A multi-wind farm collaborative prediction method based on matrix perturbation and differential privacy protection

Abstract

The application relates to a multi-wind farm collaborative prediction method based on matrix perturbation and differential privacy protection. The method provided by the application guarantees the data privacy of a model participating in a wind farm in the training and prediction stages of a vector autoregression collaborative prediction model, increases the difficulty of stealing historical wind power data of the wind farm by external attackers, a center server, other wind farms and the like, on the one hand; on the other hand, the perturbed data can still be used for wind power prediction as much as possible, and the prediction accuracy meeting the demand is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system operation and control, in particular to a multi-wind farm collaborative prediction method based on matrix perturbation and differential privacy protection. BACKGROUND

[0002] With the increasing depletion of non-renewable resources such as coal and oil and the increasing severity of energy crisis, renewable energy such as wind energy, solar energy, tidal energy and biomass energy is increasingly concerned in the world. Wind power generation is the most mature and most valuable renewable energy in renewable energy generation technology. Developing wind power is of great significance to ensuring energy security, adjusting energy structure, reducing environmental pollution and realizing sustainable development.

[0003] The intermittent nature of wind energy in nature determines that wind power has strong volatility. With the continuous increase in the number of wind farms and installed capacity, once wind power is integrated into the power grid, this power fluctuation will bring great challenges to the safe and economic operation of the power grid. Accurate prediction of wind speed and wind power in advance can alleviate the pressure of power system peak regulation and frequency regulation, and effectively improve the accommodation capacity of the power grid to wind power.

[0004] At present, the research on wind power prediction at home and abroad is more and more extensive and in-depth. Traditional prediction methods usually take the wind power prediction of a single wind farm as the goal, and only use single data information such as historical measured wind power and measured meteorological data. However, wind power not only has correlation in time, but also has correlation in space. Therefore, when predicting wind power, if the interaction between wind farms can be considered, the wind power prediction accuracy will be further improved.

[0005] However, due to commercial competition or confidentiality factors, wind farms in adjacent areas will not voluntarily share their historical power generation data, even if sharing data can improve prediction accuracy. Existing wind power prediction methods usually focus on how to improve prediction accuracy, without considering the data privacy protection of participating wind farms, which will not be conducive to data sharing to improve prediction accuracy. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide a multi-wind farm collaborative prediction method based on matrix perturbation and differential privacy protection. The method provided by the present application ensures the data privacy of the model participating wind farm in the vector autoregressive collaborative prediction model training and prediction stage, increases the difficulty of stealing wind farm historical wind power data by external attackers, center servers and other wind farms; on the other hand, the perturbed data can still be used for wind power prediction as much as possible, and the prediction accuracy meeting the demand is obtained.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A collaborative prediction method for multiple wind farms based on matrix perturbation and differential privacy protection is characterized by the following steps:

[0009] Step 1: The central server sends the local target matrix P for the Vector Autoregression (VAR) model to the n wind farms participating in the collaborative prediction. i Local covariate matrix Z i And the noise ξ to be added Pi ξ Zi The generation rules are as follows: the historical wind power data of the above n wind farms are normalized, and the normalized data is processed according to the requirements of the central server to obtain their respective local target matrices P. i With the local covariate matrix Z i The above n wind farms are denoted as: wind farm i = 1, 2, ... n;

[0010] Step 2: The central server sends the perturbation matrix of the target matrix to m third-party servers. Perturbation matrix of covariate matrix Decrypting matrix D j The generation rules and block division method; m third-party servers randomly generate blocks according to the requirements of the central server. and And obtain D j ,Will and D j After being divided into blocks as required, the perturbation matrix of the block target matrix, the perturbation matrix of the block covariate matrix, and the block decryption matrix are formed and sent to the corresponding wind farm i respectively.

