A Photovoltaic Power Prediction Method and Device Based on Localized Meteorological Interpolation

Abstract

This invention provides a photovoltaic power prediction method and apparatus based on localized meteorological interpolation, belonging to the field of photovoltaic power generation prediction technology. By using a localized meteorological interpolation model, other target sites lacking meteorological data learn the relationship between power and meteorological variables from source sites with complete meteorological data, completing localized meteorological interpolation and generating missing meteorological data. This provides complete meteorological feature input for subsequent predictions, effectively compensating for the decrease in prediction accuracy caused by missing meteorological data. By establishing a mapping relationship between power and meteorology through transfer learning rather than direct geospatial interpolation, missing meteorological data can be generated under supervised learning. This solves the problem of existing methods relying on the assumption of smooth spatial changes in meteorological data, which leads to distorted prediction results in complex terrain or drastic weather changes. This invention maintains high prediction accuracy even in complex terrain or environments with drastic meteorological field changes.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a photovoltaic power prediction method and apparatus based on localized meteorological interpolation. Background Technology

[0002] Distributed photovoltaic (PV) power generation has become an important form of promoting the rapid development of renewable energy due to its flexible deployment, low access barriers, and local consumption advantages. However, with high grid integration, its power generation exhibits more significant intermittency and volatility. This strong randomness and uncertainty can easily impact the safety and stability of the power system during large-scale grid connection. High-precision research on distributed PV power forecasting is of great significance for grid dispatching, load balancing, and renewable energy consumption. Furthermore, ultra-short-term power forecasting of distributed PV is crucial for ensuring the real-time reliability of grid power supply and the economic operation of the future electricity market.

[0003] Due to the limited capacity of individual photovoltaic (PV) sites, current research on distributed PV power prediction mainly focuses on regional prediction. Research methods are primarily divided into two categories: The first category comprises statistical and traditional machine learning and neural network models for aggregate or regional prediction scenarios, such as Autoregressive Integrated Moving Average (ARIMA), eXtreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM). These methods can uncover certain temporal patterns, but their ability to model complex spatiotemporal correlations is limited. The second category consists of deep learning methods that have emerged in recent years. These methods excel in capturing complex spatiotemporal dependencies and have gradually become a research hotspot in the field of PV prediction. Examples include Convolutional Neural Networks (CNN), Graph Neural Networks (GNN), and their improved counterparts, Graph Convolutional Networks (GCN) and Spatio-Temporal Graph Convolutional Networks (STGCN).

[0004] While the aforementioned regional distributed photovoltaic (PV) power prediction methods analyze the spatiotemporal correlation of power across multiple sites within a region, their input features generally focus on time-series power data. The fact that low-voltage distributed PV sites typically lack local meteorological monitoring equipment or forecast meteorological resources results in a lack of key objective meteorological features in regional predictions, severely limiting the improvement of site prediction performance. Furthermore, geographical and climatic differences between different distributed PV sites mean that a limited number of meteorological features cannot characterize the overall spatiotemporal coupling correlation between regional meteorological conditions, thus affecting the training effect of data-driven prediction models. Therefore, in distributed PV ultra-short-term power prediction, the lack of local meteorological monitoring data becomes a key bottleneck for further improving power prediction accuracy.

[0005] Given that existing research has performed geographic interpolation calculations on meteorological data from various distributed photovoltaic sites, the main methods include inverse distance weighting (IDW) and kriging. However, the spatial distribution of meteorological data is not uniform and smooth, and spatial interpolation methods without supervised learning will significantly decrease in accuracy when dealing with complex terrain or drastic changes in the meteorological field. Summary of the Invention

[0006] To address the issues of insufficient prediction accuracy in existing distributed photovoltaic power prediction technologies, especially distributed photovoltaic ultra-short-term power prediction technologies, due to the lack of local meteorological monitoring data, and the significant decrease in accuracy of spatial interpolation methods without supervised learning in complex terrain or drastic weather field changes, this invention provides a photovoltaic power prediction method and apparatus based on localized meteorological interpolation. This effectively compensates for the decrease in prediction accuracy caused by the lack of local meteorological data, while maintaining high prediction accuracy even in complex environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The present invention provides a photovoltaic power prediction method based on localized meteorological interpolation in its first aspect, comprising: taking distributed photovoltaic stations with complete meteorological data in a region as source stations, acquiring real meteorological data, measured photovoltaic power, and geographical location of the source stations for each time period; taking other distributed photovoltaic stations in the region lacking complete meteorological data or lacking meteorological data as target stations, acquiring measured photovoltaic power and geographical location of the target stations for each time period; based on the real meteorological data, measured photovoltaic power, and geographical location of the source stations for each time period, constructing a transfer learning model as a regional localized meteorological interpolation model using the relationship between photovoltaic power and meteorological variables; inputting the measured photovoltaic power of the target stations for each time period into the regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in the model, and obtaining generated meteorological data corresponding to the measured photovoltaic power of the target stations for each time period; constructing a merged dataset, including the measured photovoltaic power and geographical location of all stations in the region for each time period, as well as the real meteorological data of the source stations for each time period and the generated meteorological data of the target stations for each time period; establishing a photovoltaic power prediction model, inputting the merged dataset, outputting the photovoltaic power prediction sequence of all distributed photovoltaic stations for a future period, and obtaining the prediction result.

[0009] The present invention provides a preferred embodiment in its first aspect, wherein the step of constructing a transfer learning model based on the real meteorological data, measured photovoltaic power, and geographical location of the source station at various time periods, utilizing the relationship between photovoltaic power and meteorological variables as a regional localized meteorological interpolation model, inputting the measured photovoltaic power of the target station at various time periods into the regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in the model, and obtaining generated meteorological data corresponding to the measured photovoltaic power of the target station at various time periods, specifically includes: performing... After data alignment, meteorological features and corresponding power features of the source station for each time period and power features of the target station for each time period are extracted. A cross-domain mapping function is defined using transfer learning model parameters, with power features as input and meteorological features as output, to map power features to meteorological features. Meteorological features and corresponding power features of at least a portion of the source station for each time period are selected as a training set to train the cross-domain mapping function, resulting in a trained cross-domain mapping function. The power features of the target station for each time period are input into the trained cross-domain mapping function for prediction. The predicted meteorological features are generated corresponding to the power features of the target station for each time period, thereby obtaining the generated meteorological data. This preferred scheme clarifies the specific construction and implementation steps of the transfer learning model, including feature extraction, cross-domain mapping function design and training, provides an operable meteorological data generation process, and enhances the feasibility and interpretability of the method.

[0010] The present invention provides a preferred embodiment in its first aspect, wherein the step of selecting meteorological characteristics and corresponding power characteristics of at least a portion of time periods from a source station as a training set for training a cross-domain mapping function, and obtaining the trained cross-domain mapping function, involves designing a joint loss function composed of reconstruction loss and maximum mean difference loss, optimizing the joint loss function with the objective of minimizing the joint loss, iteratively updating the model parameters until the model converges, and completing the training. This preferred embodiment introduces a joint loss function composed of reconstruction loss and maximum mean difference loss, optimizes the training process of the transfer learning model, improves the consistency between the generated meteorological data and the real meteorological distribution, and enhances the reliability of the generated data.

[0011] The present invention provides a preferred embodiment in its first aspect, wherein the step of establishing a photovoltaic power prediction model, which involves inputting a merged dataset and outputting a photovoltaic power prediction sequence for all distributed photovoltaic sites within a future period, to obtain the prediction result, specifically includes: establishing an improved spatiotemporal graph convolutional network model based on a data source-aware evaluation mechanism as the prediction model; inputting the merged dataset; distinguishing between real data and generated data through the data source-aware evaluation mechanism and dynamically evaluating the credibility of the generated data; performing feature fusion based on the credibility; performing graph convolution and temporal convolution to capture the spatial and temporal dependencies between sites; and outputting a photovoltaic power prediction sequence for all distributed photovoltaic sites within a future period to obtain the prediction result. This preferred embodiment proposes an improved spatiotemporal graph convolutional network model based on a data source-aware evaluation mechanism, which can distinguish the credibility of real data and generated data, dynamically adjust the weights of generated data, and improve the model's adaptability to heterogeneous data and prediction stability.

[0012] The present invention provides a preferred embodiment in its first aspect, wherein the step of establishing an improved spatiotemporal graph convolutional network model based on a data source perception and evaluation mechanism as a prediction model, inputting a merged dataset, distinguishing between real data and generated data and dynamically evaluating the credibility of generated data through the improved spatiotemporal graph convolutional network model, and performing feature fusion based on credibility specifically includes: performing data alignment and feature extraction on the merged dataset to obtain a merged feature set containing geographical features, power features, real meteorological features, and generated meteorological features, and inputting this set into the improved spatiotemporal graph convolutional network model; splitting the merged feature set according to its source based on the improved spatiotemporal graph convolutional network model, distinguishing between real features and generated features, and assigning different weight matrices to each type of feature for each station; setting a separate dynamic gating mechanism for generated meteorological features to dynamically evaluate the credibility of generated meteorological features, and dynamically adjusting the weight matrix of generated meteorological features according to credibility, and obtaining the meteorological-geographical-power fused features for each station after weighted fusion. This preferred scheme provides a specific structural design for improving the spatiotemporal graph convolutional network model, including weight matrix block division, explicit gating units and regularization constraint mechanisms, which realizes the discriminative fusion of multi-source features and enhances the model's feature representation ability and robustness.

