A Distributed Renewable Energy Scenario Generation Method Based on Improved Group Learning

Abstract

The present invention provides a method for generating distributed renewable energy scenarios by improving group learning, including a distributed ledger platform to build a group learning network for interacting gradient parameters; each renewable energy region serves as both a blockchain node and a client of federated learning, retaining the original historical data locally without transmission to achieve data isolation, and using this data locally for neural network training; between each renewable energy region, the gradient parameters of the first layer of the neural network are interacted through the blockchain network, and the gradient parameters except the first layer are interacted through a central server. The present invention introduces group learning into scenario generation and constructs a model for generating daily scenarios of distributed renewable energy regions. The proposed improved group learning method can not only break the data islands between regions, but also significantly shorten the training time compared with group learning.

Description

Technical Field

[0001] The present invention provides a distributed renewable energy scenario generation method based on improved group learning, belonging to the technical field of electric power. Background Art

[0002] Renewable energy output is characterized by randomness. As its penetration rate increases, its impact on the economic, safe, and stable operation of the system becomes increasingly prominent. Therefore, characterizing renewable energy uncertainty is a key technology for effectively balancing the supply and demand relationship in the new power system. Different stakeholders in different renewable energy sites may have different interests, leading to data sharing challenges. However, to more effectively describe renewable energy output uncertainty and improve model performance, data from different regions must be integrated, which often poses security risks such as data leakage during data transmission.

[0003] At present, the main approach to privacy issues in distributed modeling is federated learning. The core idea of federated learning is the transfer between model parameters. The original data is always kept locally, which to a certain extent prevents the original data from being leaked during the upload process. Studies have shown that although federated learning does not exchange local data and can protect user data privacy to a certain extent, it still faces security risks such as data set privacy leakage. The literature (Zhang Zehui, Fu Yao, Gao Tiegang. Research on federated deep neural network models supporting data privacy protection [J]. Acta Automatica Sinica, 2022, 48(05): 1273-1284.) discusses in detail how malicious participants can use gradient parameters to reversely infer data set information during the training process using the federated learning method. The details are as follows:

[0004] Neural network structure such as Figure 4 As shown, x1 and x2 are input data, w i (i=1,2...8) is the weight, b i (i=1,2,3,4) is the bias term, target is the target value, the activation function is sigmoid, and the deviation between the calculated target value and the output value is LOSS.

[0005] Referring to the conclusion of this paper, in the first case, when all gradient information is not encrypted, the original data can be easily obtained through the first layer gradient parameters. In the second case, if the bias term is a random number, try the possible bias term values, repeat the brute force solution over and over again until the gradient value calculated by the inferred result is close to or equal to the actual gradient value, and the information will be leaked. That is, no matter whether the bias term is a fixed value or a random number, the original data will be inferred. In the third case, if the gradient parameter of the first layer weight is is encrypted, and the hidden layer output value Out can be obtained through the hidden layer gradient parameter h1 and Outh2 , and then combined with the first layer weight parameters, the input data x1 and x2 are obtained through formula (1).

[0006]

[0007] Renewable energy users have high requirements for data privacy, so we should try our best to find a more effective distributed privacy protection method to ensure data security.

[0008] The paper (Warnat-Herresthal S, Schultze H, Shastry KL, et al. Swarm Learning for decentralized and confidential clinical machine learning [J]. Nature, 2021, 594(7862).) proposes a swarm learning (SL) approach. Compared to federated learning, this approach eliminates the need for a central server, meaning that model parameters do not need to be centrally managed by a central server. This distributed deep learning approach, based on edge computing and blockchain technology, effectively protects data privacy and transparently records the data processing process, offering anonymity, decentralization, traceability, and tamper-resistance. Replacing the central server of federated learning with a parameter aggregation chain prevents privacy leaks during gradient parameter transmission. However, blockchain networks are slow, and uploading all neural network parameters to the blockchain for transaction execution would be time-consuming.

[0009] While group learning, leveraging blockchain technology and federated learning methods, can effectively improve data privacy protection issues that traditional centralized modeling lacks, training time is prohibitively long. Therefore, it is necessary to find a method that can accelerate the generation and training of distributed renewable energy scenarios while effectively protecting privacy. Summary of the Invention

[0010] To address the above issues, the present invention proposes a distributed renewable energy scenario generation method based on improved group learning:

[0011] To address the shortcomings of technology 1, blockchain technology is added to prevent federated learning from leaking the privacy of original data.

