A power load prediction method and system based on privacy protection distributed clustering

Abstract

This invention discloses a method and system for power load forecasting based on privacy-preserving distributed clustering. The method first extracts load features to form a local dataset, then performs an iterative clustering process. In each iteration, each power user performs local clustering allocation to generate several clusters. Secret partitions are calculated based on the local sum and number of load features in each cluster. The cluster centers are updated based on the secret partitions, and local clustering allocation is re-executed until the clustering termination condition is met, generating the final cluster. A corresponding LSTM load forecasting model is constructed for each cluster to perform power load forecasting. This invention achieves high-quality user segmentation and accurate load forecasting while protecting the privacy of user power load data. The system can operate normally even if some users do not participate in the calculation, making it suitable for power sales companies to perform user group segmentation and refined load forecasting in the power market environment.

Description

Technical Field

[0001] This invention relates to the fields of electricity market, load forecasting, and privacy computing, and in particular to a method and system for electricity load forecasting based on privacy-preserving distributed clustering. Background Technology

[0002] Cluster analysis, as a crucial preprocessing step in load forecasting, groups users with similar electricity consumption characteristics together, thereby improving the accuracy of subsequent forecasting models. However, traditional clustering methods require direct access to detailed user electricity consumption data, such as time-of-use consumption and load curve characteristics. This data directly reflects users' production and operational activities and electricity consumption patterns, constituting core business secrets. In the context of increasingly fierce competition in the electricity market, users' demand for data privacy protection is growing stronger, while electricity retailers struggle to collect raw electricity consumption data. Therefore, there is an urgent need for a method that can provide high-precision clustering results while protecting user electricity load data privacy, laying the foundation for subsequent deep learning-based load forecasting. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a method and system for power load forecasting based on privacy-preserving distributed clustering, which can accurately group large users without obtaining their original electricity consumption data, thus providing support for LSTM-based load forecasting models.

[0004] Technical solution: The present invention provides a power load forecasting method based on privacy-preserving distributed clustering, comprising the following steps:

[0005] The electricity users and electricity sales companies participating in the clustering calculation are identified as the central coordinators. Each electricity user extracts load characteristics based on its own historical electricity load data to form a local dataset.

[0006] Initial cluster centers are determined and distributed to power users. An iterative clustering process is performed based on the local dataset. In each iteration, each power user performs local clustering allocation to generate several clusters. Secret partitions are calculated based on the local sum of load features in each cluster and the number of load features. Cluster centers are updated based on secret partitions. Local clustering allocation is re-executed until the clustering termination condition is met, and the final clusters are generated.

[0007] Add calibration noise to the cluster centers in the final clusters;

[0008] For each cluster after adding calibration noise, a corresponding LSTM load prediction model is constructed to predict power load.

[0009] Furthermore, after updating the cluster centers based on secret sharding, it also includes power load anomaly detection and empty clustering processing;

[0010] The method for detecting abnormal power load includes: calculating the distance variance from the load feature within each cluster to the cluster center, identifying abnormal load features based on the distance variance and its corresponding threshold, and constructing a corresponding LSTM load prediction model for each abnormal load feature.

[0011] Methods for handling empty clusters include: using a strategy of splitting the largest cluster to create new cluster centers.

[0012] Furthermore, the clustering termination condition includes: the maximum change in the cluster center is less than its corresponding threshold, or the maximum number of iterations is reached.

[0013] Furthermore, the method for determining the initial cluster centers includes: selecting initial cluster centers based on the K-means++ principle, and adding random perturbations to each initial cluster center.

[0014] Furthermore, methods for selecting initial cluster centers based on the K-means++ principle include:

[0015] The grid is divided according to the numerical range of the load characteristics, and the load characteristic density distribution within each grid is calculated.

[0016] Randomly select a point from the high-density grid as the first initial cluster center;

[0017] Calculate the minimum sum of squared distances from the load features within each grid to the first initial cluster center, and select subsequent cluster centers with a probability proportional to the sum of squared distances.

