A multi-level secure data resource pool construction method for photovoltaic power prediction

Abstract

The application relates to the technical field of solar photovoltaic power generation, and particularly discloses a multi-level secure data resource pool construction method for photovoltaic power prediction. The application ensures clear and traceable data ownership through data DNA fingerprints and smart contracts, effectively protecting the legal rights and interests and data sovereignty of data subjects. In combination with hierarchical differential privacy and a federal learning architecture, cross-domain collaborative training is realized, photovoltaic power prediction accuracy is improved, mathematical privacy protection is provided, advanced attacks such as model reverse are resisted, and data "available but invisible" is achieved. Based on a Shapley value and a multi-dimensional quality evaluation dynamic incentive mechanism, high-quality data contribution is encouraged, and the overall quality and availability of the data resource pool are improved. Through attribute-based dynamic access control and secure multi-party computation, fine-grained and context-aware permission management is realized, and the safety of core parameters is protected, so that the data sharing willingness and collaborative efficiency are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of solar photovoltaic power generation technology, and specifically to a method for constructing a multi-level secure data resource pool for photovoltaic power prediction. Background Technology

[0002] With the large-scale development of the photovoltaic industry, the accuracy requirements for photovoltaic power prediction are increasing. High-precision prediction relies on the collaborative analysis of large-scale, multi-source heterogeneous data such as meteorological data, power plant operation data, and equipment parameter data. However, the current photovoltaic data sharing and application process faces many bottlenecks: traditional centralized data resource pools require all participants to upload raw data, resulting in a separation of data ownership and control. This not only infringes on the legitimate rights and interests of data subjects but also poses a risk of sensitive information leakage, making it difficult to guarantee data sovereignty. Existing data anonymization and encryption methods can only resist basic attacks and cannot cope with new privacy threats such as model reverse engineering attacks and member inference attacks, making it difficult to meet relevant compliance requirements and resulting in insufficient privacy protection measures. There is a lack of unified quality assessment standards and effective incentive mechanisms. Some participants provide inferior data to reduce costs, leading to data resource pool pollution, which seriously affects the accuracy of photovoltaic power prediction models and results in inconsistent data quality. There is a lack of mutual trust mechanisms among different photovoltaic power plants, grid companies, and equipment manufacturers, resulting in low willingness to share data, forming data silos, and failing to maximize the utilization of data value, creating barriers to cross-domain collaboration.

[0003] Existing access control mechanisms are mostly static authorizations, unable to dynamically adjust access permissions based on data contribution and usage scenarios. This leads to prominent free-riding issues, further suppressing the enthusiasm for data sharing and resulting in a lack of flexibility in permission management. Therefore, there is an urgent need to build a multi-level secure data resource pool that balances data sovereignty protection, privacy and security safeguards, data quality control, cross-domain collaboration efficiency, and dynamic incentives, providing reliable data support for photovoltaic power prediction. Summary of the Invention

[0004] This invention aims to address the core issues in existing photovoltaic data sharing schemes, such as difficulty in guaranteeing data sovereignty, insufficient privacy protection, poor data quality, difficulties in cross-domain collaboration, and lack of dynamic permission management. It aims to achieve secure sharing of data that is usable but not visible, thereby improving the accuracy of photovoltaic power prediction and the efficiency of data collaboration.

[0005] This invention proposes a multi-level secure data resource pool construction method based on a blockchain-federated learning-differential privacy (BFL-DP) triple protection architecture, achieving end-to-end data security and value release through the following schemes: A method for constructing a multi-level secure data resource pool for photovoltaic power prediction includes: S1. Based on the consortium blockchain, a data ownership confirmation and traceability system is built to generate a data DNA fingerprint for each batch of data, and data asset registration and management are realized through smart contracts. S2. Based on the data registered in S1, construct a federated learning framework to collaboratively train the heterogeneous models and heterogeneous data provided by the participants and generate a global prediction model. S3. During the federated learning training and data query process in S2, an adaptive noise injection mechanism is used to implement hierarchical differential privacy protection. S4. Based on the performance of the global prediction model generated in S2, calculate the Shapley value for each data source, and combine it with the data quality score to implement tokenization incentives. S5. Based on the data contribution and quality assessment results calculated in S4, and combined with multi-dimensional attributes, execute attribute-based dynamic access control to complete the construction of a secure data resource pool.

[0006] Preferably, S1 includes: Based on the metadata (MetaData) including data source, type, and collection time, the local sensitivity hash (SampleHash) of each data sample, and the data generation timestamp, the data DNA fingerprint (DNA(D)) ​​is calculated using the SHA-256 hash function H(). The calculation relationship is as follows: DNA (D)=H (MetaData||H (SampleHash1||...||SampleHash)||Timestamp).