[0012] In order to solve the shortcomings of the second technique, instead of putting all the gradient parameters on the chain, only the gradient of the first layer of the neural network is put on the chain. As described in formula (1), the weight parameter ω iLike the original data, it is always retained locally and not transmitted. Therefore, when the gradient parameters of the first-layer weights are encrypted, x1 and x2 cannot be obtained through Equation (1), that is, the original data will not be exposed. The dispatching center needs to generate scenarios for multiple renewable energy regions so as to better reflect the correlation of renewable energy regions. Therefore, the structure of the federated learning center server is retained, and the gradients except the first layer of the neural network are transmitted on the center server.

[0013] Therefore, the improved group learning designed in the present invention only needs to upload the gradients of the first layer to the blockchain network for training, and the gradients of the remaining layers are trained by federated learning. The improvement of group learning can not only ensure that the original data is not leaked, but also accelerate the overall training speed.

[0014] The specific technical solution is as follows:

[0015] A distributed renewable energy scenario generation method based on improved group learning, comprising the following steps:

[0016] Step 1: Set the number M of each renewable energy region participating in the improved group learning modeling, set the number N of gradient exchanges, and set the set R of the measured values of the power of each region m ;

[0017] Step 2: Process the day-ahead prediction values and the measured values and use them as the training set;

[0018] Step 3: Use the training set in Step 2, and each region performs conditional generative adversarial network (CGAN) training locally to obtain the gradients of the first layer of the generator and the discriminator respectively Gradients of the remaining layers

[0019] Step 4: The gradients of the first layer of the neural network of each renewable energy region are uploaded to the distributed ledger platform, and the parameters of the remaining layers are uploaded to the central server; calculate respectively Integrate the gradients of the first layer and the gradients of the remaining layers, and form a complete new round of initial gradients with the obtained results;

[0020] Step 5: Repeat Steps 3 and 4 until the set number N of gradient exchanges;

[0021] Step 6: Extract the model parameters of the CGAN generator, input the prediction values and q noises conforming to the standard normal distribution into the generator, and output to obtain the day-ahead scenario set.

[0022] Further, in Step 2, the day-ahead prediction values and the measured values are normalized and used as the training set.

[0023] In Step 3, each region performs CGAN training locally using the RMSprop optimization algorithm.

[0024] In step 4, the distributed ledger platform is the Hyperledger Fabric platform.

[0025] In step 4, the federated averaging algorithm is used to calculate Integrate the gradients of the first layer and the gradients of the remaining layers, and form a complete new round of initial gradients with the obtained results;

[0026] There is also provided a distributed renewable energy scenario generation device based on improved swarm learning, including a distributed ledger platform, and a swarm learning network for building interactive gradient parameters;

[0027] Each residential load area serves as both a blockchain node and a client of federated learning, retaining the original load historical data locally without transmission to achieve data isolation, and using this data locally for neural network training;

[0028] Between each residential load area, the gradient parameters of the first layer of the neural network are interacted through the blockchain network, and the gradient parameters except the first layer are interacted through the central server.

[0029] Among them, the distributed ledger platform is Hyperledger Fabric.

[0030] The neural network adopts the CGAN model. The generator of the CGAN inputs the conditional value and noise, and the discriminator inputs the generated sample, conditional value, and real sample; the load day-ahead prediction data is used as the conditional value, and the measured data is used as the real sample to be input into the CGAN for training to generate the day-ahead scenario set.

[0031] The present invention also provides a computer system, including: a processor; a memory storing executable instructions, and when the executable instructions are executed by the processor, the distributed renewable energy scenario generation method based on improved swarm learning of the computer system is implemented.

[0032] The present invention also provides a computer-readable storage medium storing executable instructions, and when the executable instructions are executed by the processor, the distributed renewable energy scenario generation method based on improved swarm learning is implemented.

