Method and device for predicting power generation of newly-built photovoltaic power station
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PETROCHINA SHENZHEN NEW ENERGY RESEARCH INSTITUTE CO LTD
- Filing Date
- 2024-11-20
- Publication Date
- 2026-05-26
Smart Images
Figure CN122092191A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method and apparatus for predicting the power generation of a newly built photovoltaic power plant. Background Technology
[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.
[0003] Most existing photovoltaic power generation prediction methods rely on sufficient historical power data samples for modeling. However, in the early stages of photovoltaic power plant construction or during the upgrade process, there is often a lack of power data, which seriously restricts the accuracy and reliability of the prediction model. Summary of the Invention
[0004] This invention provides a method for predicting the power generation of newly built photovoltaic power plants, enabling high-precision power prediction even when training samples are scarce. The method includes:
[0005] Using photovoltaic power plants as graph nodes, photovoltaic power plant power data as graph node features, and the relationships between photovoltaic power plants as graph adjacency matrices, we construct the target domain graph structure for newly built photovoltaic power plants and the source domain graph structure for historical photovoltaic power plants.
[0006] A graph neural network is used as the first spatial encoder to capture spatial information of the source domain graph structure and the target domain graph structure; the encoding parameters of the first spatial encoder are adjusted using a domain classification model, and the domain-independent features in the spatial information are extracted using the adjusted first spatial encoder; the domain classification model is trained based on the Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during the training process;
[0007] Extracting the temporal characteristics of photovoltaic power plants;
[0008] The target domain graph structure is input into the second spatial encoder, which outputs the target domain spatial features. The second spatial encoder is trained on the graph neural network based on the first preset sample data of the target domain corresponding to the newly built photovoltaic power station.
[0009] The domain-independent features, target domain spatial features, and temporal features are input into the time series regression model, which outputs the predicted power generation of the newly built photovoltaic power station. The time series regression model is trained based on the temporal features, spatial information, and historical power generation values of historical photovoltaic power stations, and then applied to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
[0010] This invention also provides a device for predicting the power generation of a newly built photovoltaic power plant, used to achieve high-precision power prediction of a newly built photovoltaic power plant under the condition of scarce training samples. The device includes:
[0011] The graph structure data construction module is used to construct the target domain graph structure of newly built photovoltaic power plants and the source domain graph structure of historical photovoltaic power plants, with photovoltaic power plants as graph nodes, photovoltaic power plant power data as graph node features, and the relationship between photovoltaic power plants as a graph adjacency matrix.
[0012] The first spatial encoder processing module is used to capture spatial information of the source domain graph structure and the target domain graph structure using a graph neural network as the first spatial encoder; adjust the encoding parameters of the first spatial encoder using a domain classification model; and extract domain-independent features from the spatial information using the adjusted first spatial encoder; the domain classification model is trained based on the Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during the training process.
[0013] The time feature extraction module is used to extract the time features of photovoltaic power plants;
[0014] The second spatial encoder processing module is used to input the target domain graph structure into the second spatial encoder and output the target domain spatial features; the second spatial encoder is trained on the graph neural network based on the first preset sample data of the target domain corresponding to the newly built photovoltaic power station.
[0015] The target domain power generation prediction module is used to input domain-independent features, target domain spatial features and temporal features into a time series regression model, and output the predicted power generation of the newly built photovoltaic power station. The time series regression model is trained based on the temporal features, spatial information and historical power generation values of historical photovoltaic power stations, and then applied to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
[0016] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for predicting the power generation of a newly built photovoltaic power station.
[0017] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the power generation of a newly built photovoltaic power station.
[0018] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for predicting the power generation of a newly built photovoltaic power station.
