Regional distributed photovoltaic probabilistic prediction method, system, device and storage medium

Abstract

A regional distributed photovoltaic probability prediction method, system, device and storage medium, the method comprises: acquiring multi-source data of a target region in a preset historical period; dividing the target region into multiple sub-regions according to a preset regional division rule; clustering and down-sampling photovoltaic data points in each sub-region to generate an initial feature tensor with N vertices; constructing adjacent edges based on the physical distance and output similarity between vertices to form a graph topology structure; inputting the graph topology structure and its corresponding initial feature tensor into a graph attention network; extracting wide-area spatio-temporal correlation features of the multiple sub-regions through the graph attention network, and fusing the wide-area spatio-temporal correlation features to output high-order features after fusion; inputting the high-order features into a multi-task learning network constructed based on deep quantile regression to synchronously output photovoltaic output probability prediction intervals of each sub-region at a set confidence level. The present application improves the accuracy and reliability of regional distributed photovoltaic output probability prediction.

Description

Technical Field

[0001] This invention belongs to the field of new energy power prediction technology, specifically relating to a regional distributed photovoltaic probability prediction method, system, equipment and storage medium. Background Technology

[0002] As an important renewable energy source, photovoltaic (PV) capacity continues to climb in the power system. Distributed PV, in particular, is expected to surpass centralized PV in installed capacity due to its lower installation costs and technical complexity, making it easier to widely adopt. However, PV power exhibits strong periodicity and morning / evening ramp-up characteristics, posing significant challenges to the safe and stable operation of the power system and power balance dispatch decisions. Furthermore, the midday backfeeding problem of distributed PV power also greatly challenges the lifespan of grid substation equipment. Currently, methods for predicting distributed PV power are mainly divided into two categories: point prediction methods and probabilistic prediction methods. Point prediction methods, such as those based on physical models, statistical learning, or deep neural networks, can provide a single estimated value but cannot reflect uncertainty. Probabilistic prediction methods can be further subdivided into parametric methods (such as quantile regression QR), non-parametric methods (such as kernel density estimation KDE), and generative models (such as variational autoencoders (VAE), generative adversarial networks (GAN), and denoising diffusion probabilistic models (DDPM). Among them, quantile regression QR achieves interval prediction by assuming the shape of the power distribution and regressing quantiles at different confidence levels; kernel density estimation KDE and methods based on historical error statistics do not require a pre-defined distribution, but have limited ability to fit the distribution of tail data; variational autoencoders (VAE), generative adversarial networks (GAN), and denoising diffusion probabilistic models (DDPM) directly generate multi-scene samples from noise, gradually enhancing their ability to fit complex distributions.

[0003] However, the existing forecasting technologies mentioned above still have significant shortcomings in addressing the refined scheduling requirements brought about by the high proportion of distributed photovoltaic (PV) grid integration. Specifically, existing forecasting methods struggle to deeply integrate widely distributed numerical weather forecast information with the complex spatiotemporal correlation characteristics between distributed PV power plants, resulting in an inaccurate grasp of the overall regional power output trend. Furthermore, the power output characteristics of distributed PV vary significantly under different meteorological scenarios (such as sunny and cloudy days), and most existing models employ a uniform architecture, making it difficult to effectively characterize this shift in power output distribution caused by changes in meteorological conditions. This makes it difficult to balance the coverage and sharpness of the probability prediction interval, failing to provide sufficient and reliable decision support for guaranteed scheduling. Summary of the Invention

[0004] The purpose of this invention is to address the problems in the prior art by providing a regional distributed photovoltaic probability prediction method, system, device, and storage medium. It automatically captures the deep correlation between distributed photovoltaic areas and meteorological elements through an adaptive attention mechanism, constructs differentiated probability prediction models for different meteorological scenarios, and sets independent task branches for each prediction area and performs joint optimization, thereby significantly improving the accuracy and reliability of regional distributed photovoltaic power output probability prediction.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a method for probabilistic prediction of regional distributed photovoltaic power generation is provided, including: Acquire multi-source data of the target area within a preset historical time period and generate a standardized data tensor set; According to the preset regional division rules, the target area is divided into multiple regions. Based on the standardized data tensor set, the photovoltaic data points in each region are clustered and downsampled to generate an initial feature tensor with N vertices. Adjacent edges are constructed based on the physical distance and output similarity between each vertex to form a graph topology structure that represents the spatial relationship between multiple regions. The graph topology and its corresponding initial feature tensor are input into the graph attention network. The graph attention network extracts the wide-area spatiotemporal correlation features of multiple regions, and the wide-area spatiotemporal correlation features are fused to output the fused high-order features. High-order features are input into a multi-task learning network based on deep quantile regression, and the probability prediction range of photovoltaic output for each region under a set confidence level is output simultaneously.

[0006] As a preferred embodiment, the step of acquiring multi-source data of the target area within a preset historical time period and generating a standardized data tensor set includes: Numerical weather forecast data, historical operation data of distributed photovoltaic power, and static information data of the target area within a preset historical period are collected; wherein, the static information data includes installed capacity information, location distribution information, historical actual power information of the dispatch terminal, and load forecast information recorded by the dispatch terminal. The multi-source data is preprocessed, and the preprocessed data is unified to a 1-hour time resolution. After alignment, a standardized data tensor set is generated.