[0013] In its first aspect, this invention provides a preferred embodiment whereby the dynamic gating mechanism specifically comprises: inputting generated meteorological features into a multilayer perceptron; the multilayer perceptron outputting the reliability of the generated meteorological features based on their closeness to real meteorological features; passing the output of the multilayer perceptron through a sigmoid function to output a gating coefficient between 0 and 1; and adjusting the weight matrix of the generated meteorological features using this gating coefficient. This preferred embodiment designs a dynamic gating mechanism based on a multilayer perceptron and a sigmoid function, achieving real-time evaluation of the reliability of the generated meteorological data and adaptive weight adjustment, further optimizing the feature fusion process.

[0014] The present invention provides a preferred embodiment in its first aspect, wherein the step of using distributed photovoltaic (PV) stations with complete meteorological data within a region as source stations to obtain real meteorological data, measured PV power, and geographical location of the source stations for each time period specifically includes: using multiple distributed PV stations with complete meteorological data within a region as multiple source stations, and obtaining real meteorological data, measured PV power, and geographical location of each source station for each time period. This preferred embodiment expands the number of source stations by using multiple stations with complete meteorological data as source stations, thereby improving the representativeness and robustness of meteorological interpolation and reducing dependence on a single source station.

[0015] The present invention provides a preferred embodiment in its first aspect, wherein the step of constructing a transfer learning model as a regional localized meteorological interpolation model based on real meteorological data, measured photovoltaic power, and geographical location of each source station at various time periods, and utilizing the relationship between photovoltaic power and meteorological variables, specifically includes: constructing a transfer learning model as a corresponding regional localized meteorological interpolation model based on real meteorological data, measured photovoltaic power, and geographical location of each source station at various time periods, and utilizing the relationship between photovoltaic power and meteorological variables. Multiple regional localized meteorological interpolation models are provided, corresponding to the number of source stations. This preferred embodiment constructs an independent localized meteorological interpolation model for each source station, forming a multi-model parallel generation system, enhancing the diversity and adaptability of the generated data.

[0016] The present invention provides a preferred embodiment in its first aspect, wherein the step of inputting the measured photovoltaic power of the target site for each time period into a regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in the model, and obtaining generated meteorological data corresponding to the measured photovoltaic power of the target site for each time period specifically includes: inputting the measured photovoltaic power of the target site for each time period into each regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in each model, and outputting multiple sets of first generated meteorological data corresponding to the measured photovoltaic power of the target site for each time period; introducing an integrated gating network to dynamically weight and integrate the first generated meteorological data output by each localized meteorological interpolation model to obtain second generated meteorological data as the final generated meteorological data for subsequent steps. This preferred embodiment introduces an integrated gating network to dynamically weight and fuse the outputs of multiple localized meteorological interpolation models, optimizing the quality of the final generated meteorological data and improving the accuracy and stability of the generated data.

[0017] In its first aspect, this invention provides a preferred embodiment of the photovoltaic power prediction method based on localized meteorological interpolation, which further includes: constructing a correction model based on acquired meteorological forecast data for a future period, and correcting the output power prediction results for all distributed photovoltaic sites for the future period. This preferred embodiment incorporates meteorological forecast data for correction based on the prediction results, further improving the real-time performance and adaptability of the prediction, and reducing prediction bias caused by sudden meteorological changes.

[0018] The present invention provides a preferred embodiment in its first aspect, wherein the step of constructing a correction model based on acquired meteorological forecast data for a future period, and correcting the output power prediction results of all distributed photovoltaic (PV) sites for that future period, specifically includes: converting the acquired meteorological forecast data for each time period into reference PV power to obtain a reference PV power prediction sequence; constructing an adaptive fusion corrector to dynamically evaluate the reliability of the PV power prediction sequence output by the improved spatiotemporal graph convolutional network model and the reference PV power prediction sequence converted from the meteorological forecast data, and performing weighted fusion, and training based on the reference PV power prediction sequences and PV power prediction sequences for historical time periods; inputting the PV power prediction sequence output by the improved spatiotemporal graph convolutional network model for the current time period and the reference PV power prediction sequence into the trained adaptive fusion corrector for correction to obtain a corrected PV power prediction sequence. This preferred embodiment constructs an adaptive fusion corrector, dynamically evaluates and fuses the model prediction results with the reference power converted from the meteorological forecast, realizes the optimized integration of multi-source information, and improves the reliability and accuracy of the final prediction results.

[0019] In a second aspect, the present invention provides a preferred embodiment of a photovoltaic power prediction device based on localized meteorological interpolation, used to execute the method comprising: a data acquisition module, used to acquire real meteorological data, measured photovoltaic power, and geographical location of distributed photovoltaic stations with complete meteorological data in the region as source stations, and to acquire measured photovoltaic power and geographical location of other distributed photovoltaic stations in the region lacking complete meteorological data or lacking meteorological data as target stations; and a localized meteorological interpolation module, used to construct transfer learning based on the real meteorological data, measured photovoltaic power, and geographical location of the source stations in each time period, utilizing the relationship between photovoltaic power and meteorological variables. The model, serving as a regional localized meteorological interpolation model, inputs the measured photovoltaic power of the target station at various time periods into the regional localized meteorological interpolation model, learns the relationship between photovoltaic power and meteorological variables in the model, and obtains generated meteorological data corresponding to the measured photovoltaic power of the target station at various time periods. The data merging module is used to construct a merged dataset, which includes the measured photovoltaic power and geographical location of all stations in the region at various time periods, as well as the real meteorological data of the source stations at various time periods and the generated meteorological data of the target stations at various time periods. The photovoltaic power prediction module is used to establish a photovoltaic power prediction model, input the merged dataset, and output the photovoltaic power prediction sequence of all distributed photovoltaic stations in the future period to obtain the prediction results.

[0020] Compared with the prior art, the present invention has the following beneficial technical effects:

[0021] This invention constructs and utilizes a localized meteorological interpolation model, enabling target stations lacking meteorological data to learn the relationship between power and meteorological variables from source stations with complete meteorological data. This completes localized meteorological interpolation, generating the missing meteorological data and achieving intelligent generation from power data to meteorological data. This provides complete meteorological feature input for subsequent forecasts, effectively compensating for the decline in forecast accuracy caused by missing meteorological data. Furthermore, this invention utilizes the relationship between photovoltaic power and meteorological variables to construct a transfer learning model as a regional localized meteorological interpolation model. By establishing a mapping relationship between power and meteorology through transfer learning rather than directly performing geospatial interpolation, missing meteorological data can be generated under supervised learning. This addresses the problem of existing methods relying on the assumption of smooth spatial changes in meteorological data, which leads to distorted prediction results in complex terrain or drastic weather changes. Therefore, this invention maintains high prediction accuracy even in complex terrain or environments with drastic meteorological field changes.

[0022] Furthermore, this invention introduces a data source awareness mechanism into the Spatiotemporal Graph Convolutional Network (STGCN) model to distinguish the credibility of real meteorological data from generated meteorological data. This avoids the negative impact of uncertainty in generated data on prediction accuracy, improves the model's adaptability to heterogeneous data, and maintains high prediction stability and accuracy even under complex weather changes. The photovoltaic power prediction method based on localized meteorological interpolation proposed in this invention uses an improved spatiotemporal graph convolutional network model with a data source awareness and evaluation mechanism as the prediction model. Through reliable data generation and data source awareness and evaluation mechanisms, it achieves model data augmentation and refined data differentiation, effectively improving the accuracy and stability of distributed photovoltaic power prediction. Attached Figure Description

[0023] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the steps of the photovoltaic power prediction method based on localized meteorological interpolation provided in Embodiment 1 of the present invention.

[0025] Figure 2 This is a detailed flowchart of the photovoltaic power prediction method based on localized meteorological interpolation provided in Embodiment 1 of the present invention;

[0026] Figure 3 This is a schematic diagram of the photovoltaic power prediction device based on localized meteorological interpolation provided in Embodiment 1 of the present invention;

[0027] Figure 4 A map showing the geographical distribution and characteristics of regional stations;

[0028] Figure 5 The graph shows the ultra-short-term power prediction curves of distributed photovoltaic clusters over a week using three data processing methods.

[0029] Figure 6 The graphs show the ultra-short-term power prediction curves of distributed photovoltaic clusters under clear weather conditions for each model.

[0030] Figure 7 The graphs show the ultra-short-term power prediction curves of distributed photovoltaic clusters under cloudy conditions for each model.

[0031] Figure 8 The graphs show the ultra-short-term power prediction curves of distributed photovoltaic clusters for each model on cloudy days.