[0019] In this embodiment of the invention, photovoltaic power plants are used as graph nodes, photovoltaic power plant power data are graph node features, and the relationships between photovoltaic power plants are graph adjacency matrices. A target domain graph structure for newly built photovoltaic power plants and a source domain graph structure for historical photovoltaic power plants are constructed. A graph neural network is used as a first spatial encoder to capture the spatial information of the source and target domain graph structures. The encoding parameters of the first spatial encoder are adjusted using a domain classification model, and the adjusted first spatial encoder extracts domain-independent features from the spatial information. The domain classification model is trained based on a Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during training. Temporal features of photovoltaic power plants are extracted. The target domain graph structure is input into a second spatial encoder, which outputs the target domain spatial features. The second spatial encoder is trained on a graph neural network based on first preset sample data of the target domain corresponding to the newly built photovoltaic power plant. The domain-independent features, target domain spatial features, and temporal features are input into a time-series regression model, which outputs the predicted power generation of the newly built photovoltaic power plant. The time-series regression model is trained based on the temporal features, spatial information, and historical power generation values of historical photovoltaic power plants, and then applied to predict the power generation of photovoltaic power plants in the target domain corresponding to the newly built photovoltaic power plant. Compared with existing technologies that rely on scarce historical power data samples and model based on small samples of transfer learning, this method effectively utilizes known data from photovoltaic power plants that have been operating for a long time, and combines graph neural networks and multi-source domain knowledge transfer technology to achieve high-precision power prediction for newly built photovoltaic power plants under the condition of scarce training samples. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be 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. In the drawings:
[0021] Figure 1 This is a flowchart illustrating the power generation prediction method for newly built photovoltaic power plants in this embodiment of the invention.
[0022] Figure 2 This is a graph showing the prediction results of the power generation prediction method for newly built photovoltaic power plants in an embodiment of the present invention;
[0023] Figure 3 This is a schematic diagram of a newly built photovoltaic power plant power generation prediction device in an embodiment of the present invention;
[0024] Figure 4 This is a schematic diagram of another newly built photovoltaic power plant power generation prediction device in an embodiment of the present invention;
[0025] Figure 5 This is a schematic diagram of the computer device structure according to an embodiment of the present invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.
[0027] First, the relevant technical terms in the embodiments of this invention will be introduced:
[0028] Graph Neural Networks (GNNs) are a general term for algorithms that use neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation.
[0029] Domain-Adversarial Training of Neural Networks (DANN) is a transfer learning method in the field of deep learning that combines the idea of adversarial training. It aims to improve the model's generalization ability on the target domain by minimizing the feature distribution difference between the source and target domains.
[0030] To achieve high-precision power prediction for newly built photovoltaic power plants under conditions of scarce training samples, this invention provides a method for predicting the power generation of newly built photovoltaic power plants. This method effectively utilizes historical data from photovoltaic power plants that have been operating for a long time, combined with graph neural networks and multi-source domain knowledge transfer technology, to achieve high-precision power prediction for newly built photovoltaic power plants under conditions of scarce training samples. Figure 1 This is a flowchart illustrating the power generation prediction method for newly built photovoltaic power plants in this embodiment of the invention. Figure 1 As shown, the method may include:
[0031] Step 101: Using photovoltaic power plants as graph nodes, photovoltaic power plant power data as graph node features, and the relationship between photovoltaic power plants as a graph adjacency matrix, construct the target domain graph structure for newly built photovoltaic power plants and the source domain graph structure for historical photovoltaic power plants;
[0032] Step 102: Using a graph neural network as the first spatial encoder, capture the spatial information of the source domain graph structure and the target domain graph structure; adjust the encoding parameters of the first spatial encoder using a domain classification model, and extract domain-independent features from the spatial information using the adjusted first spatial encoder; the domain classification model is trained based on the Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during the training process;
[0033] Step 103: Extract the temporal characteristics of the photovoltaic power station;
[0034] Step 104: Input the target domain graph structure into the second spatial encoder and output the target domain spatial features; the second spatial encoder is trained on the graph neural network based on the first preset sample data of the target domain corresponding to the newly built photovoltaic power station.
[0035] Step 105: Input the domain-independent features, target domain spatial features, and temporal features into the time series regression model, and output the predicted power generation of the newly built photovoltaic power station; the time series regression model is trained based on the temporal features, spatial information, and historical power generation values of historical photovoltaic power stations, and then applied to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
[0036] In this embodiment of the invention, photovoltaic power plants are used as graph nodes, photovoltaic power plant power data are used as graph node features, and the relationships between photovoltaic power plants are used as graph adjacency matrices to construct the target domain graph structure for newly built photovoltaic power plants and the source domain graph structure for historical photovoltaic power plants. Photovoltaic power plants and their interrelationships are modeled as graph structures to capture the spatial relationships between photovoltaic power plants using graph neural networks.
