A photovoltaic power prediction data governance method and system based on a data resource pool

Abstract

The application discloses a photovoltaic power prediction data management method and system based on a data resource pool, belongs to the field of solar photovoltaic power generation, and comprises the following steps: constructing a space-time dynamic knowledge graph based on a photovoltaic data resource pool, and performing quality detection on photovoltaic power prediction data to obtain abnormal data; introducing a physical constraint into a generative adversarial network to process the abnormal data and obtain repaired data; processing the repaired data based on a self-supervised learning method to obtain high-quality pseudo labels; adopting a hybrid architecture of a block chain and a directed acyclic graph to record the conversion operation of data entities and metadata, and obtaining a data blood relationship graph; adopting a data quality evaluation model to process the repaired and labeled data in combination with the high-quality pseudo labels and the data blood relationship graph to obtain a comprehensive quality score; and obtaining and executing an optimal repair strategy based on the data state and the comprehensive quality score. The application can realize efficient management of photovoltaic data resource pools of different scales.

Description

Technical Field

[0001] This invention belongs to the field of solar photovoltaic power generation technology, and in particular relates to a method and system for photovoltaic power prediction data governance based on a data resource pool. Background Technology

[0002] With the booming development of the photovoltaic industry, the photovoltaic power prediction data resource pool has gathered massive amounts of multi-source heterogeneous data. As the core foundation for accurate photovoltaic power prediction and optimized power plant operation, the quality of these data directly determines the effectiveness and value of related applications. However, due to the influence of multiple factors such as differences in data acquisition equipment, environmental interference, and transmission failures, photovoltaic power prediction data generally suffers from quality defects, which seriously restricts the improvement of prediction accuracy and the efficient operation of power plants. Existing data governance methods have significant limitations: In terms of quality detection, traditional rule engines rely on manually configured thresholds, can only identify known types of anomalies, and cannot adaptively capture new types of data anomalies, resulting in insufficient detection accuracy and flexibility, exhibiting a passive and lagging problem; in terms of missing value imputation, traditional methods such as mean imputation and interpolation ignore the inherent spatiotemporal correlation and physical constraints of photovoltaic data, easily introducing systemic biases and compromising the authenticity and usability of the data; outlier handling is often subjective and arbitrary, simply removing outliers may result in the loss of key information such as extreme events, while directly retaining them can interfere with subsequent analysis and affect model robustness; data lineage tracing faces difficulties, as the original source and transformation history of multi-source data are difficult to trace after a series of operations such as fusion and transformation, making it impossible to quickly locate the root cause when data quality problems occur; labeled data is severely scarce, and data quality labeling relies on manual work by experts, which is not only costly and inefficient but also has limited coverage, making it difficult to meet the needs of large-scale data governance. Therefore, there is an urgent need for a data governance method with intelligent, adaptive, and interpretable characteristics to specifically address the above problems, comprehensively improve the quality and governance efficiency of photovoltaic power prediction data, and provide strong support for the sustainable and healthy development of the photovoltaic industry. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a photovoltaic power prediction data governance method and system based on a data resource pool, thereby resolving the issues present in the prior art.

[0004] To achieve the above objectives, this invention provides a photovoltaic power prediction data governance method based on a data resource pool, comprising: A spatiotemporal dynamic knowledge graph is constructed based on the photovoltaic data resource pool. The photovoltaic power prediction data is then subjected to quality detection based on the knowledge graph to identify abnormal data. The abnormal data is processed using a generative adversarial network with introduced physical constraints to obtain repaired data; A self-supervised learning method based on contrastive predictive coding is used to process the repaired data to obtain high-quality pseudo-labels; By adopting a hybrid architecture of blockchain and directed acyclic graph, the transformation operations of data entities and metadata are recorded to obtain a traceable data lineage graph; A data quality assessment model covering six dimensions is used, and the repaired and labeled data is processed by combining the high-quality pseudo-labels and the data lineage map to obtain a comprehensive quality score; A multi-agent reinforcement learning method is adopted to process data state and comprehensive quality score from a set of preset repair strategies, and then obtain and execute the optimal repair strategy.

[0005] Optionally, the process of constructing a spatiotemporal dynamic knowledge graph based on the photovoltaic data resource pool includes: Based on the data entities in the photovoltaic data resource pool, four entity types are defined: physical entity, data entity, event entity, and rule entity. Based on the relationships between the data entities, three relationship types are defined: static relationship, dynamic relationship, and spatiotemporal relationship. Based on the dynamic relationship, a relationship strength evolution model is used to model the association strength that changes with time and environmental state, resulting in a dynamic relationship network. Based on the entities, relationships, and dynamic relationship networks, the spatiotemporal dynamic knowledge graph is constructed.