[0007] As a preferred embodiment, the method further includes the steps of performing weather classification feature analysis based on a standardized data tensor set and labeling different weather conditions as scenarios. Specifically, a weather classification recognition model is trained based on the photovoltaic output curve morphology features and corresponding gridded meteorological element features in the standardized data tensor set. The weather classification recognition model is then used to label different weather conditions as scenarios, resulting in weather scenario labels. The scenario labels include at least sunny and cloudy labels. The weather scenario labels are used as additional features of vertices in the graph topology or as conditional inputs to the multi-task learning network to guide differentiated modeling under different meteorological scenarios.

[0008] As a preferred embodiment, the step of constructing adjacent edges based on the physical distance and force similarity between vertices includes: The spatial distance between any two vertices is calculated, and the Pearson correlation coefficient between the historical photovoltaic power output sequences corresponding to any two vertices is calculated as the power output similarity. The physical distance and the similarity of output are weighted and summed. When the result is less than a preset threshold, an adjacent edge is constructed between the corresponding two vertices.

[0009] As a preferred embodiment, the graph attention network includes multiple graph attention layers. Each graph attention layer adaptively calculates the attention weights between each vertex and its neighboring vertices through an attention mechanism, and aggregates the feature information of neighboring vertices according to the attention weights to update the feature representation of the current vertex, and outputs the higher-order features.

[0010] As a preferred embodiment, the multi-task learning network sets up a prediction task branch for each region; The multi-task learning network includes a shared input layer and a feature extraction network, as well as multiple parallel task-specific output layers. Each task-specific output layer corresponds to a region and is used to output the photovoltaic power output prediction value of that region at multiple preset quantiles. The multiple preset quantiles are used to form the probability prediction interval at the set confidence level.

[0011] As a preferred embodiment, the multi-task learning network is jointly trained and optimized using a multi-task bouncing loss function, which is the sum of the quantile loss functions of each region; the quantile loss function is used to measure the deviation between the predicted value and the true value at each preset quantile.

[0012] As a preferred option, the preset regional division rules include administrative division boundary division rules or location-based boundary division rules.

[0013] Secondly, a regional distributed photovoltaic probabilistic prediction system is provided, comprising: The data acquisition module is used to acquire multi-source data of the target area within a preset historical period and generate a standardized data tensor set; The graph data construction module is used to divide the target area into multiple regions according to the preset regional division rules. Based on the standardized data tensor set, the photovoltaic data points in each region are clustered and downsampled to generate an initial feature tensor with N vertices. Adjacent edges are constructed based on the physical distance and output similarity between each vertex to form a graph topology structure that represents the spatial relationship between multiple regions. The spatial feature extraction module is used to input the graph topology and its corresponding initial feature tensor into the graph attention network, extract the wide-area spatiotemporal correlation features of multiple regions through the graph attention network, fuse the wide-area spatiotemporal correlation features, and output the fused high-order features. The multi-task probability prediction module is used to input high-order features into a multi-task learning network built on deep quantile regression, and simultaneously output the photovoltaic output probability prediction interval for each region at a set confidence level.

[0014] As a preferred embodiment, the multi-source data acquired by the data acquisition module includes numerical weather forecast data, historical operation data of distributed photovoltaic power generation, and static information data; The static information data includes installed capacity information, location distribution information, historical actual power information of the dispatch terminal, and load forecast information recorded by the dispatch terminal; The data acquisition module preprocesses the multi-source data and unifies the preprocessed data to a 1-hour time resolution. After alignment, it generates a standardized data tensor set.

[0015] As a preferred embodiment, the system also includes a weather classification module for performing weather classification feature analysis based on a standardized data tensor set and labeling different weather conditions as scenarios. The weather classification module trains a weather classification recognition model based on the photovoltaic output curve morphology features and corresponding gridded meteorological element features in the standardized data tensor set. The weather classification recognition model is then used to label different weather conditions as scenarios, resulting in weather scenario labels. These scenario labels include at least sunny and cloudy labels. The weather scenario labels are used as additional features of vertices in the graph topology or as conditional inputs to a multi-task learning network to guide differentiated modeling under different meteorological scenarios.

[0016] As a preferred embodiment, when constructing adjacent edges based on the physical distance and output similarity between vertices, the graph data construction module calculates the geographical spatial distance between every two vertices and the Pearson correlation coefficient between the photovoltaic historical output sequences corresponding to every two vertices as the output similarity; the physical distance and output similarity are weighted and summed, and when the result is less than a preset threshold, adjacent edges are constructed between the corresponding two vertices.

[0017] As a preferred embodiment, the graph attention network used by the spatial feature extraction module includes multiple graph attention layers. Each graph attention layer adaptively calculates the attention weight between each vertex and its neighboring vertices through an attention mechanism, and aggregates the feature information of neighboring vertices according to the attention weights to update the feature representation of the current vertex, and outputs the higher-order features.

[0018] As a preferred embodiment, the multi-task learning network in the multi-task probability prediction module sets up a prediction task branch for each region; The multi-task learning network includes a shared input layer and a feature extraction network, as well as multiple parallel task-specific output layers. Each task-specific output layer corresponds to a region and is used to output the photovoltaic power output prediction value of that region at multiple preset quantiles. The multiple preset quantiles are used to form the probability prediction interval at the set confidence level.