[0011] Step 3: Each wind farm i aggregates the data obtained in Step 2 from m third-party servers to obtain the perturbation matrix N of the local target matrix. Pi The perturbation matrix N of the local covariate matrix Zi and the local decryption matrix D i ; N Pi and N Zi Each is right-multiplied by its local objective matrix P i and local covariate matrix Z i The local target matrix P obtained by perturbing the matrix of each wind farm i i N Pi and local covariate matrix Z i N Zi Add noise that follows a specific distribution, and then upload it to the central server;

[0012] Step 4, the central server aggregates the data uploaded by each wind farm i in step 3 to obtain a target matrix P' and a covariant matrix Z' after matrix perturbation and noise perturbation, and then the central server estimates the coefficient matrix by using a multivariate least square method to establish a vector autoregressive prediction model for collaborative prediction of multiple wind farms;

[0013] Step 5, the central server sends the local covariant matrix generation rule to each wind farm i participating in collaborative prediction according to step 1, and each wind farm i uploads the local covariant matrix after matrix perturbation and noise perturbation according to steps 2 and 3; the central server aggregates the local covariant matrix after matrix perturbation and noise perturbation uploaded by each wind farm i, and calculates the target matrix to be decrypted according to the established prediction model. Step 6, the central server sends the target matrix P Ni after column block to each corresponding wind farm i, each wind farm i multiplies the received block target matrix P i by the local decryption matrix D Ni of itself and uploads P Ni to the central server; the central server aggregates P i uploaded by each wind farm i again to obtain the real estimated value of the target matrix . Step 7, the central server sends the real estimated value of the target matrix to each corresponding wind farm i after column block, and the wind power prediction value of each wind farm is obtained.

[0014] On the basis of the above scheme,

[0015] The normalization processing in step 1 refers to normalizing the historical wind power time series to the [0, 1] interval according to the following formula (1):

[0016]

[0017] In the formula, p t,o is the original data of the active power measured value of the wind farm at time t, p min and p max are the minimum value and the maximum value of the active power time series of the wind farm in the observation period, and p t is the active power measured value of the wind farm at time t after data normalization, i.e. 0≤p t ≤1.

[0018] The wind power time series of each wind farm i can be represented as a one-dimensional random variable, so the historical power time series of wind farms i=1, 2, …n is represented as a column vector p i , as shown in the following formula (2):

[0019] pi = [p i,1 ,p i,2 ,…Lp i,g …Lp i,N T (2);

[0020] In the above formula (2), g = 1, 2, 3…N is a data index value; N is the length of the wind power time series; T is the number of time points in the historical power time series, and the aforementioned t time point is a time point in the wind power time series;

[0021] The generation rules of the matrices P i , Z i in step 1 are shown in the following formula (3):

[0022]

[0023] In the above formula (3), i = 1, 2, …n is the number of wind farms participating in collaborative prediction, and n is the number of wind farms participating in collaborative prediction; p i,N is the historical power of each wind farm i; K is the lag order of the vector autoregressive model VAR;

[0024] The generation rules of the noise to be added ξ Pi , ξ Zi in step 1 are as follows:

[0025] ξ Pi , ξ Zi = G1(n, λ) - G2(n, λ) ∈ R T (4);

[0026] In the above formula (4), G1(n, λ) and G2(n, λ) are random variables subject to independent and identically distributed gamma distribution, λ is related to the variance of the noise, and thus the strength of differential privacy protection can be controlled.

[0027] On the basis of the above scheme, step 2 is specifically as follows:

[0028] The center server sends the perturbation matrix P The perturbation matrix Z The generation rules of the decryption matrix D j and the blocking mode to the m third-party servers; the m third-party servers respectively generate P and Z according to the requirements of the center server and obtain D j , and P and D j are blocked according to the requirements to form the blocked perturbation matrix and the blocked decryption matrix; j is the index value of the third-party server, j = 1, 2, …, m; ​

[0029] The generation rule and the blocking manner of the above-mentioned The generation rule and the blocking manner of the above-mentioned

[0030]

[0031] The generation rule and the blocking manner of the above-mentioned D j The generation rule and the blocking manner of the above-mentioned D

[0032]

[0033] The m third-party servers respectively send the blocking matrixes of and D j to each corresponding wind farm i.