[0037] In this embodiment of the invention, a graph neural network is used as the first spatial encoder to capture the spatial information of the source domain graph structure and the target domain graph structure. The encoding parameters of the first spatial encoder are adjusted using a domain classification model, and the domain-independent features in the spatial information are extracted using the adjusted first spatial encoder. The domain classification model is trained based on the DANN model, and the loss of the domain classification model is maximized during the training process.
[0038] In one embodiment, using a graph neural network as the first spatial encoder to capture spatial information of the source domain graph structure and the target domain graph structure may include: encoding the source domain graph structure and the target domain graph structure into spatial graph embedding features using the graph neural network as the first spatial encoder, and capturing the spatial information of the source domain graph structure and the target domain graph structure based on the spatial graph embedding features. Using a graph neural network as the first spatial encoder can better capture the topological structure information in graph data, thereby improving the expressive power and prediction accuracy of subsequent domain classification models.
[0039] In one embodiment, the graph structure data of the source domain and the target domain are encoded as spatial graph embedding features. These features should capture the spatial information of the input graph structure data as much as possible. The spatial information can reflect the correlation between photovoltaic power plants and may include irradiance, temperature, humidity, wind speed, etc.
[0040] In this embodiment of the invention, in order to effectively utilize the first spatial encoder to extract domain-independent features from spatial information, remove domain-related features, and retain features unrelated to a specific domain, the encoding parameters of the first spatial encoder can be adjusted using a domain classification model. The adjusted first spatial encoder is then used to extract domain-independent features from the spatial information. The domain classification model is trained based on a Domain Adversarial Neural Network (DANN) model, and its loss is maximized during training. Adjusting the encoding parameters of the first spatial encoder using the domain classification model can include: using spatial graph embedding features and corresponding domain type labels as samples, and training the domain classification model based on the DANN model. During training, by maximizing the loss of the domain classification model, the domain classification model updates the spatial graph embedding features, and the updated spatial graph embedding features are provided to the first spatial encoder to adjust its encoding parameters.
[0041] In one embodiment, maximizing the domain classification model loss during training, causing the domain classification model to update spatial graph embedding features, and providing the updated spatial graph embedding features to the first spatial encoder to adjust its encoding parameters can include: maximizing the domain classification model loss during training, causing the domain classification model to update spatial graph embedding features through gradient inversion, and backpropagating the updated spatial graph embedding features to the first spatial encoder to adjust its encoding parameters. During training, the parameters of both the first spatial encoder and the domain classification model are optimized simultaneously. However, by maximizing the domain classification model loss, an adversarial relationship exists between the first spatial encoder and the domain classification model. The optimization goal of the first spatial encoder is to maximize the loss of the domain classification model, while the optimization goal of the domain classification model is to minimize its own loss. This adversarial training process allows the first spatial encoder to gradually learn features that are difficult to distinguish between different domains, achieving domain adaptation.
[0042] Unlike existing methods for transferring data from source objects to target objects, this invention combines graph neural networks and domain adversarial networks to extract domain-independent features from spatial information, taking into account the significant differences in the spatiotemporal graph structures of photovoltaic systems in different regions and the spatiotemporal correlation of photovoltaic power plants. This enables the transfer of graph features containing generalized prediction knowledge between different photovoltaic graph structures.
[0043] In this embodiment of the invention, extracting the temporal features of a photovoltaic power station includes: using a Long Short-Term Memory (LSTM) neural network to extract the temporal features of the photovoltaic power station. This is because the LSTM neural network is a special type of recurrent neural network capable of capturing long-term dependencies in time series. In the extraction of the temporal features of a photovoltaic power station, the LSTM neural network can learn the patterns of power generation over time, including periodicity and trends, thus enabling accurate prediction of photovoltaic power. In addition, convolutional neural networks or bidirectional LSTM neural networks can also be used to extract the spatial features of the photovoltaic power station.
[0044] In one embodiment, time features may include timestamps, seasons, weather conditions, etc.
[0045] In this embodiment of the invention, the target domain graph structure is input into a second spatial encoder, which outputs the target domain spatial features. The second spatial encoder is trained on a graph neural network based on a first preset sample data of the target domain corresponding to the newly built photovoltaic power station. The second spatial encoder is trained based on a limited set of samples from the target domain, and the first preset sample data can be the target domain graph structure data and spatial features modeled by the newly built photovoltaic power station.