[0032] Figure 9 The graphs show the ultra-short-term power prediction curves of distributed photovoltaic clusters under rainy weather for each model.

[0033] Figure 10 The distribution of daily average MAPE error for each model in June;

[0034] Figure 11 A comparison chart of MAPE errors predicted by individual site modeling versus prediction using the method of Example 1 of this invention;

[0035] Figure 12 This is a flowchart illustrating the steps of the photovoltaic power prediction method based on localized meteorological interpolation provided in Embodiment 3 of the present invention.

[0036] Figure 13 This is a schematic diagram of the photovoltaic power prediction device based on localized meteorological interpolation provided in Embodiment 3 of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Example 1: Please refer to Figure 1 In one optional implementation, a photovoltaic power prediction method based on localized meteorological interpolation is provided, which is mainly implemented through the following steps:

[0039] S1. Using distributed photovoltaic stations with complete meteorological data in the region as source stations, obtain the real meteorological data, measured photovoltaic power and geographical location of the source stations for each time period. Using other distributed photovoltaic stations in the region that lack complete meteorological data or lack meteorological data as target stations, obtain the measured photovoltaic power and geographical location of the target stations for each time period.

[0040] S2. Based on the real meteorological data, measured photovoltaic power and geographical location of the source station at various time periods, a transfer learning model is constructed using the relationship between photovoltaic power and meteorological variables as a regional localized meteorological interpolation model. The measured photovoltaic power of the target station at various time periods is input into the regional localized meteorological interpolation model to learn the relationship between photovoltaic power and meteorological variables in the model, thereby obtaining generated meteorological data corresponding to the measured photovoltaic power of the target station at various time periods.

[0041] It's understandable that transfer learning is a machine learning paradigm. Its basic idea is that when the distributions of the source and target domains differ, but they share some similarity in task relevance, knowledge learned in the source domain can be transferred to the target domain through knowledge sharing, feature mapping, or parameter transfer to improve the performance of the target task. Formally, assuming the source domain... With the target domain The definitions are respectively and ,in, , Let the feature spaces of the source domain and the target domain be represented respectively. , These represent the edge distributions of the feature spaces of the source and target domains, respectively. The corresponding learning tasks are... and ,in, , The learning tasks corresponding to the source domain and the target domain are respectively. , For the label space of the source and target domains, The prediction function corresponding to the source domain. Let be the prediction function corresponding to the target domain. In traditional supervised learning, it is required that... Transfer learning allows or And through knowledge transfer The predictive performance is improved. In the process of data-driven distributed photovoltaic power prediction, when the dataset is lacking, existing feature knowledge or information can be expanded through transfer learning to enrich the dimension and quantity of input data, thereby improving the model's predictive performance.

[0042] In a preferred embodiment of the present invention, for step S2, source stations with complete meteorological data within the region are selected. A transfer learning model is constructed using the relationship between measured photovoltaic power data and real meteorological data. By minimizing the difference in feature distribution between the source and target domains, and combining the maximum mean difference (MMD) with reconstruction loss, meteorological data for the target station consistent with the power trend is generated. Please refer to... Figure 2 Specifically, step S2 is further implemented through the following steps:

[0043] Data alignment and feature extraction: S21. After aligning the real meteorological data, measured photovoltaic power and geographical location of the source site for each time period and the measured photovoltaic power and geographical location of the target site for each time period, extract the meteorological features and corresponding power features of the source site for each time period and the power features of the target site for each time period.

[0044] Constructing a cross-domain mapping function: S22. Using the parameters of the transfer learning model, define a cross-domain mapping function with power features as input and meteorological features as output, which is used to map power features to meteorological features.

[0045] Domain Adaptation Model Training: S23. Select meteorological features and corresponding power features of at least a portion of the time period of the source station as a training set to train the cross-domain mapping function and obtain the trained cross-domain mapping function; In a preferred embodiment, for the step of selecting meteorological features and corresponding power features of at least a portion of the time period of the source station as a training set to train the cross-domain mapping function and obtain the trained cross-domain mapping function, during the training process, a joint loss function consisting of reconstruction loss and maximum mean difference loss is designed, and the joint loss function is optimized with the goal of minimizing the joint loss. The model parameters are iteratively updated until the model converges, and the training is completed.

[0046] Apply the cross-domain mapping function: S24. Input the power features of the target site at each time period into the trained cross-domain mapping function for prediction.

[0047] Meteorological data generation: S25. Predict and generate meteorological features corresponding to the power characteristics of the target station at each time period, thereby obtaining the generated meteorological data.

[0048] More specifically, for each of the above sub-steps, a localized meteorological interpolation model based on transfer learning is constructed. In this process, the geographical locations of all stations are known. Localized meteorological data interpolation is achieved using meteorological data from the source stations. The closer the interpolation is to the source station, the higher its reliability; the farther away, the lower its reliability. In the meteorological data generation problem, the source domain is the photovoltaic stations with meteorological observation data (i.e., the source stations), and the target domain is other stations lacking meteorological observation data (i.e., the target stations). Let the actual meteorological data and measured photovoltaic power data (already aligned and feature extracted) of the source stations be... ,in, Indicates the source site is at the The meteorological feature vector at each time point, where d represents the feature dimension; Indicates the source site is at the The power output corresponding to each moment, i.e., the power characteristic. The target site only has power data: , Indicates the target site is at the The power output at time n is given. To generate meteorological data at the target station, this invention employs a feature mapping method based on transfer learning to construct a cross-domain mapping function, such that the target station's power output at time n is... Meteorological feature vector at each time point ,in, Indicates the parameters of the transfer learning model The defined mapping function utilizes the learned relationship between power and meteorology in the source domain to map the power data of the target station to corresponding meteorological features. Specifically, in a preferred embodiment, this can be achieved by minimizing the distribution difference between the source and target domains in the feature space. Commonly used optimization objectives are: ,in: To reconstruct the loss, and to ensure the consistency between the generated meteorological data and the meteorological distribution of the source stations; The maximum mean difference (MMD) is used to measure the difference in feature distributions between the source and target domains. For feature mapping function (i.e., the cross-domain mapping function mentioned above) ); For balance parameters; For source domain data; For target domain data.

[0049] This embodiment is for a scenario with only one source site. To provide complete and accurate meteorological observation data, it is necessary to migrate the meteorological model from the source station to other target stations. Let the meteorological characteristics of the source station be... , L represents the optimization objective. The target site power data is as follows: Through transfer learning models It can generate meteorological characteristics of the target station. , ,in, The essence of this approach is to learn the implicit mapping relationship between the meteorological data of the source station and the power of the target station, thereby generating a meteorological sequence with the characteristics of the target station. Through the above optimization, the generated meteorological data of the target station not only matches its power change trend, but also closely resembles the actual meteorological data of the source station in terms of distribution, thus improving the availability and reliability of the meteorological data.

[0050] S3. Construct a merged dataset containing measured photovoltaic power and geographical location of all stations in the region at various times, as well as real meteorological data of source stations at various times and generated meteorological data of target stations at various times.

[0051] S4. Establish a photovoltaic power prediction model, input the merged dataset, and output the photovoltaic power prediction sequence of all distributed photovoltaic sites in the future period to obtain the prediction results.

[0052] This embodiment constructs and utilizes a localized meteorological interpolation model, allowing other target stations lacking meteorological data to learn the relationship between power and meteorological variables from source stations with complete meteorological data. This achieves localized meteorological interpolation, generating missing meteorological data and effectively compensating for the decline in prediction accuracy caused by missing meteorological data. Furthermore, this invention utilizes the relationship between photovoltaic power and meteorological variables to construct a transfer learning model as a regional localized meteorological interpolation model. By establishing a mapping relationship between power and meteorology through transfer learning rather than directly performing geospatial interpolation, missing meteorological data can be generated under supervised learning. This overcomes the problem of traditional methods, which rely on the assumption of smooth spatial changes in meteorological data and are prone to distortion in complex terrain or drastic weather changes. Therefore, this invention maintains high prediction accuracy even in complex terrain or environments with drastic weather changes. Specifically, this invention utilizes the supervised learning of the power-meteorological mapping relationship from source stations and transfers it to target stations. It generates corresponding meteorological data using real-time power data from target stations, making the generated results naturally adaptable to local terrain features and able to quickly respond to sudden weather changes, breaking through the limitations of the geographical smoothness assumption in traditional spatial interpolation.

[0053] Furthermore, considering that the reliability of generated meteorological data is still lower than that of actual measured data, using real and generated data indiscriminately during training will negatively impact the reliability and stability of predictions. Therefore, a more preferred implementation method is provided. Based on step S4, an improved empty graph convolutional network model with a data source awareness and evaluation mechanism is introduced to distinguish between real and generated data and dynamically evaluate the credibility of the generated data. Specifically, the following more preferred implementation step S40 is provided for step S4:

[0054] S40. An improved spatiotemporal graph convolutional network model based on a data source awareness and evaluation mechanism is established as the prediction model. The merged dataset is input, and the data source awareness and evaluation mechanism distinguishes between real and generated data and dynamically evaluates the credibility of the generated data. After feature fusion based on credibility, graph convolution and temporal convolution are performed to capture the spatial and temporal dependencies between sites, outputting a photovoltaic power prediction sequence for all distributed photovoltaic sites within a future period, thus obtaining the prediction result. In this embodiment, the improved spatiotemporal graph convolutional network model (STGCN model) includes graph convolution modules and temporal convolution modules, which are alternately stacked to learn the spatial and temporal dependency features of multiple sites. Weight matrix partitioning and gating units are introduced in feature extraction to distinguish the credibility of real and generated data.

