CN113673159A - Renewable energy space-time scene generation method based on federal deep-generation learning - Google Patents

Renewable energy space-time scene generation method based on federal deep-generation learning Download PDF

Info

Publication number
CN113673159A
CN113673159A CN202110957064.0A CN202110957064A CN113673159A CN 113673159 A CN113673159 A CN 113673159A CN 202110957064 A CN202110957064 A CN 202110957064A CN 113673159 A CN113673159 A CN 113673159A
Authority
CN
China
Prior art keywords
model
renewable energy
generator
discriminator
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110957064.0A
Other languages
Chinese (zh)
Inventor
李扬
李嘉政
王瑞浓
王彬
韩猛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Northeast Electric Power University
Original Assignee
Northeast Dianli University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Northeast Dianli University filed Critical Northeast Dianli University
Priority to CN202110957064.0A priority Critical patent/CN113673159A/en
Publication of CN113673159A publication Critical patent/CN113673159A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/60—Protecting data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Hardware Design (AREA)
  • Artificial Intelligence (AREA)
  • Economics (AREA)
  • Computer Security & Cryptography (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • Bioethics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Biomedical Technology (AREA)
  • Geometry (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Molecular Biology (AREA)

Abstract

A renewable energy output space-time scene generation method based on federal deep-generation learning is characterized by comprising the following steps: carrying out normalization processing on the historical data of the renewable energy sources and remolding the historical data of the renewable energy sources into a 24 x 24 array matrix; designing a global least square to generate a confrontation network model, and initializing the global model in a central server; deploying the global model to each client (renewable energy power station); randomly selecting a certain number of clients, and training a local model by using historical data by the selected clients; when the set communication interval is reached, the client uploads the parameters of the local model to the central server; the central server carries out weighted average on the uploaded model parameters and takes the weighted average as a new global model parameter; the central server deploys the global model to each client again; the iteration is repeated continuously. And finally, a high-quality renewable energy scene for capturing the space-time correlation can be generated.