[0006] Optionally, the process of quality inspection of photovoltaic power prediction data based on the knowledge graph includes: Based on the spatiotemporal dynamic knowledge graph, a graph neural network is used to process the semantic information of entities and relationships and extract implicit semantic rules. Based on the semantic rules and the preset explicit rules, logical reasoning and consistency checks are performed on the photovoltaic power prediction data to identify data that violates the rules; The data that violates the rules is marked as abnormal data.

[0007] Optionally, the expression for processing the semantic information of entities and relations using a graph neural network is as follows: ; In the formula, , Let v and u be the feature vectors of nodes v and u in the k-th layer, respectively. Let v be the relationship vector between nodes v and u. Let be the weight matrix of the k-th layer, and σ be the activation function. Let v be the feature vector of node v in the kth layer. It is a set of nodes.

[0008] Optionally, the process of using a generative adversarial network with introduced physical constraints to process the anomalous data and obtain repaired data includes: A generator incorporating a spatiotemporal attention U-Net architecture is used to generate repair data based on noise and contextual data; A discriminator with a dual discrimination mechanism is used to process the repair data output by the generator, and at the same time evaluate the authenticity and physical consistency of the data. The training process of the generative adversarial network is stabilized using the Wasserstein distance.

[0009] Optionally, the physical constraints include: introducing physical constraints based on the law of conservation of energy to check the consistency of the conversion between power data and irradiance data; and introducing physical constraints based on power ramp-up rate limits to check the rationality of instantaneous changes in power data.

[0010] Optionally, the process of processing the repaired data to obtain high-quality pseudo-labels based on a self-supervised learning method of contrastive predictive coding includes: Five types of pre-training tasks are designed: time series prediction, mask reconstruction, time series comparison, transformation invariance, and cluster assignment. The self-supervised learning model is pre-trained to obtain a pre-trained model. The pre-trained model is used to process the repaired data to generate high-quality pseudo-labels.

[0011] Optionally, the process of constructing a traceable data lineage graph using a hybrid architecture of blockchain and directed acyclic graph includes: Define nodes in a directed acyclic graph as data entities, define edges in a directed acyclic graph as data transformation operations, and associate them with metadata containing timestamps, operation types, and operation parameters. The key lineage records containing the aforementioned metadata are stored in the blockchain to form an immutable storage chain; Design a bloodline query language and smart contracts to query and verify the bloodline data.

[0012] Optionally, a kinship influence analysis is performed based on the traceable kinship data. The kinship influence analysis process includes: Based on the aforementioned data lineage map, a depth-first search algorithm is used to process the data with quality problems and trace back to locate the root cause of the problem. Based on the aforementioned kinship map, a breadth-first search algorithm is used to process the changed data and propagate backwards to assess the scope of the impact.

[0013] Optionally, a data quality assessment model covering six dimensions is used to process the repaired and labeled data, specifically including: Data quality is defined using six dimensions: completeness, accuracy, consistency, timeliness, validity, and credibility. The analytic hierarchy process (AHP) is used to assign dynamic weights to each quality dimension. Based on the dynamic weights, the evaluation scores of each dimension are weighted and fused to obtain a comprehensive quality score.

[0014] Optionally, the process of employing a multi-agent reinforcement learning method includes: deploying dedicated repair strategy agents for irradiance data, power data, temperature data, and wind speed data respectively; deploying a coordinating agent to perform global optimization and conflict resolution on the decisions of each dedicated repair strategy agent; and using the QMIX algorithm to learn the cooperative strategies of each agent.