[0034] On the basis of the above-mentioned scheme, step 3 is specifically:

[0035] Each wind farm i aggregates the from the m third-party servers and to obtain the perturbation matrix N of the local target matrix, the perturbation matrix N of the local covariant matrix and the local decryption matrix D, as shown in the following formula (7): Pi Zi i

[0036]

[0037]

[0038]

[0039] Each wind farm i right multiplies the local target matrix P and the local covariant matrix Z by N and N respectively, as shown in the following formula (8): Pi Zi i i

[0040] P i N Pi ∈R T×n

[0041] Z i N Zi ∈R T×nK (8);

[0042] Each wind farm i perturbs the obtained local target matrix P i N Pi and the local covariant matrix Z i N Zi ​​​​​​​​each column of P is added with the same noise vector, uploaded to the central server, as shown in the following formula (9):

[0043] P i N Pi +[ξ Pi ,ξ Pi ,Kξ Pi ]

[0044] Z i N Zi +[ξ Zi ,ξ Zi ,…Kξ Zi ] (9);

[0045] In the above formula, [ξ Pi ,ξ Pi ,…Kξ Pi ]∈R T×n , [ξ Zi ,ξ Zi ,…Kξ Zi ]∈R T×nK .

[0046] On the basis of the above scheme, step 4 is specifically:

[0047] The central server aggregates the data uploaded by each wind farm i in step 3 to obtain the target matrix P' and the covariance matrix Z' after matrix disturbance and noise disturbance, as shown in the following formula (10):

[0048]

[0049]

[0050] In the above formula,

[0051]

[0052]

[0053] Wherein, L P (λ) and L Z (λ) are noise vector sets subject to Laplace distribution;

[0054] Further, the coefficient matrix is estimated by using multivariate least squares method, and a vector autoregressive prediction model for multi-wind farm collaborative prediction is established, as shown in the following formula (11):

[0055]

[0056] On the basis of the above scheme, step 5 is specifically:

[0057] The central server republishes the local covariate matrix Z for prediction to each wind farm i participating in the collaborative prediction, as described in step 1. i The generation rules are as follows: each wind farm i uploads its local covariate matrix Z, after matrix perturbation and noise perturbation, according to steps 2 and 3 and the requirements of the central server. i N Zi +[ξ Zi ,ξ Zi ,Kξ Zi The central server aggregates the local covariate matrices after matrix perturbation and noise perturbation according to equation (9) to generate the covariate matrix Z'. f ∈R T'×nK And obtain the estimated value of the target matrix, as shown in equation (12):

[0058]

[0059] When the influence of noise disturbance is ignored, the following equation (13) is obtained;

[0060]

[0061] In the above formula This is the true estimate of the target matrix;

[0062] The central server will use the target matrix to be decrypted obtained from equation (13) Divide into columns as P Ni Then it is sent to the corresponding wind farm i, as shown in equation (14):

[0063]

[0064] In the above formula, P Ni ∈R T'×1 (i = 1, 2, 3Ln);

[0065] Each wind farm i will decrypt its own local matrix D. i Right multiply the received block target matrix P Ni P was obtained Ni D i The data is then uploaded to the central server, which aggregates the P data uploaded by each wind farm. Ni D i Obtain the true estimate of the target matrix As shown in equations (15)-(17):

[0066]

[0067] P Ni D i ∈R T'×n (16);

[0068]

[0069] will According to the column block, a column vector formed by the wind power prediction value of each wind farm i is obtained, and the central server sends the corresponding column vector to each wind farm i, that is, the wind power prediction value of each wind farm i.

[0070] The multi-wind farm collaborative prediction method based on matrix perturbation and differential privacy protection has the following beneficial effects:

[0071] (1) The perturbation matrix is generated by multiple third-party servers and sent to each wind farm, and the overall perturbation matrix needs to aggregate all third-party server perturbation matrix information, thereby increasing the difficulty of collusion between the third-party server and the central server to steal user data.

[0072] (2) The data uploaded by each wind farm is subjected to matrix perturbation and then subtracted by the noise generated by two random variables, and the two random variables are independent and identically distributed and subject to gamma distribution. In addition, each wind farm uploads wind power data in the form of a block matrix, and only by aggregating all the data uploaded by the wind farms can effective information be further obtained.

[0073] (3) The local noise is aggregated by the central server and subject to Laplace distribution, thereby satisfying the Laplace mechanism in differential privacy and further enhancing the privacy protection effect.