[0046] In this embodiment of the invention, domain-independent features, target domain spatial features, and temporal features are input into a time-series regression model, which outputs the predicted power generation of newly built photovoltaic power plants. The time-series regression model is trained based on the temporal features, spatial information, and historical power generation values of historical photovoltaic power plants, and then applied to predict the power generation of photovoltaic power plants in the target domain corresponding to newly built photovoltaic power plants. Through multiple iterations of training and parameter adjustment, the model can learn general feature representations of the data. These representations not only capture the underlying structure of the data but can also be shared across different domains, thereby improving the model's generalization ability and prediction accuracy. For example, Figure 2 This is a graph showing the prediction results of the power generation prediction method for newly built photovoltaic power plants in an embodiment of the present invention.
[0047] In this embodiment of the invention, before inputting the domain-independent features, target domain spatial features, and temporal features into the time-series regression model, the method further includes: fusing the temporal features and spatial information of historical photovoltaic power plants; using the fused temporal features and spatial information, as well as the historical power generation values of historical photovoltaic power plants, to train a first time-series regression model for the source domain corresponding to the historical photovoltaic power plants; constructing a second time-series regression model for the target domain corresponding to the newly built photovoltaic power plant based on the first time-series regression model; the model architecture of the first time-series regression model and the second time-series regression model is the same; the second time-series regression model is used to predict the power generation of the photovoltaic power plants in the target domain corresponding to the newly built photovoltaic power plants; inputting the domain-independent features, target domain spatial features, and temporal features into the time-series regression model includes: inputting the domain-independent features, target domain spatial features, and temporal features into the second time-series regression model.
[0048] For example, time features such as timestamps, seasons, and weather conditions of photovoltaic power plants are extracted and fused with spatial information to form a fused feature set. Simultaneously, historical power generation values of photovoltaic power plants are prepared as true labels. The fused feature set is used as input to the time-series regression model. The model calculates prediction results based on the input feature set and current parameters, obtains the true power generation values corresponding to the input feature set from historical data, uses a loss function to calculate the difference between the model's prediction results and the true labels, calculates the gradient of the model parameters based on the gradient information of the loss function, and uses an optimization algorithm to update the model parameters based on the gradient information to reduce losses and improve model prediction performance.
[0049] In one embodiment, after training a first time-series regression model, a second time-series regression model is constructed based on the first time-series regression model for the target domain corresponding to the newly built photovoltaic power station. The model architecture of the first time-series regression model and the second time-series regression model is the same. The second time-series regression model is used to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
[0050] For example, the model parameters of the second time series regression model are initialized using the model parameters of the first time series regression model. The data features of the first time series regression model are then transferred to the second time series regression model, realizing knowledge transfer from the source domain to the target domain. Through transfer learning, the second time series regression model can learn more general feature representations, which are useful for different tasks. Therefore, the second time series regression model can exhibit stronger generalization ability when applied to new datasets.
[0051] To further utilize the knowledge learned from the source domain, this embodiment of the invention uses second preset sample data corresponding to the target domain of the newly built photovoltaic power station to optimize the model parameters after initialization of the second time-series regression model. The second preset sample data consists of a small amount of power generation data from the target domain. Unlike existing fine-tuning methods that directly "concatenate" source and target domain data for model training, this embodiment considers the differences in feature distribution between the source and target domains, enabling the transfer of graph features containing generalized prediction knowledge between different photovoltaic graph structures. This is because power generation data may be affected by various factors such as weather conditions and equipment status, which may have unique distributions and characteristics in the target domain. Therefore, fine-tuning the model using target domain data allows the model to better adapt to changes in the target domain data, thereby improving model stability.
[0052] In this embodiment of the invention, after outputting the predicted power generation of the newly built photovoltaic power station, the model parameters of the second time series regression model can be optimized by using the actual power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
[0053] In summary, the power generation prediction method for newly built photovoltaic power plants in this embodiment of the invention captures spatial information by leveraging the advantages of graph neural networks in modeling inter-photovoltaic correlations, and combines it with domain adversarial algorithms to train the model. This effectively achieves simultaneous modeling of spatiotemporal correlations and domain-invariant features. By integrating multiple modules such as spatial coding, domain classification, and temporal regression, this embodiment of the invention constructs a complete prediction framework that can maintain high power prediction accuracy even in scenarios with scarce training samples.