[0018] Furthermore, the method for each power user to perform local clustering assignment to generate several clusters includes: calculating the weighted Euclidean distance between the load characteristics and the cluster centers, and adding controllable noise to the weighted Euclidean distance; assigning the load characteristics to the category corresponding to the cluster center with the smallest weighted Euclidean distance.

[0019] Furthermore, secret shards are calculated based on the local sum and number of load features in each cluster. Updating the cluster centers based on these secret shards involves updating the cluster centers using the Shamir secret sharing scheme, which includes four stages: secret distribution, secret verification, secret aggregation, and secret reconstruction. This includes the following steps:

[0020] Construct local sums and polynomials for the number of load features to share load features, respectively, and calculate secret fragments. Distribute the secret fragments to power users. Each power user performs secret verification and secret aggregation after receiving the secret fragment. Reconstruct the secret based on the aggregation structure by interpolation and calculate new cluster centers.

[0021] The power load forecasting system based on privacy-preserving distributed clustering described in this invention includes:

[0022] The load feature extraction unit is used to identify the electricity users and electricity sales companies participating in the clustering calculation as the central coordinator. Each electricity user extracts load features based on its own historical electricity load data to form a local dataset.

[0023] The clustering allocation unit is used to determine the initial cluster centers and distribute them to power users. It performs an iterative clustering process based on the local dataset. In each iteration, each power user performs local clustering allocation to generate several clusters. It calculates secret partitions based on the local sum of load features in each cluster and the number of load features. It updates the cluster centers based on the secret partitions and re-executes local clustering allocation until the clustering termination condition is met, generating the final clusters.

[0024] A differential privacy processing unit is used to add calibration noise to the cluster centers in the final clustering;

[0025] The load forecasting unit is used to build a corresponding LSTM load forecasting model for each cluster after adding calibration noise, and to perform power load forecasting.

[0026] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the privacy-preserving distributed clustering-based power load forecasting method.

[0027] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the privacy-preserving distributed clustering-based power load forecasting method.

[0028] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows:

[0029] (1) By combining the K-means++ initialization strategy with secret sharing technology, the clustering quality and algorithm convergence speed are improved while protecting the privacy of user power load data, providing a more accurate user grouping basis for subsequent LSTM load forecasting.

[0030] (2) The (v,n) threshold secret sharing mechanism is adopted, which provides the fault tolerance capability of the system. Even if some users are offline or submit incorrect data, the system can still operate normally, which is of great significance to large users in the power system.

[0031] (3) Introducing a zero-knowledge proof verification mechanism can effectively prevent malicious users from submitting incorrect data and ensure the reliability and security of clustering results;

[0032] (4) By using a specially designed power load feature extraction and weighted distance calculation mechanism, the sensitivity and accuracy of clustering to power load characteristics are improved;

[0033] (5) The clustering results are seamlessly integrated with LSTM deep learning prediction, achieving the goal of accurate load prediction for different user groups while protecting user privacy. Attached Figure Description

[0034] Figure 1 This is a flowchart of the power load forecasting method according to an embodiment of the present invention.

[0035] Figure 2 This is a diagram of the power load forecasting system architecture according to an embodiment of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] like Figure 1 As shown, the power load forecasting method based on privacy-preserving distributed clustering includes the following steps.

[0038] Step 1, System Initialization. Identify the n electricity users and electricity sales companies participating in the clustering calculation as the central coordinator, and set the number of clusters K, the secret sharing threshold v (v≤n), and the maximum number of iterations T. max And the convergence threshold ε.

[0039] Step 2: Load data feature extraction and preparation. For each user u i (i∈{1,2,...,n}) Based on its own historical power load data, the local dataset X is formed by extracting electricity consumption characteristics. i ={x i1 ,x i2 ,...,x im}, where each feature vector x ih ∈R d Let represent the h-th electricity consumption feature vector of the i-th user. This vector is a d-dimensional real number vector and includes features such as the shape of the daily load curve, peak-to-valley ratio, load factor, and peak-hour electricity consumption ratio. Here, m indicates that the i-th user has m features in total.

