Power user profiling method considering privacy protection and demand response

Abstract

The application discloses a power user portrait method considering privacy protection and demand response, and mainly comprises the following steps: firstly, collecting historical load data and historical temperature and humidity data of a building, and extracting load characteristics of the building; secondly, determining the optimal number of user portraits by minimizing the overlap degree; thirdly, a smart scheduling terminal in the building performs local calculation and introduces zero error, and the calculation result is used as an initial value of a distributed privacy protection AAC algorithm; finally, according to the initial value of the distributed privacy protection AAC algorithm, the user parameter value is updated, and a typical user portrait and a typical user portrait type to which the building belongs are obtained through convergence. The application can analyze the user portrait according to the user electricity habit without collecting specific load data of the user, and the user portrait can be used in a demand response system, so that the user with different electricity habits can formulate a targeted electricity strategy, the response capability of the building load is better explored, and the user privacy is protected.

Description

Technical Field

[0001] This invention relates to the field of power system load characteristic analysis, specifically to a method for profiling power users that considers privacy protection and demand response. Background Technology

[0002] Demand Response (DR), a power demand-side management method, refers to the process by which building load controllers, upon receiving compensation signals from the power supplier to reduce load or increase dynamic electricity prices when the reliability of the power system is threatened (e.g., excessive load peaks), change users' habitual electricity consumption patterns to reduce or shift the electricity load during a certain period, thereby responding to the demand for power supply stability and ensuring grid stability.

[0003] Traditional demand response technology mainly uses optimization algorithms, which plan the load actions of users in the power system. Since this planning method only aims to reduce the load fluctuation of the system without considering the differences in users' electricity consumption habits, introducing typical user profiles into demand response technology can improve user comfort. Different demand response strategies can be customized for different typical users.

[0004] Meanwhile, traditional centralized user profiling algorithms require the collection of user load data, which raises concerns about user privacy. The system needs to aggregate this data to a server, which then performs clustering analysis before sending the results back to the requesting side. This approach has two drawbacks: first, such a massive amount of data places extremely high demands on user communication devices; second, user privacy cannot be adequately guaranteed. Summary of the Invention

[0005] This invention relates to a user profiling algorithm. The purpose of this invention is to provide a distributed privacy-preserving user profiling algorithm that solves the privacy leakage problem caused by the need to collect user load data in traditional centralized user profiling algorithms.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A user profiling method for electricity users, considering both privacy protection and demand response, is based on a Gaussian mixture model. The method uses a distributed privacy-preserving AAC algorithm to estimate the parameters of the Gaussian mixture model, obtaining user profiles for typical buildings and the typical user profile categories to which the buildings belong, without collecting user privacy data.

[0008] S1 collects historical load data and historical temperature and humidity data of buildings, extracts load characteristic data of buildings based on historical load data, extracts comfort characteristic data of buildings based on historical temperature and humidity data, and uses Gaussian mixture model to cluster the load characteristic data and comfort characteristic data of buildings.

[0009] S2 determines the optimal number of user profiles by minimizing the average overlap index of the Gaussian mixture model;

[0010] S3 buildings perform local calculations, and to protect user privacy, zero-sum errors are introduced into the local calculation results.

[0011] S4 uses the AAC algorithm to estimate the parameters of the Gaussian mixture model. To protect user privacy, the local calculation results, which introduce zero-sum error, are used as the initial values ​​of user parameters in the distributed privacy-preserving AAC algorithm. The user parameter values ​​are updated based on the distributed privacy-preserving AAC algorithm until the user parameter values ​​converge. The converged user parameter values ​​are then used to calculate the typical user profile and the typical user profile category to which the building belongs.

[0012] The historical load data in S1 is the net power exchange between each building and the power grid, and the historical temperature and humidity data includes indoor temperature data, indoor humidity data, and outdoor temperature data.

