A distributed estimation method and system for community-level photovoltaic power generation

Abstract

The present invention discloses a community-level post-meter photovoltaic power generation distributed estimation method and system, which relates to the technical field of post-meter photovoltaic power generation estimation. The method comprises: obtaining meteorological data of each community and net load power data of photovoltaic power generation users in the community, inputting them into a corresponding personalized local estimation model, and estimating photovoltaic power generation power; the training process of the model is: obtaining photovoltaic power generation related historical data of each community, and constructing a local data set; taking each community as a client, adopting a federated meta-learning method, and collaboratively training the estimation models of each client through multiple rounds of communication; wherein, in each round of communication, a differential privacy mechanism based on a privacy budget adaptive allocation strategy is adopted to add noise to the gradient uploaded by the client, and a gradient estimation method based on similarity is adopted during uploading to estimate the gradient of the client with communication failure. The present invention realizes higher precision and more stable post-meter photovoltaic power generation estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of post-meter photovoltaic power generation estimation, and in particular to a community-level post-meter photovoltaic power generation distributed estimation method and system. Background Art

[0002] At present, distributed photovoltaic systems (DPVSs) have been favored by more and more users due to their low installation cost and easy operation, and distributed photovoltaic power generation has developed rapidly. However, most small-capacity distributed photovoltaic power generation systems are installed behind the meter (Behind-the-Meter, BTM), and photovoltaic power generation information is not measured separately. Utilities / companies can only obtain net load data, where net load is defined as actual load minus photovoltaic power generation. In other words, distributed photovoltaic power generation is invisible to the front side of the meter (distribution network side). The unobservability of distributed photovoltaic power generation will bring a series of problems to the distribution system, such as unacceptable net load prediction errors, inaccurate reliability analysis, and obstacles to the design of optimal power outage restoration plans. Therefore, it is necessary to study the power estimation of photovoltaic power generation behind the meter and separate distributed photovoltaic power generation from net load.

[0003] At present, the estimation of photovoltaic power generation after the meter is mainly divided into model-driven methods and data-driven methods. Among them, the model-driven method refers to inferring the physical parameters of the photovoltaic power generation system based on the data of smart meters, and then estimating the photovoltaic power generation power at any time by building a physical model of the photovoltaic power generation system. However, this model-driven method is difficult to simulate the factors affecting the photovoltaic power generation power such as weather type, cloud movement, shading, and photovoltaic panel pollution, resulting in low estimation accuracy of the model-driven method. In contrast, the data-driven method does not rely on the physical model of the photovoltaic power generation system, but models the photovoltaic power generation power after the meter based on smart meter data and meteorological data. The commonly used data-driven methods include context-supervised source separation methods, dictionary learning methods, machine learning methods, etc. Among them, machine learning methods such as random forests and neural networks have been applied in the estimation of photovoltaic power generation after the meter and achieved good estimation performance. Furthermore, since machine learning methods have high requirements on the quantity and quality of data, they usually adopt centralized methods for learning and training, that is, they require all users' smart meter data and meteorological data to be uploaded to a central service platform for training.

[0004] As people pay more attention to privacy and security, considering that the training process of centralized methods is likely to cause privacy issues, such as the leakage of information such as the community's electricity consumption pattern and location, current research focuses more on distributed estimation methods, that is, estimation methods in which all data is kept locally. Among them, federated learning has achieved better results in load power forecasting, non-intrusive load monitoring and other fields due to its distributed learning capabilities. In the federated learning method, each client trains the model on a local device and then uploads the model parameters to the central server. It does not need to share the original data, thereby reducing the risk of privacy leakage.

[0005] However, when applying the federated learning method to the community-level post-meter photovoltaic power generation estimation scenario, the following problems still exist:

[0006] 1) Community heterogeneity: There are obvious differences in the geographical environment and photovoltaic power generation system configuration of different communities, resulting in unique photovoltaic power generation characteristics in each community. Traditional federated learning methods only train a single global model, which is difficult to adapt to the photovoltaic power generation characteristics of each community.

[0007] 2) Communication failure. Due to unstable communication links and network congestion, communication failure often occurs during federated learning, which will cause some clients to be unable to upload model parameters to the server in time. These clients are called unavailable clients. The lack of model parameter information from unavailable clients will lead to a decline in global model performance and reduced estimation accuracy.