[0019] As a preferred embodiment, the multi-task learning network in the multi-task probability prediction module is jointly trained and optimized using a multi-task bouncing ball loss function, which is the sum of the quantile loss functions of each region; the quantile loss function is used to measure the deviation between the predicted value and the true value at each preset quantile.

[0020] As a preferred embodiment, when the graph data construction module divides the target area into multiple regions according to preset regional division rules, the preset regional division rules include administrative division rules or location distribution boundary rules.

[0021] Thirdly, an electronic device is provided, including a processor and a memory, wherein the processor is configured to execute a computer program stored in the memory to implement the regional distributed photovoltaic probability prediction method.

[0022] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the regional distributed photovoltaic probability prediction method.

[0023] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects: The proposed method for probabilistic prediction of regional distributed photovoltaic (PV) power generation involves acquiring multi-source data of a target region within a preset historical time period and generating a standardized data tensor set. Based on this, the target region is divided into multiple zones according to preset regional division rules. Using the standardized data tensor set, PV data points within each zone are clustered and downsampled to generate an initial feature tensor with N vertices. Adjacency edges are constructed based on the physical distance and output similarity between vertices, forming a graph topology representing the spatial relationships between multiple zones. This abstracts the dispersed PV zones into vertices and their relationships in a graph neural network, laying the foundation for the fusion of wide-area spatial information. The fundamental approach involves inputting the graph topology and its corresponding initial feature tensors into a graph attention network (GAT) to extract wide-area spatiotemporal correlation features of multiple regions under weather conditions and outputting fused high-order features. This utilizes the graph attention mechanism to adaptively mine the deep nonlinear correlation between meteorological elements and photovoltaic (PV) output, effectively overcoming the shortcomings of traditional methods in capturing complex spatiotemporal dependencies. The high-order features are then input into a multi-task learning network based on deep quantile regression, simultaneously outputting the PV output probability prediction interval for each region at a set confidence level. This allows for independent modeling and joint optimization of each region, balancing regional differences and overall generalization. In summary, the proposed regional distributed PV probability prediction method significantly improves the accuracy and reliability of regional distributed PV output probability prediction, effectively solving the technical problems of existing technologies that fail to fully integrate wide-area spatiotemporal features and have poor adaptability to multiple meteorological scenarios.

[0024] Furthermore, numerical weather prediction data, historical operating data of distributed photovoltaic power generation, and static information data are collected, and the time resolution of the photovoltaic power output curve and gridded meteorological elements is unified to one hour later to generate a standardized data tensor set. This invention, through targeted resolution unification and alignment processing of multi-source heterogeneous data, not only ensures the matching and consistency of various data on the time scale but also preserves the complete information of each data source. This provides a high-quality, standardized data foundation for subsequent weather classification identification and GAT spatial feature extraction, ensuring the reliability and accuracy of the entire prediction process.

[0025] Furthermore, the static information data includes installed capacity information, location distribution information, historical actual power information and load forecast information recorded by the dispatching terminal. By introducing the existing multi-dimensional static information of the dispatching terminal as an auxiliary data source, the information dimensions of the data tensor set are enriched without increasing the additional collection cost. This provides the model with a more comprehensive description of the regional photovoltaic characteristics, which helps to improve the prediction model's ability to understand the overall power output characteristics of the region, thereby improving the prediction accuracy.

[0026] Furthermore, a weather classification model is trained based on photovoltaic power output curves and gridded meteorological elements, and different weather conditions such as sunny and cloudy days are labeled. This invention jointly models the dynamic characteristics of photovoltaic power output with the numerical characteristics of meteorological elements, making the weather classification results more consistent with the actual needs of photovoltaic power prediction. Different scene labels can provide clear scene prior knowledge for subsequent GAT feature extraction and multi-task prediction models, guiding the model to extract key features in a targeted manner under different meteorological conditions, thereby significantly improving the model's adaptability and prediction accuracy under multiple scenarios.

[0027] Furthermore, by using physical distance and similarity as the basis for constructing adjacent edges, this invention clarifies the construction criteria for adjacent edges in the graph topology. By simultaneously considering the proximity of geographical locations and the similarity of photovoltaic power output behavior, it can more scientifically and comprehensively reflect the true spatial correlation strength between different areas, making the constructed graph topology more consistent with the physical mechanism of distributed photovoltaic regional prediction, thereby ensuring the effectiveness and accuracy of subsequent graph attention network feature extraction.

[0028] Furthermore, this invention employs a graph attention network (GAT) to capture wide-area high-dimensional feature associations in numerical weather prediction, constructs high-order fusion features and embeds them into vertices, and utilizes an attention mechanism to adaptively focus on meteorological information sources that are more relevant to the prediction target. This enables the extraction and fusion of deep high-dimensional features in wide-area numerical weather prediction, further enhancing the model's ability to characterize regional photovoltaic power output characteristics under complex meteorological conditions and effectively improving prediction accuracy.

