Photovoltaic base guarantee power prediction method based on hierarchical graph neural network and joint loss function

Abstract

The invention discloses a photovoltaic base guarantee power prediction method based on a hierarchical graph neural network and a joint loss function, and relates to the technical field of new energy power generation prediction. The method comprises the following steps: firstly, acquiring historical actually measured irradiance, predicted irradiance and geographical location information of each power station in the photovoltaic base, and dividing the photovoltaic base into a plurality of sub-regions by adopting a clustering method after normalization processing and feature construction so as to realize refined expression of a complex space structure. And then constructing a local adjacency matrix and a global adjacency matrix based on correlation between the interior of the sub-region and the regions, and constructing a hierarchical graph neural network comprising a local graph convolutional layer and a global graph convolutional layer to realize joint modeling of a local space-time relationship and a global evolution rule of photovoltaic output. And further introducing a joint loss function, performing end-to-end training on the model by using the divided training set, and obtaining a next-day guarantee power prediction result of the photovoltaic base on a test set.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction technology, and more specifically, to a method for predicting the guaranteed power of photovoltaic bases based on hierarchical graph neural networks and joint loss functions. Background Technology

[0002] Photovoltaic power generation is significantly affected by weather events, exhibiting strong spatiotemporal fluctuations and uncertainties. In large-scale centralized grid-connected scenarios, especially under meteorological conditions such as sudden cloud formations, sandstorms, or prolonged overcast skies, the overall output of photovoltaic systems may experience a rapid decline or prolonged periods of low power, leading to a sharp reduction in power generation and consequently triggering operational risks such as grid supply-demand imbalances, voltage fluctuations, and frequency exceedances. Therefore, there is an urgent need for a reliable power prediction method that can accurately characterize the power supply reliability of photovoltaic bases under different operating conditions, providing a forward-looking decision-making basis for dispatching operations and supporting the safe and efficient grid connection of large-scale photovoltaic bases.

[0003] Currently, existing research mainly focuses on optimizing the accuracy of point-to-point power prediction or interval prediction based on statistical methods. Research on guaranteed power prediction for large-scale photovoltaic (PV) bases is still lacking. Furthermore, current models primarily model PV power plants. When facing large-scale PV bases with complex terrain and multiple coupled meteorological sub-regions, traditional models struggle to capture the power output evolution patterns at different spatiotemporal scales, making it difficult for prediction results to support power balance decisions in day-ahead dispatching.

[0004] Therefore, there is an urgent need for a photovoltaic power prediction method based on hierarchical graph neural networks and joint loss functions to solve these problems. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problems mentioned in the background section, and to provide a photovoltaic power prediction method based on hierarchical graph neural networks and joint loss functions, comprising the following steps: Step S10: Based on the distribution of irradiance resources and geographical features, the photovoltaic base is divided into multiple sub-regions to achieve spatial fine-grained modeling of the photovoltaic power generation system under complex geographical conditions. Step S20: Construct a hierarchical graph neural network to mine the spatiotemporal correlation characteristics of photovoltaic power at the "local-global" scale, and capture the correlation and evolution patterns within and between regions. Step S30: Construct a joint loss function that comprehensively considers prediction accuracy and reliability, balancing the power concentration of power supply guarantee with the success rate of guarantee. Step S40: Train the model based on the training set and test and verify the model performance on the test set to obtain the next day's guaranteed power prediction result for the photovoltaic base.

[0006] As a preferred technical solution of the present invention, step S10 includes the following sub-steps: Step S101: Obtain historical measured irradiance, predicted irradiance, latitude, longitude and altitude information for each power station in the photovoltaic base; Step S102: Normalize the historical irradiance sequence, predicted irradiance sequence, latitude, longitude and altitude information, and construct a comprehensive feature vector based on the normalized features; Step S103: The K-Means algorithm is used to cluster the comprehensive feature vectors, and power plants with similar irradiance resources and geographical environment are divided into the same sub-region.

[0007] As a preferred technical solution of the present invention, step S20 includes the following sub-steps: Step S201: Calculate the correlation between any two stations in the same region based on the historical power sequence of the photovoltaic power station, and construct the sub-region adjacency matrix accordingly.

[0008] As a preferred technical solution of the present invention, step S20 further includes: A global adjacency matrix is ​​constructed based on the correlation between power stations across regions to describe the global relational structure within the entire photovoltaic base.

[0009] As a preferred technical solution of the present invention, the hierarchical graph neural network in step S20 includes: Local graph convolution is driven by the sub-region adjacency matrix to extract spatiotemporal correlation features within the region; By using a global adjacency matrix to drive global graph convolution to fuse cross-regional correlation information, we can model the overall evolution of photovoltaic bases.