[0008] 3) Malicious privacy attacks. Although the federated learning method only uploads model parameters instead of original data, it has stronger privacy protection capabilities than the centralized method. However, attackers can still restore some sensitive information through model parameters and infer whether the obtained data comes from a specific party, thereby leaking user electricity usage data. Summary of the invention

[0009] In order to address the shortcomings of the above-mentioned prior art, the present invention provides a community-level post-meter photovoltaic power distributed estimation method and system, which takes each community as a client, adopts a federated meta-learning method as the main architecture, and collaboratively trains each client. In this process, a gradient estimation method for coping with communication failure is embedded to mitigate the negative impact of communication failure. A differential privacy mechanism based on a privacy budget adaptive allocation strategy is adopted to deal with malicious privacy attacks, thereby achieving more accurate and stable distributed estimation of community-level post-meter photovoltaic power.

[0010] In a first aspect, the present invention provides a method for distributed estimation of community-level post-meter photovoltaic power generation power.

[0011] A distributed estimation method for community-level photovoltaic power generation after the meter, comprising:

[0012] Obtain meteorological data for each community and net load power data of photovoltaic power generation users in the community;

[0013] The acquired data is input into the personalized local estimation model corresponding to the community to estimate the photovoltaic power generation power of each community; wherein the training process of the personalized local estimation model includes:

[0014] Obtain historical data related to photovoltaic power generation in each community and build a local data set;

[0015] Each community is regarded as a client, and the federated meta-learning method is used to collaboratively train the photovoltaic power estimation model of each client through multiple rounds of communication to obtain a personalized local estimation model for each community. In each round of communication, the differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, and the similarity-based gradient estimation method is used during uploading to estimate the gradient of the client with communication failure.

[0016] In a second aspect, the present invention provides a community-level distributed estimation system for post-meter photovoltaic power generation.

[0017] A distributed estimation system for community-level photovoltaic power generation after the meter, comprising:

[0018] A data acquisition module is used to obtain meteorological data of each community and net load power data of photovoltaic power generation users in the community;

[0019] The photovoltaic power generation estimation module is used to input the acquired data into the personalized local estimation model corresponding to the community to estimate the photovoltaic power generation of each community; wherein the training process of the personalized local estimation model includes:

[0020] Obtain historical data related to photovoltaic power generation in each community and build a local data set;

[0021] Each community is regarded as a client, and the federated meta-learning method is used to collaboratively train the photovoltaic power estimation model of each client through multiple rounds of communication to obtain a personalized local estimation model for each community. In each round of communication, the differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, and the similarity-based gradient estimation method is used during uploading to estimate the gradient of the client with communication failure.

[0022] In a third aspect, the present invention further provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps of the method described in the first aspect are completed.

[0023] In a fourth aspect, the present invention further provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the method described in the first aspect.

[0024] One or more of the above technical solutions have the following beneficial effects:

[0025] 1. The present invention provides a community-level post-meter photovoltaic power distributed estimation method and system, which takes each community as a client, adopts a federated meta-learning method as the main architecture, and collaboratively trains each client to obtain initial model parameters with strong adaptability for each client. After local training, a personalized photovoltaic power estimation model suitable for each client can be quickly obtained. In addition, a gradient estimation method for coping with communication failure is embedded in this federated meta-learning process to reduce the negative impact of communication failure, and a differential privacy mechanism based on a privacy budget adaptive allocation strategy is adopted to deal with malicious privacy attacks. Through the above method, the problems of community heterogeneity, communication failure, and malicious privacy attacks in the existing community-level post-meter photovoltaic power estimation based on the federated learning method are solved, and a more accurate and stable distributed estimation of community-level post-meter photovoltaic power is achieved.