[0040] Load characteristics include time-domain characteristics, frequency-domain characteristics, and enterprise attribute characteristics. Time-domain characteristics include daily load curve shape characteristics, peak-to-valley ratio, load factor, peak-hour electricity consumption ratio, and load growth rate. Frequency-domain characteristics include power spectral density characteristics, main frequency components, statistical characteristics, electricity consumption statistical distribution characteristics, periodic indicators, and volatility indicators. Enterprise attribute characteristics include industry category (after coding), scale indicators, and energy consumption pattern coding.

[0041] Step 3: Securely initialize cluster centers.

[0042] Step 3.1: The user securely submits load statistics information.

[0043] Step 3.2: The electricity sales company selects the initial cluster centers {μ1,μ2,...μ1} based on the K-means++ principle. j ..,μ K}, where μ j Let represent the j-th cluster center, where each cluster center corresponds to one category. The specific steps are as follows:

[0044] (a) Statistical information of the user based on the secret shared security aggregation of the power load feature space, including the numerical range of load features in each dimension and the load feature density distribution after dividing the feature space into grids;

[0045] (b) The electricity sales company randomly selects a point as the first cluster center μ1;

[0046] (c) For j = 2 to K: The user calculates the minimum sum of squared distances from the load feature points in each grid to the existing cluster centers based on secret sharing security. The electricity sales company selects the next cluster center μ with a probability proportional to the sum of squared distances. j .

[0047] Step 3.3: Add a random perturbation to each initial cluster center to obtain the perturbed cluster centers μ′. j Where μ′ j =μ j +δ j δ j It follows a mean of 0 and a variance of σ. 2 The Gaussian distribution.

[0048] Step 3.4: The electricity sales company will add the perturbated initial cluster centers {μ′1,μ′2,...,μ′}. K It is securely distributed to all electricity users.

[0049] Step 4: Iterative clustering process. In each iteration t (t=1,2,...,T) max Perform the following operations in ).

[0050] Step 4.1: Each user performs local clustering assignment.

[0051] (a) Calculate the local load characteristic vector x ih Weighted Euclidean distances between cluster centers:

[0052]

[0053] in, Indicates user ui The weighted Euclidean distance between the h-th feature and the j-th cluster center. Indicates user u i The l-th dimension component of the h-th eigenvector, Let w represent the l-th dimension of the j-th cluster center, d represent the total dimension of the feature vector, and w represent the total dimension of the feature vector. l The weight of the l-th feature is dynamically adjusted based on the importance of the power load feature, i.e., it is automatically calculated based on the contribution of the feature to the accuracy of load forecasting, and satisfies...

[0054] (b) To prevent the accurate distance calculation result L ij Sensitive user information is deduced, and controllable noise is added to the distance to obtain the noise-added result.

[0055]

[0056] Where ε ij It follows a mean of 0 and a variance of η. 2 The Gaussian distribution.

[0057] (c) Assign each feature vector to the category corresponding to the cluster center with the smallest distance.

[0058] (d) Calculate the j-th category set C j Sum of medium load eigenvectors ij With quantity N ij :

[0059]

[0060] N ij =|x∈X i c(x) = j|;

[0061] Where c(x) = j represents the feature vector x being assigned to category j, X i Indicates user u i The set of load feature vectors, N ij Indicates user u i The number of load feature vectors belonging to category j, S ij Indicate u i The sum of all load feature vectors belonging to category j.

[0062] Step 4.2: Cluster center update based on threshold secret sharing. For each category j∈{1,2,...,K}, perform:

[0063] Step 4.2.1: Secret distribution.