[0055] Understandably, the STGCN model, or Spatiotemporal Graph Convolutional Network, is particularly suitable for modeling data with both spatial and temporal dependencies. The STGCN model combines graph convolution and temporal convolution, learning spatial and temporal features jointly through multiple stacked layers. At each layer, graph convolution is first applied to capture spatial relationships, followed by temporal convolution to capture temporal relationships. Finally, the learned spatial and temporal features are merged for use in the next layer or the final prediction task. The basic structure of STGCN can be described as follows:

[0056] ;

[0057] in, It is the first The node feature matrix of the layer; It is the first The weight matrix of the layered graph convolutional layer; It is the first Layer-time convolution kernel; It is an activation function; It is the first The adjacency matrix of the node feature matrix of the layer; It is the first Layer node feature matrix.

[0058] Spatial graph convolution: The goal of graph convolution is to extract features from graph-structured data, fusing the feature information of each node with the information of its neighboring nodes, and capturing spatial dependencies by defining the adjacency relationships in the graph structure. For a given graph, let the set of nodes be... ,in, Let be the number of nodes, where each node represents a photovoltaic power station; the set of edges between nodes is . The edges represent the spatial relationships between power stations. The adjacency matrix of the graph. It describes the connection relationships between nodes, where Represents a node With nodes The strength of the relationship between them. In graph convolutional networks, the graph convolution operation is defined by the following formula:

[0059] ;

[0060] in, It is the first The node feature matrix of the layer, Where N is the number of nodes, For the first The feature dimensions of the layer; It is a normalized adjacency matrix. ,in It is a degree matrix. It is a normalized adjacency matrix. ; It is the first The learnable weight matrix of the layer, ,in For the first The feature dimensions of the layer It is an activation function.

[0061] Temporal convolution: Temporal convolution captures temporal trends and periodicity by performing sliding convolution on historical data at different time steps. Convolutional kernels at different time steps can learn different temporal patterns, helping the model better capture short-term and long-term temporal dependencies. It is assumed that the features of each node at each time step... There is observation data available on both sides. The goal of temporal convolution is to pass through a temporal convolution filter. This is used to extract patterns of change in time series data. The formula for temporal convolution is:

[0062] ;

[0063] in: It is a time step Upper The node feature matrix of the layer, ; It is a time step Upper The node feature matrix of the layer; It is a time step Upper The node feature matrix of the layer; It is the first The weights of each convolutional kernel at different times. , For the first Layer node feature dimensions For the first Layer node feature dimensions; Indicates the convolution operation; This refers to the number of convolutional kernels. Based on two convolutional modules, the STGCN model can effectively learn joint spatial-temporal features from the input data. In regional distributed photovoltaic ultra-short-term power prediction, the geographical location of regional sites, meteorological data, and measured power will be used together as the input dataset to construct the STGCN model, achieving high-precision power prediction.

[0064] This invention constructs an improved STGCN model based on a data source awareness and evaluation mechanism. Based on a merged dataset composed of real and generated data, this invention innovatively proposes a perception and evaluation mechanism for different data sources within the STGCN model. This mechanism identifies differences in data credibility, refines the reliability of data information, and constructs a prediction model with stronger interpretability and robustness. In a preferred embodiment of this invention, in step S4, an improved spatiotemporal graph convolutional network model based on a data source awareness and evaluation mechanism is established as the prediction model. The merged dataset is input, and the improved spatiotemporal graph convolutional network model distinguishes between real and generated data and dynamically evaluates the credibility of the generated data. Feature fusion is then performed based on the credibility, specifically through the following steps:

[0065] S401. Based on a spatiotemporal graph convolutional network containing spatial convolutional layers (i.e., graph convolution), temporal convolutional layers (i.e., time-series convolution), and fully connected layers, a spatiotemporal graph structure including weight matrix partitioning, explicit gating units, and regularization constraints is constructed and placed before the spatial convolutional layers. This yields an improved spatiotemporal graph convolutional network model based on a data source-aware evaluation mechanism, which serves as the prediction model. Please refer to [reference needed]. Figure 2 .

[0066] S402. After data alignment and feature extraction of the merged dataset, a merged feature set containing geographical features, power features, real meteorological features and generated meteorological features is obtained, and then input into the improved spatiotemporal graph convolutional network model.

[0067] S403. Based on the improved spatiotemporal graph convolutional network model, the merged feature set is split according to its source, distinguishing between real features and generated features, and different weight matrices are assigned to each type of feature of each site.

[0068] Specifically, in improving the spatiotemporal graph convolutional network model, weight matrix segmentation is added during the feature mapping stage, allowing input features to be split and weighted according to their source. Let any site... The input feature vector is:

[0069] ;

[0070] in, Indicates geographical features, Indicates historical power characteristics, It represents the true meteorological characteristics, while To generate meteorological features. It should be noted that the target station only has generated meteorological features, with zero actual meteorological features; the source station only has actual meteorological features, with zero generated meteorological features. Accordingly, the feature transformation can be written as:

[0071] ;

[0072] in, It is the comprehensive feature vector of station j after feature transformation, where the addition represents the fusion of multiple features; All are learnable weight matrices that control the contributions of features from different sources to distinguish between real and generated data. They are, in order: geographic location feature weight matrix, historical power feature weight matrix, real meteorological feature weight matrix, and generated meteorological feature weight matrix. Represents element-wise product; For gating coefficients. Unlike traditional STGCN, the weight matrix block structure in the improved spatiotemporal graph convolutional network model can clearly distinguish the contributions of real meteorological data and generated meteorological data at the parameter level.

[0073] S404. A dynamic gating mechanism is set up separately for the generated meteorological features to dynamically evaluate the credibility of the generated meteorological features, and the weight matrix of the generated meteorological features is dynamically adjusted according to the credibility. After weighted fusion, the meteorological-geographic-power fusion features of each station are obtained. In a preferred embodiment of the present invention, the dynamic gating mechanism specifically involves: inputting the generated meteorological features into a multilayer perceptron; the multilayer perceptron outputting the credibility of the generated meteorological features according to the closeness between the generated meteorological features and the real meteorological features; outputting a gating coefficient between 0 and 1 through a sigmoid function; and adjusting the weight matrix of the generated meteorological features through the gating coefficient.

[0074] Specifically, when improving the spatiotemporal graph convolutional network model, display gating units and gating coefficients are set. It is the core of the data source awareness mechanism. If it were directly set as a fixed hyperparameter, for example... This can reduce the contribution of fixed weights to the generated meteorological data. However, fixed weights cannot adapt to differences in different scenarios, resulting in poor robustness and generalization. Therefore, this invention further designs a learnable gating mechanism. Specifically, It can be defined as:

[0075] ;

[0076] in, For the site In time The generation of meteorological characteristic data, Use the Sigmoid function, ensuring that the output range is (0,1); This is a multilayer perceptron used to dynamically estimate the reliability of generated meteorological data based on its distribution characteristics. In this case, the model can adaptively adjust the weights of generated data from different stations and at different times during training. When the generated data distribution is highly consistent with the true distribution, It will approach 1; conversely, it will reduce its effect to reduce noise interference.

[0077] S405. After feature fusion based on credibility, graph convolution and temporal convolution are performed to capture the spatial and temporal dependencies between sites, outputting the photovoltaic power prediction sequence for all distributed photovoltaic sites within a future period, thus obtaining the prediction result. More specifically, in this sub-step, after introducing a data source awareness mechanism, the update formula for graph convolution is rewritten as:

[0078] ;

[0079] in, For the first The node feature matrix of the layer, For activation function, This is the processed graph adjacency matrix. For the first The learnable weight matrix of the layer, This is the feature matrix after multi-source fusion. It already includes multi-source features processed through weight matrix partitioning and gating mechanisms. Subsequently, temporal convolution is used to model the sequence dimension:

[0080] ;

[0081] in, The kernel size is [size]. These are the parameters for the temporal convolution kernel. By alternating and stacking spatial and temporal convolutions, the model can learn the spatiotemporal dependencies of distributed photovoltaic power while explicitly distinguishing data sources.

[0082] In a preferred embodiment, the regularization constraint mechanism is as follows: during the training of the improved spatiotemporal graph convolutional network model, a regularization term for the weight matrix of the generated meteorological features is introduced into the loss function to suppress the excessive growth of the generated meteorological features. Specifically, the regularization term for the weight matrix of the generated meteorological features is... Apply constraints:

[0083] ;

[0084] in, To measure the photovoltaic power, For the site in Photovoltaic power forecast at that time is the regularization coefficient. This loss function not only constrains the prediction error but also mitigates the negative impact of generated data on the training process.