[0010] As a preferred technical solution of the present invention, step S30 includes: Step S301: Introduce a mean square error term to constrain the overall prediction accuracy; Step S302: Construct a guarantee index based on confidence interval coverage and confidence interval width, and combine it with mean square error to form a joint loss function, and perform multi-objective collaborative optimization using a weighted approach.

[0011] As a preferred technical solution of the present invention, step S40 includes the following sub-steps: Step S401, construct a complete dataset DataSet based on the photovoltaic predicted irradiance set, the total power set of photovoltaic base and the historical output set of each power station at 35040 time points throughout the year, and divide the training set and test set in chronological order.

[0012] As a preferred technical solution of the present invention, step S40 includes the following sub-steps: Step S402: Using the historical predicted irradiance set and historical power output set in the training set as model inputs and the historical total photovoltaic power output as model outputs, the hierarchical graph neural network is trained end-to-end, and the model parameters are updated through the backpropagation algorithm.

[0013] As a preferred technical solution of the present invention, step S40 further includes: step S403, inputting the predicted irradiance set and the power output set of the test set into the trained graph neural network to obtain the photovoltaic guaranteed power prediction result for the next day.

[0014] As a preferred technical solution of the present invention, step S40 further includes: comparing the predicted guaranteed power sequence with the actual photovoltaic output in the test set, and verifying the accuracy and guarantee capability of the model with deviation index and confidence interval index.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes the historical irradiance sequences, predicted irradiance sequences, and geographical location information of each power station within a photovoltaic base. Through comprehensive feature construction and regional clustering in steps S10–S103, it effectively expresses the complex spatial structure of the photovoltaic base. This allows the inherent patterns of different power stations in terms of spatial resource conditions, irradiance differences, and geographical factors to be described in a unified feature space. This avoids the model fitting difficulties caused by geographical heterogeneity in traditional methods and improves the rationality and stability of data grouping in subsequent modeling processes.

[0016] This invention utilizes a hierarchical graph neural network constructed in steps S20–S202 to simultaneously incorporate the local adjacency structure within sub-regions and the global adjacency structure between regions into the modeling process. This enables the model to uncover the spatiotemporal relationships between photovoltaic (PV) power plants at both the local and global scales. Local graph convolution effectively captures the synchronous changes exhibited by power plants within the same region under similar weather, terrain, and resource conditions, while global graph convolution further extracts the collaborative evolution patterns across spatial ranges between regions, thus achieving a more comprehensive description of the overall power output behavior of the PV base. This structure overcomes the difficulty of traditional single-layer graph neural networks in handling large-scale bases and wide regional spans, improving the model's adaptability to complex spatial structures.

[0017] This invention incorporates both prediction accuracy and reliability metrics into the optimization objective through a joint loss function constructed in steps S30–S302. Mean squared error ensures the model's fitting ability at the overall prediction level, while confidence interval coverage and interval width characterize the model's stability and reliability in terms of reliable output. A weighted approach achieves multi-objective collaborative optimization, enabling the model to maintain overall prediction accuracy while focusing more on risk mitigation capabilities under low-output scenarios. This mechanism effectively improves the photovoltaic backup power prediction's ability to cope with abnormal weather conditions, providing a more reliable and secure decision-making basis for actual power dispatch.

[0018] In summary, the region partitioning method, hierarchical graph neural network structure, and joint loss function design proposed in this invention work together to enable the model to establish complex spatial relationships between photovoltaic power plants while balancing prediction accuracy and assurance capabilities. This makes it more applicable to prediction tasks at the photovoltaic base level and suitable for assurance analysis and scheduling optimization scenarios for large-scale photovoltaic clusters. Attached Figure Description

[0019] Figure 1 This is a flowchart of a photovoltaic power prediction method based on hierarchical graph neural networks and joint loss functions. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with embodiments and appendices. Figure 1 The present invention will be further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited to these embodiments. Equivalent modifications made by those skilled in the art without departing from the principles of the present invention should fall within the protection scope of the present invention.

[0022] This embodiment takes a photovoltaic base as an example, which contains n photovoltaic power stations (site numbered as follows). Each power station has a year-round (365 days) historical and forecast record of irradiance with a resolution of 15 minutes, containing 96 time points per day and 35,040 time points per year. Therefore, each power station has an annual irradiance time series of 35,040 points.

[0023] In step S10, step S101 is first executed to obtain the irradiance resources of each power station location within the photovoltaic base, including the measured irradiance sequence collected every 15 minutes throughout the year for each power station. ,in Each measurement corresponds to a time point every 15 minutes, totaling 35,040 time points, covering 365 days. Simultaneously, predicted irradiance sequences for each power station at the same temporal resolution were acquired. ,in Furthermore, the longitude of each power station's location was obtained. ,latitude and altitude Among them, latitude and longitude information is used to reflect geographic spatial location, and altitude information is used to reflect the impact of changes in air quality, atmospheric pressure and light conditions on irradiance.