[0026] 2. The present invention adopts a federated meta-learning method to quickly generate a personalized model for each community to provide a more accurate photovoltaic power estimation result; the gradient estimation method can enhance the robustness to communication failures, and when communication fails, it can still provide estimation accuracy and convergence performance similar to normal communication conditions; by adopting a differential privacy mechanism based on a privacy budget adaptive allocation strategy, while responding to malicious privacy attacks, the privacy budget of each client and the privacy budget of each round of the client can be flexibly adjusted to reduce the impact of differential privacy noise on model performance, thereby effectively improving the efficiency of model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0028] Figure 1 This is a flow chart of training a personalized local estimation model for each community in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] It should be noted that the following detailed descriptions are exemplary only, are intended to describe specific embodiments, are intended to provide further explanation of the present invention, and are not intended to limit exemplary embodiments according to the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art to which the present invention belongs. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0030] Embodiment 1

[0031] This embodiment provides a distributed estimation method for community-level photovoltaic power generation after the meter, which specifically includes:

[0032] Obtain meteorological data for each community and net load power data of photovoltaic power generation users in the community;

[0033] The acquired data is input into the personalized local estimation model corresponding to the community to estimate the photovoltaic power generation power of each community;

[0034] The training process of the personalized local estimation model includes:

[0035] Obtain historical data related to photovoltaic power generation in each community and build a local data set;

[0036] Each community is regarded as a client, and the federated meta-learning method is used to collaboratively train the photovoltaic power estimation model of each client through multiple rounds of communication to obtain a personalized local estimation model for each community. In each round of communication, the differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, and the similarity-based gradient estimation method is used during uploading to estimate the gradient of the client with communication failure.

[0037] In this embodiment, for the community-level post-meter photovoltaic power generation estimation scenario, a federated meta-learning method is used for collaborative training to quickly generate a personalized local estimation model for each community to provide more accurate photovoltaic power generation estimation results and effectively solve the problems of community heterogeneity, communication failure, and malicious privacy attacks that exist when using the federated meta-learning method for training. Figure 1 As shown, the training process of the personalized local photovoltaic power estimation model provided in this embodiment is introduced in more detail through the following content.

[0038] Step S1, obtaining photovoltaic power generation related historical data of each community and constructing a local data set, wherein the photovoltaic power generation related historical data includes the historical meteorological data of the local community, the historical net load power data of observable photovoltaic power generation users in the local community, and the historical photovoltaic power generation power data.

[0039] Specifically, the meteorological data of each community and the smart meter data of all photovoltaic power generation users in the community are collected and processed. In each community, photovoltaic power generation users can be divided into two sets according to whether a separate smart meter is installed for the photovoltaic power generation system behind the meter: (1) Set 1: observable photovoltaic power generation users, defined as FO, whose photovoltaic power generation power is measured separately by smart meters. Therefore, for FO users, photovoltaic power generation power, load power and net load power are all known; (2) Set 2: unobservable photovoltaic users, defined as NO, whose only net load power data is measured. Therefore, for NO users, photovoltaic power generation power is unknown. The smart meter data of all photovoltaic power generation users (including photovoltaic power generation power, load power and net load power data) are collected in the local data concentrator in the community and uploaded to the central service platform through the public network.

[0040] In this embodiment, the ultimate goal of the community-level post-meter photovoltaic power generation estimation is to obtain the photovoltaic power generation of all photovoltaic power generation users in the community, that is, the sum of the photovoltaic power generation of FO users and NO users. Therefore, in this embodiment, the known photovoltaic power generation related historical data of FO users in the community is obtained through the local data concentrator in the community, and a neural network estimation model is trained using the known historical data, so as to use the estimation model to estimate the unknown post-meter photovoltaic power generation in the community.

[0041] Specifically, for the FO users in the i-th community, the meteorological data, net load power data, and photovoltaic power data are all known. Therefore, the photovoltaic power estimation problem after the table can be considered as a supervised machine learning task. The local data concentrator in the community is used to obtain relevant historical data. With historical meteorological data and historical net load power data as input features and historical photovoltaic power as the target output, a neural network model can be trained, as shown in formula (1):

[0042] PV FO,i =f FO,i (w FO,i ,T FO,i ,Net FO,i ) (1)

[0043] Among them, f FO,i represents the photovoltaic power generation power estimation model of the FO user in the i-th community; w FO,i Represents the model parameters of the neural network, including the weights and biases of neurons in each layer; T FO,i represents the set of historical meteorological data of the FO user in the i-th community, which includes total radiation, direct radiation, diffuse radiation, temperature, relative humidity and wind speed; PV FO,i and NetFO,i They respectively represent the historical photovoltaic power generation data and historical net load power data of the FO user in the i-th community.