[0064] For each user ui :

[0065] (a) Constructing a system for sharing load characteristics and S ij The (t-1)th degree polynomial:

[0066] f ij (x)=S ij +a i1 x+a i2 x 2 +...+a i,t-1 x t-1 ;

[0067] Among them, a i1 ,a i2 ,...,a i,t-1 User u i From the finite field F p The coefficients are randomly selected from the data.

[0068] (b) Construct the number N for sharing data points ij The (t-1)th degree polynomial:

[0069] g ij (x)=N ij +b i1 x+b i2 x 2 +...,+b i,t-1 x t-1 ;

[0070] Where b i1 ,b i2 ,...,b i,t-1 User u i From the finite field F p The coefficients are randomly selected from the data.

[0071] (c) Calculate the secret fragment and distribute it to all users:

[0072] For each user u i (i∈1,2,...,n): Calculate S ij The fragment f ij (r i ), calculate N ij fragment g ij (r i Secure transmission fragment pairs (f) ij (r i ),g ij (r i ()) to the remaining users. Where r i For user u i A unique identifier, usually a non-zero integer, f ij (ri ) represents user u i Load characteristic value S ij At point r i The fragment value at the location, g ij (r i ) represents user u i Number of data points N ij At point r i The fragment value at that location.

[0073] Step 4.2.2: Secret verification.

[0074] For each user u i :

[0075] (a) Generate a set of categories C j Zero-knowledge proof of π ij ,prove:

[0076] (i)f ij (0)=S ij And S ij It is an effective load characteristic and;

[0077] (ii)g ij (0)=N ij And N ij It is a non-negative integer;

[0078] (iii)f ij (r i ) and g ij (r i ) is a correctly calculated fragment;

[0079] (b) will prove π ij Send along with fragments;

[0080] (c) Each user verifies the received fragment and proof: if the verification is successful, the fragment is accepted; if the verification fails, the fragment is marked as invalid and reported to the electricity sales company.

[0081] Step 4.2.3: Secret aggregation.

[0082] For each user u i :

[0083] (a) Valid fragments received by the aggregation:

[0084]

[0085] in, Indicates user u i The collected category set C j The result of the load characteristic value segmentation aggregation Indicates user u i The collected category set C j The result of data point aggregation by data point number.

[0086] (b) Send the aggregation results to the electricity sales company

[0087] Step 4.2.4: Secret Reconstruction.

[0088] Electricity sales company implementation:

[0089] (a) Collect aggregate results from at least v valid users.

[0090] Assume the set of valid users is U = {u1, u2, ..., u...} Q}, where Q represents the number of valid users, satisfying Q≥v;

[0091] (b) Reconstructing the secret S using Lagrange interpolation j and N j :

[0092]

[0093] Where λ q These are the Lagrange coefficients, calculated using the following formula:

[0094] Where q and p are different indices of valid users, q and p ∈ {1,2,...,Q}, r p User u p The unique identifier, r q User u q A unique identifier.

[0095] (c) Calculate μ j New cluster center If N j <θ min (θ min If the minimum cluster size threshold is preset, the cluster is marked as an anomaly.

[0096] Step 4.3: Power load anomaly detection and handling.

[0097] (a) Detection of load anomalies:

[0098] (i) For the j-th category set C j Calculate the internal load characteristic vector x load The variance of the distance to the cluster center σ j :

[0099]

[0100] (ii) Recognition distance exceeds μ j +ασ j The load characteristics are abnormal points (α is an adjustable parameter);

[0101] (iii) Special handling of outliers, such as establishing an independent abnormal load model.

[0102] (b) Empty clustering:

[0103] If a certain category set C j N j If the value is 0, then the strategy of splitting the largest cluster is used: find the cluster C with the most members. max Choose C max The load feature point furthest from the center is used as the center to create a new cluster center to replace the empty cluster.

[0104] Step 4.4: Termination condition check.

[0105] (a) Calculate the maximum change in cluster centers:

[0106] (b) If Δ < ε or the maximum number of iterations T has been reached max If the cluster centers are not found, the algorithm terminates; otherwise, the cluster centers are updated. (For all j), return to step 4.1 and continue iterating.