[0074] (4) The target matrix to be decrypted obtained by the central server according to the established prediction model needs to be divided into blocks according to the column and sent to each wind farm, and each wind farm right multiplies the received block target matrix with the local decryption matrix, and then uploads it to the central server. The central server can obtain the target matrix estimate value by aggregating the data of each wind farm again, and then obtain the wind power prediction value of each wind farm. In the whole process, each wind farm can only obtain the data related to itself, thereby increasing the difficulty of stealing the historical power data of other wind farms by a certain wind farm.

[0075] (5) The sharing of historical data between spatially adjacent wind farms can improve the wind power prediction accuracy, and the method combines matrix perturbation and noise perturbation to protect the wind farm data privacy during model training and prediction, thereby promoting the data sharing between wind farms. BRIEF DESCRIPTION OF DRAWINGS

[0076] The present application has the following drawings:

[0077] Figure 1 A flowchart of a multi-wind farm collaborative prediction method based on matrix perturbation and differential privacy protection;

[0078] Figure 2 A multi-wind farm collaborative prediction method structure framework based on matrix perturbation and differential privacy protection. DETAILED DESCRIPTION

[0079] The application will be described in further detail below with reference to the drawings.

[0080] Step 1: The central server sends the local target matrix, the local covariant matrix of the vector autoregressive model VAR, and the generation rule of the noise to be added to the wind farms participating in the collaborative prediction. Then, each wind farm normalizes the historical wind power data, processes the normalized data according to the requirements of the central server, and obtains the local target matrix and the local covariant matrix.

[0081] The central server sends information to the n wind farms participating in the collaborative prediction, and the information includes the lag order K of the vector autoregressive model (VAR), the wind power time series range and its length N for model training.

[0082] The n wind farms normalize the original historical wind power time series to the [0, 1] interval according to the following formula:

[0083]

[0084] In the formula, p t,o is the original data of the active power measured value of the wind farm at t time, p min and p max are the minimum value and the maximum value of the wind farm active power time series in the observation period, respectively, and p t is the active power measured value of the wind farm at t time after data normalization, that is, 0≤p t ≤1.

[0085] The wind power time series of each wind farm can be represented as a one-dimensional random variable, so the historical power time series of the wind farm i = 1, 2, … n is represented as a column vector p i :

[0086] p i = [p i,1 , p i,2 , … p i,g … p i,N ] T (2)

[0087] In the formula, g = 1, 2, 3, … N is the data index value.

[0088] VAR takes the historical data of all wind farms in the observation area as input and takes the wind power output of a specific target wind farm in the area as output. Taking a single wind farm i as an example, the VAR model at t time can be represented as:

[0089]

[0090] where are the autoregressive coefficients, representing the influence of the power data of wind farm f lagging k steps on the output of wind farm i at time t (f = 1, 2,..., n); ε i,t is a Gaussian white noise with mean 0.

[0091] The VAR model of all wind farms at time t (t > K) can be expressed as:

[0092]

[0093] where p t = [p 1,t , p 2,t ,..., Kp n,t ] ∈ R n , is a row vector composed of the historical power data of wind farm i, B i ∈ R K×n is a is the corresponding local coefficient matrix, and ε t ∈ R n is a Gaussian white noise vector at time t. B i and ε t are composed as follows:

[0094]

[0095]

[0096] ε t = [ε 1,t , ε 2,t ,..., Kε n,t ] (5)

[0097] There are N-K equations of type (4) for a wind power time series of length N, denoted as T = N-K. The VAR model at a single time t in equation (4) is extended to T times, expressed in matrix form as:

[0098] P = ZB + E (6)

[0099] where P ∈ R T×n is called the target matrix of the VAR model, Z ∈ R T×nK is called the covariate matrix of the VAR model; B ∈ R nK ×n is called the coefficient matrix of the VAR model, composed of the local coefficient matrices in equation (5); E ∈ R T×nThe error matrix is called VAR model. In order to facilitate data block aggregation and representation, the formula (6) is divided into blocks according to the data provided by different wind farms, so that the formula (6) can be rewritten as:

[0100]

[0101] P i , Z i , B i , E i (i = 1, 2, … n) are the local target matrix, local covariate matrix, unit coefficient matrix and unit error matrix of the wind farm i respectively, and their definitions are as follows:

[0102]

[0103]

[0104]

[0105]

[0106] In fact, according to the block matrix multiplication property, we can get:

[0107]

[0108] The goal of establishing and training the VAR model is to estimate the coefficient matrix , i.e.