[0015] This invention also provides a photovoltaic power prediction data governance system based on a data resource pool, used to implement the method, the system comprising: The knowledge graph module is used to build and maintain spatiotemporal dynamic knowledge graphs, define entity types and relation types, and perform spatiotemporal dynamic modeling. The quality detection module is used to generate semantic quality detection rules based on the spatiotemporal dynamic knowledge graph, and to use the rules to detect photovoltaic power prediction data and identify abnormal data. The data repair module is used to repair photovoltaic power prediction data based on the identified abnormal data, using a generative adversarial network that incorporates physical constraints. The self-supervised labeling module is used to train a model based on the repaired large-scale unlabeled photovoltaic data using a contrastive predictive coding self-supervised learning framework to generate pseudo-labels for data quality. The bloodline tracing module is used to construct a traceable data bloodline graph based on data operation records, using a hybrid architecture of blockchain and directed acyclic graph; The quality assessment module is used to generate a comprehensive quality score based on the repaired data using a multi-dimensional quality assessment model that includes completeness, accuracy, consistency, timeliness, effectiveness, and credibility. The decision engine module is used to select and execute the optimal repair strategy from multiple repair strategies based on the data state and the comprehensive quality score, using a multi-agent reinforcement learning decision engine.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention achieves novel adaptive anomaly identification for photovoltaic (PV) data by constructing a spatiotemporal dynamic knowledge graph, significantly improving detection accuracy and flexibility compared to traditional rule engines. The use of a physical constraint generative adversarial network for data repair ensures that the repaired data conforms to physical laws such as energy conservation, solving the problem of system bias introduced by traditional methods. Pseudo-labels generated based on self-supervised learning reduce labeling requirements by over 90%, greatly reducing manual costs. A hybrid architecture combining blockchain and directed acyclic graphs enables complete traceability and tamper-proof verification of data lineage, solving the data source tracing problem. A six-dimensional quality assessment model provides a comprehensive and objective quality evaluation system, and a multi-agent reinforcement learning decision engine can intelligently select the optimal repair strategy. This solution improves the anomaly detection accuracy of PV power prediction data to over 95%, achieves 98% physical consistency of repaired data, and improves data governance efficiency by approximately 60%, providing a high-quality data foundation for accurate PV power prediction. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a diagram illustrating the overall architecture of the intelligent data governance system for the photovoltaic power prediction data governance method and system based on a data resource pool, as described in this invention. Figure 2 This is a flowchart illustrating the spatiotemporal dynamic knowledge graph construction process of the photovoltaic power prediction data governance method and system based on a data resource pool, as described in an embodiment of the present invention. Figure 3 This is a schematic diagram of the blockchain-DAG hybrid lineage graph structure of the photovoltaic power prediction data governance method and system based on data resource pool according to an embodiment of the present invention; Figure 4 This is a structural diagram of a six-dimensional data quality assessment model for a photovoltaic power prediction data governance method and system based on a data resource pool, as described in an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the working principle of the multi-agent reinforcement learning decision engine of the photovoltaic power prediction data governance method and system based on a data resource pool, as described in an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] Example 1.

[0021] This invention aims to solve the technical problems existing in photovoltaic data resource pools, such as passive and lagging quality detection, simple and crude filling of missing values, subjective and arbitrary handling of outliers, difficulty in tracing data lineage, and serious lack of labeled data, and provides a comprehensive and intelligent photovoltaic power prediction data governance solution.

[0022] like Figure 1 As shown, this embodiment provides a photovoltaic power prediction data governance method based on a data resource pool. This method proposes a data intelligent governance method and system based on a three-in-one architecture of knowledge graph-generative adversarial network-self-supervised learning (KG-GAN-SSL). The specific technical solution is as follows: (i) Knowledge graph-driven semantic data quality inspection. A spatiotemporal dynamic knowledge graph is constructed based on the photovoltaic data resource pool. The photovoltaic power prediction data is then subjected to quality inspection based on the knowledge graph to identify anomalous data.

[0023] Step S1: As Figure 2 As shown, a spatio-temporal dynamic knowledge graph (STDKG) in the photovoltaic field is constructed, and semantic quality detection rules are automatically learned from the graph using a graph neural network to achieve semantic data quality detection.

[0024] Furthermore, the process of constructing a spatiotemporal dynamic knowledge graph based on the photovoltaic data resource pool includes: defining four entity types—physical entities, data entities, event entities, and rule entities—based on the data entities in the photovoltaic data resource pool; defining three relationship types—static relationships, dynamic relationships, and spatiotemporal relationships—based on the relationships between data entities; modeling the association strength changing with time and environmental conditions using a relationship strength evolution model based on dynamic relationships to obtain a dynamic relationship network; and constructing a spatiotemporal dynamic knowledge graph based on entities, relationships, and the dynamic relationship network. A specific implementation method in this embodiment includes the following steps: Ontology layer construction: Defines four types of entities: physical entities (photovoltaic modules, inverters, combiner boxes, transformers, etc.), data entities (irradiance, temperature, wind speed, power, voltage, current, etc.), event entities (faults, maintenance, cleaning, shading, etc.), and rule entities (physical constraints, business rules, quality standards, etc.); Defines three types of relationships: static relationships (installed at, belong to, connected to, etc.), dynamic relationships (affect(t), related(t), constraint(t), etc.), and spatiotemporal relationships (prior to, simultaneous with, cause, etc.).

[0025] Spatiotemporal dynamic modeling: Considering the evolution of relationship strength over time, a formula is used. Modeling is performed, in which, Let E(t) represent the initial strength of the relationship, t represent time, and E(t) represent the environmental state, for example, the temperature-power relationship strengthens in summer and weakens in winter.

[0026] Furthermore, the process of quality inspection of photovoltaic power prediction data based on knowledge graphs includes: processing the semantic information of entities and relationships using graph neural networks based on spatiotemporal dynamic knowledge graphs to extract implicit semantic rules; performing logical reasoning and consistency checks on photovoltaic power prediction data based on semantic rules and preset explicit rules to identify data that violates the rules; and marking data that violates the rules as abnormal data. As a specific implementation of this embodiment, the process of semantic quality rule extraction includes: Automatically generate quality inspection rules from the knowledge graph, such as "IF (irradiance > 1000 W / m)". 2 "AND(power < 0.1 P_rated) THEN anomaly"; simultaneously, complex semantic rules are learned through a graph neural network (GNN), the formula is: It automatically discovers hidden rules, improving the accuracy of anomaly detection.