Description

Renewable energy space-time scene generation method based on federal deep-generation learning
Technical Field
The invention relates to a renewable energy scene generation method based on a federal learning and deep generation type model, and belongs to the technical field of application of artificial intelligence in a power system.
Background
The development and utilization of renewable energy sources (such as wind power and photovoltaic) are inevitable choices for solving environmental pollution and ensuring sustainable supply of energy sources. However, due to its inherent uncertainty and intermittency, the access of renewable energy sources has presented significant challenges to the planning and operation of today's power systems. The influence of uncertainty on the operation of the power grid caused by how to deal with the power generation of renewable energy sources is a difficult point which has to be faced in the current economic optimization research of the power system. The scene analysis method is a method for processing uncertainty, and a technology for representing renewable energy power generation by generating a group of typical time series scenes which accord with the electricity output statistical characteristics of renewable energy, so that the uncertainty optimization problem of the renewable energy is converted into a certainty problem. By using the generated renewable energy output scenario, power system operators can make reasonable decisions taking into account the uncertainty in the renewable energy power generation output.
A great deal of research has been conducted on methods for generating renewable energy output scenes in different aspects. Unfortunately, most of the previous work has focused on model-based scene generation methods that must probabilistically model renewable energy output. For example, the monte carlo method must assume that the wind and solar irradiance satisfy a certain probability distribution and then randomly sample the output scene from the assumed distribution. According to the method based on the Copula function, firstly, Copula function parameters are obtained through parameter estimation, and then a renewable energy output scene which accords with the function can be obtained through random sampling in the function. These methods rely on statistical assumptions of prior, cumbersome modeling processes and low quality of the generated scene.
With the development of artificial intelligence, many machine learning and deep learning methods are applied to the field of scene generation. Such as artificial neural networks, variational autoencoders, and generative countermeasure networks, are also used for the generation of renewable energy scenarios. The generation of countermeasure Networks (GANs) is paid attention to by researchers as a powerful distributed grid model, and the GANs can learn the potential distribution of real data in an unsupervised manner without explicit modeling, so as to generate high-quality scene data. The main idea is that two deep neural networks are trained simultaneously through a two-person zero-sum game: a generator and a discriminator. The generator network converts the noise distribution into artificial data by learning the distribution mapping relationship of known historical data, and trains discriminators to distinguish the generated data from the historical data.
However, all of the above methods require all of the data for each wind/solar power plant to be transmitted to a data center with a high performance computing cluster to generate a renewable energy scenario. Such centralized data processing methods will face the following problems: 1) these methods rely on a central workstation with powerful computing power and memory space; 2) frequent data exchange between the renewable energy power plant and the central workstation will inevitably cause large communication overhead; 3) data stored in a centralized form is vulnerable to data leakage and network attacks; 4) most importantly, some renewable energy data owners, such as independent system operators, are reluctant to share data with others in practical applications due to data privacy reasons. Data privacy may refer to business sensitive data for grid-connected renewable energy power plants with commercial competition, or personal data for homes with renewable energy technology. Currently, special laws, such as general data protection regulations, network security laws and data security laws, implemented in the european union, are promulgated in many countries to regulate the management and use of data. Therefore, a distributed scene generation method capable of solving the above problems is urgently required.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, provide a scientific and reasonable method for effectively solving the problems of data security and communication overhead caused by generating a plurality of renewable energy power station output scenes, generate a higher-quality renewable energy output scene in a distributed privacy protection mode, and provide an advanced uncertain variable processing method for the planning and operation problems of a power system considering wind and light uncertainty.
The purpose of the invention is realized by the following technical scheme: a renewable energy scene generation method based on federal deep-generation learning is provided.