[0109]

[0110] At the beginning of step 1, the central server will send the local target matrix and local covariate matrix generation rules in formula (8) to the wind farms, and then each wind farm will process the normalized historical wind power data according to the requirements, so as to obtain the local target matrix P i and the local covariate matrix Z i of each wind farm.

[0111] Step 2: The central server sends the perturbation matrix of the target matrix, the perturbation matrix of the covariate matrix, the generation rule of the decryption matrix and the block method of the two perturbation matrices and the decryption matrix to each third-party server. Multiple third-party servers randomly generate two perturbation matrices according to the requirements of the central server, and obtain the decryption matrix. After the two perturbation matrices and the decryption matrix are divided into blocks according to the requirements, the block perturbation matrix and the block decryption matrix are formed, and they are sent to each wind farm.

[0112] Firstly, the third-party servers randomly generate an invertible perturbation matrix of elements between 0 and 1 / m, and block them according to the requirements of the central server. The perturbation matrix of the target matrix and the perturbation matrix of the covariant matrix generated by the jth (j = 1, 2, 3…m) third-party server are respectively:

[0113]

[0114] In order to eliminate the influence of the perturbation matrix, the inverse matrix of the perturbation matrix of the target matrix is needed, which is called the decryption matrix. The decryption matrix generated by the jth third-party server is:

[0115]

[0116] The m third-party servers respectively send the block matrix to the corresponding wind farms.

[0117] At the beginning of step 2, the central server sends the generation rules of the perturbation matrix of the target matrix and the perturbation matrix of the covariant matrix and the decryption matrix, as well as the block method of the two perturbation matrices and the decryption matrix, to each third-party server. Here, multiple third-party servers generate perturbation matrices and send them to wind farms in blocks, so that it is difficult for a single wind farm and the central server to obtain the overall perturbation matrix N P or N Z , and thus it is impossible to eliminate the influence of the perturbation matrix to obtain the original wind power data.

[0118] Step 3, each wind farm aggregates all the third-party server data to obtain two local perturbation matrices and a local decryption matrix, respectively right multiplies the two local perturbation matrices by the local target matrix and the local covariant matrix. Finally, each wind farm adds noise subject to a specific distribution to the local target matrix and the local covariant matrix after matrix perturbation, and uploads them to the central server.

[0119] When data privacy is not considered, the parameters that the wind farm needs to upload to the central server are the local target matrix P i and the local covariant matrix Z i , and in the privacy protection scenario, the method mainly makes each wind farm upload the local target matrix and the local covariant matrix after matrix perturbation and noise perturbation to the central server:

[0120] (1) Matrix perturbation:

[0121] The wind farm i aggregates all the block perturbation matrices and block decryption matrices of the third-party servers to obtain the perturbation matrix of the local target matrix and the perturbation matrix of the local covariant matrix, as well as the local decryption matrix:

[0122]

[0123]

[0124]

[0125] The wind farm i right multiplies its local target matrix and local variable matrix by the perturbation matrix respectively: P i N Pi ∈R T×n

[0126] Z i N Zi ∈R T×nK (14)

[0127] In fact, according to the block matrix multiplication property, the aggregation of each local target matrix and local variable matrix after matrix perturbation in formula (14) is:

[0128]

[0129]

[0130] And for the decryption matrix:

[0131]

[0132] (2) Noise perturbation

[0133] Differential privacy is a strict security model that can be mathematically proven, and its idea is to add noise to the original through a certain mechanism to interfere with the original, so that the privacy-sensitive data is distorted, while the overall statistical law of the data set remains unchanged, achieving a compromise between security and data availability. Differential privacy is essentially a noise perturbation mechanism, and this method aims to add noise L(λ) that satisfies the Laplace distribution to the target matrix PN P and the variable matrix ZN Z after matrix perturbation in formula (15) using the Laplace mechanism. Considering that the data of each wind farm is distributedly stored in each wind farm, this paper uses a distributed differential privacy model to independently add noise that satisfies differential privacy to the local target matrix and local variable matrix after matrix perturbation of each wind farm.