[0044] Furthermore, since the two types of photovoltaic users (FO users and NO users) in the same community have similar geographical environments and similar photovoltaic system physical parameters, FO users and NO users are likely to have similar photovoltaic power generation characteristics. To this end, this embodiment uses the trained photovoltaic power generation power estimation model f FO,i To approximate the total photovoltaic power estimation model f of the i-th community i .

[0045] Finally, a local training set is constructed according to the above input features and target outputs, and the community local estimation model is trained using the local training set. Based on the model, an accurate estimation of the photovoltaic power generation power of the i-th community at time t (the time can be estimated every 15 / 30 minutes) is achieved.

[0046] Step S2: Take each community as a client, adopt the federated meta-learning method, and through multiple rounds of communication, collaboratively train the photovoltaic power estimation model of each client to obtain a personalized local estimation model for each community. In each round of communication, a differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, and a similarity-based gradient estimation method is used during uploading to estimate the gradient of the client that failed to communicate.

[0047] Step S2.1, taking each community as a client, adopting the federated meta-learning method, through multiple rounds of communication, collaboratively training the photovoltaic power estimation model of each client, and obtaining the final updated initial model parameters, that is, obtaining the initial model parameters of the neural network with strong adaptability.

[0048] This embodiment adopts a federated meta-learning method, taking each community as a client. Assuming there are n clients in total, each client communicates with a central service platform, which can be called a central server. The federated meta-learning method consists of R rounds of communication processes. The specific execution process steps of each round are as follows:

[0049] 1) The server sends the latest neural network initial model parameters to each client. The neural network initial model parameters of the rth round are represented as w r .

[0050] 2) Each client trains the local model based on the initial model parameters sent by the central server and the local data set, including the meta-training process and the meta-testing process. The following takes the rth round client i as an example to introduce the local training process.

[0051] In the rth round of communication, client i samples the support set from the training set. and queryset Both the support set and the query set contain input features and target outputs.

[0052] During meta-training, the initial model parameters w r As the initial value of the model parameter w, training is performed on the support set, and the model parameters are updated by one-step gradient descent. The loss function of the meta-training process is shown in formula (2), and the gradient descent process is shown in formula (3):

[0053]

[0054] Where x, y represent the input features and output targets on the support set of client i in round r, respectively; f(w, x) represents the neural network model of client i in round r; w represents the model parameters of the neural network; η1 represents the learning rate of the gradient descent process, Representation function The gradient of the neural network model after the meta-training process is expressed as

[0055] During the meta-testing process, As the initial value of the model parameter w, the test is performed on the query set. The loss function of the meta-test process is shown in formula (4). At this time, the gradient of the loss function is As shown in formula (5):

[0056]

[0057]

[0058] When the meta-test process is completed, the gradient of the loss function obtained by training is Upload to the central server.

[0059] 3) The server receives gradient information from all clients According to the gradient, the initial model parameter w of the neural network r Update, the update process is shown in formula (6):

[0060]

[0061] Among them, w r+1 is the updated initial model parameter of the neural network, n is the number of clients, η2 represents the learning rate during the parameter update process, represents the gradient of client i in round r.

[0062] Repeat steps 1)-3) for multiple rounds of communication and training until the updated neural network initial model parameters converge or the specified number of communication rounds is reached. The final updated w R These are the initial parameters of the neural network model with the strongest adaptability.

[0063] Step S2.2: In each round of communication, a differential privacy mechanism based on a privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client to achieve enhanced privacy protection.

[0064] Specifically, the specific implementation process of the differential privacy mechanism includes clipping the gradient uploaded by the client and then adding noise. In this embodiment, for each client, the gradient of the client is clipped in turn, and random noise is generated based on the Laplace mechanism, and the random noise is added to the gradient of the client, and then the gradient with added noise is uploaded to the server.