[0107] Step 5: Differential privacy enhancement processing.

[0108] Step 5.1: Add calibration noise to the final cluster centers:

[0109]

[0110] in: These are the cluster centers after adding noise, Δf is the sensitivity, representing the maximum potential impact of a load feature point on the cluster centers, and ε is the sensitivity. DP It's a privacy budget, σ DP These are noise parameters;

[0111] Step 5.2, return the cluster centers after differential privacy protection.

[0112] Step 6: Construct an LSTM load prediction model based on clustering results.

[0113] Step 6.1, for each cluster C j Electricity retailers collect anonymized historical load data from users in this category;

[0114] Step 6.2: Construct a dedicated LSTM load prediction model M for each cluster. j The LSTM load forecasting model includes a data preprocessing layer for data normalization, time-series feature extraction, and encoding of external factors (such as temperature, date type, etc.).

[0115] The LSTM network structure consists of an input layer, multiple LSTM layers, an attention mechanism layer, and an output layer. The input layer is used to receive historical load sequences and external features; the multiple LSTM layers are used to capture the long-term and short-term dependencies of the load; the attention mechanism layer is used to focus on important historical periods; and the output layer is used to predict the load in future periods.

[0116] Model optimization and adjustment methods include: automatically adjusting network hyperparameters based on user load characteristics within clusters; using early shutdown to prevent overfitting; and periodically updating to adapt to changes in electricity consumption patterns.

[0117] Step 6.3: When a new user joins, they are assigned to the nearest cluster j based on their submitted anonymized features, and the corresponding prediction model M is used. j Perform load forecasting.

[0118] The privacy-preserving distributed clustering-based power load forecasting method in this embodiment is applied to the scenario where power sales companies cluster and forecast the power load characteristics of large users. The steps are as follows.

[0119] S101, System initialization. In this embodiment, n=50 (large industrial users), K=8 (electricity consumption characteristic classification), v=35, T max For example, let's take ε = 30 and ε = 0.0005.

[0120] S102, Load Data Feature Extraction and Preparation. Each electricity user extracts a 30-dimensional feature vector based on their historical electricity load data, including daily load curve shape characteristics at 24 time points, peak-to-valley ratio, load factor, peak-hour electricity consumption ratio, and load volatility indicators. Users retain the original load data and only use the extracted features in subsequent calculations.

[0121] S103, Securely initialize cluster centers. This step includes:

[0122] S103.1, users securely submit load statistics, including the numerical range of load characteristics and grid density information, through secret sharing. Specifically, the 30-dimensional feature space is divided into grid cells, each user calculates the distribution of local load characteristics in each grid, and the global grid density is securely aggregated through Shamir's (v,n) threshold secret sharing scheme.

[0123] S103.2, the electricity sales company selects initial cluster centers based on the K-means++ principle. First, based on the density distribution, the first cluster center μ1 is selected from the high-density region; then, subsequent cluster centers are selected iteratively to keep the new cluster centers as far away from the existing centers as possible, so as to better cover the characteristics of different types of electricity loads.

[0124] S103.3, Add a random perturbation to each initial cluster center: μ′ j =μ j +δ j δ j Follows N(0,σ) 2 The distribution is such that the value of σ is small enough not to affect the clustering performance.

[0125] S103.4, the electricity sales company will set the initial cluster centers {μ′1,μ′2,...,μ′ after disturbance}. K Distribute to all electricity users.

[0126] S104, Iterative clustering process. Executed in each iteration:

[0127] S104.1, each user performs local clustering assignment. Calculate the weighted Euclidean distance, where the weights are set based on the contribution of the load feature to the prediction accuracy. Add controlled noise, assign data points, and calculate the local sum and quantity S for each load feature class. ij and N ij .

[0128] S104.2, Cluster center update based on threshold secret sharing. For each category j, perform:

[0129] S104.2.1, Secret Distribution. Each user constructs a polynomial f. ij (x) and g ij (x), compute secret fragments and distribute them securely.