[0013] The load characteristics of the building extracted from historical load data in S1 include: power factor, average load power, maximum load power, peak load power on holidays, peak load power on weekdays, and load power gradient.

[0014] The step S1, which extracts the building's comfort characteristics based on its historical temperature and humidity data, includes two steps: First, the indoor comfort index of the building at different times is calculated based on the building's historical temperature and humidity data, using the following formula:

[0015] I C =1.8T in +32-(0.55-0.55H)(1.8T in -T out -26)

[0016] Among them, I C T represents the indoor comfort index. in Indicates indoor temperature, T out H represents indoor temperature, and H represents indoor humidity;

[0017] Secondly, the average indoor comfort level on weekdays and the average indoor comfort level on holidays are extracted based on the indoor temperature comfort index data of the building.

[0018] In S1, a Gaussian mixture model is used to cluster the building's load characteristic data and building's comfort characteristic data. The probability distributions of the load characteristic data and comfort characteristic data are as follows:

[0019]

[0020] Where K represents the number of user load feature profiles, and parameter K needs to be determined before using Gaussian mixture model clustering; p(X) represents the probability distribution function of the building load curve features; X represents the load feature and comfort feature data; N j (x|μ j , ∑ j ) represents the j-th component in the Gaussian mixture model, which represents a K-dimensional Gaussian distribution. Each component corresponds to a portrait of a certain type of building. ∑ j The covariance matrix of the Gaussian mixture model, μ j In a Gaussian mixture model, the mean parameter refers to all the mean parameters of the j-th component. The mean parameter represents the mean of the class represented by that component. π j The j-th mixing coefficient in the Gaussian mixture model.

[0021] In step S2, the optimal number of user profiles is determined by minimizing the average overlap index of the Gaussian mixture model. The Gaussian mixture model overlap is defined as follows:

[0022]

[0023] Among them, OLR(N) i N j ) refers to the two Gaussian mixture model components N i and N j The degree of overlap between them, X saddle X is a saddle point in the probability distribution. submax It is the lower peak in the probability distribution, and p(X) represents the probability distribution function of the building characteristics;

[0024] Iterate through the number of user profiles from 2 to N-1, calculate the average overlap for different numbers of users, and select the number of users with the smallest average overlap as the optimal number of user profiles:

[0025]

[0026] Where K represents the number of user profiles and N represents the number of buildings within the power system.

[0027] In S3, the buildings are calculated locally, as follows: First, the posterior probability of the k-th building is calculated.

[0028]

[0029] Where, γ n，i，j Let π represent the posterior probability of the k-th building, where K represents the number of user profiles. i N represents the i-th mixing coefficient in the Gaussian mixture model. i (x n |μ i , ∑ i ) represents the load curve characteristic x of the i-th building. n The probability of the i-th component in a Gaussian mixture model;

[0030] Secondly, the local calculation result of the k-th building is obtained based on the posterior probability of the k-th building:

[0031]

[0032] Among them, s i，j h represents the local calculation result for the k-th building without mean parameter correction. i，j z represents the local calculation result of the k-th building after correction for the mean parameter. i，j Let x represent the estimated posterior probability of the k-th building. n This represents the nth feature data. This refers to the estimated value of all mean parameters of the k-th component in the Gaussian mixture model; the local computation results of all buildings are used as the initial values ​​of user parameters for the distributed privacy-preserving AAC algorithm.

[0033] The zero-sum error in S3 is introduced in the following way:

[0034]

[0035] θ i (t)=δ i (t)-δ i (t-1)

[0036] Among them, y i (t) represents the user parameter value without introducing zero-sum error. This represents the user parameter value after introducing zero-sum error, containing three sub-parameters. The introduced zero-sum error meets the following conditions:

[0037]

[0038] Where, θ i (t) represents a perturbation with zero-sum characteristics, δ i (t) represents noise with an exponentially decreasing amplitude.