[0065] The clipping operation and noise adding operation of the gradient of client i in the rth round are shown in equations (7) and (8) respectively:

[0066]

[0067] in, represents the gradient of client i in round r, represents the gradient of client i in the rth round after clipping, C represents the clipping threshold, and the gradient is limited to [-C, C] through clipping operation; represents the gradient of client i in round r after adding noise, n D The random noise generated by the Laplace mechanism (Laplace) for each client, the Laplace distribution is shown in formula (9):

[0068]

[0069] That is, the added random noise n D Satisfies Laplace distribution, and μ = 0, b = Δs / ε. ε represents the privacy budget, Δs is the sensitivity, and the calculation formula is shown in formula (10):

[0070]

[0071] Where C represents the clipping threshold and |D| represents the number of samples used by the client for training.

[0072] Furthermore, in view of the heterogeneity of community data and the possibility of communication failure, this embodiment adopts a privacy budget adaptive allocation strategy that takes personalization and communication into consideration to balance the privacy protection level and model training efficiency.

[0073] The above-mentioned personalized privacy budget adaptive allocation strategy means that in the community-level post-meter photovoltaic power generation estimation problem, the photovoltaic power generation time series of some communities are highly random and difficult to estimate, while the photovoltaic power generation time series of some communities are relatively regular and easy to estimate. For this reason, this embodiment introduces approximate entropy to describe the complexity of photovoltaic power generation time series of different communities, and allocates different privacy budgets to each client to achieve personalized differential privacy protection.

[0074] Specifically, based on the historical photovoltaic power generation time series in the client's local data set, the approximate entropy of the time series is calculated by comparing the similarity of subsequences in different time windows to measure the complexity of the time series. The community's photovoltaic power generation time series {X(t), t=1,2,…,n} is used as input, and the approximate entropy E of the community's photovoltaic power generation time series is used as output. The approximate entropy calculation method includes:

[0075] ① Taking m as the window width, divide the time series into k subsequences, where the i-th subsequence is X i (t)={x(i),x(i+1),…,x(i+m-1)};

[0076] ② Calculate the Manhattan distance d between each subsequence and all k subsequences ij =|X i (t)-X j (t)|1;

[0077] ③Define threshold F, statistics and subsequence X i The ratio of the number of subsequences whose distance to (t) is greater than F to the total number of subsequences k is denoted as r: 0.1-0.25, SD: standard deviation of subsequences;

[0078] ④ Calculate the logarithmic mean of k subsequences

[0079] ⑤Change the window width m to m+1 and repeat steps ①-④;

[0080] ⑥ Calculate approximate entropy

[0081] Through the above method, the approximate entropy is calculated, and the larger the approximate entropy, the more complex the photovoltaic power generation time series is, which can provide more useful information for model training, and the more difficult it is to estimate. Therefore, this embodiment considers allocating more privacy budget to clients with larger approximate entropy, that is, according to the approximate entropy of the client, combined with the preset global privacy budget, the privacy budget of each round of the client is obtained. Specifically, assuming that the global privacy budget is ε, the approximate entropy of client i is ApEn i , then the privacy budget of client i is as shown in formula (11):

[0082] ε i =ApEn i log2ε (11)

[0083] Furthermore, the privacy budget adaptive allocation strategy of communication is considered: in the traditional differential privacy mechanism, a fixed privacy budget is arranged for each client in each round. However, due to communication failure, the client did not upload the gradient in some communication rounds, which means that the privacy budget of client i in the round of communication failure is saved. For this reason, in this embodiment, these saved privacy budgets are allocated to subsequent communication rounds: if the client communication fails in the current round, the privacy budget of the current round is saved, and the privacy budget of the current round is superimposed on the privacy budget of the subsequent round, and the privacy budget of the subsequent round is updated. In this way, the training efficiency of the model is further improved.

[0084] Specifically, at the beginning of training, assume that the privacy budget for each round is ε 0 , a total of R rounds of communication are performed. When the client fails to communicate in round r1, the privacy budget of this round is saved, and the privacy budget of all subsequent rounds is updated to ε1, as shown in formula (12):

[0085]

[0086] When communication failure occurs at r i round, the original privacy budget of the current round is ε i-1 , then the privacy budget of the subsequent round becomes ε i , as shown in formula (13):

[0087]

[0088] Step S2.3: In each round of communication, a similarity-based gradient estimation method is used during uploading to estimate the gradient of the client with communication failure.

