observed at each timestamp and These are the aggregate weights that control the influence of L1 and L2 paradigms, and their specific calculation method is as follows: in, t This represents the number of iteration rounds. σ It is a balancing coefficient used to balance the L1 norm and L2 norm.

8. The method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning according to claim 2, characterized in that, Step S7 includes the following steps: S701. The data recovery module (23) receives the instruction to update the latent feature vector. S702. The objective function is transformed as follows: Calculate its gradient, which is done as follows: S703. Traverse all observed data and update the local latent feature vector. . S704. Repeat steps S702 and S703 until the local latent feature vector change is within the set threshold range.

9. The method for privacy protection and recovery of campus power data based on spatiotemporal federated implicit feature learning according to claim 2, characterized in that: The method for updating the latent feature matrix data using federated learning gradient data in step S9 is as follows: traverse the latent feature dimensions, according to... Update.

Images