[0039] In step S4, the user parameter values ​​are updated using the distributed privacy-preserving AAC algorithm. The update method for the user parameter values ​​is as follows:

[0040]

[0041] Among them, y i (t+1) represents the update result of the user parameter value after t+1 updates. This represents the weight value in the i-th row and i-th column of the Metropolis weight matrix. This represents the weight value in the i-th row and j-th column of the Metropolis weight matrix. This indicates that the local calculation result is updated for the tth time after introducing zero-sum error, and the parameter values ​​are updated according to the above update method until the user parameter values ​​converge.

[0042] The following method accelerates the update of the Metropolis weight matrix:

[0043] W * =(1+α)W-αI

[0044] Among them, W * Let W represent the accelerated metropolis weight matrix, I represent the identity matrix, and α represent the update parameters.

[0045] The weight values ​​of the Metropolis weight matrix W are calculated as follows:

[0046]

[0047] Where, γ i and γ j Let W represent the posterior probability of building i and building i'. i，j This represents the weight value in the i-th row and j-th column of the Metropolis weight matrix;

[0048] The parameters of the Gaussian mixture model are calculated using the convergent user parameter values ​​as follows:

[0049]

[0050]

[0051]

[0052] Where, ∑ j The covariance matrix of the Gaussian mixture model, μ j In a Gaussian mixture model, the mean parameter refers to all the mean parameters of the j-th component. The mean parameter represents the mean of the class represented by that component. π j The j-th mixing coefficient in the Gaussian mixture model.

[0053] This invention, based on Gaussian mixture models and focusing on individual buildings, proposes a privacy-preserving user profiling method for demand response systems. This method introduces zero-sum error into the user's local computation, using the results of this local computation with zero-sum error as the initial value for a distributed privacy-preserving AAC algorithm. The algorithm then derives a typical user profile for the building. Specifically, this method can analyze typical user profiles based on users' energy consumption and comfort habits within a building without collecting user load data. These user profiles can be used in demand response systems to help users with different electricity consumption habits develop targeted electricity consumption strategies, thereby better understanding the building's load response capabilities. Attached Figure Description

[0054] Figure 1 The specific process framework of the privacy-preserving user profiling method established by this invention is shown. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example:

[0057] In this embodiment, nine buildings participate in the demand response within the system, mainly including a medium-sized office building, a fast food restaurant, a mobile retail department store, a shopping mall, and five medium-sized multi-family residential buildings. All buildings are located in the same prefecture-level city and within the same climate zone. Previous load data spans five years, with the time unit being hours. Figure 1 This is a flowchart illustrating a user profiling method that considers privacy protection in a demand response system. For example... Figure 1 As shown, it includes the following steps:

[0058] Step S1: Collect historical load data and historical temperature and humidity data of the building. Extract the load characteristic data of the building based on the historical load data and extract the comfort characteristic data of the building based on the historical temperature and humidity data. Use a Gaussian mixture model to cluster the load characteristic data and comfort characteristic data of the building.

[0059] The present invention classifies the time information in the dataset as follows: weekdays, time type 1, 2, 3, 4, 5; weekends, time type 6, 7; holidays, time type 8.

[0060] The load characteristic data and comfort characteristic data of the building in this embodiment are shown in the table below.

[0061] Table 1. Building load characteristics and comfort characteristics data for a certain prefecture-level city.

[0062]

[0063] Step S2: Determine the optimal number of user profiles by minimizing the average overlap index of the Gaussian mixture model.

[0064] Based on the overlap calculation formula, the number of user profiles was traversed from categories 2 to 8, and the overlap of different user profiles was obtained as shown in the table below:

[0065] Table 2. Building load characteristics and comfort characteristics data for a certain prefecture-level city.

[0066] Number of user profiles 2 3 4 5 6 7 8 Average overlap 0.69 0.41 0.71 0.54 0.89 0.72 0.87

[0067] Based on the above, the optimal number of user profiles is 3.