[0130] S104.2.2, Secret Verification. Using the Bulletproofs zero-knowledge proof system, a proof π is generated. ij And verify the validity of the received fragments.

[0131] S104.2.3, Secret Aggregation. Aggregation Effective Fragments. and

[0132] S104.2.4, Secret Reconstruction. The electricity company reconstructs S using the results from at least v = 35 valid users via Lagrange interpolation. j and N j Calculate the new cluster centers

[0133] S104.3, Power Load Anomaly Detection and Handling. Identify abnormal load points, such as atypical power consumption behaviors caused by temporary production shutdowns or equipment failures, and use a splitting strategy to handle empty clusters.

[0134] S104.4, Termination Condition Check. Calculate the maximum change Δ. If Δ < ε or the maximum number of iterations is reached, terminate; otherwise, update the cluster centers and continue iterating.

[0135] S105, Differential Privacy Enhancement Process. Calibration noise is added to the final cluster centers, and the processed result is returned.

[0136] S106, Construction of an LSTM load forecasting model based on clustering results. A dedicated LSTM forecasting model is built for each cluster, including a data preprocessing layer, a multi-layer LSTM structure, an attention mechanism layer, and an output layer. Each model is optimized for the load characteristics of that user group, such as the periodic patterns of industrial enterprises and the temperature sensitivity of commercial users.

[0137] like Figure 2 As shown, the privacy-preserving distributed clustering system for power load forecasting described in this invention includes the following modules.

[0138] The system initialization module 101 is used to determine the set of electricity users, electricity sales companies, and related parameters involved in the calculation;

[0139] The load feature extraction module 102 is used to extract feature vectors from the raw power load data;

[0140] The secure clustering center initialization module 103 is used to securely select initial clustering centers based on the K-means++ principle and secret sharing technology.

[0141] The local clustering assignment module 104 is used to calculate the weighted distance and assign the load features to the nearest cluster centers;

[0142] Secret distribution module 105 is used to construct polynomials and generate secret fragments;

[0143] Secret verification module 106 is used to generate and verify zero-knowledge proofs;

[0144] Secret aggregation module 107 is used to aggregate received valid fragments;

[0145] Secret reconstruction module 108 is used to reconstruct the secret using Lagrange interpolation and update the cluster centers;

[0146] The load anomaly detection module 109 is used to identify and process abnormal load characteristics;

[0147] Differential privacy processing module 110 is used to add calibration noise to the final result to enhance privacy protection;

[0148] The LSTM load forecasting module 111 is used to build specialized forecasting models for different user groups based on clustering results.

[0149] In practical applications, modules 102 and 104 are deployed on the user side, modules 101, 103, 108, 109, 110 and 111 are deployed on the electricity sales company side, and modules 105, 106 and 107 are distributed among all participants, exchanging data through a secure communication channel.

[0150] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the privacy-preserving distributed clustering-based power load forecasting method.

[0151] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the power load forecasting method based on privacy-preserving distributed clustering.

[0152] The computer-readable storage medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage devices, magnetic disk storage devices or other magnetic storage devices, flash memory or any other medium that can be used to store program code in the form of instructions or data structures and is accessible by a computer.

[0153] The processor is used to execute a computer program stored in memory to implement the various steps in the methods described in the above embodiments.