[0029] Furthermore, this invention constructs a unified multi-task bouncing loss function for each prediction region and jointly optimizes it, enabling prediction tasks in different regions to collaborate and share beneficial information during training. At the same time, by utilizing the natural adaptability of the bouncing loss function to quantile regression, the accuracy and reliability of probability prediction intervals at each confidence level are ensured, thereby improving the generalization ability and robustness of the entire regional prediction system.

[0030] Furthermore, the regional division rules of the present invention include administrative division rules or location distribution boundary division rules, which can flexibly adapt to the needs of different management granularities and actual application scenarios, enhance the operability and applicability of the method, and facilitate the engineering deployment of distributed photovoltaic forecasting in different regions.

[0031] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 Flowchart of the regional distributed photovoltaic probability prediction method according to an embodiment of the present invention; Figure 2 The probability prediction results of distributed photovoltaic power generation in partition 01 according to an embodiment of the present invention are shown in the figure. Figure 3 A graph showing the probability prediction results for all distributed photovoltaic systems in the entire region according to an embodiment of the present invention; Figure 4(a) Evaluation index of quantile deviation of distributed photovoltaic probability prediction in an embodiment of the present invention; Figure 4(b) Evaluation index of confidence interval deviation for distributed photovoltaic probability prediction in an embodiment of the present invention; Figure 4(c) Evaluation index of confidence interval width for distributed photovoltaic probability prediction in an embodiment of the present invention; Figure 4(d) Evaluation index of bouncing loss fraction corresponding to the quantile of distributed photovoltaic probability prediction in the embodiment of the present invention; Figure 4(e) Winkler loss score evaluation index of the confidence interval corresponding to the distributed photovoltaic probability prediction in the embodiment of the present invention; Figure 4(f) Box plot of loss fraction distribution of distributed photovoltaic probability prediction bouncing ball in an embodiment of the present invention. Detailed Implementation

[0034] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0035] Please see Figure 1 This invention proposes a regional distributed photovoltaic probability prediction method, which mainly includes: S1. Obtain multi-source data of the target area within a preset historical time period and generate a standardized data tensor set; S2. Divide the target area into multiple regions according to the preset regional division rules. Based on the standardized data tensor set, cluster and downsample the photovoltaic data points in each region to generate an initial feature tensor with N vertices. Construct adjacent edges based on the physical distance and output similarity between each vertex to form a graph topology structure that represents the spatial relationship between multiple regions. S3. Input the graph topology and its corresponding initial feature tensor into the graph attention network, extract the wide-area spatiotemporal correlation features of multiple regions through the graph attention network, fuse the wide-area spatiotemporal correlation features, and output the fused high-order features. S4. Input the high-order features into a multi-task learning network constructed based on deep quantile regression, and simultaneously output the photovoltaic output probability prediction interval for each region under a set confidence level.

[0036] In one possible implementation, step S1 collects numerical weather forecast data, historical operation data of distributed photovoltaic power generation, and static information data of the target area within a preset historical period; wherein, the static information data includes installed capacity information, location distribution information, historical actual power information of the dispatch terminal, and load forecast information, etc., recorded by the dispatch terminal. Furthermore, the collected multi-source data is preprocessed, and the preprocessed data is unified to a 1-hour time resolution. After alignment, a standardized data tensor set is generated.

[0037] In one possible implementation, the method further includes the step of performing weather classification feature analysis based on a standardized data tensor set and scene labeling for different weather conditions. Specifically, in this step, a weather classification recognition model is trained based on the morphological features of the photovoltaic output curve in the standardized data tensor set and the corresponding gridded meteorological element features; the weather classification recognition model is used to label different weather conditions to obtain weather scene labels; the scene labels include at least sunny and cloudy labels, and different numbers are assigned to different weather conditions such as sunny and cloudy; the weather scene labels are used as additional features of vertices in the graph topology or as conditional inputs to the multi-task learning network to guide differentiated modeling under different meteorological scenarios.

[0038] In one possible implementation, when constructing adjacent edges based on the physical distance and power output similarity between vertices in step S2, the geographical spatial distance between each pair of vertices is calculated, and the Pearson correlation coefficient between the photovoltaic historical power output sequences corresponding to each pair of vertices is calculated as the power output similarity; the physical distance and power output similarity are weighted and summed, and when the result is less than a preset threshold, adjacent edges are constructed between the corresponding two vertices.

[0039] In one possible implementation, the graph attention network in step S3 includes multiple graph attention layers. Each graph attention layer adaptively calculates the attention weight between each vertex and its neighboring vertices through an attention mechanism, and aggregates the feature information of the neighboring vertices according to the attention weights to update the feature representation of the current vertex, and outputs the higher-order features.

[0040] In one possible implementation, the multi-task learning network in step S4 sets up a prediction task branch for each region; the multi-task learning network includes a shared input layer and feature extraction network, as well as multiple parallel task-specific output layers, each of the task-specific output layers corresponding to one region, used to output the photovoltaic power output prediction value of the region under multiple preset quantiles, the multiple preset quantiles being used to form the probability prediction interval under the set confidence level.

[0041] Furthermore, the multi-task learning network is jointly trained and optimized using a multi-task bouncing loss function, which is the sum of the quantile loss functions of each region; the quantile loss function is used to measure the deviation between the predicted value and the true value at each preset quantile.