[0085] Through the above embodiments, the regional distributed photovoltaic ultra-short-term power prediction method based on localized meteorological interpolation and improved spatiotemporal graph convolutional network proposed in this invention achieves model data enhancement and refined data differentiation through reliable data generation and data source perception and evaluation mechanisms, effectively improving the accuracy and stability of distributed photovoltaic power prediction.

[0086] Please refer to Figure 3 Corresponding to the embodiments of the method of the present invention, a photovoltaic power prediction device based on localized meteorological interpolation is used to execute the method and mainly consists of the following parts:

[0087] Data acquisition module 1 is used to acquire real meteorological data, measured photovoltaic power and geographical location of distributed photovoltaic stations with complete meteorological data in the region as source stations, and to acquire measured photovoltaic power and geographical location of other distributed photovoltaic stations in the region that lack complete meteorological data or lack meteorological data as target stations.

[0088] The localized meteorological interpolation module 2 is used to construct a transfer learning model based on the real meteorological data, measured photovoltaic power and geographical location of the source station at each time period, and to use the relationship between photovoltaic power and meteorological variables as the regional localized meteorological interpolation model. The measured photovoltaic power of the target station at each time period is input into the regional localized meteorological interpolation model, and the relationship between photovoltaic power and meteorological variables in the model is learned to obtain the generated meteorological data corresponding to the measured photovoltaic power of the target station at each time period.

[0089] Data merging module 3 is used to construct a merged dataset, which includes the measured photovoltaic power and geographical location of all stations in the region at various times, as well as the real meteorological data of the source stations at various times and the generated meteorological data of the target stations at various times.

[0090] The photovoltaic power prediction module 4 is used to establish a photovoltaic power prediction model. It takes the merged dataset as input and outputs the photovoltaic power prediction sequence of all distributed photovoltaic sites in the future period to obtain the prediction results.

[0091] To verify the effectiveness of the above embodiments of the present invention, examples will be used to illustrate the following:

[0092] The experimental data used in this embodiment includes actual power data collected from seven distributed photovoltaic (PV) stations in a certain area, and meteorological data collected from one local meteorological monitoring station. The data sampling interval is 15 minutes, with 96 data points per day, spanning from July 1, 2023 to June 30, 2024. The data from July 1, 2023 to May 31, 2024 is used as the training set, and the data from June 1, 2024 to June 30, 2024 is used as the test set. The geographical distribution of the stations is shown in the attached figure. Figure 4 As shown, the red markers indicate stations with both photovoltaic power and meteorological data, i.e., source stations. The black markers indicate stations with only photovoltaic power data and no meteorological data, i.e., target stations.

[0093] To objectively evaluate prediction performance, this invention employs three metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The calculation formulas are as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] In the formula: n is the total number of stations; The actual value; These are predicted values.

[0098] The experimental platform consisted of a Windows 11 x64 operating system, an AMD R9-78445HX processor, and Python language, with the deep learning framework built using PyTorch. To comprehensively verify the effectiveness of the proposed method, numerical examples were analyzed from two dimensions. First, the effectiveness of the proposed localized meteorological interpolation method was evaluated by comparing it with the following data processing methods:

[0099] Dataset comparison method 1: All sites use only measured photovoltaic power data, and do not use meteorological data.

[0100] Dataset comparison method 2: All stations used measured photovoltaic power data and meteorological data from station 1.

[0101] The method of this invention (localized meteorological interpolation): Station 1 (source station) uses measured photovoltaic power data and real meteorological data, while the other stations use measured photovoltaic power data and meteorological data generated using the method of this invention.

[0102] Secondly, we selected several typical artificial intelligence prediction methods, such as XGBoost, LSTM, GNN, and STGCN, to compare the model algorithms. At the same time, we analyzed the prediction effects of multi-site prediction methods based on the improved STGCN model and single-site modeling.

[0103] In terms of parameter settings, the improved STGCN model proposed in this invention employs two layers of temporal convolutional blocks, with 256 and 64 filters in the first and second layers, respectively; the graph convolutional part uses two improved graph convolutional layers with 256 and 64 units, respectively, and the final output layer has 128 units and uses a Dropout rate of 0.3; the optimizer uses Adam with a learning rate of 1e-4, a loss function of MSE, a batch size of 32, and 100 training epochs; the XGBoost has 256 trees, a maximum depth of 8, a learning rate of 0.01, and a subsample to feature subsample ratio of 0.5; the LSTM has three hidden layers with 64 units per layer, and the intermediate layers retain the sequence output; the GNN model uses two unit layers with 128 and 64 units, respectively; the STGCN model has the same parameters as the improved STGCN model; the batch size, training epochs, and learning units of all models are consistent with the method of this invention. The same input and output settings are used, i.e., the input sequence length is 16 and the output sequence length is 16.

[0104] Among the three data processing methods, Method 1 relies solely on measured photovoltaic power data from each site, making it difficult to accurately reflect power fluctuations caused by sudden weather changes. It has the highest overall error level, with an RMSE of 192.48 kW, a MAE of 96.54 kW, and a MAPE of 8.68%. Method 2 incorporates meteorological data from Site 1 as common meteorological data across all sites, providing additional support from environmental characteristics compared to Method 1. This results in a 3.42 kW reduction in RMSE, a 1.81 kW reduction in MAE, and a 1.89% reduction in MAPE compared to Method 1. Compared to these two methods, the localized meteorological interpolation method proposed in this invention achieves the best results, reducing the RMSE to 183.60 kW, the MAE to 92.25 kW, and the MAPE to a further 4.04%, achieving optimal performance across all three indicators. This comparative result fully demonstrates that the method of this invention can further improve the model's prediction accuracy compared to the previous two methods.

[0105] Figure 5The ultra-short-term power forecast curves for distributed photovoltaic clusters from June 20, 2024 to June 26, 2024 are presented. Figure 5 In the graph, the vertical axis represents power in kW, and the horizontal axis represents the date. Figure 5 Three data processing methods / models are presented, showing the ultra-short-term power prediction curves of distributed photovoltaic clusters over a week, and comparing them with the actual values. The figures show the prediction results of Method 1, Method 2, and the method proposed in this embodiment, along with the actual values. Region I (first enlarged area in the figure): The prediction results of all three methods lag behind the actual values ​​to some extent. The prediction value using the localized meteorological interpolation method fits the actual value better. In contrast, Method 1 and Method 2 show significant deviations in prediction, especially in the prediction of power peaks, failing to accurately capture the fluctuations in the actual data. Region II (second enlarged area in the figure): Power exhibits large peak fluctuations, which the localized meteorological interpolation method accurately captures. Method 1 and Method 2, however, fail to accurately predict this rapid rise and fall, showing significant deviations. Region III (third enlarged area in the figure): The localized meteorological interpolation method continues to track these fluctuations well, maintaining excellent prediction accuracy and a high degree of consistency with the actual data in the details of power changes. Method 1 and Method 2 can capture the general trend, but lack accuracy for subtle power changes. Figure 5 Based on the performance of the three magnified graphs, the method of this invention exhibits the best fitting effect among all the comparative methods. By combining localized meteorological interpolation with improvements to the STGCN model, the accuracy of ultra-short-term power prediction for distributed photovoltaic systems can be significantly improved.

[0106] To further compare the prediction performance of different algorithm models, the evaluation metrics for the prediction results of each model under different weather conditions were verified. The improved spatiotemporal graph convolutional network (STGCN) showed the best prediction performance under all weather conditions, with both RMSE and MAPE significantly outperforming other comparative models. Under sunny conditions, the improved STGCN achieved an RMSE of 182.45kW and a MAPE of 3.64%. Under cloudy and overcast conditions, the improved STGCN showed even greater advantages, with a MAPE of only 4.25% under cloudy conditions and only 4.71% under overcast conditions, far lower than the levels of XGBoost and LSTM exceeding 9%. In rainy weather, where power fluctuations are more drastic, the RMSE was 164.27kW, and the MAPE was only 3.56%. In comparison, XGBoost has a relatively high overall error and is difficult to adapt to power fluctuations caused by sudden weather changes; LSTM has a low RMSE under some weather conditions, but a high MAPE, indicating insufficient control over relative error; GNN performs reasonably well in RMSE, but its MAPE is still significantly higher than that of the improved STGCN, indicating insufficient prediction stability; the original STGCN is better than the previous models in capturing spatiotemporal dependencies, but it is still affected by the uncertainty of generated data.

[0107] Comprehensive analysis leads to the conclusion that the improved STGCN method proposed in this invention not only achieves the lowest error on an overall average level, but also exhibits stable and excellent prediction performance in different weather scenarios, fully verifying its advanced nature and practical value in distributed photovoltaic ultra-short-term power prediction tasks.