[0024] Subsequently, in step S102, to eliminate the dimensional differences between different power stations, the irradiance time series and geographical location information are normalized respectively. First, the measured irradiance is normalized. Compared with predicted irradiance Maximum value normalization method is used: The normalized irradiance values ​​are within the range [0,1], avoiding bias in subsequent feature modeling caused by differences in absolute irradiance intensity. Then, latitude, longitude, and altitude are normalized using constants: Where 8848 represents the height of Mount Everest, used as the upper bound for altitude normalization. After completing the above normalization process, a comprehensive feature vector is constructed in step S103 to characterize the solar resource characteristics and spatial differences of each power station: In the formula, They are numbered The maximum, mean, and variance of the normalized irradiance of the photovoltaic sites. It is a constructed comprehensive feature vector.

[0025] Subsequently, the number of clusters was set to 4, and... Four samples were randomly selected as cluster centers. The photovoltaic power stations were divided into four sub-regions using the K-Means clustering algorithm, and the clustering expression is as follows: In the formula, Indicates the sub-region number, This indicates the sub-region number to which the power station may belong. Indicates sample It may belong to the cluster center. Indicates belonging to a subregion Clusters, samples belong to This means that the power station and Cluster Center It is more similar to other cluster centers and is therefore assigned to a subregion. Within the clusters below, the algorithm iteratively adjusts the cluster centers to ensure that all samples are at Euclidean distances from their respective cluster centers. The minimum size is used to divide all power stations into four sub-regions.

[0026] After classification, the photovoltaic base is divided into multiple sub-regions with similar light resources and geographical environment characteristics. In step S20, a hierarchical graph neural network is constructed to extract local correlations within the regions and global correlations across the entire base. First, step S201 is executed to obtain normalized historical photovoltaic power data for each power station. Based on the output correlation, an adjacency matrix is ​​constructed between power plants within a sub-region and between sub-regions.

[0027] according to To characterize the similarity between the normalized historical output of various power plants, the formula is used, where if the power plant Power station j belongs to the same sub-region , then is used to represent both Correlation between them: power plants and power station The average photovoltaic output is respectively and ;in and Representing power stations and power station The historical power output at time t. If they belong to different sub-regions, the correlation is 0. Then, a first-level adjacency matrix is ​​constructed based on the power output correlation between power stations. Second-level adjacency matrix In the formula This indicates the number of power stations within the photovoltaic base.

[0028] Step S202: Design a two-layer graph neural network: the bottom layer uses a graph convolutional network to capture the spatiotemporal relationship between power stations within a sub-region, and the upper layer aggregates global features through graph pooling to mine information on the "local-global" co-evolutionary pattern.

[0029] In the formula, This represents the extracted local spatiotemporal correlation information. This represents the set of predicted irradiance for all power stations within the photovoltaic base for the next day. The convolution kernel parameter matrix for extracting local information. Let be the degree matrix of the nodes. It is a first-order adjacency matrix. It is an identity matrix.

[0030] Based on this, we will continue to analyze global spatiotemporal correlation information. Extract information: In the formula, This represents the convolution kernel parameter matrix for extracting global information.

[0031] In the formula, This represents element-wise multiplication. For the corresponding parameters; The forecast results are to ensure the photovoltaic power output for the next day.

[0032] Further, step S30 includes: Step S301, introducing a joint optimization objective in model training, and constructing a loss function containing two core elements: one is the root mean square error term, to ensure the overall accuracy of the prediction; the other is the actual guarantee coverage with high confidence, to improve the guarantee capability for low output risk: firstly, constructing a root mean square error term that considers the error between the predicted value and the true value. : In the formula, To predict and guarantee power, This represents the total actual power output of the large-scale photovoltaic power plant.

[0033] Subsequently, an error term considering predictive safeguards is constructed. : In the formula, σ is the sigmoid function.

[0034] Step S302: Integrate the core elements and construct a loss function for multi-objective collaborative optimization. In the formula, α is a dynamically adjusted parameter of the model training loss function.

[0035] After completing step S30, proceed to step S40 for model training and validation. Step S401 first involves constructing a complete historical dataset. It includes the predicted irradiance set for 35,040 time points throughout the year, the total power set of photovoltaic bases, and the historical power output set of each power station: The first 28,800 points were then used as the training set, and the last 6,240 points were used as the test set. The training set was constructed as follows: The test set is constructed as follows: Similarly, the historical predicted irradiance of each photovoltaic power station is aggregated from the first 28,800 data points. Total power of photovoltaic base Historical output data of each photovoltaic power station As a training set The historical predicted irradiance of each photovoltaic power station is aggregated from the first 6240 data points. Total power of photovoltaic base Historical output data of each photovoltaic power station As a test set .