[0007] Preferably, S2 includes: Knowledge distillation is used to transfer knowledge from the global teacher model to the client student model. The loss function is: , Where LKD represents knowledge distillation loss, which is a comprehensive loss function that combines classification loss and KL divergence loss; LCE represents cross-entropy loss. α The weight coefficient (ranging from 0 to 1) is used to balance the proportions of cross-entropy loss and KL divergence loss in the total loss; LKL represents divergence loss, used to measure the difference between the student model output distribution and the teacher model output distribution, thus enabling knowledge transfer. Output for the client-side student model; zt Output for the global teacher model; T For temperature parameters (softening probability distribution). s For the Softmax function; For non-independent and identically distributed data, adaptive aggregation weights are calculated based on the amount of data from each client and the Jensen-Shannon divergence between their data distribution and the global data distribution.

[0008] Preferably, S3 includes: The total privacy budget is dynamically allocated based on the current task sensitivity score and the preset sensitivity threshold. e total ; Based on the allocated privacy budget, Laplace noise is injected during record-level data queries, Gaussian noise is injected during gradient-level federated learning uploads, and exponential mechanism noise is injected when output-level prediction results are published.

[0009] Preferably, in step S4, the method for calculating the Shapley value for each data source includes: The contribution φᵢ of each data source i is calculated using the model performance v(S) trained on a subset S of all data sources, and the formula is as follows: , in, ᵢ The Shapley value is the data source. v ( S (For using a subset of data) S The performance of the trained model N This represents the total amount of data.

[0010] Preferably, in step S4, the method for implementing tokenization incentives includes: According to Shapley value ᵢ and data quality score Qᵢ The number of tokens to be issued, Tokenᵢ, is calculated using the token issuance formula: , in, α Weighted by contribution level; β For quality weights; t This is the token settlement cycle number, used to mark the time dimension or calculation cycle of token issuance; Q i This represents the data quality score.

[0011] Preferably, S5 includes: The definition includes multidimensional attributes such as subject attributes, object attributes, environmental attributes, and behavioral attributes; Based on multidimensional attributes, access control decisions are made through the XACML policy decision engine. The decision results include: allow, deny, not applicable, and indeterminate (uncertain).

[0012] Preferably, during the data query and model interaction process, a secure multi-party computation protocol is used to protect core parameters.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention ensures clear and traceable data ownership through data DNA fingerprinting and smart contracts, effectively protecting the legitimate rights and sovereignty of data subjects. Combining hierarchical differential privacy and a federated learning architecture, it provides mathematically provable privacy protection while achieving cross-domain collaborative training and improving the accuracy of photovoltaic power prediction, resisting advanced attacks such as model reverse engineering, and achieving data that is "usable but not visible." A dynamic incentive mechanism based on Shapley values ​​and multi-dimensional quality assessment encourages the contribution of high-quality data, improving the overall quality and availability of the data resource pool. Through attribute-based dynamic access control and secure multi-party computation, it achieves fine-grained, context-aware permission management and protects the security of core parameters, thereby significantly improving data sharing willingness and collaborative efficiency, providing reliable, secure, and high-quality data support for photovoltaic power prediction. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a diagram illustrating the overall architecture of the multi-level secure data resource pool for the method and system for constructing a multi-level secure data resource pool for photovoltaic power prediction, as described in this invention. Figure 2 The flowchart of the adaptive heterogeneous federated learning algorithm for the construction method and system of multi-level secure data resource pool for photovoltaic power prediction in this embodiment of the invention is shown below. Figure 3 This invention relates to an attribute-based dynamic access control decision tree for a multi-level secure data resource pool construction method and system for photovoltaic power prediction. Figure 4 This invention relates to a method and system for constructing a multi-level secure data resource pool for photovoltaic power prediction, which employs a hybrid scheme of attribute-based homomorphic encryption and secret sharing. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1: S1. Based on the consortium blockchain, a data ownership confirmation and traceability system is built to generate data DNA fingerprints for each batch of data, and data asset registration and management are realized through smart contracts.

[0019] The system framework corresponding to this step is as follows: Figure 1 As shown. A data ownership verification and traceability system based on a consortium blockchain is constructed. This embodiment employs data DNA fingerprinting technology to generate a unique identifier for each batch of data. The generation formula is: DNA (D)=H (MetaData||H (SampleHash1||...||SampleHash)||Timestamp), MetaData contains metadata such as data source, type, and collection time; SampleHashᵢ is the Local Sensitive Hash (LSH) of the data sample; H() uses the SHA-256 hash function; and Timestamp is the data generation timestamp.