[0054] This invention also provides a photovoltaic power plant power generation prediction device, as described in the following embodiments. Since the principle behind this device's solution is similar to that of the photovoltaic power plant power generation prediction method, its implementation can refer to the implementation of the photovoltaic power plant power generation prediction method; repeated details will not be elaborated further.
[0055] Figure 3 This is a schematic diagram of a photovoltaic power generation prediction device in an embodiment of the present invention, such as... Figure 3 As shown, the device includes:
[0056] The graph structure data construction module 301 is used to construct the target domain graph structure of newly built photovoltaic power plants and the source domain graph structure of historical photovoltaic power plants, with photovoltaic power plants as graph nodes, photovoltaic power plant power data as graph node features, and the relationship between photovoltaic power plants as a graph adjacency matrix.
[0057] The first spatial encoder processing module 302 is used to capture spatial information of the source domain graph structure and the target domain graph structure using a graph neural network as the first spatial encoder; adjust the encoding parameters of the first spatial encoder using a domain classification model; and extract domain-independent features from the spatial information using the adjusted first spatial encoder; the domain classification model is trained based on the Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during the training process.
[0058] The time feature extraction module 303 is used to extract the time features of the photovoltaic power station;
[0059] The second spatial encoder processing module 304 is used to input the target domain graph structure into the second spatial encoder and output the target domain spatial features; the second spatial encoder is trained on the graph neural network based on the first preset sample data of the target domain corresponding to the newly built photovoltaic power station.
[0060] The target domain power generation prediction module 305 is used to input domain-independent features, target domain spatial features and temporal features into a time series regression model and output the predicted power generation of the newly built photovoltaic power station. The time series regression model is trained based on the temporal features, spatial information and historical power generation values of historical photovoltaic power stations, and then applied to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
[0061] In one embodiment, the first spatial encoder processing module 302 is specifically used for:
[0062] Using a graph neural network as the first spatial encoder, the source domain graph structure and the target domain graph structure are encoded into spatial graph embedding features, and the spatial information of the source domain graph structure and the target domain graph structure is captured based on the spatial graph embedding features.
[0063] Using spatial graph embedding features and corresponding domain type labels as samples, a domain classification model is trained based on a DANN model. During training, the loss of the domain classification model is maximized to update the spatial graph embedding features. The updated spatial graph embedding features are then provided to the first spatial encoder to adjust its encoding parameters.
[0064] In one embodiment, the first spatial encoder processing module 302 is specifically used for:
[0065] During training, the loss of the domain classification model is maximized, and the domain classification model updates the spatial graph embedding features through gradient reversal. The updated spatial graph embedding features are then backpropagated to the first spatial encoder so that the first spatial encoder can adjust its encoding parameters.
[0066] In one embodiment, the time feature extraction module 303 is specifically used for:
[0067] Long Short-Term Memory (LSTM) neural networks are used to extract the temporal characteristics of photovoltaic power plants.
[0068] Figure 4 This is a structural block diagram of another newly built photovoltaic power plant power generation prediction device in an embodiment of the present invention, such as... Figure 4 As shown, in one embodiment, Figure 4 The newly constructed photovoltaic power plant power generation prediction device shown may also include:
[0069] First time series regression model training module 401: Before inputting domain-independent features, target domain spatial features and time features into the time series regression model, it fuses the time features and spatial information of historical photovoltaic power plants, and uses the fused time features and spatial information, as well as the historical power generation value of historical photovoltaic power plants, to train the first time series regression model of the source domain corresponding to the historical photovoltaic power plants.
[0070] Second time-series regression model construction module 402: used to construct a second time-series regression model for the target domain corresponding to the newly built photovoltaic power station based on the first time-series regression model; the model architecture of the first time-series regression model and the second time-series regression model is the same; the second time-series regression model is used to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station;
[0071] The target domain power generation prediction module 305 is specifically used to input the domain-independent features, target domain spatial features and time features into the second time series regression model.