Firstly, designing a global Least square generation countermeasure network (LSGANs) model for a scene generation task, wherein the model can generate a renewable energy scene by a data-driven method; then, the federal Learning (Federa Learning) technology is adopted, so that the renewable energy power stations do not need to train the data centrally, but each power station performs distributed training on the local LSGANs model by using the historical data of the power station, and the invention is named as a federal least square generation countermeasure network (Fed-LSGAN). In the training process, each power station only needs to transmit the model parameters to the central server, so that the communication overhead of data transmission is reduced, and the safety of data is protected. The method comprises the following specific steps:
1) carrying out normalization processing on historical data of renewable energy sources;
2) reshaping the historical data into a 24 x 24 array;
3) initializing a global LSGANs model in a central server;
4) deploying the global LSGANs model to each client (renewable energy power station);
5) randomly selecting a certain number of clients, and training a local model by using historical data by the selected clients;
6) when the set communication interval is reached, the client uploads the parameters of the local model to the central server;
7) the central server carries out weighted average on the uploaded model parameters and takes the weighted average as a new global model parameter;
8) the central server deploys the global model to each client again;
9) repeat 5) -8) until a set number of exercises.
In the step 1), max-min normalization processing needs to be performed on the historical data so as to eliminate the influence of data dimension, and the calculation formula is as follows:
Figure BDA0003220806010000031
wherein x is historical output data, xmaxAnd xminThe maximum and minimum values of the historical data are respectively.
In the step 2), the historical data is reshaped into a matrix array of 2424.
In the step 3), the global model generates a countermeasure network (LSGANs) model for Least square. LSGANs are an improved version of GANs, all of which are unsupervised generative models inspired by game theory. GANs include two differentiable deep neural networks: a discriminator and a generator. The generator network maps random noise to generated samples by learning the potential distribution of historical data, the discriminator judges whether input data is real historical data or generated data as far as possible, and two-person zero-sum game training is carried out on the generator and the discriminator. Theoretically, when nash equilibrium is reached, the two networks reach an optimal solution, and the generator can generate false and spurious samples through random noise.
In the traditional GANs training process, a batch of random noise vectors z-PZIs fed to a generator which maps from the noise space to the data space G (z; theta)g) Wherein G (z; thetag) Obey a distribution P for the samples generated by the generatorg. At the same time, from the real data x-PdOr generated data G (z; theta)g) Is sent to a discriminator, which is typically a binary classifier, whose goal is to identify where the data came from. The output of the network of discriminators can be expressed as
Figure BDA0003220806010000032
After the targets of the generator and the arbiter are clarified, we need to define a loss function to update their deep neural networks, respectively. For a given discriminator, the generator wishes to increase the probability output p of the discriminator on the generated samplefakeSince a larger discriminator output means that the sample is more realistic. For a given generator, the discriminator is minimizing pfakeWhile seeking prealIs maximized. In other words, a smaller generator loss LGIndicating that the manifolds of the generated data and the real data are very close, with a small discriminator loss LDThe values indicate that the arbiter has a strong ability to distinguish which distribution the data comes from. Therefore, the loss functions of these two neural networks can be defined as
Figure BDA0003220806010000041
Figure BDA0003220806010000042
Where E represents the expected value.
A gambling relationship needs to be established between the generators and the discriminators in order that both networks can be trained simultaneously. It is necessary to construct a function that can combine (3) and (4), the cost function V of the max-min game modelGANs(G, D) can be represented as:
Figure BDA0003220806010000043
however, in the conventional GANs, the Jensen-Shannon divergence (JS divergence) is optimized essentially, and the cross entropy is taken as a loss function, which causes the problems of mode collapse, gradient disappearance, unstable training and the like of the conventional GANs model. To solve this problem, the present invention employs a least squares generation countermeasure network. Replacing the cross entropy loss function with a least square loss function, wherein the expression is as follows:
Figure BDA0003220806010000044
where a and b are labels of the generated sample and the real sample, respectively, and c denotes a label that the generator desires the arbiter to decide on the generated sample. LSGANs use the least square loss function as the loss function of the discriminator and the generator, and by punishing samples far away from the decision boundary, more gradients can be provided, and the problem of gradient disappearance is relieved. At the same time, such a loss function is advantageous for generating higher quality samples, since it penalizes false samples and "pulls" them to the decision boundary.