[0033] The present invention proposes a distributed renewable energy scenario generation method based on improved swarm learning, which introduces swarm learning into the scenario generation field for the first time, can better construct a model for generating the day-ahead scenarios of distributed renewable energy regions, and the generated day-ahead scenario results are consistent. The improved swarm learning method proposed by the present invention can not only break the data islands between regions, but also significantly shorten the training time compared with swarm learning. Description of the Drawings

[0034] Figure 1 For the improved group learning renewable energy regional framework of the present invention;

[0035] Figure 2 For the conditional generative adversarial network structure diagram of the present invention;

[0036] Figure 3 For the day-ahead scenario generation process of renewable energy regions based on improved group learning of the present invention;

[0037] Figure 4 Neural network structure of the prior art;

[0038] Figure 5(a) shows the day-ahead scenario set of wind power output power generated with the wind power of the Brussels region in Belgium as the modeling object in the embodiment;

[0039] Figure 5(b) shows the day-ahead scenario set of wind power output power generated with the wind power of the Flemish region in Belgium as the modeling object in the embodiment;

[0040] Figure 5(c) shows the day-ahead scenario set of wind power output power generated with the wind power of the Walloon region in Belgium as the modeling object in the embodiment;

[0041] Figure 6(a) shows the autocorrelation coefficient of the Brussels region in the embodiment;

[0042] Figure 6(b) shows the partial autocorrelation coefficient of the Brussels region in the embodiment;

[0043] Figure 6(c) shows the autocorrelation coefficient of the Flemish region in the embodiment;

[0044] Figure 6(d) shows the partial autocorrelation coefficient of the Flemish region in the embodiment;

[0045] Figure 6(e) shows the autocorrelation coefficient of the Walloon region in the embodiment;

[0046] Figure 6(f) shows the partial autocorrelation coefficient of the Walloon region in the embodiment. Detailed implementation manners

[0047] The specific technical solutions of the present invention will be described in conjunction with the accompanying drawings.

[0048] Taking the wind power of the three regions of Brussels, Flanders, and Wallonia in Belgium as the modeling object, that is, the number of each renewable energy region participating in the improved group learning modeling is 3. Set the number of gradient exchange rounds to 100. Take the day-ahead prediction values and actual measurement values of the regional wind power from January 1, 2019 to December 31, 2019 as the data set. Select the day-ahead prediction values and measured values of the wind power on 285 days in 2019 as the training set for model training. Using the above training set, each region uses as Figure 2The conditional generative adversarial network structure shown is used for CGAN training locally with the RMSprop optimization algorithm. The gradients of the first layer of the neural network in each renewable energy region are uploaded to the distributed ledger platform, and the parameters of the remaining layers are uploaded to the central server. The federated averaging algorithm is used to calculate the new round of initial gradients, and this cycle is repeated until the set number of gradient exchange rounds, which is 100 times. The distributed ledger platform selects Hyperledger Fabric. Finally, the predicted values and measured values on January 1, 2019 in the test set are used as test samples for model verification. The parameters of the CGAN generator model are extracted, and the predicted values and 300 noises that conform to the standard normal distribution are input into the generator, and the day-ahead scenario set shown in Figure 5 is output. The day-ahead scenario sets of the wind power output power generated in the three regions are shown by the light gray dotted lines respectively.

[0049] As Figure 1 shown, the framework of the renewable energy region based on improved swarm learning. Under this framework, Hyperledger Fabric is used as the distributed ledger platform to build a swarm learning network for interacting gradient parameters. Different renewable energy regions act as different nodes on the blockchain, and the certificate authority endorses their true identities to establish a trust basis among the renewable energy regions; the consensus mechanism ensures that different regions have the same ledger; the smart contract provides parameter operation services to all renewable energy regions, and the transactions are traceable and immutable. Different renewable energy regions locally retain the original historical data of the renewable energy output without transmission, realizing data isolation, which is the first protection of the original data, and neural network training is carried out locally using this data. Each region interacts the gradients of the first layer of the neural network through the blockchain distributed ledger platform and interacts the gradients except the first layer through the central server. Compared with swarm learning, it can effectively accelerate the model training speed. Compared with federated learning, it prevents the original data from being inferred using the gradient information after being maliciously captured during the gradient transmission process, which is the second protection of the original data.

[0050] The neural network adopts the CGAN model. Compared with the original GAN, CGAN has a better generalization effect on specified types of data samples. The input of the generator of CGAN is the conditional value and noise, and the input of the discriminator is the generated sample, conditional value, and real sample. The day-ahead prediction data is used as the conditional value, and the measured data is used as the real sample to be input into CGAN for training to generate the day-ahead scenario set. The structures of the generator and discriminator adopted are as Figure 2 shown.