[0072] In one embodiment, the second time-series regression model construction module 402 is specifically used for:
[0073] The model parameters of the second time series regression model are initialized using the model parameters of the first time series regression model, and the data features of the first time series regression model are transferred to the second time series regression model.
[0074] In one embodiment, the second time-series regression model construction module 402 is further configured to:
[0075] The model parameters of the second time-series regression model are optimized by using the second preset sample data of the target domain corresponding to the newly built photovoltaic power station.
[0076] In one embodiment, the second time-series regression model construction module 402 is further configured to:
[0077] The model parameters of the second time-series regression model are optimized by using the actual photovoltaic power generation of the target domain corresponding to the newly built photovoltaic power station.
[0078] Based on the aforementioned inventive concept, such as Figure 5As shown, the present invention also proposes a computer device 500, including a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 530, it implements the above-mentioned method for predicting the power generation of newly built photovoltaic power plants.
[0079] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the power generation of a newly built photovoltaic power station.
[0080] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for predicting the power generation of a newly built photovoltaic power station.
[0081] In summary, in this embodiment of the invention, a target domain graph structure for a newly built photovoltaic power station and a source domain graph structure for historical photovoltaic power stations are constructed by using photovoltaic power stations as graph nodes, photovoltaic power station power data as graph node features, and the relationship between photovoltaic power stations as a graph adjacency matrix. A graph neural network is used as the first spatial encoder to capture the spatial information of the source domain graph structure and the target domain graph structure. The encoding parameters of the first spatial encoder are adjusted using a domain classification model, and the domain-independent features in the spatial information are extracted using the adjusted first spatial encoder. The domain classification model is trained based on a Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during the training process. Temporal features of photovoltaic power stations are extracted. The target domain graph structure is input into a second spatial encoder, and the target domain spatial features are output. The second spatial encoder is trained on a graph neural network based on the first preset sample data of the target domain corresponding to the newly built photovoltaic power station. The domain-independent features, target domain spatial features, and temporal features are input into a time-series regression model, and the predicted power generation of the newly built photovoltaic power station is output. The time-series regression model is trained based on the temporal features, spatial information, and historical power generation values of historical photovoltaic power stations, and is applied to predict the power generation of photovoltaic power stations in the target domain corresponding to the newly built photovoltaic power station. Compared with existing technologies that rely on scarce historical power data samples and model based on small samples of transfer learning, this method effectively utilizes historical data from photovoltaic power plants that have been operating for a long time, and combines graph neural networks and multi-source domain knowledge transfer technology to achieve high-precision power prediction for newly built photovoltaic power plants under the condition of scarce training samples.
[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0086] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting the power generation of a newly built photovoltaic power station, characterized in that, include: Using photovoltaic power plants as graph nodes, photovoltaic power plant power data as graph node features, and the relationships between photovoltaic power plants as graph adjacency matrices, we construct the target domain graph structure for newly built photovoltaic power plants and the source domain graph structure for historical photovoltaic power plants. Using a graph neural network as the first spatial encoder, spatial information of the source domain graph structure and the target domain graph structure is captured; The encoding parameters of the first spatial encoder are adjusted using a domain classification model, and the adjusted first spatial encoder is used to extract domain-independent features from the spatial information. The domain classification model is trained based on the Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during the training process. Extracting the temporal characteristics of photovoltaic power plants; The target domain graph structure is input into the second spatial encoder, which outputs the target domain spatial features. The second spatial encoder is obtained by training a graph neural network based on the first preset sample data of the target domain corresponding to the newly built photovoltaic power station. The domain-independent features, target domain spatial features, and temporal features are input into the time series regression model, which outputs the predicted power generation of the newly built photovoltaic power station. The time series regression model is trained based on the temporal features, spatial information, and historical power generation values of historical photovoltaic power stations, and then applied to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
2. The method as described in claim 1, characterized in that, Using a graph neural network as the first spatial encoder, spatial information of the source domain graph structure and the target domain graph structure is captured, including: Using a graph neural network as the first spatial encoder, the source domain graph structure and the target domain graph structure are encoded into spatial graph embedding features, and the spatial information of the source domain graph structure and the target domain graph structure is captured based on the spatial graph embedding features. The encoding parameters of the first spatial encoder are adjusted using a domain classification model, including: Using spatial graph embedding features and corresponding domain type labels as samples, a domain classification model is trained based on a DANN model. During training, the loss of the domain classification model is maximized to update the spatial graph embedding features. The updated spatial graph embedding features are then provided to the first spatial encoder to adjust its encoding parameters.