In the step 4), the central server distributes the global LSGANs model to all clients (renewable energy power stations) participating in training together, and each client has a local model of the client.
In the step 5), when training is started, a certain proportion of clients are randomly selected in each global iteration to perform local training once. In training, noise randomly sampled from Gaussian distribution is sent to a generator, scene samples are generated through a series of neural networks, and real historical samples are sampled from historical data; the two are sent to a discriminator, and the discriminator judges whether the input sample comes from the sample generated by the generator or is true. The gradient is then calculated by the loss function defined in step 3, and the model parameters are updated using a gradient-based optimization algorithm. In summary, for the discriminator
Figure BDA0003220806010000045
For generators
Figure BDA0003220806010000051
Wherein α is the learning rate; adam is an optimizer;
Figure BDA0003220806010000052
and
Figure BDA0003220806010000053
classification is as the gradient of the discriminator and generator functions. ThetadAnd thetagRespectively, the parameters of the generator and the discriminator neural network model.
In the step 6), the local client model is continuously trained by using the historical data of the local client model. When the set communication interval is reached, all clients transmit the local LSGANs model parameters to the central server without uploading all historical data.
In the step 7), the central server carries out weighted average on the uploaded model parameters
Figure BDA0003220806010000054
Wherein
Figure BDA0003220806010000055
And
Figure BDA0003220806010000056
respectively as the generator and discriminator model parameters of the global model; n is a radical ofeIs the selected client set.
In the step 8), the global LSGANs model of the central server is deployed into the local models of the clients again.
Figure BDA0003220806010000057
In the step 9), repeating the steps 5-8 until the set training times are reached:
the invention discloses a renewable energy space-time scene generation method based on federal deep-generation learning. Fed-LSGAN, a federal study, enables our invention to generate scenes in a privacy-preserving manner without sacrificing the quality of the generated scenes by transmitting model parameters rather than all data; and the depth generation model based on the LSGANs can generate scenes conforming to historical data distribution by capturing the space-time characteristics of renewable energy sources. The proposed Fed-LSGAN can successfully perform renewable energy scene generation and this approach is superior to existing centralized approaches.
Drawings
FIG. 1 is a schematic diagram of a basic generative countermeasure architecture.
FIG. 2 is a flow framework of a renewable energy space-time scene generation method based on federated deep generative learning according to the present invention.
Fig. 3 is a diagram illustrating the variation of the loss function value during the training process.
FIG. 4 is a schematic view of a wind power output scenario generated by the present invention.
FIG. 5 is a schematic view of a photovoltaic output scenario generated using the present invention.
FIG. 6 is a schematic diagram of a generated spatial correlation scene.
Fig. 7 is a schematic diagram of the impact of federal learning parameters on the proposed method.
Detailed Description
The preferred embodiments will be described in detail below with reference to the accompanying drawings. It should be emphasized that the following description is merely exemplary in nature and is not intended to limit the scope of the invention or its application.
Fig. 1 is a schematic diagram of a basic structure of a generation countermeasure network, and is composed of two deep neural networks of a generator and an arbiter. The working principle is as follows: the generator samples random noise from a certain known noise distribution and sends the random noise into a generator network, and scene samples are generated through a series of neural networks; meanwhile, randomly extracting renewable energy output samples from historical data, and sending real samples and generated samples to a discriminator; the discriminator is responsible for judging where the sent sample comes; the generator and arbiter network is then updated by the defined loss function. Through continuous training, when Nash equilibrium is achieved, the generator is optimal, and a high-quality fake and spurious renewable energy output scene can be generated at the moment.
Referring to fig. 2, a renewable energy space-time scene generation method based on federal learning and generation countermeasure network is provided. Firstly, designing a global LSGANs model for a scene generation task, wherein the model can generate a renewable energy scene by a data-driven method; and then, a federal learning technology is adopted to prevent all renewable energy power stations from transmitting historical data to a central workstation, and each power station performs distributed training on local LSGANs models by using the historical data of the power station. In the training process, each power station only needs to transmit the model parameters to the central server, so that the communication overhead of data transmission is reduced, and the safety of data is protected. The method comprises the following specific steps:
1) normalizing historical data of wind and light output; in the step 1), max-min normalization processing needs to be carried out on the historical data so as to eliminate the influence of data dimension.
2) Reshaping the historical data into a 24 x 24 matrix; in the step 2), the historical data of the time series is reshaped into N matrix arrays of 24 × 24.
3) Constructing a global LSGANs model; in the step 3), the global model is an LSGANs model. LSGANs are an improved version of GANs, all of which are unsupervised generative models inspired by game theory. GANs include two differentiable deep neural networks: a discriminator and a generator. The generator network maps random noise to generated samples by learning the potential distribution of historical data, the discriminator judges whether input data is real historical data or generated data as far as possible, and two-person zero-sum game training is carried out on the generator and the discriminator. Theoretically, when nash equilibrium is reached, the two networks reach an optimal solution, and the generator can generate false and spurious samples through random noise.
4) Distributing the global model to each client; in the step 4), the central server distributes the global LSGANs model to all clients (renewable energy power stations) participating in training together, and each client has a local model of the client.
5) And randomly selecting the client to carry out local training. In the step 5), when training is started, a certain proportion of clients are randomly selected in each global iteration to perform local training once. In training, noise randomly sampled from Gaussian distribution is sent to a generator, scene samples are generated through a series of neural networks, and real historical samples are sampled from historical data; the two are sent to a discriminator, and the discriminator judges whether the input sample comes from the sample generated by the generator or is true. Finally, the generator and the discriminator model parameters are updated through a loss function.
6) And uploading the model parameters to a central server. In the step 6), the local client model is continuously trained by using the historical data of the local client model. When the set communication interval is reached, all clients transmit the local LSGANs model parameters to the central server without uploading all historical data.
7) And the central server performs parameter aggregation. In the step 7), the central server performs weighted average processing on the uploaded model parameters.
8) The central server redistributes the global model to the clients. In the step 8), the global LSGANs model of the central server is deployed into the local models of the clients again.
9) And repeating the steps. In the step 9), repeating the steps 5-8 until the set training times are reached:
FIG. 3 shows the variation of the loss function of each client in the whole global model training process. As can be seen from the figure, the loss value drops rapidly at the beginning of training and then rises vertically every 100 global iterations. This is because the server distributes the aggregation parameters to each client after the client communicates with the central server, which can result in a dramatic increase in client loss values. Eventually, the discriminator loss value for each client will be close to 0, indicating that the distribution of the generated scenes is close enough to the true data distribution.
FIG. 4 is a wind power output scenario generated by the method of the present invention. It can be seen that the centroid of the generated scene after clustering is almost consistent with the centroid of the real sample, and the generated output curve conforms to the historical condition. According to the autocorrelation coefficient, the generated scene can capture the time series characteristic of the wind power output. As can be seen from the normalized error map, the error of the scene generated by the method is very small.
FIG. 5 is a photovoltaic output scenario generated using the method of the present invention. Consistent with the analysis of the wind power output, the method can be known to generate high-quality photovoltaic scenes.
Fig. 6 is a generated spatial signature scenario. By using the global model to generate all renewable energy output scenarios, we calculate the pearson correlation coefficients between scenarios of different plant sites. This figure illustrates that the Fed-LSGAN can capture not only the temporal characteristics of the renewable energy sources, but also the spatial characteristics between multiple sites. More importantly, it can protect the data privacy of each renewable power plant and reduce communication overhead when training the model.
Fig. 7 is a graph that verifies the robustness of the proposed inventive method for changing federal learning settings, such as the synchronization interval K and the client engagement rate E. As can be seen from the figure, for the synchronization interval K, under the condition that E is fixed, the performance of the proposed inventive method under different synchronization intervals K is basically unchanged, which proves that the Fed-LSGAN has good robustness for different synchronization intervals; regarding the client engagement rate E, different engagement rates E have no effect on the scene generation quality of the training model, but may affect the convergence speed of the scheme in a given time interval.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (10)