Claims

1. A power load forecasting method based on privacy-preserving distributed clustering, characterized in that, Includes the following steps: The electricity users and electricity sales companies participating in the clustering calculation are identified as the central coordinators. Each electricity user extracts load characteristics based on its own historical electricity load data to form a local dataset. Initial cluster centers are determined and distributed to power users. An iterative clustering process is performed based on the local dataset. In each iteration, each power user performs local clustering allocation to generate several clusters. Secret partitions are calculated based on the local sum of load features in each cluster and the number of load features. Cluster centers are updated based on secret partitions. Local clustering allocation is re-executed until the clustering termination condition is met, and the final clusters are generated. Add calibration noise to the cluster centers in the final cluster. Where, μ j It is the cluster center in the final cluster. These are the cluster centers after adding noise, Δf is the sensitivity, representing the maximum potential impact of a load feature point on the cluster centers, and ε is the sensitivity. DP It's a privacy budget, σ DP These are noise parameters; For each cluster after adding calibration noise, a corresponding LSTM load forecasting model is constructed to predict power load. Secret shards are calculated based on the local sum and number of load features in each cluster. Updating cluster centers based on these secret shards involves updating the cluster centers using the Shamir secret sharing scheme, which includes four stages: secret distribution, secret verification, secret aggregation, and secret reconstruction. The steps are as follows: Construct local sums and polynomials for the number of load features to share load features, respectively, and calculate secret fragments. Distribute the secret fragments to power users. Each power user performs secret verification and secret aggregation after receiving the secret fragment. Reconstruct the secret based on the aggregation structure by interpolation and calculate new cluster centers.

2. The power load forecasting method based on privacy-preserving distributed clustering according to claim 1, characterized in that, After updating the cluster centers based on secret partitioning, it also includes power load anomaly detection and empty clustering processing; The method for detecting abnormal power load includes: calculating the distance variance from the load feature within each cluster to the cluster center, identifying abnormal load features based on the distance variance and its corresponding threshold, and constructing a corresponding LSTM load prediction model for each abnormal load feature. Methods for handling empty clusters include: using a strategy of splitting the largest cluster to create new cluster centers.

3. The power load forecasting method based on privacy-preserving distributed clustering according to claim 1, characterized in that, The clustering termination conditions include: the maximum change in the cluster centers is less than their corresponding threshold, or the maximum number of iterations is reached.

4. The power load forecasting method based on privacy-preserving distributed clustering according to claim 1, characterized in that, The method for determining the initial cluster centers includes: Initial cluster centers are selected based on the K-means++ principle, and random perturbations are added to each initial cluster center.

5. The power load forecasting method based on privacy-preserving distributed clustering according to claim 4, characterized in that, Methods for selecting initial cluster centers based on the K-means++ principle include: The grid is divided according to the numerical range of the load characteristics, and the load characteristic density distribution within each grid is calculated. Randomly select a point from the high-density grid as the first initial cluster center; Calculate the minimum sum of squared distances from the load features within each grid to the first initial cluster center, and select subsequent cluster centers with a probability proportional to the sum of squared distances.

6. The power load forecasting method based on privacy-preserving distributed clustering according to claim 1, characterized in that, The methods for each electricity user to perform local clustering assignment to generate several clusters include: Calculate the weighted Euclidean distance between the load feature and the cluster center, and add controllable noise to the weighted Euclidean distance; assign the load feature to the category corresponding to the cluster center with the smallest weighted Euclidean distance.

7. A power load forecasting system based on privacy-preserving distributed clustering according to the method of claim 1, characterized in that, include: The load feature extraction unit is used to identify the electricity users and electricity sales companies participating in the clustering calculation as the central coordinator. Each electricity user extracts load features based on its own historical electricity load data to form a local dataset. The clustering allocation unit is used to determine the initial cluster centers and distribute them to power users. It performs an iterative clustering process based on the local dataset. In each iteration, each power user performs local clustering allocation to generate several clusters. It calculates secret partitions based on the local sum of load features in each cluster and the number of load features. It updates the cluster centers based on the secret partitions and re-executes local clustering allocation until the clustering termination condition is met, generating the final clusters. A differential privacy processing unit is used to add calibration noise to the cluster centers in the final clustering; The load forecasting unit is used to build a corresponding LSTM load forecasting model for each cluster after adding calibration noise, and to perform power load forecasting.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the power load forecasting method based on privacy-preserving distributed clustering according to any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the power load forecasting method based on privacy-preserving distributed clustering according to any one of claims 1-6.

Images