3. The method as described in claim 2, characterized in that, During training, the domain classification model loss is maximized, causing the domain classification model to update the spatial graph embedding features. These updated spatial graph embedding features are then provided to the first spatial encoder, enabling the first spatial encoder to adjust its encoding parameters, including: During training, the loss of the domain classification model is maximized, and the domain classification model updates the spatial graph embedding features through gradient reversal. The updated spatial graph embedding features are then backpropagated to the first spatial encoder so that the first spatial encoder can adjust its encoding parameters.
4. The method as described in claim 1, characterized in that, Extracting the temporal characteristics of photovoltaic power plants, including: Long Short-Term Memory (LSTM) neural networks are used to extract the temporal characteristics of photovoltaic power plants.
5. The method as described in claim 1, characterized in that, Before inputting domain-independent features, target domain spatial features, and temporal features into the time series regression model, the following steps are also included: The temporal and spatial characteristics of historical photovoltaic power plants are fused together. Using the fused temporal and spatial characteristics, as well as the historical power generation values of historical photovoltaic power plants, the first time series regression model of the source domain corresponding to the historical photovoltaic power plants is trained. A second time-series regression model is constructed based on the first time-series regression model for the target domain corresponding to the newly built photovoltaic power plant. The model architecture of the first time-series regression model and the second time-series regression model is the same. The second time-series regression model is used to predict the power generation of the photovoltaic power plant in the target domain corresponding to the newly built photovoltaic power plant. Domain-independent features, target domain spatial features, and temporal features are input into the time series regression model, including: The domain-independent features, target domain spatial features, and temporal features are input into the second time-series regression model.
6. The method as described in claim 5, characterized in that, Based on the first time-series regression model, a second time-series regression model is constructed for the target domain corresponding to the newly built photovoltaic power station, including: The model parameters of the second time series regression model are initialized using the model parameters of the first time series regression model, and the data features of the first time series regression model are transferred to the second time series regression model.
7. The method as described in claim 6, characterized in that, The second time-series regression model for the target domain corresponding to the newly built photovoltaic power plant, based on the first time-series regression model, also includes: The model parameters of the second time-series regression model are optimized by using the second preset sample data of the target domain corresponding to the newly built photovoltaic power station.
8. The method as described in claim 7, characterized in that, After outputting the predicted power generation of newly built photovoltaic power plants, it also includes: The model parameters of the second time-series regression model are optimized by using the actual photovoltaic power generation of the target domain corresponding to the newly built photovoltaic power station.
9. A device for predicting the power generation of a newly built photovoltaic power station, characterized in that, include: The graph structure data construction module is used to construct the target domain graph structure of newly built photovoltaic power plants and the source domain graph structure of historical photovoltaic power plants, with photovoltaic power plants as graph nodes, photovoltaic power plant power data as graph node features, and the relationship between photovoltaic power plants as a graph adjacency matrix. The first spatial encoder processing module is used to capture spatial information of the source domain graph structure and the target domain graph structure using a graph neural network as the first spatial encoder. The encoding parameters of the first spatial encoder are adjusted using a domain classification model, and the adjusted first spatial encoder is used to extract domain-independent features from the spatial information. The domain classification model is trained based on the Domain Adversarial Neural Network (DANN) model, and the loss of the domain classification model is maximized during the training process. The time feature extraction module is used to extract the time features of photovoltaic power plants; The second spatial encoder processing module is used to input the target domain graph structure into the second spatial encoder and output the target domain spatial features. The second spatial encoder is obtained by training a graph neural network based on the first preset sample data of the target domain corresponding to the newly built photovoltaic power station. The target domain power generation prediction module is used to input domain-independent features, target domain spatial features and temporal features into a time series regression model, and output the predicted power generation of the newly built photovoltaic power station. The time series regression model is trained based on the temporal features, spatial information and historical power generation values of historical photovoltaic power stations, and then applied to predict the power generation of the photovoltaic power station in the target domain corresponding to the newly built photovoltaic power station.
10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.
11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.
12. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.