[0068] Step S3: The building performs local calculations. At the same time, to protect user privacy, zero-sum errors are introduced into the local calculation results.

[0069] Step S4: To protect user privacy, the parameters of the Gaussian mixture model are estimated using a distributed privacy-preserving AAC algorithm. The local computation results, which introduce zero-sum error, are used as the initial values ​​of the user parameters in the distributed privacy-preserving AAC algorithm. The user parameter values ​​are updated based on the distributed privacy-preserving AAC algorithm until the user parameter values ​​converge. The converged user parameter values ​​are then used to calculate the typical user profile and the typical user profile category to which the building belongs.

[0070] The final user load profile of the building, obtained based on the user parameters calculated by the distributed privacy-preserving AAC algorithm, is as follows:

[0071] Table 3. Building Clustering Results

[0072] category result Cluster 1 1，2，9 Cluster 2 5，7 Cluster 3 3，4，6，8

[0073] Table 4 User Load Characteristics Profile

[0074] Typical User 1 Typical User 2 Typical User 3 Load factor characteristics 2.77 1.01 0.18 Average power characteristics 1.31 2.00 0.22 Maximum power characteristics 1.49 1.99 0.10 Average power characteristics during holidays 0.50 1.27 2.00 Weekday average power characteristics 0.22 0.54 1.22 Load power slope characteristics 0.21 0.40 0.31 Peak characteristics of weekday comfort index 0.43 0.44 0.72 Peak characteristics of comfort index during holidays 1.22 1.12 1.50

[0075] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method of power consumer profiling considering privacy protection and demand response, characterized in that, The user portrait method is based on a Gaussian mixture model, and parameters of the Gaussian mixture model are estimated by using a distributed privacy protection AAC algorithm to obtain a typical user portrait of a typical building and a typical user portrait category to which the building belongs without collecting user privacy, and the method comprises the following steps: S1: collecting historical load data and historical temperature and humidity data of the building, extracting load feature data of the building according to the historical load data, extracting comfort feature data of the building according to the historical temperature and humidity data of the building, and clustering the load feature data of the building and the comfort feature data of the building by using a Gaussian mixture model; S2: determining an optimal user portrait quantity by minimizing an average overlap index of the Gaussian mixture model; S3: performing local calculation by the building, and introducing zero and error in a local calculation result to protect user privacy; S4: estimating parameters of the Gaussian mixture model by using an AAC algorithm, taking the local calculation result with the introduced zero and error as an initial value of a user parameter in the distributed privacy protection AAC algorithm, updating the user parameter value based on the distributed privacy protection AAC algorithm until the user parameter value converges, and calculating a typical user portrait and a typical user portrait category to which the building belongs by using the converged user parameter value; In the S4, the user parameter value is updated based on the distributed privacy protection AAC algorithm, and the user parameter value is updated in the following manner: wherein y i (t+1) represents the update result of the user parameter value t+1 times update, represents the weight value of the i-th row and the i-th column of the accelerated Metropolis weight matrix, represents the weight value of the i-th row and the j-th column of the accelerated Metropolis weight matrix, represents the local calculation result after introducing the zero-sum error, and the parameter value is updated according to the above update mode until the user parameter value converges; The Metropolis weight matrix is updated in the following manner: W * = (1 + a)W - aI where W * denotes an accelerated Metropolis weight matrix, W denotes a Metropolis weight matrix, I denotes an identity matrix, and a denotes an update parameter; The weight value of the Metropolis weight matrix W is calculated in the following manner: where γ i and γ j denote the posterior probabilities of the ith and jth buildings, W i，j denotes the weight value of the ith row and jth column of the Metropolis weight matrix; The Gaussian mixture model parameters are calculated by using the converged user parameter value in the following manner: where ∑ j denotes the covariance matrix of the Gaussian mixture model, μ j denotes all mean parameters of the jth component in the Gaussian mixture model, π j denotes the jth mixing coefficient in the Gaussian mixture model.