[0027] in, , They are nodes v , u In the k The feature vector of the layer, For nodes v and u Relationship vectors For the first k The weight matrix of the layer, σ For activation function, Let v be the feature vector of node v in the kth layer. It is a set of nodes.

[0028] (ii) Generative Adversarial Network-Driven Data Repair. A generative adversarial network with introduced physical constraints is used to process abnormal data and obtain repaired data.

[0029] Furthermore, the data inpainting process driven by the generative adversarial network includes: using a generator with a spatiotemporal attention U-Net architecture to generate inpainted data based on noise and contextual data; using a discriminator with a dual discrimination mechanism to process the inpainted data output by the generator, while simultaneously evaluating the authenticity and physical consistency of the data; and using Wasserstein distance to stabilize the training process of the generative adversarial network. As a specific implementation of this embodiment, the following steps are included: Step S2: A Physics-Constrained GAN (PC-GAN) is used to intelligently repair missing and anomalous data. The discriminator simultaneously checks the authenticity and physical consistency of the data. The PC-GAN consists of a generator with a spatiotemporal attention U-Net architecture and a dual discriminator.

[0030] Generator Architecture: The generator G adopts a spatiotemporal attention U-Net architecture, including an encoder, a bottleneck layer, and a decoder, which learns to generate high-quality data from noise and context. The formula is as follows: Where z is noise, x_context is context data, and c is conditional information (time, weather, etc.); the encoder is used to extract multi-scale spatiotemporal features, the bottleneck layer realizes spatiotemporal attention fusion, and the decoder reconstructs missing data layer by layer.

[0031] Physical Constraint Discriminator: Discriminator D employs a dual discrimination mechanism, not only determining the authenticity of data but also checking physical consistency. The expression is as follows: ,in Used to determine the authenticity of data. Used to check physical consistency; physical constraint loss is expressed by the formula. The calculation involves the first term representing the energy conservation constraint, the second term representing the ramp rate constraint, and λ1 and λ2 representing weighting coefficients. P For power data, η For conversion efficiency, I For irradiance data, The rate of change of power, is the maximum ramp rate threshold, and 1() is the indicator function.

[0032] Adversarial training strategy: Employing a Wasserstein distance-stabilized training process, the generator loss function is... The discriminator loss function is GP is the gradient penalty term, which prevents mode collapse.

[0033] (III) Data labeling driven by self-supervised learning. Based on the self-supervised learning method of contrastive predictive coding, the repaired data is processed to obtain high-quality pseudo-labels.

[0034] Furthermore, the process of obtaining high-quality pseudo-labels from data annotation driven by self-supervised learning includes: designing five types of pre-training tasks—temporal prediction, mask reconstruction, temporal comparison, transform invariance, and cluster assignment—to pre-train the self-supervised learning model and obtain a pre-trained model; using the pre-trained model, processing the repaired data to generate high-quality pseudo-labels. As a specific implementation method in this embodiment, the following steps are included: Step S3: Based on the self-supervised framework of Contrastive Predictive Coding (CPC), high-quality data annotations are automatically generated, reducing annotation requirements by 90%.

[0035] Pre-training task design: Five types of self-supervised pre-training tasks are designed, namely temporal prediction (predicting data in the next k steps), mask reconstruction (randomly masking 15% of the data and reconstructing it), temporal comparison (positive sample pairs have greater similarity than negative sample pairs), transformation invariance (representation remains unchanged after data augmentation), and cluster assignment (unsupervised clustering to learn data patterns), and corresponding loss functions are designed for each.

[0036] The formula for time series prediction loss is: ; The formula for mask reconstruction loss is: ; The formula for time series comparison loss is: ; The formula for transformation invariance loss is: ; The clustering assignment loss formula is: ; in, for t + k Data at any given time For data at time t and earlier, x masked For the concealed data, f For reconstruction function, These represent the feature representations of the data at the corresponding time points, and sim is the similarity function. τ For temperature parameters, T For data augmentation operations, q i Assign probabilities to clusters. p i Let be the target distribution probability.

[0037] Pseudo-label generation: High-confidence pseudo-labels are generated using a pre-trained model. The formula is as follows: ,reserve The sample; the false labels are optimized using the confirmation bias correction method, the formula is as follows: ,in y expert Data was annotated by a small number of experts.

[0038] Active learning sampling: Prioritize labeling samples with high uncertainty and high representativeness. The sampling score is calculated using the formula Score(x)=U(x)・R(x), where U(x) is uncertainty and R(x) is representativeness.