[0042] In one possible implementation, the preset regional division rules in step S2 include administrative division boundary division rules or location-based distribution boundary division rules, and a clustering downsampling method is used for the region to construct an initial feature tensor with N vertices, represented as N×Feature_Num (number of vertices × number of features).

[0043] First, the regional distributed photovoltaic probability prediction method proposed in this embodiment of the invention addresses the problem that conventional distributed photovoltaic probability prediction methods cannot fully integrate the complex spatiotemporal features of wide-area data. It adopts a graph attention network, which uses the graph attention mechanism to adaptively allocate weights and adjust the graph topology to extract the spatiotemporal features of wide-area numerical weather forecasts.

[0044] Furthermore, the regional distributed photovoltaic probability prediction method proposed in this embodiment of the invention addresses the problem of significant differences in the overall output characteristics of regional distributed photovoltaic under multiple meteorological scenarios by introducing meteorological scenario classification labels, constructing and learning models for multiple scenarios separately, and extracting key features of different scenarios for targeted learning and modeling.

[0045] Taking distributed photovoltaic (PV) power generation in a certain region as the research object, the method of this invention is applied to the probability prediction of distributed PV power generation in this region. The region is divided into two distributed PV areas for probability prediction. The prediction of various indicators of distributed PV power generation in the whole region is evaluated. Since PV power generation is not available at night, the nighttime portion is excluded from the statistics.

[0046] Among them, the quantile deviations corresponding to the upper and lower limits of the intervals of 95%, 90%, and 85% are all no more than 5%, the interval deviations are no more than 10%, the average interval width is no more than 35% of the installed capacity, the bouncing loss of the upper and lower limit quantiles of each interval is about 0.01, and the Winkler loss is less than 0.35.

[0047] As can be seen from the evaluation indicators above, the method of the present invention has a relatively stable prediction effect on the probability prediction of regional distributed photovoltaic. Under the current condition of a lack of distributed photovoltaic site information, relying solely on numerical weather forecasts and historical distributed photovoltaic data from the dispatch terminal to give probability prediction results can serve as a method for evaluating the probability characteristics of regional distributed photovoltaic power generation sites.

[0048] The evaluation indicators are calculated as follows: 1) Quantile deviation :

[0049] in, It is the empirical percentile rank, calculated using the following formula:

[0050] The indicative function is defined as follows:

[0051] It is the target quantile. It is the first The true value of each time series It is the first Each time sequence Quantile predictions This represents the total number of time series.

[0052] Quantile deviation can be used to evaluate the prediction performance of the corresponding quantile. A deviation greater than 0 indicates that the prediction is too high, while a deviation less than 0 indicates that the prediction is too low.

[0053] 2) Interval deviation :

[0054] in, It is the experience coverage rate, calculated using the following formula:

[0055] These are the upper and lower limit quantiles of the corresponding confidence interval.

[0056] Interval bias can be used to evaluate the predictive performance of probability prediction confidence intervals. A bias greater than 0 indicates that the prediction interval width is too large and the interval prediction is too conservative. Conversely, a bias less than 0 indicates that the prediction interval width is too small and some tail distributions are not fully considered by the model.

[0057] 3) Loss of the ball :

[0058] The smaller the ball's loss, the better the prediction effect.

[0059] 4) Winkler score:

[0060] The Winkler score is a loss score; the lower the score, the less information is lost in the probability prediction result, and the better the prediction effect.

[0061] Please see Figure 2 as well as Figure 3 , Figure 2 The results of the distributed photovoltaic probability prediction for partition 01 are presented. Figure 2 For photovoltaic power output, a normalization factor of 1.05 times the maximum output value was used for normalization. Figure 3 This presents the probabilistic prediction results for all distributed photovoltaic systems in the entire region, using the same method as... Figure 2 Using the same normalization method, a comparison shows that, for the entire region, due to the wider coverage of meteorological forecast information, the probability prediction interval for the two uphill climbs in the morning and evening is more able to include the actual output within the interval.

[0062] The statistical indicators for distributed photovoltaic power generation in the region are shown in Figures 4(a) to 4(f): Figure 4(a) shows the distribution of quantile deviation under different quantile conditions. Quantile deviation is used to evaluate the prediction effect of the corresponding quantile; a deviation greater than 0 indicates that the prediction is too high, and a deviation less than 0 indicates that the prediction is too low. As can be seen from the figure, under different quantile conditions, the quantile deviation is controlled within a small range, indicating that the method of the present invention does not systematically overestimate or underestimate the prediction of each quantile and has good unbiasedness.

[0063] Figure 4(b) shows the distribution of interval bias under different confidence interval conditions. Interval bias is used to evaluate the predictive performance of probability prediction confidence intervals. A bias greater than 0 indicates that the prediction interval width is too large and the prediction is too conservative, while a bias less than 0 indicates that the prediction interval width is too small and some tail distributions are not fully considered in the model. As can be seen from the figure, the interval bias remains at a low level at all confidence levels, indicating that the prediction interval constructed by the method of this invention achieves a good balance between coverage and sharpness.

[0064] Figure 4(c) shows the percentage of average interval width to installed capacity under different confidence intervals. The average interval width reflects the degree of uncertainty quantification in the probabilistic prediction interval; the smaller the interval width, the more accurate and certain the prediction. As can be seen from the figure, the average interval width does not exceed 35% of the installed capacity, indicating that the method of the present invention can provide a relatively compact prediction interval while ensuring coverage, providing probabilistic information with clear reference value for scheduling decisions.