[0020] Data DNA fingerprints have three main characteristics: uniqueness (collision probability < 2-256), verifiability (authenticity can be verified without the original data), and tamper resistance (minor changes to the data can lead to completely different DNA).

[0021] Deploying smart contracts enables data asset management. Smart contracts include functions such as data registration and ownership transfer. The mapping relationship is mapping (bytes32=>DataAsset) public assets, supporting full lifecycle traceability and transactions of data.

[0022] S2. Based on the data registered in S1, a federated learning framework is constructed to collaboratively train the heterogeneous models and data provided by the participants, generating a global prediction model. The algorithm flow is as follows: Figure 2 As shown.

[0023] (1) Model heterogeneity handling: This embodiment supports different participants using different model architectures such as CNN, LSTM, and Transformer, and achieves collaborative training through knowledge distillation. The loss function is: , Where LKD represents knowledge distillation loss, which is a comprehensive loss function that combines classification loss and KL divergence loss; LCE represents cross-entropy loss. αThe weight coefficient (ranging from 0 to 1) is used to balance the proportions of cross-entropy loss and KL divergence loss in the total loss; LKL represents divergence loss, used to measure the difference between the student model output distribution and the teacher model output distribution, thus enabling knowledge transfer. Output for the client-side student model; zt Output for the global teacher model; T For temperature parameters (softening probability distribution). s For the Softmax function; (2) Data heterogeneous processing: For non-independent and identically distributed (Non-IID) data, an adaptive aggregation weight is introduced, calculated using the following formula: , in, n i For the client i The amount of data; N The total data volume is represented by JS(), which is the Jensen-Shannon divergence (measuring the difference between local and global data distributions). Clients with large data distribution differences have their weights automatically reduced to improve the convergence of the global model. P i Indicates the client i The probability of the local data distribution; P global Indicates the probability of the global data distribution; w i Indicates the client i Adaptive aggregation weights.

[0024] S3. During the federated learning training and data query process in S2, an adaptive noise injection mechanism is used to implement hierarchical differential privacy protection.

[0025] Adaptive Noise Injection (ANI) mechanism is used to achieve mathematically provable privacy protection. The specific steps are as follows: (1) Dynamic allocation of privacy budget: The total privacy budget is dynamically allocated based on task sensitivity and query complexity. e The total allocation formula is: , in, St This represents the sensitivity score for the current task. k To adjust the parameters; S threshold is the sensitivity threshold; The privacy budget allocated to the current task / query round is dynamically adjusted based on the sensitivity of the task; The preset total privacy budget (global privacy threshold) controls the upper limit of privacy loss throughout the entire process, preventing excessive leakage due to multiple rounds of operations.

[0026] (2) Layered noise injection: Record level: Laplace noise is added during the sampling of raw data, the formula is: , Where, Δ f For the sensitivity of the query function; The data after injecting Laplace noise; The original photovoltaic data is used; Lap() is the Laplace distribution function, used to generate noise that meets differential privacy requirements. The privacy budget allocated to record-level operations is part of the total budget. e A subset of total.

[0027] Gradient level: Gaussian noise is added when uploading gradients in federated learning. The formula is: Where γ is the noise standard deviation, C The gradient clipping threshold; The gradient of the model after injecting Gaussian noise; The gradient of the original model is computed locally on the client side; The privacy budget allocated to gradient-level operations is e A subset of total.

[0028] Output stage: When publishing the prediction results, add exponential noise, the formula is: , in, u ( D , o ) represents the output utility function; Δ u Sensitivity to utility function; The privacy budget allocated to output-level operations is e A subset of total; The prediction result is o The probability of; The total privacy loss satisfies the advanced combinatorial theorem, namely: , in, This is the preset total privacy budget.

[0029] Ensure that privacy protection does not exceed the limits.

[0030] S4. Based on the performance of the global prediction model generated in S2, calculate the Shapley value for each data source and combine it with the data quality score to implement tokenization incentives.

[0031] (1) Establish a dynamic data quality assessment and fair incentive mechanism: Shapley value data estimation: Based on game theory, the marginal contribution of each data source to the model performance is calculated using the following formula: , in, ᵢ For data source k Shapley value; v ( S (For using a subset of data) S The performance of the trained model N Given the total amount of data, the Shapley value satisfies the three axioms of efficiency, symmetry, and zero-element property.