[0039] (iv) Data lineage tracing. A hybrid architecture of blockchain and directed acyclic graph is adopted to record the transformation operations of data entities and metadata, thereby obtaining a traceable data lineage graph.

[0040] Furthermore, the process of tracing the data lineage graph includes: defining nodes in the directed acyclic graph as data entities, defining edges in the directed acyclic graph as data transformation operations, and associating them with metadata containing timestamps, operation types, and operation parameters; storing key lineage records containing metadata in the blockchain to form an immutable storage chain; and designing a lineage query language and smart contracts to query and verify the data lineage graph. As a specific implementation of this embodiment, the following steps are included: Step S4: As Figure 3 As shown, a blockchain-DAG hybrid data lineage graph is constructed to support bidirectional data traceability and smart contract verification, enabling full lifecycle data traceability.

[0041] Directed Acyclic Graph (DAG) lineage modeling: Model the data flow as a DAG, where nodes represent data entities (raw, cleaned, merged, feature-engineered, etc.), edges represent data transformation operations (collection, cleaning, aggregation, computation, etc.), and metadata includes timestamps, operation types, operation parameters, and quality scores; design a lineage query language, such as "TRACE × FROM source WHEREquality<0.8 DEPTH 5".

[0042] Blockchain Immutable Storage: Key lineage information is stored on the blockchain, with the following block structure: Block={header:{prev_hash,timestamp,merkle} root},body:{lineage records Automatic verification is achieved through smart contracts, which include the verify function and the audit function, used to verify data lineage and audit data operations, respectively.

[0043] Bloodline influence analysis: Depth-first search (DFS) is used to trace back and locate the root cause of the problem. The formula is Root Cause =DFS backward (faultydata stop condition Breadth-first search (BFS) is used for backpropagation to assess the impact range, using the formula Impact. Set =BFS forward (modified data depth limit ).

[0044] (v) Multidimensional data quality assessment. A data quality assessment model covering six dimensions was adopted, and high-quality pseudo-labels and data lineage maps were used to process the repaired and labeled data to obtain a comprehensive quality score.

[0045] Furthermore, the processing of the repaired and labeled data specifically includes: defining data quality using six dimensions: completeness, accuracy, consistency, timeliness, validity, and credibility; assigning dynamic weights to each quality dimension using the analytic hierarchy process (AHP); and weighting and fusing the evaluation scores of each dimension based on the dynamic weights to obtain a comprehensive quality score. As a specific implementation method of this embodiment, the following steps are included: Step S5: As Figure 4 As shown, a six-dimensional data quality assessment model (6DDQ) is established to achieve a comprehensive and objective assessment of data quality.

[0046] Quality dimensions are defined as follows: completeness, accuracy, consistency, timeliness, effectiveness, and credibility. Each dimension uses a corresponding calculation formula.

[0047] Completeness: ,in For the amount of missing data, N_total represents the total amount of data. accuracy: Where n is the number of data samples, xi For real data, For predictive data; consistency: Where Nconflict is the amount of conflicting data and Ncheck is the amount of checked data; Timeliness: , where λ is the attenuation coefficient and Δt is the data delay; Validity: , where Nvalid is the amount of data that passed the physical constraint test; Credibility: ,in To ensure the credibility of the data source, For historical credibility, To ensure the credibility of the consensus; Overall Quality Score: The weights of each dimension are determined using the Analytic Hierarchy Process (AHP). The formula for the overall quality score is as follows: , where Σ wᵢ =1; The weights are dynamically adjusted according to the application scenario. For example, the weight of timeliness is increased in real-time prediction scenarios, the weight of accuracy is increased in offline training scenarios, and the weight of credibility is increased in compliance audit scenarios.

[0048] Quality evolution monitoring: Real-time monitoring of quality trends, the formula is as follows ,when It triggers quality degradation alarms in a timely manner; and displays the data quality status in real time through a quality score visualization dashboard.

[0049] (vi) Intelligent Data Repair Decision Engine. A multi-agent reinforcement learning method is adopted to process data based on data status and comprehensive quality score, obtain and execute the optimal repair strategy from a set of preset repair strategies.

[0050] Step S6: As Figure 5 As shown, a multi-agent reinforcement learning decision engine is constructed to comprehensively consider the repair effect, computational cost and time constraints, and automatically select the optimal repair strategy.

[0051] State space definition: State vector S = [missing rate, outlier rate, data distribution, contextual information, historical repair results].

[0052] Action space definition: Optional repair strategies include interpolation (linear, spline, Kriging, etc.), prediction (LSTM, Prophet, WaveNet, etc.), generation (GAN, VAE, Flow, etc.), neighborhood methods (KNN, weighted average, etc.), and deletion methods (directly removing low-quality data).