[0051] The distributed renewable energy scenario generation method based on improved swarm learning has a process as Figure 3 shown.

[0052] Step 1: Set the number M of renewable energy regions participating in the improved group learning modeling, set the number N of gradient exchanges, and set the set R of measured power values for each region. m ;

[0053] Step 2: Normalize the day-ahead prediction values and measured values and use them as the training set.

[0054] Step 3: Use the training set in Step 2, and each region conducts conditional generative adversarial network (CGAN) training locally to obtain the first-layer gradients of the generator and discriminator respectively. Gradients of the remaining layers

[0055] Step 4: The first-layer gradients of the neural network for each renewable energy region are uploaded to the distributed ledger platform, and the parameters of the remaining layers are uploaded to the central server; use the federated averaging algorithm to calculate Integrate the first-layer gradients and the gradients of the remaining layers, and form a complete new round of initial gradients with the obtained results.

[0056] Step 5: Repeat Steps 3 and 4 until the set number N of gradient exchanges.

[0057] Step 6: Extract the CGAN generator model parameters, input the prediction values and q noises that conform to the standard normal distribution into the generator, and output to obtain the day-ahead scenario set.

[0058] As Figures 5(a) to 5(c) shown, the wind power of the three regions of Brussels, Flanders, and Wallonia in Belgium is used as the modeling object. The day-ahead prediction values and actual measurement values of the regional wind power from January 1, 2019 to December 31, 2019 are used as the data set. The prediction values and measured values on January 1, 2019 in the test set are used as test samples for model verification. The day-ahead scenario sets of the wind power output generated by the three regions are respectively shown as the light gray dotted lines in the following figures.

[0059] The present invention uses the improved group learning and group learning methods to train the model, and the generated day-ahead scenario results are consistent. The training situation is shown in the following table.

[0060]

[0061] The output trend of the day-ahead scenario set of wind power generated based on the improved group learning is basically the same as the prediction value, without obvious abnormal fluctuations, and the measured value can be well enveloped within the generated scenario set.

[0062] Figure 6 is the autocorrelation and partial autocorrelation coefficient diagram obtained from the generated day-ahead scenario set.

[0063] Figures 6(a) to 6(f)The box plot of the correlation coefficient shows the box plots of the correlation coefficients of the three regional scenarios generated by this method. The asterisk points represent the correlation coefficients of the measured wind power output values. It can be seen from the figure that the coefficients of the measured values in each region can be included in the generated scenarios. The fluctuation trends of the correlation coefficients of the generated scenario sets in the three wind power regions are similar, which is related to the proximity of the regions and the use of federated learning for training. From the autocorrelation coefficient plots in Figures 6(a), 6(c), and 6(e), it can be observed that for the wind farms in the three regions of Belgium, when the lag time is 1-6 hours, there is a certain correlation, and the autocorrelation coefficient decreases as the time interval increases. The partial correlation coefficient plots in Figures 6(b), 6(d), and 6(f) show that at a lag time of 1 hour, the partial correlation coefficient is large, indicating strong partial correlation. When the time interval is greater than 1 hour, the partial correlation coefficient between the output points decreases, that is, the partial correlation weakens. From the above analysis, it can be concluded that the variation characteristics of the autocorrelation coefficient and the partial correlation coefficient conform to the characteristics of renewable energy output.

[0064] A distributed renewable energy scenario generation device based on improved swarm learning is also provided, including a distributed ledger platform, and a swarm learning network for interacting gradient parameters is built.

[0065] Each residential load area acts as both a blockchain node and a client of federated learning, retaining the original load historical data locally without transmission to achieve data isolation, and using these data locally for neural network training.

[0066] Among the residential load areas, the first-layer gradient parameters of the neural network are interacted through the blockchain network, and the gradient parameters except the first layer are interacted through the central server.

[0067] Among them, the distributed ledger platform is Hyperledger Fabric.

[0068] The neural network adopts the CGAN model. The generator of the CGAN inputs the conditional value and noise, and the discriminator inputs the generated sample, conditional value, and real sample; the load day-ahead prediction data is used as the conditional value, and the measured data is used as the real sample to input into the CGAN for training to generate the day-ahead scenario set.