2. The method of claim 1, wherein, In the S1, the historical load data is the net exchange power of each building and the power grid, and the historical temperature and humidity data includes indoor temperature data, indoor humidity data and outdoor temperature data.

3. The method of claim 1, wherein, In the S1, the load features are extracted from the historical load data, and the load features include a power factor, a load power average value, a load power maximum value, a holiday load power peak value, a weekday load power peak value and a load power slope.

4. The method of claim 1, wherein, In the S1, the comfort features of the building are extracted from the historical temperature and humidity data of the building in two steps: first, the indoor comfort indexes of the building at different times are calculated according to the historical temperature and humidity data of the building, and the formula is as follows I C = 1.8T in + 32 - (0.55 - 0.55H) (1.8T in - T out - 26) where I C represents the indoor comfort index, T in represents the indoor temperature, T out represents the indoor temperature, H represents the indoor humidity; Second, the weekday indoor comfort average value and the holiday indoor comfort average value of the building are extracted according to the indoor temperature comfort index data of the building.

5. The method of claim 1, wherein, In the S1, the load feature data and the comfort feature data of the building are clustered by using the Gaussian mixture model, and the probability distribution of the load feature data and the comfort feature data is as follows: wherein K represents the number of user load feature profiles, the parameter K needs to be determined before using Gaussian mixture model clustering, p(X) represents a probability distribution function of building load curve features, X represents load feature and comfort feature data, N j (x|μ j ,∑ j ) represents the jth component in the Gaussian mixture model, which represents a K-dimensional Gaussian distribution, each component corresponds to a profile of a certain type of building,∑ j is the covariance matrix of the Gaussian mixture model, μ j is the mean parameter of the jth component in the Gaussian mixture model, the mean parameter represents the mean of the category represented by the component, π j is the jth mixing coefficient in the Gaussian mixture model.

6. The method of claim 1, wherein, In the S2, the optimal user portrait quantity is determined by minimizing the average overlap index of the Gaussian mixture model, and the overlap index of the Gaussian mixture model is defined as follows: where OLR(N i , N j ) refers to the overlap between two Gaussian mixture model components N i and N j , X saddle is the saddle point in the probability distribution, X submax is the lower peak in the probability distribution, and p(X) represents the probability distribution function of the building feature; From 2 to N-1, the average overlap under different user quantities is calculated, and the user quantity with the minimum average overlap is selected as the optimal user portrait quantity: Wherein, K represents the user portrait quantity, and N represents the number of buildings in the power system.

7. The method of claim 1, wherein, The building in the S3 performs local calculation, and the method is as follows: first, the posterior probability of the kth building is calculated where γ n,i,j represents the posterior probability of the kth building, K represents the number of user portraits, π i represents the i-th mixing coefficient in the Gaussian mixture model, N i (x n | μ i ,∑ i ) represents the load curve feature x n of the i-th building; and the probability of the i-th component in the Gaussian mixture model; Secondly, the local calculation result of the kth building is obtained according to the posterior probability of the kth building: where s i，j denotes the local calculation result of the kth building without mean parameter correction, h i,j denotes the local calculation result of the kth building after mean parameter correction, z i，j denotes the estimation result of the posterior probability of the kth building, x n denotes the n-th feature data, denotes the estimation value of all mean parameters of the kth component in the Gaussian mixture model; the local calculation result of all buildings is used as the initial value of the user parameter of the distributed privacy protection AAC algorithm.

8. The method of claim 1, wherein, The introduction mode of the zero-sum error in the S3 is as follows: θ i (t) = δ i (t) - δ i (t - 1) where y i (t) denotes the user parameter value without the introduced zero-sum error, denotes the user parameter value with the introduced zero-sum error, which contains three sub-parameters, and the introduced zero-sum error satisfies the following conditions: where θ i (t) represents a disturbance with zero and characteristic, δ i (t) represents a noise with exponentially decreasing amplitude.

Images