[0053] Reward function design: Taking into account the repair effect, computational cost, and time, the reward function is as follows: , where α, β, and γ are weighting coefficients, ΔQ is the quality improvement value, Cost is the computation cost, and Time is the time consumed.

[0054] Multi-agent collaboration: Dedicated agents are deployed for different data types, such as Agent_Irradiance (focusing on meteorological data repair) and Agent_Power (focusing on power data repair), and Agent_Coordination is set up (responsible for global optimization and conflict resolution); the QMIX algorithm is used to learn the collaboration strategy, the formula is as follows: Where f is a monotonically increasing function, Q 1 to Qn Each intelligent agent Q Value function, s1 to s n These represent the states of each agent, a1 to a2. n These represent the actions of each intelligent agent.

[0055] As a specific implementation method of this embodiment, this embodiment addresses the governance of a provincial-level photovoltaic data resource pool. The scenario description includes: This provincial-level photovoltaic data resource pool aggregates data from over 200 photovoltaic power stations across the province, with a total installed capacity of 5GW and a cumulative data volume of 50TB. Data sources include local data acquisition equipment at photovoltaic power stations, weather stations, and third-party data service providers. The data quality is uneven, with a large amount of missing and abnormal data, complex data lineage, and severely insufficient labeled data.

[0056] Knowledge Graph: Construct a spatiotemporal dynamic knowledge graph for the photovoltaic field containing 350,000 entities and 1.2 million relationships, covering physical entities, data entities, event entities, rule entities, and various relationships; PC-GAN: It adopts a network structure with a 6-layer generator and a 5-layer discriminator, and is trained based on Wasserstein distance; Self-supervised learning: Pre-trained based on 1 billion samples, jointly optimizing five types of tasks: temporal prediction, mask reconstruction, temporal comparison, transformation invariance, and cluster assignment; Kinship Graph: Adopts a consortium blockchain 4-node architecture, storing 20 million kinship records through DAG; Quality assessment: Based on a six-dimensional data quality assessment model, the analytic hierarchy process (AHP) was used to determine the initial weights of each dimension, with completeness weighted at 0.2, accuracy weighted at 0.25, consistency weighted at 0.15, timeliness weighted at 0.1, validity weighted at 0.2, and credibility weighted at 0.1. Decision Engine: Deploying the Agent 辐照度 Agent 功率 Agent 温度 Agent 风速 Agent 协调五类智能体 Co-training is performed based on the QMIX algorithm.

[0057] As another specific implementation of this embodiment, this embodiment focuses on the governance of real-time data from a 100MW power plant. The scenario description includes: This 100MW photovoltaic power plant needs to process real-time data streams with a data transmission rate of 5000 records per second. It requires millisecond-level data quality detection and repair to support 15-minute ultra-short-term photovoltaic power prediction, placing extremely high demands on the real-time performance and accuracy of data governance.

[0058] Streaming knowledge graph update: An incremental update strategy is adopted to update relationship strength in real time with an update latency of <100ms, ensuring the timeliness of quality detection rules; Online GAN inference: Quantization acceleration of PC-GAN models, GPU-based inference, single data repair time <20ms; Distributed lineage records: Asynchronous writing of lineage records is implemented using Kafka message queues to ensure that real-time data stream processing is not affected, while guaranteeing the complete storage of lineage information; Edge agent deployment: Dedicated agents are deployed at local edge nodes of the power plant to enable on-site data repair, reduce network transmission latency, and improve real-time processing capabilities.

[0059] In summary, this invention constructs a spatiotemporal dynamic knowledge graph and automatically learns semantic quality detection rules through graph neural networks; it employs physical constraint-based adversarial networks to repair missing and anomalous data, ensuring the physical consistency of the repaired data; it generates high-quality pseudo-labels based on contrastive predictive coding through self-supervised learning, reducing annotation requirements; it constructs a blockchain-DAG hybrid data lineage graph to achieve bidirectional data traceability; it establishes a six-dimensional data quality assessment model for comprehensive and objective quality evaluation; and it uses multi-agent reinforcement learning to automatically select the optimal repair strategy. This invention solves the problems of passive quality detection, poor data repair effects, high annotation costs, and difficult lineage tracing in existing photovoltaic power prediction data governance, improving data quality and prediction accuracy, and can be widely applied to data governance in the photovoltaic and related energy fields.

[0060] This embodiment also provides a photovoltaic power prediction data governance system based on a data resource pool, including: Knowledge Graph Module: Used to build and maintain spatiotemporal dynamic knowledge graphs, define entity types and relation types, and perform spatiotemporal dynamic modeling; Quality detection module: Based on semantic rules automatically generated from knowledge graphs, it detects data anomalies through graph neural networks; Data Repair Module: Employs physical constraint-based adversarial networks to intelligently repair missing and abnormal data; Self-supervised labeling module: Generates high-quality pseudo-labels by comparing the predictive encoding self-supervised learning framework; Bloodline tracing module: It uses a blockchain-DAG hybrid architecture to store bloodline information and supports bloodline query, verification and impact analysis; Quality assessment module: Based on a six-dimensional data quality assessment model, it calculates a comprehensive quality score and monitors the quality evolution trend in real time; Decision engine module: Employs multi-agent reinforcement learning to automatically select the optimal data repair strategy.