[0069] In addition, according to the embodiments of the present disclosure, the above-described method can be implemented as a computer software program. For example, the embodiments of the present disclosure include a computer program product, which includes a computer program tangibly embodied on a computer-readable medium, and the computer program includes program code for executing the above method.

[0070] According to the embodiments of the present disclosure, a computer architecture as described above can be used to implement the method according to the embodiments of the present disclosure, or multiple computer architectures as described above can cooperate with each other to implement the method according to the embodiments of the present disclosure.

[0071] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a part of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, as well as combinations of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0072] The units or modules described in the embodiments of the present disclosure can be implemented in software or in programmable hardware. The described units or modules can also be provided in a processor, and the names of these units or modules do not, in some cases, constitute a limitation to the units or modules themselves.

[0073] As another aspect, the present disclosure also provides a computer-readable storage medium, which can be the computer-readable storage medium included in the device described in the above embodiments; or it can be a computer-readable storage medium that exists separately and is not assembled into the device. The computer-readable storage medium stores one or more programs, and the one or more programs are used by one or more processors to execute the methods described in the present disclosure.

[0074] The above description is only a preferred embodiment of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the present disclosure.

Claims

1. A distributed renewable energy scenario generation method based on improved swarm learning, characterized in that, It includes the following steps: Step 1: Set the number of each renewable energy region participating in the improved group learning modeling , set the gradient exchange rounds, and set the set of measured power values for each region ; Step 2: Process the day-ahead prediction value and the measured value and use them as the training set; Step 3: Using the training set in Step 2, each region performs conditional generative adversarial network (CGAN) training locally to obtain the first-layer gradients of the generator and discriminator respectively , , and the gradients of the remaining layers , ; Step 4: The gradients of the first layer of the neural network in each renewable energy region , are uploaded to the distributed ledger platform, and the parameters of the remaining layers , are uploaded to the central server; calculate , , , , integrate the gradients of the first layer and the gradients of the remaining layers, and form a complete new round of initial gradients with the obtained results; Step 5: Repeat Steps 3 and 4 until the set number of gradient exchange rounds ; Step 6: Extract the CGAN generator model parameters, input the predicted value and q noises that conform to the standard normal distribution into the generator, and output the day-ahead scenario set.

2. The distributed renewable energy scenario generation method based on improved swarm learning according to claim 1, wherein, In Step 2, the day-ahead prediction value and the measured value are normalized and used as the training set.

3. A distributed renewable energy scenario generation method based on improved swarm learning according to claim 1, characterized in that, In Step 3, each region uses the RMSprop optimization algorithm to perform CGAN training locally.

4. A distributed renewable energy scenario generation method based on improved swarm learning according to claim 1, characterized in that, In Step 4, the distributed ledger platform is the Hyperledger Fabric platform.

5. A distributed renewable energy scenario generation method based on improved swarm learning according to claim 1, characterized in that In step 4, the federated averaging algorithm is used to calculate , , , , integrate the gradients of the first layer and the gradients of the remaining layers, and form the obtained results into a complete new round of initial gradients.

6. A distributed renewable energy scenario generation device based on improved swarm learning is used to implement a distributed renewable energy scenario generation method based on improved swarm learning as described in claim 1, and is characterized in that It includes a distributed ledger platform and constructs a group learning network for interacting gradient parameters; Each residential load region acts as both a blockchain node and a client of federated learning, retaining the original load historical data locally without transmission to achieve data isolation, and using this data locally for neural network training; Between each residential load region, the gradient parameters of the first layer of the neural network are interacted through the blockchain network, and the gradient parameters other than the first layer are interacted through the central server.

7. The distributed renewable energy scenario generation device based on improved swarm learning according to claim 6, characterized in that, The distributed ledger platform described is Hyperledger Fabric.

8. The distributed renewable energy scenario generation device based on improved swarm learning according to claim 6, characterized in that, The neural network uses the CGAN model. The generator of the CGAN inputs the conditional value and noise, and the discriminator inputs the generated sample, conditional value, and real sample; the day-ahead prediction data of the load is used as the conditional value, and the measured data is used as the real sample to be input into the CGAN for training to generate the day-ahead scenario set.

9. A computer system, characterized in that, It includes one or more processors and one or more memories. The memories store computer-executable instructions, and when the computer-executable instructions are executed by the processors, the method according to any one of claims 1 to 5 is implemented.

10. A computer-readable storage medium, characterized in that, It includes a memory, the memory stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images