[0036] Step S40 includes: Step S402, performing end-to-end training of the model based on the training set. As input to the model, The output of the model is used for model training. This yields the photovoltaic guaranteed power prediction model model(x), where x represents the model's input.

[0037] Step S403: Test the model performance on the test set based on the trained model. As input to the model.

[0038] The output yields the guaranteed power prediction sequence for the photovoltaic power base: In the formula, This ensures the power prediction results for the photovoltaic power base based on the final 6240 data points. Subsequently, the prediction results can be used as a basis for... Compared with actual photovoltaic output Compare them.

[0039] In summary, the photovoltaic base power prediction method based on hierarchical graph neural network and joint loss function of this invention explores the spatiotemporal correlation features of "local-global" by constructing a hierarchical graph neural network architecture, and uses a loss function that considers the concentration of power supply and the success rate of power supply to make targeted adjustments to the prediction model, thus providing information support for the coordinated and optimized operation of large-scale photovoltaic bases.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A photovoltaic plant power guarantee prediction method based on hierarchical graph neural network and joint loss function, characterized in that, The method comprises the following steps: Step S10, based on the irradiance resource distribution and geographical features, the photovoltaic base is divided into multiple sub-regions, realizing the spatial fine modeling of the photovoltaic power generation system under complex geographical conditions; Step S20, a hierarchical graph neural network is constructed to mine the spatio-temporal correlation features of photovoltaic power under the "local-global" scale, and capture the correlation and evolution law between regions and within regions; Step S30, a joint loss function considering prediction accuracy and reliability is constructed to balance the concentration of power supply guarantee power and the success rate of guarantee; Step S40, based on the training set, the model is trained, and the model is tested and the performance is verified on the test set, and the next day guarantee power prediction result of the photovoltaic base is obtained.

2. The photovoltaic power plant guaranteed power prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, Step S10 comprises the following sub-steps: Step S101, the historical measured irradiance, predicted irradiance, latitude and longitude and altitude information of each power station in the photovoltaic base are obtained; Step S102, the historical irradiance sequence, predicted irradiance sequence, latitude and longitude and altitude information are normalized, and a comprehensive feature vector is constructed based on the normalized features; Step S103, the K-Means algorithm is used to cluster the comprehensive feature vector, and the power stations with similar irradiance resources and geographical environment are divided into the same sub-region.

3. The photovoltaic power plant guaranteed power prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, Step S20 comprises the following sub-steps: Step S201, the correlation between any two stations in the same region is calculated according to the historical power sequence of the photovoltaic power station, and a sub-region adjacency matrix is constructed accordingly.

4. The photovoltaic power plant power guarantee prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, Step S20 further comprises: Based on the correlation between the power stations across regions, a global adjacency matrix is constructed to describe the global correlation structure within the overall range of the photovoltaic base.

5. The photovoltaic power plant power prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, The hierarchical graph neural network in step S20 comprises: Local graph convolution is driven by the sub-region adjacency matrix to extract the spatio-temporal correlation features within the region; Global graph convolution is driven by the global adjacency matrix to fuse the cross-region correlation information, and realize the modeling of the overall evolution law of the photovoltaic base.

6. The photovoltaic power plant power prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, Step S30 comprises: Step S301, the mean square error term is introduced to constrain the overall prediction accuracy; Step S302, a guarantee index based on the confidence interval coverage rate and the confidence interval width is constructed, and a joint loss function is constructed together with the mean square error, and a weighted method is used for multi-objective collaborative optimization.

7. The photovoltaic power plant power prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, Step S40 comprises the following sub-steps: Step S401, based on the photovoltaic prediction irradiance set of 35040 time points throughout the year, the total power set of the photovoltaic base and the historical output set of each power station, a complete dataset DataSet is constructed, and the training set and the test set are divided in time sequence.

8. The photovoltaic power plant power guarantee prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, Step S40 comprises the following sub-steps: Step S402, the historical prediction irradiance set and the historical output set in the training set are used as the input of the model, and the historical photovoltaic total output is used as the output of the model, and the hierarchical graph neural network is trained end to end, and the model parameters are updated through the back propagation algorithm. 9.The photovoltaic power plant guaranteed power prediction method based on hierarchical graph neural network and joint loss function of claim 1, wherein, Step S40 further comprises: Step S403, the prediction irradiance set and the output set in the test set are input into the trained graph neural network, and the photovoltaic guarantee power prediction result of the future one day is obtained.

10. The photovoltaic power plant power prediction method based on hierarchical graph neural network and joint loss function according to claim 1, characterized in that, The step S40 further comprises: comparing the predicted guaranteed power sequence with the actual photovoltaic output in the test set to verify the accuracy and guarantee capacity of the model by a deviation index and a confidence interval index.

Images