[0061] This invention provides an intelligent governance method and system for photovoltaic power prediction data based on a three-in-one architecture of knowledge graph-generative adversarial network-self-supervised learning (KG-GAN-SSL). Addressing the challenge of quality governance for massive, multi-source, heterogeneous data in photovoltaic data resource pools, it achieves intelligent processing of the entire process of data quality detection, repair, annotation, traceability, and evaluation. Compared to traditional governance methods that rely on manual rule configuration, simple fill-in repair, and manual annotation, this invention offers higher governance accuracy, efficiency, and adaptability, significantly reducing governance costs. It provides high-quality data support for accurate photovoltaic power prediction and optimized power plant operation, and has significant implications for the intelligent development of the photovoltaic industry. This invention first constructs a spatiotemporal dynamic knowledge graph (STDKG) for the photovoltaic field, and automatically learns semantic quality detection rules through graph neural networks to achieve adaptive identification of novel anomalies. Second, it proposes a Physically Constrained Generative Adversarial Network (PC-GAN) to incorporate physical constraints such as energy conservation and ramp rate into the data repair process, ensuring the authenticity and consistency of the repaired data. Third, based on a contrastive predictive coding (CPC) self-supervised framework, it designs five types of pre-training tasks to generate high-quality pseudo-labels, significantly reducing the need for manual annotation. Simultaneously, it constructs a blockchain-DAG hybrid data lineage graph to achieve rapid tracing of data sources and transformation history. A six-dimensional data quality assessment model (6DDQ) is established to achieve comprehensive and objective evaluation and dynamic monitoring of data quality. Finally, a multi-agent reinforcement learning decision engine automatically selects the optimal data repair strategy, balancing repair effectiveness and cost. This invention can achieve efficient governance of photovoltaic data resource pools of different scales (provincial-level massive data / power station-level real-time streaming data). Compared with traditional data governance methods, it achieves breakthrough improvements in anomaly detection accuracy, physical consistency of data repair, annotation cost control, and traceability efficiency, providing a brand-new intelligent solution for photovoltaic industry data governance.

[0062] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A photovoltaic power prediction data governance method based on a data resource pool, characterized in that, Includes the following steps: A spatiotemporal dynamic knowledge graph is constructed based on the photovoltaic data resource pool. The photovoltaic power prediction data is then subjected to quality detection based on the knowledge graph to identify abnormal data. The abnormal data is processed using a generative adversarial network with introduced physical constraints to obtain repaired data; A self-supervised learning method based on contrastive predictive coding is used to process the repaired data to obtain high-quality pseudo-labels; By adopting a hybrid architecture of blockchain and directed acyclic graph, the transformation operations of data entities and metadata are recorded to obtain a traceable data lineage graph; A data quality assessment model covering six dimensions is used, and the repaired and labeled data is processed by combining the high-quality pseudo-labels and the data lineage map to obtain a comprehensive quality score; A multi-agent reinforcement learning method is adopted to process data state and comprehensive quality score from a set of preset repair strategies, and then obtain and execute the optimal repair strategy.

2. The photovoltaic power prediction data governance method based on a data resource pool according to claim 1, characterized in that, The process of constructing a spatiotemporal dynamic knowledge graph based on a photovoltaic data resource pool includes: Based on the data entities in the photovoltaic data resource pool, four entity types are defined: physical entity, data entity, event entity, and rule entity. Based on the relationships between the data entities, three relationship types are defined: static relationship, dynamic relationship, and spatiotemporal relationship. Based on the dynamic relationship, a relationship strength evolution model is used to model the association strength that changes with time and environmental state, resulting in a dynamic relationship network. Based on the entities, relationships, and dynamic relationship networks, the spatiotemporal dynamic knowledge graph is constructed.

3. The photovoltaic power prediction data governance method based on a data resource pool according to claim 2, characterized in that, The process of quality inspection of photovoltaic power prediction data based on the knowledge graph includes: Based on the spatiotemporal dynamic knowledge graph, a graph neural network is used to process the semantic information of entities and relationships and extract implicit semantic rules. Based on the semantic rules and the preset explicit rules, logical reasoning and consistency checks are performed on the photovoltaic power prediction data to identify data that violates the rules; The data that violates the rules is marked as abnormal data.