[0108] Figure 6 , Figure 7 , Figure 8 , Figure 9 The graphs show a comparison of the ultra-short-term power prediction curves for each model under sunny, cloudy, overcast, and rainy conditions. As can be seen from the graphs, the power output fluctuation of the regional distributed photovoltaic (PV) cluster is relatively small under sunny conditions, and the overall curve is smoother. Most models can track the changes in regional distributed PV power output well, and the method of this invention best matches the actual power curve. Under cloudy and rainy conditions, all models exhibit some degree of lag in their predictions. In comparison, the improved STGCN model shows less lag in its prediction results and demonstrates the best trend-following performance during the ultra-short-term prediction period. Especially at power abrupt changes and in areas of fluctuation, the prediction results are smoother and more accurate compared to other methods. Figures 6 to 9 In the graph, the vertical axis represents power in kW, and the horizontal axis represents time.

[0109] Figure 10The data shows the prediction errors of different models over a one-month period, with the Enhanced STGCN model showing a lower MAPE value in most cases compared to other methods. Traditional LSTM and GNN models, when handling multi-site meteorological data and spatiotemporal correlations, failed to fully utilize the reliability of the generated data, resulting in larger prediction errors in some time periods. The XGBoost model performed stably, but its prediction accuracy was relatively low when facing complex spatiotemporal dependencies. Overall, the superior performance of the Enhanced STGCN model validates its potential in multi-site photovoltaic power prediction, especially its significant advantages in handling large-scale meteorological data and spatiotemporal relationships.

[0110] Appendix Figure 11 The results show a comparison of MAPE errors between standalone modeling (independent STGCN model) and the improved STGCN model method used in this embodiment under different site conditions. The vertical axis represents the MAPE value, and the horizontal axis represents the site. The figure shows that, under all site conditions, the method of localized meteorological interpolation and the improved spatiotemporal map convolutional network has significant advantages over standalone modeling, with generally lower MAPE errors.

[0111] Specifically, the standalone modeling method exhibited significant error fluctuations across stations 1 to 7, particularly at stations 3 and 6, where the MAPE reached 7.71% and 7.62%, respectively, demonstrating a high level of error. However, the method incorporating localized meteorological interpolation and an improved spatiotemporal map convolutional network showed lower errors across all stations, especially at stations 1 and 4, where the MAPE decreased to 3.37% and 3.74%, respectively. This method, based on localized meteorological interpolation and the improved spatiotemporal map convolutional network, demonstrated higher prediction accuracy across all stations, effectively reducing error fluctuations and providing more stable and accurate ultra-short-term power forecasts compared to standalone modeling.

[0112] This invention addresses the problem of distributed photovoltaic (PV) power prediction, specifically the challenges posed by the lack of meteorological data for most sites and the difficulty of traditional prediction models in fully utilizing multi-source information. It proposes a multi-local meteorological site PV power prediction method based on localized meteorological interpolation and an improved STGCN (Significant Localized Geological Context-Based Network). First, a transfer learning-based interpolation method is proposed: to address the lack of meteorological observations for most distributed PV power plants, a mapping relationship between source site data and target site power and geographic information is used to complete meteorological features, enhancing the diversity and effectiveness of the prediction model's input. Second, an improved STGCN model with data source awareness is designed: weight matrix partitioning and gating units are introduced into the STGCN framework to model the contributions of real and generated meteorological data respectively. The improved STGCN simultaneously captures spatial correlation and temporal dependence features, ensuring stable prediction performance even with inconsistent input data quality. The feasibility and superiority of the method are verified on a real dataset.

[0113] Example 2: Example 1 primarily relies on a single source site for knowledge transfer. If the source site data is inaccurate, noisy, or insufficiently representative, the generated meteorological data for the entire region will exhibit systematic bias. When the climate within a region is complex and the terrain is highly undulating, the meteorological model from a single source site cannot represent the true situation of all target sites. Therefore, considering such situations, this invention provides a preferred solution by introducing multi-source transfer learning. Instead of relying on a single source site, it selects multiple reasonably distributed and high-quality data sites within the region as source sites. Specifically, multi-source domain adaptation technology can be used to train a basic mapping function for each source site, and then an integrated gating network can be used to dynamically select or weight the knowledge from multiple source sites to generate data for the target sites.

[0114] Specifically, based on Example 1, the step S1 in Example 1, "using distributed photovoltaic stations with complete meteorological data in the region as source stations, and obtaining the real meteorological data, measured photovoltaic power and geographical location of the source stations at various time periods", specifically includes: using multiple distributed photovoltaic stations with complete meteorological data in the region as multiple source stations, and obtaining the real meteorological data, measured photovoltaic power and geographical location of each source station at various time periods;

[0115] The step S2 in Example 1, "based on the real meteorological data, measured photovoltaic power and geographical location of each source station at each time period, constructing a transfer learning model using the relationship between photovoltaic power and meteorological variables as a regional local meteorological interpolation model", specifically includes: based on the real meteorological data, measured photovoltaic power and geographical location of each source station at each time period, constructing a transfer learning model using the relationship between photovoltaic power and meteorological variables as a corresponding regional local meteorological interpolation model (or referred to as a "single-source local meteorological interpolation model"). There are multiple regional local meteorological interpolation models, corresponding to the number of source stations.

[0116] The step S2 in Example 1, which involves "inputting the measured photovoltaic power of the target station at each time period into the regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in the model, and obtaining generated meteorological data corresponding to the measured photovoltaic power of the target station at each time period," specifically includes: inputting the measured photovoltaic power of the target station at each time period into the regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in each model, and outputting multiple sets of first generated meteorological data corresponding to the measured photovoltaic power of the target station at each time period; introducing an integrated gating network to dynamically weight and integrate the first generated meteorological data output by each localized meteorological interpolation model to obtain second generated meteorological data as the final generated meteorological data for subsequent steps.

[0117] More specifically, the following preferred steps are adopted for steps S1 and S2 in Embodiment 1:

[0118] S10. Using multiple distributed photovoltaic sites with complete meteorological data in the region as multiple source sites, obtain the real meteorological data, measured photovoltaic power and geographical location of each source site at each time period;

[0119] S20. Based on the real meteorological data, measured photovoltaic power, and geographical location of each source site at each time period, a transfer learning model is constructed using the relationship between photovoltaic power and meteorological variables as a single-source localized meteorological interpolation model. The measured photovoltaic power of the target site at each time period is input into each single-source localized meteorological interpolation model to learn the relationship between photovoltaic power and meteorological variables in each model, and output multiple sets of first generated meteorological data corresponding to the measured photovoltaic power of the target site at each time period. Thus, the target site obtains multiple sets of first generated meteorological data under the reference of different source sites. Then, an integrated gating network is introduced to dynamically weight and integrate the first generated meteorological data output by each single-source localized meteorological interpolation model to obtain the second generated meteorological data as the final generated meteorological data for subsequent steps.

[0120] More specifically, the following procedures can be followed when implementing S10 and S20:

[0121] Building a model pool: Assuming there are K high-quality source sites in the region, a transfer learning model G is independently trained for each source site k. k That is, the single-source localized meteorological interpolation model.

[0122] Generate candidate data: for the power data Y of the target site t They are then input into each single-source localized meteorological interpolation model G. k In the process, K sets of different generated meteorological data were obtained. .

[0123] Weighted fusion using a meta-gated network: Train a lightweight neural network as the meta-gated network, specifically a multilayer perceptron (MLP). The input to this network is the power data Y of the target site. t and geographical location L oct The output of this network is a K-dimensional weight vector, which is normalized through a Softmax layer to satisfy... , These are the dynamic weight coefficients assigned by the meta-gated network to the transfer learning model of the k-th source site. The meta-network learns which source site's judgment should be trusted more when the target site is in this power mode.

[0124] Weighted integration: The final generated meteorological data consists of all single-source localized meteorological interpolation models G. k Weighted average of the output: The data generated in this way in Example 2 has a much higher stability and reliability than the results generated from a single source site, thus making the final output photovoltaic power prediction results for the target site more accurate.

[0125] This embodiment acquires historical meteorological and photovoltaic power data from multiple source sites to construct a multi-source transfer learning model, fully utilizing cross-regional meteorological correlations and power output patterns. By introducing a meta-gated network to achieve dynamic weight allocation, the model adaptively selects the optimal source site contribution under different weather conditions. This method not only improves the spatial consistency of the generated meteorological data but also enhances its adaptability to complex terrain and local climates, significantly improving the accuracy of meteorological interpolation at the target site and thus optimizing photovoltaic power prediction performance. Simultaneously, this method effectively mitigates the risks associated with abnormal or missing data from a single source site, enhancing the model's robustness.

[0126] Example 3: Please refer to Figure 12The photovoltaic power prediction method based on localized meteorological interpolation further includes: S5. Constructing a correction model based on acquired meteorological forecast data for a future period, and correcting the output power prediction results for all distributed photovoltaic sites for the future period. In a preferred embodiment, S5 is specifically implemented through the following steps:

[0127] S51. Convert the obtained meteorological forecast data for each time period into reference photovoltaic power to obtain the reference photovoltaic power prediction sequence. This sub-step is specifically implemented through the following process:

[0128] S511. Data Source Acquisition and Alignment: Obtain the ultra-short-term power prediction values ​​P for the region from the trained improved STGCN model. stgcn We acquire high spatiotemporal resolution numerical weather prediction (NWP) data in real time from meteorological services. Key elements include total solar irradiance (GHI), ambient temperature, wind speed, and cloud cover. We then time-align the data to ensure that the forecast time step of the NWP data is perfectly aligned with the STGCN forecast output (e.g., every 15 minutes for the next 0-4 hours).