[0134] Using the infinite divisibility of the Laplace distribution, it can be decomposed into the form of n independent and identically distributed sums:

[0135]

[0136] where G1(n, λ) and G2(n, λ) are random variables following independent gamma distributions, λ is related to the variance of noise and can further control the strength of differential privacy protection. Then define the random noise column vector ξ Pi Zi = G1(n, λ) - G2(n, λ) ∈ R T , the local objective matrix and the local covariance matrix of each wind farm after matrix perturbation are respectively:

[0137]

[0138] where [ξ Pi , ξ Pi , Kξ Pi ] ∈ R T×n , [ξ Zi , ξ Zi , Kξ Zi ] ∈ R T×nK .

[0139] Step 4: The central server aggregates the data uploaded by all wind farms to obtain the objective matrix and the covariance matrix after matrix perturbation and noise perturbation, and then estimates the coefficient matrix by using the multivariate least square method to establish a vector autoregressive prediction model for multi-wind farm collaborative prediction.

[0140] According to the above formula (15) (17), the central server aggregates the data of n wind farms to obtain:

[0141]

[0142]

[0143] where:

[0144]

[0145]

[0146] Further, the multivariate least square method is used to estimate the coefficient matrix:

[0147]

[0148] In fact, if no noise is added to the objective matrix and the covariance matrix, the coefficient matrix estimate after matrix perturbation can be obtained according to formula (10):

[0149]

[0150] ​Step 5: The central server sends the local co-variable matrix generation rule to each wind farm participating in the collaborative prediction according to step 1, and the wind farm uploads the local co-variable matrix after disturbance according to the requirement of the central server. After aggregating all the wind farm data, the central server calculates the target matrix to be decrypted according to the established prediction model. The central server sends the target matrix to be decrypted to the corresponding wind farm after being divided into blocks by column, and each wind farm right multiplies the received block target matrix by the local decryption matrix and uploads it to the central server. The central server aggregates the data uploaded by each wind farm to obtain the true estimate value of the target matrix, and sends the target matrix to the corresponding wind farm after being divided into blocks by column, which is the wind power prediction value of each wind farm.

[0151] In the prediction stage, the central server reissues the range and length of the wind power time series required for prediction, and the wind farm still uploads the local co-variable matrix after matrix disturbance and noise disturbance according to the above rule. After being aggregated by formula (18), the co-variable matrix is generated, denoted as Z' f ∈R T'×nK Then, the estimate value of the target matrix is obtained according to formula (7) (21):

[0152]

[0153] When the influence of noise disturbance is ignored, the simultaneous equations (6) (10) (19) (22) can obtain an approximate result:

[0154]

[0155] In the formula is the true estimate value of the target matrix. In order to eliminate the influence of the disturbance matrix N P , a natural idea is to right multiply the inverse matrix of N P , that is, the decryption matrix D. However, D is distributed stored in each wind farm, so the block matrix method is still used here. The obtained in formula (24) is divided by column to obtain:

[0156]

[0157] In the formula Ni ∈R T'×1 (i=1, 2, 3Ln), according to the properties of block matrix multiplication:

[0158]

[0159] Then the central server sends P Ni to the wind farm i, and the wind farm i right multiplies D i on P Ni , that is:

[0160] P Ni D i ∈R T'×n (27)

[0161] The wind farm i sends P Ni D i to the center server, and the center server aggregates the data of each wind farm to obtain:

[0162]

[0163] The column is blocked as:

[0164]

[0165] In the formula, P fi ∈R T'×1 (i=1,2,3Ln), that is, a column vector composed of the wind power prediction value of each wind farm is obtained, and finally the center server publishes the wind power prediction data of each wind farm.

[0166] In this way, the disturbance matrix and the decryption matrix are always mastered by each wind farm after being blocked in the whole wind power prediction process, and each wind farm can only access the data related to itself. The center server mainly aggregates and publishes data, and the data it can access are all disturbed. It is worth noting that due to the influence of noise disturbance, according to formula (24), the final wind power prediction accuracy will decrease compared with that without adding noise.