[0065] Figure 4(d) shows the trend of the bouncing loss score under different quantile conditions. As can be seen from the figure, the bouncing loss score increases with the increase of the quantile, which conforms to the basic law of quantile regression—the closer the quantile is to the tail of the distribution (such as the 95th percentile), the greater the prediction difficulty and the higher the corresponding loss value. Under each quantile condition, the bouncing loss score remains at a low level and changes smoothly without drastic fluctuations or abnormal jumps, indicating that the method of the present invention can maintain stable and accurate prediction performance at different quantiles, verifying the reliability and robustness of the method of the present invention in probabilistic prediction quantile regression tasks. Furthermore, the bouncing loss score shown in the figure is approximately 0.01 at some typical quantiles, indicating that the deviation between the predicted value and the true value of each quantile regression is small, and the prediction effect is excellent.

[0066] Figure 4(e) shows the distribution of Winkler loss scores at different confidence levels. Winkler loss is a scoring function that comprehensively considers the coverage and width of the prediction interval; a lower score indicates less information loss and better prediction performance. As can be seen from the figure, the Winkler loss is below 0.35, indicating a low level. This suggests that the probability prediction interval provided by the method of this invention maintains a reasonable interval width while covering the true value, resulting in minimal overall loss of prediction information and excellent prediction performance.

[0067] Figure 4(f) is a box plot of the bouncing loss score, used to show the distribution and dispersion of the bouncing loss score under different grouping conditions. The horizontal axis of the figure is the group label (such as different confidence configurations or different area prediction tasks), and the vertical axis is the bouncing loss score. Each box in the figure reflects the central tendency and dispersion of the bouncing loss score through the interquartile range. The horizontal line inside the box represents the median, and the upper and lower edges of the box correspond to the upper and lower quartiles, respectively. The upper and lower beards extend to the extreme values ​​within the non-outlier range, and points outside the beard range are marked as outliers. As can be seen from the figure, the overall height of each box is relatively low, indicating that the bouncing loss score is concentrated in the lower value range under different grouping conditions, with small dispersion. The median line is relatively central or slightly lower in each box, indicating that the bouncing loss of most prediction results is lower than the mean, and the prediction accuracy is good. The number of outliers is limited and the deviation is not large, indicating that the method of the present invention can maintain stable and reliable prediction performance in most prediction tasks, with only slight fluctuations in a very few cases, and strong overall robustness.

[0068] Figures 4(d) and 4(f) jointly evaluate the bouncing loss performance of the method of the present invention from two different dimensions: Figure 4(d) depicts the trend of bouncing loss with quantiles, reflecting the systematic performance of the model at different quantiles; Figure 4(f) depicts the distribution of bouncing loss under different groupings, reflecting the consistency and stability of the model's prediction results. Figures 4(d) and 4(f) fully present the excellent performance of the method of the present invention in terms of both quantile prediction accuracy and overall prediction robustness.

[0069] The above indicators collectively demonstrate that the method of this invention has good probabilistic prediction performance at 95%, 90%, and 85% confidence levels, and can provide a reliable technical means for assessing the probabilistic characteristics of regional distributed photovoltaic power generation by relying solely on numerical weather forecasts and historical data from the dispatching terminal.

[0070] Another embodiment of the present invention also proposes a regional distributed photovoltaic probabilistic prediction system, comprising: The data acquisition module is used to acquire multi-source data of the target area within a preset historical period and generate a standardized data tensor set; The graph data construction module is used to divide the target area into multiple regions according to the preset regional division rules. Based on the standardized data tensor set, the photovoltaic data points in each region are clustered and downsampled to generate an initial feature tensor with N vertices. Adjacent edges are constructed based on the physical distance and output similarity between each vertex to form a graph topology structure that represents the spatial relationship between multiple regions. The spatial feature extraction module is used to input the graph topology and its corresponding initial feature tensor into the graph attention network, extract the wide-area spatiotemporal correlation features of multiple regions through the graph attention network, fuse the wide-area spatiotemporal correlation features, and output the fused high-order features. The multi-task probability prediction module is used to input high-order features into a multi-task learning network built on deep quantile regression, and simultaneously output the photovoltaic output probability prediction interval for each region at a set confidence level.

[0071] In one possible implementation, the multi-source data acquired by the data acquisition module includes numerical weather forecast data, historical operation data of distributed photovoltaic power generation, and static information data; The static information data includes installed capacity information, location distribution information, historical actual power information of the dispatch terminal, and load forecast information recorded by the dispatch terminal.

[0072] In one possible implementation, the data acquisition module preprocesses the multi-source data and unifies the preprocessed data to a 1-hour time resolution, then generates a standardized data tensor set after alignment.

[0073] In one possible implementation, a weather classification module is further included, used to perform weather classification feature analysis based on a standardized data tensor set and to label different weather conditions as scenarios. The weather classification module trains a weather classification recognition model based on the photovoltaic output curve morphology features and corresponding gridded meteorological element features in the standardized data tensor set. The weather classification recognition model is used to label different weather conditions as scenarios to obtain weather scenario labels. The scenario labels include at least sunny day labels and cloudy day labels. The weather scenario labels are used as additional features of vertices in the graph topology or as conditional inputs to a multi-task learning network to guide differentiated modeling under different meteorological scenarios.