4. The photovoltaic power prediction data governance method based on a data resource pool according to claim 3, characterized in that, The expression for processing the semantic information of entities and relations using a graph neural network is as follows: ； In the formula, , Let v and u be the feature vectors of nodes v and u in the k-th layer, respectively. Let v be the relationship vector between nodes v and u. Let σ be the weight matrix of the k-th layer, and σ be the activation function. Let v be the feature vector of node v in the kth layer. It is a set of nodes.

5. The photovoltaic power prediction data governance method based on a data resource pool according to claim 1, characterized in that, The process of using a generative adversarial network with introduced physical constraints to process the anomalous data and obtain repaired data includes: A generator incorporating a spatiotemporal attention U-Net architecture is used to generate repair data based on noise and contextual data; A discriminator with a dual discrimination mechanism is used to process the repair data output by the generator, and at the same time evaluate the authenticity and physical consistency of the data. The training process of the generative adversarial network is stabilized using the Wasserstein distance.

6. The photovoltaic power prediction data governance method based on a data resource pool according to claim 5, characterized in that, The physical constraints include: introducing physical constraints based on the law of conservation of energy to check the consistency of the conversion between power data and irradiance data; and introducing physical constraints based on power ramp rate limits to check the rationality of instantaneous changes in power data.

7. The photovoltaic power prediction data governance method based on a data resource pool according to claim 1, characterized in that, The process of processing the repaired data to obtain high-quality pseudo-labels using a self-supervised learning method based on contrastive predictive coding includes: Five types of pre-training tasks are designed: time series prediction, mask reconstruction, time series comparison, transformation invariance, and cluster assignment. The self-supervised learning model is pre-trained to obtain a pre-trained model. The pre-trained model is used to process the repaired data to generate high-quality pseudo-labels.

8. The photovoltaic power prediction data governance method based on a data resource pool according to claim 1, characterized in that, The process of constructing a traceable data lineage graph using a hybrid architecture of blockchain and directed acyclic graph includes: Define nodes in a directed acyclic graph as data entities, define edges in a directed acyclic graph as data transformation operations, and associate them with metadata containing timestamps, operation types, and operation parameters. The key lineage records containing the aforementioned metadata are stored in the blockchain to form an immutable storage chain; Design a bloodline query language and smart contracts to query and verify the bloodline data.

9. The photovoltaic power prediction data governance method based on a data resource pool according to claim 8, characterized in that, Based on the traceable kinship data, a kinship influence analysis is performed. The process of the kinship influence analysis includes: Based on the aforementioned data lineage map, a depth-first search algorithm is used to process the data with quality problems and trace back to locate the root cause of the problem. Based on the aforementioned kinship map, a breadth-first search algorithm is used to process the changed data and propagate backwards to assess the scope of the impact.

10. The photovoltaic power prediction data governance method based on a data resource pool according to claim 1, characterized in that, The data quality assessment model, covering six dimensions, is used to process the repaired and labeled data, specifically including: Data quality is defined using six dimensions: completeness, accuracy, consistency, timeliness, validity, and credibility. The analytic hierarchy process (AHP) is used to assign dynamic weights to each quality dimension. Based on the dynamic weights, the evaluation scores of each dimension are weighted and fused to obtain a comprehensive quality score.

11. The photovoltaic power prediction data governance method based on a data resource pool according to claim 1, characterized in that, The process of using a multi-agent reinforcement learning method includes: deploying dedicated repair strategy agents for irradiance data, power data, temperature data, and wind speed data respectively; deploying a coordinating agent to perform global optimization and conflict resolution on the decisions of each dedicated repair strategy agent; and using the QMIX algorithm to learn the cooperative strategies of each agent.

12. A photovoltaic power prediction data governance system based on a data resource pool, characterized in that, The system for implementing the method of claim 1, wherein the system comprises: The knowledge graph module is used to build and maintain spatiotemporal dynamic knowledge graphs, define entity types and relation types, and perform spatiotemporal dynamic modeling. The quality detection module is used to generate semantic quality detection rules based on the spatiotemporal dynamic knowledge graph, and to use the semantic quality detection rules to detect photovoltaic power prediction data and identify abnormal data. The data repair module is used to repair photovoltaic power prediction data based on the identified abnormal data, using a generative adversarial network that incorporates physical constraints. The self-supervised labeling module is used to train a model based on the repaired large-scale unlabeled photovoltaic data using a contrastive predictive coding self-supervised learning framework to generate pseudo-labels for data quality. The bloodline tracing module is used to construct a traceable data bloodline graph based on data operation records, using a hybrid architecture of blockchain and directed acyclic graph; The quality assessment module is used to generate a comprehensive quality score based on the repaired data using a multi-dimensional quality assessment model that includes completeness, accuracy, consistency, timeliness, effectiveness, and credibility. The decision engine module is used to select and execute the optimal repair strategy from multiple repair strategies based on the data state and the comprehensive quality score, using a multi-agent reinforcement learning decision engine.

Images