[0032] (2) Data quality assessment: The data is comprehensively scored from three dimensions: completeness, accuracy, and timeliness, to obtain a quality score. Qi ( t ).

[0033] This embodiment uses tokenization incentives, issuing data tokens based on Shapley values ​​and quality scores, using the following formula: , in, α Weighted by contribution level; β The tokens are weighted by quality; they can be used to exchange for computing power resources, purchase prediction services, or redeem revenue sharing, forming a positive incentive cycle. t This is the token settlement cycle number; Qi This represents the data quality score.

[0034] S5. Based on the data contribution and quality assessment results calculated in S4, and combined with multi-dimensional attributes, execute attribute-based dynamic access control to complete the construction of a secure data resource pool.

[0035] The decision logic of access control policies is as follows: Figure 3 As shown, the security protection mechanism for core parameters can be found in [link to relevant documentation]. Figure 4 The example shown is a hybrid scheme combining homomorphic encryption and secret sharing.

[0036] This step proposes an attribute-based dynamic access control (ABDAC) model, specifically including: (1) Definition of multidimensional attributes: Access permissions are determined based on four types of attributes: subject attributes (identity, role, reputation score, contribution), object attributes (data sensitivity level, data type, data source), environmental attributes (time, location, network environment), and behavioral attributes (operation type, access frequency). (2) Strategy Decision Engine: The decision engine is built using XACML (eXtensible Access Control Markup Language) policy language. The decision results include four categories: Permit, Deny, NotApplicable, and Indeterminate. It supports multi-condition combination judgment and access rate limit to realize a dynamic mechanism of "the more you contribute, the greater your privileges".

[0037] This embodiment proposes an attribute-based dynamic access control (ABDAC) model, specifically including: (1) Definition of multidimensional attributes: Access permissions are determined based on four types of attributes: subject attributes (identity, role, reputation score, contribution), object attributes (data sensitivity level, data type, data source), environmental attributes (time, location, network environment), and behavioral attributes (operation type, access frequency). (2) Strategy Decision Engine: The decision engine is built using XACML (eXtensible Access Control Markup Language) policy language. The decision results include four categories: Permit, Deny, NotApplicable, and Indeterminate. It supports multi-condition combination judgment and access rate limit to realize a dynamic mechanism of "the more you contribute, the greater your privileges".

[0038] Example 2: Scenario Description: my country has selected 15 photovoltaic companies from three provinces and cities to jointly build a data resource pool, involving 1,200 distributed power stations. The data types include meteorological data, component operation data, inverter parameters, etc., and it is necessary to achieve secure data sharing and joint forecasting across enterprises. System Configuration: Consortium blockchain: It adopts the Hyperledger Fabric architecture and deploys 4 consensus nodes, which are responsible for data DNA registration, smart contract execution and transaction traceability; Federated learning: 15 clients (local nodes of each enterprise), 1 central server (responsible for global model aggregation), supporting hybrid model architectures of CNN and LSTM; Differential privacy: Total privacy budget ε = 3.0 (strong privacy protection level). k =0.8, Sthreshold=0.6; Data scale: 2.3TB of historical data accumulated, with a daily increase of 200GB. Data quality assessment indicators include completeness (weight 0.3), accuracy (weight 0.5), and timeliness (weight 0.2).

[0039] Example 3: Scenario characteristics: A large-scale photovoltaic base in three provinces of my country, with individual power plants ranging from 100 to 500 MW, and high data sensitivity (including core equipment operating parameters and grid dispatch data), requiring cross-provincial data collaboration and ultra-short-term power forecasting; Implementation Highlights: Secure computation: The core model parameters are protected by a 3-out-of-5 Shamir secret sharing scheme, ensuring that sensitive information is not leaked even if a single node is compromised; Heterogeneous adaptation: Supports the coexistence of three models: CNN, LSTM, and Transformer, and achieves knowledge synergy through knowledge distillation; Access control is tiered: data is divided into three levels: public, internal, and confidential. Access permissions are dynamically assigned based on contribution and reputation score. Confidential data is only accessible to the top 10 contributing companies. Additional incentive: The top 3 contributors in Shapley value will receive an additional 20% computing power resource reward.

[0040] In summary, this invention proposes a triple-protection architecture based on blockchain-federated learning-differential privacy (BFL-DP), integrating six core technologies: data DNA fingerprinting for ownership verification, adaptive heterogeneous federated learning, hierarchical differential privacy protection, Shapley value incentives, dynamic attribute access control, and secure multi-party computation. Blockchain ensures data sovereignty and full lifecycle traceability; federated learning supports cross-entity heterogeneous collaborative computation; and differential privacy provides mathematically provable privacy protection. Combined with data quality assessment and dynamic incentive mechanisms, this achieves secure sharing of data that is usable but not visible. This invention improves the accuracy of photovoltaic power prediction and the security of data sharing, reduces communication overhead, and can be widely applied in scenarios such as distributed photovoltaic data alliances and large-scale photovoltaic bases, providing technical support for the marketization of data elements in the energy internet.