[0074] In one possible implementation, when the graph data construction module constructs adjacent edges based on the physical distance and output similarity between each vertex, it calculates the geographic spatial distance between each pair of vertices and the Pearson correlation coefficient between the photovoltaic historical output sequences corresponding to each pair of vertices as the output similarity; it performs a weighted summation of the physical distance and output similarity, and when the result is less than a preset threshold, it constructs an adjacent edge between the corresponding two vertices.

[0075] In one possible implementation, the graph attention network used by the spatial feature extraction module includes multiple graph attention layers. Each graph attention layer adaptively calculates the attention weight between each vertex and its neighboring vertices through an attention mechanism, and aggregates the feature information of neighboring vertices according to the attention weights to update the feature representation of the current vertex, and outputs the higher-order features.

[0076] In one possible implementation, the multi-task learning network in the multi-task probability prediction module sets up a prediction task branch for each region. The multi-task learning network includes a shared input layer and a feature extraction network, as well as multiple parallel task-specific output layers. Each task-specific output layer corresponds to a region and is used to output the photovoltaic power output prediction value of that region at multiple preset quantiles. The multiple preset quantiles are used to form the probability prediction interval at the set confidence level.

[0077] In one possible implementation, the multi-task learning network in the multi-task probability prediction module is jointly trained and optimized using a multi-task bouncing ball loss function, which is the sum of the quantile loss functions of each region; the quantile loss function is used to measure the deviation between the predicted value and the true value at each preset quantile.

[0078] In one possible implementation, when the graph data construction module divides the target area into multiple regions according to a preset regional division rule, the preset regional division rule includes administrative division rules or location distribution boundary rules.

[0079] Another embodiment of the present invention also proposes an electronic device, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the regional distributed photovoltaic probability prediction method.

[0080] Another embodiment of the present invention also proposes a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the regional distributed photovoltaic probability prediction method.

[0081] The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals. For ease of explanation, the above content only shows the parts related to the embodiments of the present invention; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. This computer-readable storage medium is non-transitory and can be stored in storage devices formed by various electronic devices, enabling the execution process described in the method of the embodiments of the present invention.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for probabilistic prediction of regional distributed photovoltaic power, characterized in that, include: Acquire multi-source data of the target area within a preset historical time period and generate a standardized data tensor set; According to the preset regional division rules, the target area is divided into multiple regions. Based on the standardized data tensor set, the photovoltaic data points in each region are clustered and downsampled to generate an initial feature tensor with N vertices. Adjacent edges are constructed based on the physical distance and output similarity between each vertex to form a graph topology structure that represents the spatial relationship between multiple regions. The graph topology and its corresponding initial feature tensor are input into the graph attention network. The graph attention network extracts the wide-area spatiotemporal correlation features of multiple regions, and the wide-area spatiotemporal correlation features are fused to output the fused high-order features. High-order features are input into a multi-task learning network based on deep quantile regression, and the probability prediction range of photovoltaic output for each region under a set confidence level is output simultaneously.

2. The regional distributed photovoltaic probabilistic prediction method according to claim 1, characterized in that, The step of acquiring multi-source data of the target area within a preset historical time period and generating a standardized data tensor set includes: Numerical weather forecast data, historical operation data of distributed photovoltaic power, and static information data of the target area within a preset historical period are collected; wherein, the static information data includes installed capacity information, location distribution information, historical actual power information of the dispatch terminal, and load forecast information recorded by the dispatch terminal. The multi-source data is preprocessed, and the preprocessed data is unified to a 1-hour time resolution. After alignment, a standardized data tensor set is generated.

3. The regional distributed photovoltaic probabilistic prediction method according to claim 1, characterized in that, It also includes the steps of performing weather classification feature analysis based on a standardized data tensor set and labeling different weather conditions as scenarios; specifically, a weather classification recognition model is trained based on the photovoltaic output curve morphology features and corresponding gridded meteorological element features in the standardized data tensor set; the weather classification recognition model is used to label different weather conditions as scenarios to obtain weather scenario labels; the scenario labels include at least sunny day labels and cloudy day labels; the weather scenario labels are used as additional features of vertices in the graph topology or as conditional inputs to the multi-task learning network to guide differentiated modeling under different meteorological scenarios.

4. The regional distributed photovoltaic probabilistic prediction method according to claim 1, characterized in that, The steps for constructing adjacent edges based on the physical distance and output similarity between vertices include: The spatial distance between any two vertices is calculated, and the Pearson correlation coefficient between the historical photovoltaic power output sequences corresponding to any two vertices is calculated as the power output similarity. The physical distance and the similarity of output are weighted and summed. When the result is less than a preset threshold, an adjacent edge is constructed between the corresponding two vertices.

5. The regional distributed photovoltaic probabilistic prediction method according to claim 1, characterized in that, The graph attention network includes multiple graph attention layers. Each graph attention layer adaptively calculates the attention weight between each vertex and its neighboring vertices through an attention mechanism, and aggregates the feature information of the neighboring vertices according to the attention weights to update the feature representation of the current vertex, and outputs the higher-order features.