[0129] S512. NWP Power Conversion: Converts the meteorological elements of NWP into reference photovoltaic power generation P. nwp This is to allow for comparison and integration with STGCN's predicted values ​​on the same scale. Specifically, a photovoltaic power generation physical model is used for the conversion. In one optional implementation, the calculation formula is as follows:

[0130] ;

[0131] In the formula, and Let be the efficiency and total area of ​​the photovoltaic module, respectively, and be a system constant. To determine the effective irradiance; considering the influence of the solar incident angle, the original GHI is corrected as follows: ,in, The angle of incidence refers to the angle between the sunlight and the normal to the photovoltaic panel plane; The operating temperature of the solar panel is determined by the ambient temperature provided by the NWP. Wind speed Estimation with irradiance GHI: a and b are empirical coefficients, determined based on the specific installation method (roof mounting or ground mounting); T stc The standard test condition temperature is, for example, 25°C. This is the power temperature coefficient.

[0132] S52. Construct an adaptive fusion corrector to dynamically evaluate the reliability of the photovoltaic power prediction sequence output by the improved spatiotemporal graph convolutional network model and the reference photovoltaic power prediction sequence converted from weather forecast data, and perform weighted fusion. The corrector is trained based on the reference photovoltaic power prediction sequence and the photovoltaic power prediction sequence from historical time periods. In this step, an adaptive weighted fusion rule based on uncertainty is designed. The core of this rule is to assign weights to the STGCN prediction and the NWP prediction, with the weights inversely proportional to their respective uncertainties. That is, the lower the uncertainty of the prediction source, the higher its weight. The specific implementation process of this step is as follows:

[0133] S521. Calculate the uncertainty of the STGCN and NWP forecasts respectively. The uncertainty can be calculated using historical error statistics: the forecast error (e.g., root mean square error RMSE) of the forecast source over a recent period is used as an estimate of the uncertainty. More specifically, for STGCN, the uncertainty calculation is similar for NWP: collect the predicted and actual measured values ​​of STGCN over a recent period (e.g., the past 24 hours). Calculate the absolute or squared error for each forecast time step (e.g., the next 1 hour, 2 hours, etc.). Then, the statistic of the error for each forecast time step (e.g., standard deviation, variance) can be calculated as the uncertainty for that time step. Alternatively, for simplicity, the average uncertainty of the entire forecast sequence can be calculated. In a preferred embodiment, considering that the uncertainty will vary under different weather conditions, a sliding window can be used to calculate the error variance for each time step.

[0134] S522. Calculate the weight of each prediction source based on the uncertainty.

[0135] For each prediction time step t, the weights of STGCN and NWP are calculated as follows:

[0136] ;

[0137] ;

[0138] In the formula, STGCN prediction weights at time t. Let U be the NWP prediction weight at time t. stgcn Let U be the uncertainty of the STGCN prediction at time t; nwp Let be the uncertainty of the NWP prediction at time t. Here, considering that we want larger uncertainties to have smaller weights, we use the uncertainty of the other party as the numerator.

[0139] S523. The final prediction is obtained using a weighted average, as shown in the following formula:

[0140] ;

[0141] S53. Input the photovoltaic power prediction sequence output by the improved spatiotemporal graph convolutional network model in the current time period and the reference photovoltaic power prediction sequence into the trained adaptive fusion corrector for correction, to obtain the corrected photovoltaic power prediction sequence. The specific implementation process is as follows:

[0142] Real-time input: The ultra-short-term power prediction sequence output by the improved STGCN model at the current time. The corresponding time period is the converted NWP reference power prediction sequence. Real-time ground meteorological observation data, including irradiance, cloud cover, wind speed, temperature, and precipitation, is used to determine the current weather conditions and trigger corresponding fusion rules. Below are some common weather conditions and their triggering conditions:

[0143] (1) Rapidly Changing Scenarios: When the weather changes rapidly, such as with drastic changes in cloud cover or violent fluctuations in irradiance, STGCN may react with a lag. In this case, the weight of NWP should be increased. The trigger condition is that the standard deviation of irradiance exceeds a threshold within the last 15 minutes, for example, 100 W / m². Alternatively, the cloud cover change rate, for example, cloud cover changes exceeding 50% within the past 30 minutes. The fusion rule is to increase the weight of NWP, for example, .

[0144] (2) Stable high irradiance scenarios: Under clear, cloudless conditions with stable irradiance, STGCN can usually make accurate predictions and should be given a higher weight. The triggering condition is that the current irradiance is higher than the threshold, for example, 600 W / m², and the irradiance fluctuation is very small in the last 30 minutes, for example, the standard deviation is less than 50 W / m², and the cloud cover is less than 20%. The fusion rule is to reduce the weight of NWP by 1, for example, .

[0145] (3) Complex multi-cloud scenarios: Cloud cover is abundant and fluctuates, requiring a balance of STGCN and NWP weights. The trigger condition is cloud cover between 20% and 80% and moderate irradiance fluctuations, such as a standard deviation between 50-100 W / m². The fusion rule is a balanced weighting, for example, .

[0146] (4) Extreme weather events: such as heavy rain, hail, etc. In these cases, NWP may provide a more accurate trend, but the power may drop sharply, requiring special handling. The triggering condition is precipitation (rain / snow) or extremely low irradiance, such as below 100 W / m², and a continuous decrease. The fusion rules mainly rely on NWP, for example... It can also trigger extreme weather handling sub-models.

[0147] Based on real-time meteorological data, the current weather type is identified, and preset adaptive fusion rules are invoked to dynamically adjust the fusion weights of STGCN and NWP. Under clear and stable weather conditions, STGCN is given a higher weight to highlight its ability to capture short-term fluctuations; under cloudy or abrupt weather conditions, the weight of NWP is increased to enhance the overall stability of the forecast.

[0148] Perform correction: For each future time t+k within the prediction time range, calculate the adaptive weight for that time according to the fusion rule established in S52. and .

[0149] Execute weighted fusion: .

[0150] Imposing physical constraints: The final fused predictions are checked and corrected. A nighttime zero-value constraint is introduced, forcing the predicted power to be set to zero when the solar altitude angle is below a threshold. A theoretical limit constraint is also introduced to ensure... Not exceeding the power station's theoretical maximum installed capacity By introducing a ramp rate constraint, the rate of power change at adjacent time points is smoothed to avoid physically impossible large power jumps.

[0151] Please refer to Figure 13 Corresponding to the embodiments of the method of the present invention, a photovoltaic power prediction device based on localized meteorological interpolation is used to execute the method and mainly consists of the following parts:

[0152] Data acquisition module 1 is used to acquire real meteorological data, measured photovoltaic power and geographical location of distributed photovoltaic stations with complete meteorological data in the region as source stations, and to acquire measured photovoltaic power and geographical location of other distributed photovoltaic stations in the region that lack complete meteorological data or lack meteorological data as target stations.

[0153] The localized meteorological interpolation module 2 is used to construct a transfer learning model based on the real meteorological data, measured photovoltaic power and geographical location of the source station at each time period, and to use the relationship between photovoltaic power and meteorological variables as the regional localized meteorological interpolation model. The measured photovoltaic power of the target station at each time period is input into the regional localized meteorological interpolation model, and the relationship between photovoltaic power and meteorological variables in the model is learned to obtain the generated meteorological data corresponding to the measured photovoltaic power of the target station at each time period.

[0154] Data merging module 3 is used to construct a merged dataset, which includes the measured photovoltaic power and geographical location of all stations in the region at various times, as well as the real meteorological data of the source stations at various times and the generated meteorological data of the target stations at various times.

[0155] Module 4, which is used to establish a photovoltaic power prediction model, takes the merged dataset as input, and outputs the photovoltaic power prediction sequence of all distributed photovoltaic sites in the future period to obtain the prediction results;

[0156] Correction module 5 is used to construct a correction model based on the acquired meteorological forecast data for a future period, and to correct the output power prediction results for all distributed photovoltaic sites for the future period. Please refer to the above description in this embodiment for the specific correction process.

[0157] This embodiment improves the prediction results output by the STGC model by incorporating meteorological forecast data, effectively reducing prediction bias caused by changes in meteorological conditions and enhancing prediction accuracy. The corrected model dynamically adjusts the power values ​​in the prediction sequence by combining key parameters such as temperature, irradiance, and cloud cover from the meteorological forecast, exhibiting stronger adaptability, especially during periods of abrupt weather changes. The final output is a highly reliable, corrected photovoltaic power prediction result, providing strong support for grid dispatch.