[0167] The contents not described in detail in the specification belong to the prior art known to those skilled in the art.​

Claims

1. A method for multi-wind farm collaborative forecasting based on matrix perturbation and differential privacy protection, characterized in that, Comprising the following steps: Step 1, the central server sends the local target matrix P of the vector autoregressive model VAR to n wind farms participating in collaborative prediction i , the local covariate matrix Z i , and the generation rule of the noise to be added ξ Pi , ξ Zi ; the n wind farms normalize the historical wind power data, process the normalized data according to the requirements of the central server, and obtain the local target matrix P i and the local covariate matrix Z i of each wind farm; the n wind farms are denoted as: wind farm i = 1, 2, … n; Step 2, the center server sends the perturbation matrix of the target matrix to m third-party servers The perturbation matrix of the covariant matrix The decryption matrix D j The generation rule and the block mode; m third-party servers respectively generate and and obtain D j , and and D j are blocked according to the requirement to form the perturbation matrix of the block target matrix, the perturbation matrix of the block covariant matrix and the block decryption matrix and are respectively sent to each corresponding wind farm i; Step 3, each wind farm i will get the local target matrix disturbance matrix N after the third party server data from m aggregation Pi , local covariance matrix disturbance matrix N Zi and local decryption matrix D i ; N Pi and N Zi right multiply its local target matrix P i and local covariance matrix Z i ; each wind farm i will get the local target matrix P i N Pi and local covariance matrix Z i N Zi plus noise subject to a certain distribution, upload to the central server; Step 4, the central server aggregates the data uploaded by each wind farm i obtained in step 3 to obtain a target matrix P' and a covariant matrix Z' after matrix perturbation and noise perturbation, and then the central server estimates the coefficient matrix using a multivariate least square method to establish a vector autoregressive prediction model for multi-wind farm collaborative prediction. Step 5, the center server re-sends the local covariate matrix generation rule to each wind farm i participating in the collaborative prediction according to step 1, each wind farm i uploads the local covariate matrix after matrix perturbation and noise perturbation according to steps 2 and 3 as required by the center server; the center server aggregates the local covariate matrix uploaded by each wind farm i after matrix perturbation and noise perturbation, and calculates the target matrix to be decrypted according to the established prediction model The center server sends the target matrix to be decrypted to each wind farm i Block by column into P Ni Send to each corresponding wind farm i, each wind farm i multiplies the received block target matrix P i Right and uploads to the center server Ni ​ The central server aggregates the P uploaded by each wind farm i again Ni D i The real estimated value of the target matrix is obtained The After being divided by column, the data is sent to each corresponding wind farm i, and the wind power prediction value of each wind farm is obtained.

2. The method of claim 1, wherein the method is characterized in that: The normalization processing in step 1 refers to normalizing the historical wind power time series to the interval [0, 1] according to the following formula (1): where p t,o is the original data of the measured active power of the wind farm at time t, p min and p max are the minimum and maximum values of the time series of the active power of the wind farm in the observation period, respectively, p t is the measured active power of the wind farm at time t after data normalization, i.e., 0≤p t ≤1; The wind power time series of each wind farm i can be represented as a one-dimensional random variable, and thus the historical power time series of wind farms i = 1, 2,... n are represented as column vectors p i As shown in the following formula (2): p i = [p i,1 , p i,2 , Lp i,g Lp i,N ] T (2) ; In the above formula (2), g = 1, 2, 3…N is a data index value; N is the length of the wind power time series; T is the number of time points in the historical power time series, and the aforementioned t time point is a time point in the wind power time series. The matrices P described in step 1 i , Z i The generation rule of Z is shown in the following equation (3): In the above formula (3), i = 1, 2, … n is the number of wind farms participating in collaborative prediction, and n is the number of wind farms participating in collaborative prediction; p i,N is the historical power of each wind farm i; K is the lag order of the vector autoregressive model VAR. The noise ξ to be added in step 1 Pi , ξ Zi The generation rule is: ξ Pi ,ξ Zi = G1(n, λ) - G2(n, λ) e R T (4); In the above formula (4), G1(n, λ) and G2(n, λ) are random variables subject to independent and identically distributed gamma distribution, λ is related to the variance of the noise, and thus the strength of differential privacy protection can be controlled.