[0041] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for constructing a multi-level secure data resource pool for photovoltaic power prediction, characterized in that, include: S1. Based on the consortium blockchain, a data ownership confirmation and traceability system is built to generate a data DNA fingerprint for each batch of data, and data asset registration and management are realized through smart contracts. S2. Based on the data registered in S1, construct a federated learning framework to collaboratively train the heterogeneous models and heterogeneous data provided by the participants and generate a global prediction model. S3. During the federated learning training and data query process in S2, an adaptive noise injection mechanism is used to implement hierarchical differential privacy protection. S4. Based on the performance of the global prediction model generated in S2, calculate the Shapley value for each data source, and combine it with the data quality score to implement tokenization incentives. S5. Based on the data contribution and quality assessment results calculated in S4, and combined with multi-dimensional attributes, execute attribute-based dynamic access control to complete the construction of a secure data resource pool.

2. The method for constructing a multi-level secure data resource pool for photovoltaic power prediction according to claim 1, characterized in that, S1 includes: Based on the metadata (MetaData) including data source, type, and collection time, the local sensitivity hash (SampleHash) of each data sample, and the data generation timestamp, the data DNA fingerprint (DNA(D)) ​​is calculated using the SHA-256 hash function H(). The calculation relationship is as follows: DNA (D)=H (MetaData||H (SampleHash1||...||SampleHash)||Timestamp).

3. The method for constructing a multi-level secure data resource pool for photovoltaic power prediction according to claim 1, characterized in that, S2 includes: Knowledge distillation is used to transfer knowledge from the global teacher model to the client student model. The loss function is: ， Among them, L KD The knowledge distillation loss represents a comprehensive loss function that combines classification loss and KL divergence loss; L CE Represents cross-entropy loss; α L represents the weighting coefficient, used to balance the proportions of cross-entropy loss and KL divergence loss in the total loss; KL This represents divergence loss, used to measure the difference between the output distribution of the student model and the output distribution of the teacher model, thus enabling knowledge transfer. Output for the client-side student model; z t Output for the global teacher model; T For temperature parameters, σ For the Softmax function; For non-independent and identically distributed data, adaptive aggregation weights are calculated based on the amount of data from each client and the Jensen-Shannon divergence between their data distribution and the global data distribution.

4. The method for constructing a multi-level secure data resource pool for photovoltaic power prediction according to claim 1, characterized in that, S3 includes: The total privacy budget is dynamically allocated based on the current task sensitivity score and the preset sensitivity threshold. ε total ; Based on the allocated privacy budget, Laplace noise is injected during record-level data queries, Gaussian noise is injected during gradient-level federated learning uploads, and exponential mechanism noise is injected when output-level prediction results are published.

5. The method for constructing a multi-level secure data resource pool for photovoltaic power prediction according to claim 1, characterized in that, In step S4, the method for calculating the Shapley value for each data source includes: The contribution φᵢ of each data source i is calculated using the model performance v(S) trained on a subset S of all data sources, and the formula is as follows: ， in, φᵢ The Shapley value is the data source. v ( S (For using a subset of data) S The performance of the trained model N This represents the total amount of data.

6. The method for constructing a multi-level secure data resource pool for photovoltaic power prediction according to claim 5, characterized in that, In S4, the method for implementing tokenization incentives includes: According to Shapley value φᵢ and data quality score Qᵢ The number of tokens to be issued, Tokenᵢ, is calculated using the token issuance formula: ， in, α Weighted by contribution level; β For quality weights; t This is the token settlement cycle number, used to mark the time dimension or calculation cycle of token issuance; Q i This represents the data quality score.

7. The method for constructing a multi-level secure data resource pool for photovoltaic power prediction according to claim 1, characterized in that, S5 includes: The definition includes multidimensional attributes such as subject attributes, object attributes, environmental attributes, and behavioral attributes; Based on multidimensional attributes, access control decisions are made through the XACML policy decision engine. The decision results include: allow, deny, not applicable, and indeterminate (uncertain).

8. The method for constructing a multi-level secure data resource pool for photovoltaic power prediction according to claim 1, characterized in that, During data query and model interaction, a secure multi-party computation protocol is used to protect core parameters.

Images