[0089] In this embodiment, for all clients in the same round, clients with successful communication are regarded as available clients, and clients with failed communication are regarded as unavailable clients. A gradient estimation method for coping with communication failure is adopted to automatically identify the similarity between any two clients, and then the gradient of the available client that is most similar to the currently unavailable client is used to replace it, so as to obtain an estimate of the gradient of the currently unavailable client.

[0090] Furthermore, identifying the similarity between clients means: Assume that in round r, the set of unavailable clients is U r , the set of available clients is A r , then the central server can obtain A rGradient information of all clients in the set. At this time, the normalized cosine distance of the gradients of two available clients is used to measure the similarity between the two clients, that is, the similarity score of the available clients i and j in round r As shown in formula (14):

[0091]

[0092] in, express and The vector product of Represents the gradient The similarity score ranges from (0,1). When the similarity score approaches 1, it means that the two clients are more similar.

[0093] However, the similarity score calculated by a single communication round cannot provide accurate similarity information. Therefore, this embodiment uses the average similarity score to describe and evaluate the similarity between two clients i and j. That is, for two clients in round r, the similarity between two clients is evaluated based on the rounds before round r in which two clients are available at the same time. Average similarity score between clients i and j in round r As shown in formula (15):

[0094]

[0095] in, Indicates the number of rounds before the rth round that clients i and j are available simultaneously.

[0096] Furthermore, the gradient of the unavailable client is estimated: in the r+1th round of communication, it is assumed that client i is unavailable, that is, the server has not received the gradient uploaded by client i for more than the specified time. At this time, the central server uses the gradient of the available client j that is most similar to client i instead, as shown in formula (16):

[0097]

[0098] in, represents the gradient of available client j, represents the estimated value of the gradient for the unavailable client i.

[0099] Furthermore, since the above process is carried out during the training of the neural network model, the noise can be automatically filtered out through training, so the above process does not need to consider the impact of the gradient noise added in step S2.2 on the gradient similarity calculation. Through the above step S2.3, the central server can obtain the gradient information from all clients in the rth round.

[0100] Step S2.4: Based on the final updated initial model parameters, after this training, a personalized local estimation model of the client can be quickly generated.

[0101] Through the above steps, the initial model parameters of the neural network with strong adaptability can be obtained. Based on the parameters, after local training, a personalized model adapted to the photovoltaic power generation characteristics of each client can be quickly obtained.

[0102] After obtaining a personalized local estimation model adapted to the photovoltaic power generation characteristics of each client, the meteorological data of each community and the net load power data of photovoltaic power generation users in the community are obtained, and the acquired data are input into the personalized local estimation model corresponding to the community to accurately estimate and output the photovoltaic power generation power of each community.

[0103] Embodiment 2

[0104] This embodiment provides a community-level photovoltaic power generation distributed estimation system, including:

[0105] A data acquisition module is used to obtain meteorological data of each community and net load power data of photovoltaic power generation users in the community;

[0106] The photovoltaic power generation estimation module is used to input the acquired data into the personalized local estimation model corresponding to the community to estimate the photovoltaic power generation of each community; wherein the training process of the personalized local estimation model includes:

[0107] Obtain historical data related to photovoltaic power generation in each community and build a local data set;

[0108] Each community is regarded as a client, and the federated meta-learning method is used to collaboratively train the photovoltaic power estimation model of each client through multiple rounds of communication to obtain a personalized local estimation model for each community. In each round of communication, the differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, and the similarity-based gradient estimation method is used during uploading to estimate the gradient of the client with communication failure.

[0109] Embodiment 3

[0110] This embodiment provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, the steps in the distributed estimation method of community-level post-meter photovoltaic power generation power as described above are completed.

[0111] Embodiment 4

[0112] This embodiment also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps in the community-level post-meter photovoltaic power distribution estimation method as described above are completed.

[0113] The steps involved in the above embodiments 2 to 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0114] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0115] The above description is only a preferred embodiment of the present invention. Although the specific implementation mode of the present invention is described in conjunction with the accompanying drawings, it is not a limitation of the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the protection scope of the present invention.