3. The method of claim 2, wherein: Step 2 is specifically: The center server sends the perturbation matrix of the target matrix to m third-party servers The perturbation matrix of the covariant matrix The decryption matrix D j The generation rule and the block mode; m third-party servers randomly generate and and obtain D j , and and D j are blocked to form the block perturbation matrix and the block decryption matrix; j is the index value of the third-party server, j = 1, 2, …, m; The above The generation rule and the block manner are shown in the following formula (5): The above D j The generation rule and the block mode of the above D are shown in the following expression (6). The m third-party servers respectively send and D j the block matrix to each corresponding wind farm i.

4. The method of claim 3, wherein: Step 3 is specifically: The wind farms i aggregate the results of step 2 from m third party servers and The disturbance matrix N of the local target matrix obtained after Pi the disturbance matrix N of the local covariant matrix Zi and the local decryption matrix D i as shown in equation (7) below: Each wind farm i multiplies its local target matrix P Pi and local covariant matrix Z Zi by its local target matrix P i and local covariant matrix Z i respectively, as shown in the following equation (8): P i N Pi ∈R T×n Z i N Zi ∈R T×nK (8); Each wind farm i will get the local target matrix P after matrix perturbation i N Pi and the local covariant matrix Z i N Zi Each column of P and Z is added with the same noise vector, uploaded to the central server, as shown in the following equation (9): P i N Pi +[ξ Pi ,ξ Pi ,...ξ Pi ] Z i N Zi +[ξ Zi ,ξ Zi ,...ξ Zi ] (9) In the above formula, [ξ Pi ,ξ Pi ,...ξ Pi ] ∈ R T×n , [ξ Zi ,ξ Zi ,...ξ Zi ] ∈ R T×nK .

5. The method of claim 4, wherein: Step 4 is specifically: After the central server aggregates the data uploaded by each wind farm i in step 3, the target matrix P' and the covariate matrix Z' after matrix perturbation and noise perturbation are obtained, as shown in the following formula (10): In the above formula, wherein L P (λ) and L Z (λ) are sets of noise vectors subject to Laplace distribution. Further, the multiple least squares method is used to estimate the coefficient matrix and a vector autoregressive prediction model for the collaborative prediction of multiple wind farms is established, as shown in the following equation (11).

6. The method of claim 5, wherein: Step 5 is specifically: The central server reissues the local covariate matrix Z for prediction to each wind farm i participating in the collaborative prediction according to step 1 i The generation rule is that each wind farm i uploads the local covariate matrix Z after matrix perturbation and noise perturbation according to steps 2 and 3 according to the requirements of the central server i N Zi +[ξ Zi ,ξ Zi ,...ξ Zi ] The central server aggregates the local covariate matrix after matrix perturbation and noise perturbation according to formula (9) to generate the covariate matrix Z' f ∈R T'×nK And get the estimated value of the target matrix as shown in the following formula (12): When the influence of noise perturbation is ignored, the following formula (13) is obtained: In the above formula is the true estimate of the target matrix; The center server sends the target matrix to be decrypted obtained in formula (13) to each corresponding wind farm i, as shown in the following formula (14): The center server sends the target matrix to be decrypted obtained in formula (13) to each corresponding wind farm i, as shown in the following formula (14): Ni The center server sends the target matrix to be decrypted obtained in formula (13) to each corresponding wind farm i, as shown in the following formula (14): P in the above formula Ni ∈R T'×1 (i = 1, 2, 3 Ln) Each wind farm i will multiply the received partial target matrix P i with the local decryption matrix D Ni , obtaining P Ni D i and upload to the central server, which aggregates the P Ni D i received from each wind farm i, obtaining the real estimate of the target matrix as shown in the following equations (15)-(17): P Ni D i ∈R T'×n (16) The wind power prediction value of each wind farm i is obtained by dividing the wind power prediction value of each wind farm i by the number of wind farms in the wind farm cluster, and the wind power prediction value of each wind farm i is obtained by dividing the wind power prediction value of each wind farm i by the number of wind farms in the wind farm cluster. The column vector is sent to each wind farm i by the central server, and the wind power prediction value of each wind farm i is sent to each wind farm i by the central server.

Images