[0158] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0159] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on localized meteorological interpolation, characterized in that, include: Using distributed photovoltaic sites with complete meteorological data in the region as source sites, we obtain the real meteorological data, measured photovoltaic power and geographical location of the source sites at various time periods. Using other distributed photovoltaic sites in the region that lack complete meteorological data or lack meteorological data as target sites, we obtain the measured photovoltaic power and geographical location of the target sites at various time periods. Based on real meteorological data, measured photovoltaic power and geographical location of the source site at various time periods, a transfer learning model is constructed using the relationship between photovoltaic power and meteorological variables as a regional localized meteorological interpolation model. The measured photovoltaic power of the target site at various time periods is input into the regional localized meteorological interpolation model to learn the relationship between photovoltaic power and meteorological variables in the model, thereby obtaining generated meteorological data corresponding to the measured photovoltaic power of the target site at various time periods. Construct a merged dataset that includes measured photovoltaic power and geographical location of all stations in the region at various time periods, as well as real meteorological data of source stations at various time periods and generated meteorological data of target stations at various time periods; Based on a spatiotemporal graph convolutional network (SPCRN) containing spatial convolutional layers, temporal convolutional layers, and fully connected layers, a spatiotemporal graph structure including weight matrix partitioning, explicit gating units, and regularization constraint mechanisms is constructed and placed before the spatial convolutional layers. This yields an improved SPCRN model based on a data source-aware evaluation mechanism, which serves as the prediction model. The merged dataset is aligned and features are extracted to obtain a merged feature set containing geographical features, power features, real meteorological features, and generated meteorological features. This merged feature set is then input into the improved SPCRN model. Based on the improved SPCRN model, the merged feature set is split according to its source, distinguishing between real and generated features. Different weight matrices are assigned to each type of feature for each site. A dynamic gating mechanism is set separately for generated meteorological features to dynamically evaluate their credibility, and the weight matrix of the generated meteorological features is dynamically adjusted based on the credibility. After weighted fusion, the meteorological-geographical-power fusion features for each site are obtained. Graph convolution and temporal convolution are performed to capture the spatial and temporal dependencies between sites, outputting the photovoltaic power prediction sequence for all distributed photovoltaic sites within a future period, thus obtaining the prediction results.

2. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 1, characterized in that, The steps involved, based on real meteorological data, measured photovoltaic power, and geographical location of the source station at various time periods, constructing a transfer learning model using the relationship between photovoltaic power and meteorological variables as a regional localized meteorological interpolation model, inputting the measured photovoltaic power of the target station at various time periods into the regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in the model, and obtaining the generated meteorological data corresponding to the measured photovoltaic power of the target station at various time periods, specifically include: After aligning the real meteorological data, measured photovoltaic power and geographical location of the source site and the measured photovoltaic power and geographical location of the target site for each time period, the meteorological characteristics and corresponding power characteristics of the source site and the power characteristics of the target site for each time period are extracted. A cross-domain mapping function is defined using the parameters of the transfer learning model, taking power features as input and meteorological features as output, to map power features to meteorological features; Meteorological characteristics and corresponding power characteristics of at least a portion of the time period of the source station are selected as the training set to train the cross-domain mapping function and obtain the trained cross-domain mapping function. The power features of the target site at different time periods are input into the trained cross-domain mapping function for prediction; The meteorological characteristics corresponding to the power characteristics of the target station at various time periods are predicted and generated, thereby obtaining the generated meteorological data.

3. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 2, characterized in that, The step of selecting meteorological features and corresponding power features of at least a portion of the time period of the source station as a training set for training the cross-domain mapping function and obtaining the trained cross-domain mapping function involves designing a joint loss function consisting of reconstruction loss and maximum mean difference loss during the training process, optimizing the joint loss function with the goal of minimizing the joint loss, iteratively updating the model parameters until the model converges, and completing the training.

4. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 1, characterized in that, The dynamic gating mechanism is as follows: the generated meteorological features are input into the multilayer perceptron. The multilayer perceptron outputs the credibility of the generated meteorological features based on the degree of similarity between the generated meteorological features and the real meteorological features. The output of the multilayer perceptron is output as a gating coefficient between 0 and 1 through the Sigmoid function. The weight matrix of the generated meteorological features is adjusted through the gating coefficient.

5. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 1, characterized in that, The step of using distributed photovoltaic stations with complete meteorological data in the region as source stations to obtain real meteorological data, measured photovoltaic power and geographical location of the source stations at various time periods specifically includes: using multiple distributed photovoltaic stations with complete meteorological data in the region as multiple source stations to obtain real meteorological data, measured photovoltaic power and geographical location of each source station at various time periods.

6. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 5, characterized in that, The step of constructing a transfer learning model as a regional localized meteorological interpolation model based on the real meteorological data, measured photovoltaic power and geographical location of each source station at each time period specifically includes: constructing a transfer learning model as a corresponding regional localized meteorological interpolation model based on the real meteorological data, measured photovoltaic power and geographical location of each source station at each time period, and utilizing the relationship between photovoltaic power and meteorological variables. There are multiple regional localized meteorological interpolation models, corresponding to the number of source stations.

7. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 6, characterized in that, The step of inputting the measured photovoltaic power of the target station at each time period into the regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in the model, and obtaining generated meteorological data corresponding to the measured photovoltaic power of the target station at each time period specifically includes: inputting the measured photovoltaic power of the target station at each time period into the regional localized meteorological interpolation model, learning the relationship between photovoltaic power and meteorological variables in each model, and outputting multiple sets of first generated meteorological data corresponding to the measured photovoltaic power of the target station at each time period; introducing an integrated gating network to dynamically weight and integrate the first generated meteorological data output by each localized meteorological interpolation model to obtain second generated meteorological data as the final generated meteorological data for subsequent steps.

8. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 1, characterized in that, Also includes: Based on the acquired meteorological forecast data for the future period, a correction model is constructed to correct the output power prediction results for all distributed photovoltaic sites for the future period.

9. The photovoltaic power prediction method based on localized meteorological interpolation according to claim 8, characterized in that, The step of constructing a correction model based on acquired meteorological forecast data for a future period, and correcting the output power prediction results for all distributed photovoltaic sites for that future period, specifically includes: The obtained meteorological forecast data for each time period are converted into reference photovoltaic power to obtain the reference photovoltaic power prediction sequence; An adaptive fusion corrector is constructed to dynamically evaluate the reliability of the photovoltaic power prediction sequence output by the improved spatiotemporal graph convolutional network model and the reference photovoltaic power prediction sequence converted from meteorological forecast data, and to perform weighted fusion. The model is then trained based on the reference photovoltaic power prediction sequence and the photovoltaic power prediction sequence from historical time periods. The photovoltaic power prediction sequence output by the improved spatiotemporal graph convolutional network model in the current time period and the reference photovoltaic power prediction sequence are input into the trained adaptive fusion corrector for correction to obtain the corrected photovoltaic power prediction sequence.

10. A photovoltaic power prediction device based on localized meteorological interpolation, used to perform the method described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire real meteorological data, measured photovoltaic power and geographical location of distributed photovoltaic stations with complete meteorological data in the region as source stations, and to acquire measured photovoltaic power and geographical location of other distributed photovoltaic stations in the region that lack complete meteorological data or lack meteorological data as target stations. The localized meteorological interpolation module is used to construct a transfer learning model based on the real meteorological data, measured photovoltaic power and geographical location of the source station at various time periods. This model is used as a regional localized meteorological interpolation model by utilizing the relationship between photovoltaic power and meteorological variables. The measured photovoltaic power of the target station at various time periods is input into the regional localized meteorological interpolation model, and the relationship between photovoltaic power and meteorological variables in the model is learned to obtain the generated meteorological data corresponding to the measured photovoltaic power of the target station at various time periods. The data merging module is used to build a merged dataset, which includes the measured photovoltaic power and geographical location of all stations in the region at various times, as well as the real meteorological data of the source stations at various times and the generated meteorological data of the target stations at various times. The photovoltaic power prediction module is used to construct a spatiotemporal graph structure, including weight matrix partitioning, explicit gating units, and regularization constraints, based on a spatiotemporal graph convolutional network containing spatial convolutional layers, temporal convolutional layers, and fully connected layers. This structure is placed before the spatial convolutional layer to obtain an improved spatiotemporal graph convolutional network model based on a data source-aware evaluation mechanism, which serves as the prediction model. The merged dataset is aligned and features are extracted to obtain a merged feature set containing geographical features, power features, real meteorological features, and generated meteorological features, which is then input into the improved spatiotemporal graph convolutional network model. Based on the improved spatiotemporal graph convolutional network model, the merged feature set is split according to its source, distinguishing between real and generated features. Different weight matrices are assigned to each type of feature for each site. A separate dynamic gating mechanism is set for the generated meteorological features to dynamically evaluate their credibility, and the weight matrix of the generated meteorological features is dynamically adjusted based on the credibility. After weighted fusion, the meteorological-geographical-power fused features for each site are obtained. Graph convolution and temporal convolution are performed to capture the spatial and temporal dependencies between sites, outputting the photovoltaic power prediction sequence for all distributed photovoltaic sites within a future period, thus obtaining the prediction results.

Images