1. A renewable energy output space-time scene generation method based on federal deep-generation learning is characterized by comprising the following steps:
1) carrying out normalization processing on historical data of renewable energy sources;
2) reshaping the historical data into a 24 x 24 array;
3) initializing a global LSGANs model in a central server;
4) deploying the global LSGANs model to each client (renewable energy power station);
5) randomly selecting a certain number of clients, and training a local model by using historical data by the selected clients;
6) when the set communication interval is reached, the client uploads the parameters of the local model to the central server;
7) the central server carries out weighted average on the uploaded model parameters and takes the weighted average as a new global model parameter;
8) the central server deploys the global model to each client again;
9) repeat 5) -8) until a set number of exercises.
2. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: the step 1)) needs to carry out max-min normalization processing on the historical data so as to eliminate the influence of data dimension.
3. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: and 2) reshaping the time series historical data samples of each unit into a 24 x 24 matrix array.
4. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: in the step 3), the global model generates a confrontation network (LSGANs) model for Least square. LSGANs are an improved version of GANs, all of which are unsupervised generative models inspired by game theory. GANs include two differentiable deep neural networks: a discriminator and a generator. The generator network maps random noise to generated samples by learning the potential distribution of historical data, the discriminator judges whether input data is real historical data or generated data as far as possible, and two-person zero-sum game training is carried out on the generator and the discriminator. Theoretically, when nash equilibrium is reached, the two networks reach an optimal solution, and the generator can generate false and spurious samples through random noise. The modeling process is as follows:
in the traditional GANs training process, a batch of random noise vectors z-PZIs fed to a generator which maps from the noise space to the data space G (z; theta)g) Wherein G (z; thetag) Obey a distribution P for the samples generated by the generatorg. At the same time, from the real data x-PdOr generated data G (z; theta)g) Is sent to a discriminator, which is typically a binary classifier, whose goal is to identify where the data came from. The output of the network of discriminators can be expressed as
Figure FDA0003220806000000021
After the targets of the generator and the arbiter are clarified, we need to define a loss function to update their deep neural networks, respectively. For a given discriminator, the generator wishes to increase the probability output p of the discriminator on the generated samplefakeSince a larger discriminator output means that the sample is more realistic. For a given generator, the discriminator is minimizing pfakeWhile seeking prealIs maximized. In other words, a smaller generator loss LGIndicating that the manifolds of the generated data and the real data are very close, with a small discriminator loss LDThe values indicate that the arbiter has a strong ability to distinguish which distribution the data comes from. Therefore, the loss functions of these two neural networks can be defined as
Figure FDA0003220806000000022
Figure FDA0003220806000000023
Where E represents the expected value.
A gambling relationship needs to be established between the generators and the discriminators in order that both networks can be trained simultaneously. It is necessary to construct a function that can combine (2) and (3), the cost function V of the max-min game modelGANs(G, D) can be represented as:
Figure FDA0003220806000000024
however, in the conventional GANs, the Jensen-Shannon divergence (JS divergence) is optimized essentially, and the cross entropy is taken as a loss function, which causes the problems of mode collapse, gradient disappearance, unstable training and the like of the conventional GANs model. To solve this problem, the present invention employs a least squares generation countermeasure network. Replacing the cross entropy loss function with a least square loss function, wherein the expression is as follows:
Figure FDA0003220806000000025
where a and b are labels of the generated sample and the real sample, respectively, and c denotes a label that the generator desires the arbiter to decide on the generated sample. LSGANs use the least square loss function as the loss function of the discriminator and the generator, and by punishing samples far away from the decision boundary, more gradients can be provided, and the problem of gradient disappearance is relieved. At the same time, such a loss function is advantageous for generating higher quality samples, since it penalizes false samples and "pulls" them to the decision boundary.
5. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: in the step 4), the central server distributes the global LSGANs model to all clients (renewable energy power stations) participating in training together, and each client has a local model of the client.
6. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: in the step 5), when training is started, a certain proportion of clients are randomly selected in each global iteration to perform local training once. In training, noise randomly sampled from Gaussian distribution is sent to a generator, scene samples are generated through a series of neural networks, and real historical samples are sampled from historical data; the two are sent to a discriminator, and the discriminator judges whether the input sample comes from the sample generated by the generator or is true. The gradient is then calculated by the loss function defined in step 3, and the model parameters are updated using a gradient-based optimization algorithm. In summary, for the discriminator
Figure FDA0003220806000000031
For generators
Figure FDA0003220806000000032
Wherein α is the learning rate; adam is an optimizer;
Figure FDA0003220806000000033
and
Figure FDA0003220806000000034
classification is as the gradient of the discriminator and generator functions. ThetadAnd thetagRespectively, the parameters of the generator and the discriminator neural network model.
7. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: in the step 6), the local client model is continuously trained by using the historical data of the local client model. When the set communication interval is reached, all clients transmit the local LSGANs model parameters to the central server without uploading all historical data.
8. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: in the step 7), the central server carries out weighted average on the uploaded model parameters
Figure FDA0003220806000000035
Wherein
Figure FDA0003220806000000036
And
Figure FDA0003220806000000037
respectively as the generator and discriminator model parameters of the global model; n is a radical ofeIs the selected client set.
9. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: in the step 8), the global LSGANs model of the central server is deployed into the local models of the clients again.
Figure FDA0003220806000000041
10. The method for generating renewable energy output space-time scene based on federal deep-generation learning as claimed in claim 1, wherein: in the step 9), the steps 5) to 8) are repeated until the set training times are reached.
CN202110957064.0A 2021-08-19 2021-08-19 Renewable energy space-time scene generation method based on federal deep-generation learning Pending CN113673159A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110957064.0A CN113673159A (en) 2021-08-19 2021-08-19 Renewable energy space-time scene generation method based on federal deep-generation learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110957064.0A CN113673159A (en) 2021-08-19 2021-08-19 Renewable energy space-time scene generation method based on federal deep-generation learning