6. The regional distributed photovoltaic probabilistic prediction method according to claim 1, characterized in that, The multi-task learning network sets up a prediction task branch for each region; The multi-task learning network includes a shared input layer and a feature extraction network, as well as multiple parallel task-specific output layers. Each task-specific output layer corresponds to a region and is used to output the photovoltaic power output prediction value of that region at multiple preset quantiles. The multiple preset quantiles are used to form the probability prediction interval at the set confidence level.

7. The regional distributed photovoltaic probabilistic prediction method according to claim 6, characterized in that, The multi-task learning network is jointly trained and optimized using a multi-task bouncing loss function, which is the sum of the quantile loss functions of each region. The quantile loss function is used to measure the deviation between the predicted value and the true value at each preset quantile.

8. The regional distributed photovoltaic probabilistic prediction method according to claim 1, characterized in that, The preset regional division rules include administrative division boundary division rules or location-based boundary division rules.

9. A regional distributed photovoltaic probabilistic prediction system, characterized in that, include: The data acquisition module is used to acquire multi-source data of the target area within a preset historical period and generate a standardized data tensor set; The graph data construction module is used to divide the target area into multiple regions according to the preset regional division rules. Based on the standardized data tensor set, the photovoltaic data points in each region are clustered and downsampled to generate an initial feature tensor with N vertices. Adjacent edges are constructed based on the physical distance and output similarity between each vertex to form a graph topology structure that represents the spatial relationship between multiple regions. The spatial feature extraction module is used to input the graph topology and its corresponding initial feature tensor into the graph attention network, extract the wide-area spatiotemporal correlation features of multiple regions through the graph attention network, fuse the wide-area spatiotemporal correlation features, and output the fused high-order features. The multi-task probability prediction module is used to input high-order features into a multi-task learning network built on deep quantile regression, and simultaneously output the photovoltaic output probability prediction interval for each region at a set confidence level.

10. The regional distributed photovoltaic probabilistic prediction system according to claim 9, characterized in that, The multi-source data acquired by the data acquisition module includes numerical weather forecast data, historical operation data of distributed photovoltaic power generation, and static information data. The static information data includes installed capacity information, location distribution information, historical actual power information of the dispatch terminal, and load forecast information recorded by the dispatch terminal; The data acquisition module preprocesses the multi-source data and unifies the preprocessed data to a 1-hour time resolution. After alignment, it generates a standardized data tensor set.

11. The regional distributed photovoltaic probabilistic prediction system according to claim 9, characterized in that, It also includes a weather classification module, used to perform weather classification feature analysis based on a standardized data tensor set and to label different weather conditions as scenarios. The weather classification module trains a weather classification recognition model based on the photovoltaic output curve morphology features and corresponding gridded meteorological element features in the standardized data tensor set. The weather classification recognition model is used to label different weather conditions as scenarios, resulting in weather scenario labels. The scenario labels include at least sunny day labels and cloudy day labels. The weather scenario labels are used as additional features of vertices in the graph topology or as conditional inputs to a multi-task learning network to guide differentiated modeling under different meteorological scenarios.

12. The regional distributed photovoltaic probabilistic prediction system according to claim 9, characterized in that, When constructing adjacent edges based on the physical distance and power output similarity between each vertex, the graph data construction module calculates the geographical spatial distance between each pair of vertices and the Pearson correlation coefficient between the photovoltaic historical power output sequences corresponding to each pair of vertices as the power output similarity. The physical distance and the similarity of output are weighted and summed. When the result is less than a preset threshold, an adjacent edge is constructed between the corresponding two vertices.

13. The regional distributed photovoltaic probabilistic prediction system according to claim 9, characterized in that, The graph attention network used in the spatial feature extraction module includes multiple graph attention layers. Each graph attention layer adaptively calculates the attention weight between each vertex and its neighboring vertices through an attention mechanism, and aggregates the feature information of neighboring vertices according to the attention weights to update the feature representation of the current vertex, and outputs the higher-order features.

14. The regional distributed photovoltaic probabilistic prediction system according to claim 9, characterized in that, The multi-task learning network in the multi-task probability prediction module sets up a prediction task branch for each region. The multi-task learning network includes a shared input layer and a feature extraction network, as well as multiple parallel task-specific output layers. Each task-specific output layer corresponds to a region and is used to output the photovoltaic power output prediction value of that region at multiple preset quantiles. The multiple preset quantiles are used to form the probability prediction interval at the set confidence level.

15. The regional distributed photovoltaic probabilistic prediction system according to claim 14, characterized in that, The multi-task learning network in the multi-task probability prediction module is jointly trained and optimized using a multi-task bouncing ball loss function, which is the sum of the quantile loss functions of each region. The quantile loss function is used to measure the deviation between the predicted value and the true value at each preset quantile.

16. The regional distributed photovoltaic probabilistic prediction system according to claim 9, characterized in that, When the graph data construction module divides the target area into multiple regions according to preset regional division rules, the preset regional division rules include administrative division rules or location distribution boundary division rules.

17. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the regional distributed photovoltaic probabilistic prediction method as described in any one of claims 1 to 8.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the regional distributed photovoltaic probabilistic prediction method as described in any one of claims 1 to 8.

Images