Claims

1. A method for distributed estimation of community-level photovoltaic power generation, characterized in that: include: Obtain meteorological data for each community and net load power data of photovoltaic power generation users in the community; The acquired data is input into the personalized local estimation model corresponding to the community to estimate the photovoltaic power generation power of each community; wherein the training process of the personalized local estimation model includes: Obtain historical data related to photovoltaic power generation in each community and build a local data set; Each community is regarded as a client, and the federated meta-learning method is used to collaboratively train the photovoltaic power estimation model of each client through multiple rounds of communication to obtain a personalized local estimation model for each community. In each round of communication, a differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, and a similarity-based gradient estimation method is used during uploading to estimate the gradient of the client that failed to communicate. Each community is regarded as a client, and each client communicates with the central server. Federated meta-learning method is used for multiple rounds of communication to collaboratively train the estimation models of each client, including: Each client trains the local model based on the initial model parameters sent by the central server and the local data set, and uploads the loss function gradient after training; After receiving the gradients uploaded by all clients, the central server updates the initial model parameters according to the gradients and sends the updated initial model parameters again; After multiple rounds of communication and training, until the updated initial model parameters converge or reach the preset communication rounds, the final updated initial model parameters are obtained. After local training based on the parameters, the personalized local estimation model of each client can be quickly obtained; A differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, including: For each client, the client's gradient is clipped in turn, and random noise is generated based on the Laplace mechanism, and the random noise is added to the client's gradient, and then the gradient with added noise is uploaded to the server; wherein the added random noise satisfies the Laplace distribution, and the Laplace distribution is determined based on the preset clipping threshold, the number of samples used by the client for training, and the privacy budget; The privacy budget calculation for each client in each communication round includes: According to the historical photovoltaic power generation time series in the client's local data set, the approximate entropy of the time series is calculated by comparing the similarities of subsequences in different time windows to measure the complexity of the time series; Based on the client's approximate entropy and the preset global privacy budget, the privacy budget of the client in each round is obtained; If the client communication fails in the current round, the privacy budget of the current round is saved, and the privacy budget of the current round is added to the privacy budget of the subsequent round to update the privacy budget of the subsequent round; The gradient estimation method based on similarity is used to estimate the gradient of the client with communication failure, including: For all clients in the same round, the clients with successful communication are regarded as available clients, and the clients with failed communication are regarded as unavailable clients. The gradient of the available client most similar to the current unavailable client is used to replace it, and the gradient estimate of the current unavailable client is obtained.

2. The method for distributed estimation of community-level post-meter photovoltaic power generation according to claim 1, characterized in that: The historical data related to photovoltaic power generation include the historical meteorological data of the local community, the historical net load power data of observable photovoltaic power generation users in the local community, and the historical photovoltaic power generation power data.

3. The method for distributed estimation of community-level post-meter photovoltaic power generation according to claim 1, characterized in that: The average similarity score is used to evaluate the similarity between two clients, including: For each pair of clients in round r, the similarity between the two clients is evaluated by calculating the number of rounds in which the two clients are available simultaneously in the rounds before the rth round.

4. A community-level photovoltaic power generation distributed estimation system, using the community-level photovoltaic power generation distributed estimation method according to any one of claims 1 to 3, characterized in that: include: A data acquisition module is used to obtain meteorological data and net load power data of each distributed photovoltaic power generation community; The photovoltaic power generation estimation module is used to input the acquired data into the personalized local estimation model corresponding to the community, estimate and output the photovoltaic power generation of each community; wherein the training process of the personalized local estimation model includes: Obtain historical data related to photovoltaic power generation in each community and build a local data set; Each community is regarded as a client, and the federated meta-learning method is used to collaboratively train the photovoltaic power estimation model of each client through multiple rounds of communication to obtain a personalized local estimation model for each community. In each round of communication, the differential privacy mechanism based on the privacy budget adaptive allocation strategy is used to add noise to the gradient uploaded by the client, and the similarity-based gradient estimation method is used during uploading to estimate the gradient of the client with communication failure.

5. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of a method for distributed estimation of community-level post-meter photovoltaic power generation are completed as described in any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of a community-level post-meter photovoltaic power distribution estimation method as described in any one of claims 1-3.

Images