Publications (1)

Publication Number Publication Date
CN113673159A true CN113673159A (en) 2021-11-19

Family

ID=78544345

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110957064.0A Pending CN113673159A (en) 2021-08-19 2021-08-19 Renewable energy space-time scene generation method based on federal deep-generation learning

Country Status (1)

Country Link
CN (1) CN113673159A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114338628A (en) * 2022-03-17 2022-04-12 军事科学院系统工程研究院网络信息研究所 Nested meta-learning method and system based on federated architecture

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109904878A (en) * 2019-02-28 2019-06-18 西安交通大学 A kind of windy electric field electricity-generating timing simulation scenario building method
CN111553587A (en) * 2020-04-26 2020-08-18 中国电力科学研究院有限公司 New energy scene generation method and system based on confrontation learning model
CN111950868A (en) * 2020-07-28 2020-11-17 国网电力科学研究院有限公司 Comprehensive energy system load scene generation method based on generation countermeasure network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109904878A (en) * 2019-02-28 2019-06-18 西安交通大学 A kind of windy electric field electricity-generating timing simulation scenario building method
CN111553587A (en) * 2020-04-26 2020-08-18 中国电力科学研究院有限公司 New energy scene generation method and system based on confrontation learning model
CN111950868A (en) * 2020-07-28 2020-11-17 国网电力科学研究院有限公司 Comprehensive energy system load scene generation method based on generation countermeasure network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
BANGZHOU XIN: "PRIVATE FL-GAN:DIFFERENTIAL PRIVACY SYNTHETIC DATA GENERATION BASED ON FEDERATED LEARNING", 2020 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING(ICASSP), 9 April 2020 (2020-04-09), pages 2927 - 2931 *
YANG LI等: "Privacy-preserving Spatiotemporal Scenario Generation of Renewable Energies: A Federated Deep Generative Learning Approach", IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, vol. 18, no. 4, 20 July 2021 (2021-07-20), pages 1 - 11, XP011896426, DOI: 10.1109/TII.2021.3098259 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114338628A (en) * 2022-03-17 2022-04-12 军事科学院系统工程研究院网络信息研究所 Nested meta-learning method and system based on federated architecture
CN114338628B (en) * 2022-03-17 2022-06-03 军事科学院系统工程研究院网络信息研究所 Nested meta-learning method and system based on federated architecture

Similar Documents

Publication Publication Date Title
Assaf et al. Explainable deep neural networks for multivariate time series predictions.
CN107770263B (en) safe access method and system for Internet of things terminal based on edge calculation
Chen et al. POBA-GA: Perturbation optimized black-box adversarial attacks via genetic algorithm
Yuan et al. Intrusion detection for smart home security based on data augmentation with edge computing
CN109639479B (en) Network traffic data enhancement method and device based on generation countermeasure network
CN103530620A (en) Method for identifying bird nest on electric transmission line tower
CN110889603A (en) Power system economic dispatching method considering wind power correlation based on PCA-Copula theory
CN110490659B (en) GAN-based user load curve generation method
CN109214119A (en) Bridge Earthquake Resistance Design method based on response surface model
CN108537133A (en) A kind of face reconstructing method based on supervised learning depth self-encoding encoder
Badr et al. A novel evasion attack against global electricity theft detectors and a countermeasure
CN113392919A (en) Federal attention DBN cooperative detection system based on client selection
CN112232488A (en) Deep learning and data driving-based new energy output scene generation method
CN113673159A (en) Renewable energy space-time scene generation method based on federal deep-generation learning
CN109660522B (en) Deep self-encoder-based hybrid intrusion detection method for integrated electronic system
CN112465184A (en) Cloud energy storage system control method of small-sample generation type counterstudy network
CN115859344A (en) Secret sharing-based safe sharing method for data of federal unmanned aerial vehicle group
CN111967577B (en) Energy Internet scene generation method based on variation self-encoder
CN114724245A (en) CSI-based incremental learning human body action identification method
Xia et al. A Novel Hybrid Model for Short‐Term Wind Speed Forecasting Based on Twice Decomposition, PSR, and IMVO‐ELM
CN113515890A (en) Renewable energy day-ahead scene generation method based on federal learning
CN110351241B (en) GWA (global warming environment) optimization-based industrial network DDoS (distributed denial of service) intrusion detection system classification method
CN113872524A (en) Method and device for positioning inefficient group strings and computer storage medium
Peng et al. Stochastic scenario generation for wind power and photovoltaic system based on CGAN
CN114363464B (en) Method and system for preventing fraud information from spreading

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination