A multi-level distributed photovoltaic power prediction method
By employing a multi-level distributed photovoltaic power prediction method, historical power sequences are reconstructed using gridding and spatiotemporal correlation, and combined with a 3D-CNN model, the problem of data inconsistency in distributed photovoltaic systems is solved, achieving high-precision and robust global power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID JIANGSU ELECTRIC POWER CO LTD RESEARCH INSTITUTE
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-12
Smart Images
Figure CN122203198A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy power prediction technology, and in particular to a multi-level distributed photovoltaic power prediction method. Background Technology
[0002] In recent years, distributed photovoltaic (PV) power generation has developed rapidly due to its proximity to loads, short construction period, and flexible resource utilization, gradually forming a power source characterized by multi-point grid connection and local consumption. The output of distributed PV is influenced by various meteorological factors, including solar irradiance, cloud cover, ambient temperature, wind speed and direction, and humidity, exhibiting significant randomness, volatility, and intermittency. When the penetration rate of distributed PV is high, its output fluctuations can adversely affect the voltage qualification rate of the distribution network, power flow distribution, and the execution of dispatch plans. Therefore, accurate prediction of distributed PV output is of great significance.
[0003] Existing photovoltaic power prediction technologies typically focus on individual power plants, building prediction models based on historical power data and meteorological observations or numerical weather prediction data, and then summarizing the prediction results from individual plants to obtain regional prediction results. However, distributed photovoltaic systems are characterized by small individual capacity, a large number of sites, and scattered spatial distribution. At the site level, due to differences in data acquisition and communication conditions, the input data for prediction models generally suffers from problems such as missing power measurements, incomplete meteorological data, inconsistent sampling frequencies, and numerous outliers. This makes traditional modeling methods that rely on complete historical samples and on-site meteorological observations difficult to apply directly, or require significant investment in data governance and maintenance costs.
[0004] On the other hand, regional or higher-level forecasts often require comprehensive consideration of the spatial correlation between meteorological factors and the spatiotemporal coupling between different grids or stations. Existing technologies often use interpolation of numerical weather prediction results or extrapolation of a few meteorological stations to obtain irradiance information for the target area. However, due to the limited spatial resolution of numerical weather prediction and the sparse and insufficient representativeness of station observations, the above methods cannot simultaneously meet the requirements of full coverage and high resolution at the distributed station cluster scale, thus affecting the accuracy of grid-level and regional power predictions.
[0005] Therefore, there is an urgent need for a solution that can make full use of the spatiotemporal correlation of the photovoltaic power generation system and multi-source meteorological information to predict distributed photovoltaic power, even when distributed photovoltaic data is incomplete, so as to improve prediction accuracy and engineering usability. Summary of the Invention
[0006] This invention provides a multi-level distributed photovoltaic power prediction method, which can effectively solve the problems in the background technology.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-level distributed photovoltaic power prediction method includes the following steps: S10. Determine the area scope based on the administrative boundary or dispatch / distribution network management boundary of the target area. Perform matrix-based grid subdivision of the target area under a unified coordinate system to generate grid cells with unique numbers. Adaptively set the spatial resolution of the grid cells. Determine the representative spatial positioning point and validity of each grid cell. Output the set of grid cells covering the target area and the center positioning point. Map all distributed photovoltaic sites in the target area to the corresponding grid cells. Construct grid equivalent sites in each grid cell and calculate the equivalent installed capacity of the grid equivalent sites. S20. Utilize the spatiotemporal correlation of adjacent grids or nearby complete stations to reconstruct the historical power curves of the target grid that is completely black or semi-complete in the correction information, and generate a complete historical power sequence. S30. First, determine the global similar day candidate set at the regional scale. Then, for each grid cell, based on the obtained complete historical power sequence, perform a second screening of the global similar day candidate set at the grid scale to form a grid-level similar day set. Perform quality control and weight allocation on the grid-level similar day set to obtain weather type-specific similar day sets for each grid cell. S40. Using the day-grid field as the sample unit, samples are extracted from the set of similar days of different weather types to construct a 3D-CNN input tensor containing grid spatial dimension, time dimension and variable dimension; the power signal obtained by mapping a few complete stations is used as the supervision signal, the loss is calculated by using a masking mechanism, the similarity score of similar days is converted into sample weights and introduced into the loss function to train the model, and the grid-level photovoltaic power prediction result is output by using the trained model.
[0008] Furthermore, in step S10, the spatial resolution of the grid cell can be adaptively set according to the density of distributed photovoltaic sites, the spatial variation scale of the meteorological field, and the operational forecast scale, specifically as follows: A finer grid is used when distributed photovoltaic sites are dense or the spatial gradient of cloud clusters is large; a coarser grid is used when distributed photovoltaic sites are sparse or the business focuses more on regional scale aggregation results.
[0009] Furthermore, in step S10, the representative spatial positioning point is the center point of the grid cell.
[0010] Further, in step S10, the validity determination of the mesh cell is specifically as follows: When the overlap area between a grid cell and the region boundary is greater than or equal to a set threshold, the grid cell is retained. When the overlap area between a grid cell and the region boundary is less than a set threshold, the grid cell is trimmed, merged, or removed.
[0011] Furthermore, in step S10, when the distributed photovoltaic site is mapped to the corresponding grid cell, a consistency check is performed by combining the distribution network topology information or a distance threshold.
[0012] Furthermore, in step S10, the equivalent installed capacity is the sum of the installed capacity of each distributed photovoltaic site within the grid unit, and includes the installed capacity statistics of sites with completely blacked-out information.
[0013] Furthermore, in step S20, the historical power data of the target grid and adjacent grids / nearby complete stations are preprocessed. The preprocessing includes timestamp alignment, deduplication of duplicate records, and uniformity of sampling intervals. Reference objects are selected based on grid adjacency, spatial distance, distribution network topology and / or historical correlation, and a power sequence mapping relationship between the target grid and the reference objects is established. The power of missing time periods in the completely black or semi-complete target grid is reconstructed, and outliers are identified and corrected. Physical constraint consistency correction is performed on the reconstructed power sequence to generate a complete historical power sequence.
[0014] Furthermore, establishing the power sequence mapping relationship includes: First, the historical power sequences of the target grid and reference objects are normalized according to the equivalent installed capacity. Then, the weights of the reference objects are determined based on the historical correlation within the overlapping effective time period. Finally, the power of the target grid in the missing time period is reconstructed by using correlation weighting, distance weighting, or a combination of both. After normalization, the sum of the weights of each reference object is 1.
[0015] Furthermore, the identification and correction of outliers includes at least: physical boundary detection based on power non-negativity constraints and capacity upper limit constraints, abrupt change detection based on ramp rate threshold, and anomaly detection based on statistical outlier criteria. Among them, the identified anomalies are corrected by neighborhood interpolation, reference object mapping value replacement or weighted fusion, and the corrected power sequence is subject to continuity constraints.
[0016] Further, in step S30, a global similarity candidate set is selected at the regional scale, specifically as follows: Using the overall weather background of the target area as a constraint, and utilizing numerical weather forecast data for the target day and meteorological observations or irradiance data from a few complete stations in the area, a global feature vector characterizing the regional-scale irradiance conditions is constructed. By combining the historical sample database, a global feature vector of the same dimension is constructed for each historical day, and the global similarity between the historical day and the target day is calculated by the sample entropy. The top few historical days with the highest similarity are selected as the global similar day candidate set.
[0017] Furthermore, the calculation process in determining the global similar day candidate set is as follows: Let the global feature time series of the target date and a certain historical date be as follows: Let the embedding dimension be... The threshold is Then the vector distance is defined as: Calculate length and length Match ratio: The final calculation result of the sample entropy is as follows: Therefore, sample entropy The smaller the value, the more similar the overall evolution of the historical date and the target date. The smallest front One historical date is used as a candidate set of globally similar dates.
[0018] Furthermore, in step S30, a secondary grid-level screening is performed at the grid scale to form a grid-level similar day set. The specific process is as follows: Using the grid center point as a reference position, the historical power curve or normalized power curve corresponding to the grid and its local meteorological / irradiance characteristics are extracted from the global similar day candidate set to construct a local feature vector to describe the grid power output pattern. Calculate the local similarity between the target date and each historical date, and select the historical dates with the highest similarity to form a grid-level similar date set for the grid.
[0019] Furthermore, the local feature vector preferably includes intraday key morphological indicators of the power curve, such as peak power and peak time, rising / falling slope, midday dip degree, fluctuation intensity, number and amplitude of climbing events, etc., while combining the irradiance sequence morphological characteristics of the grid as constraints.
[0020] Furthermore, in step S30, the quality control specifically involves removing historical days with obvious power outages, system failures, or data anomalies; the weight allocation is determined based on local similarity, with historical days having higher weights for higher similarity; and the weather type classification includes three categories: sunny, cloudy, and rainy.
[0021] Furthermore, in step S40, the variable dimensions include numerical weather forecasting elements such as cloud cover, shortwave radiation correlation, temperature, humidity, and wind, as well as their derived features. The time dimension includes the forecast lead time sequence and the historical evolution sequence; The grid space dimension is a row × column matrix of grid cells.
[0022] Furthermore, in step S40, the masking mechanism specifically calculates the loss only for grid locations covered by supervised signals; The introduction of sample weights makes high-similarity samples contribute more to the model parameter update than low-similarity samples, forming a training strategy dominated by similarity days and supported by expanded samples.
[0023] The technical solution of this invention can achieve the following technical effects: This invention transforms the prediction problem of a large number of discrete stations into a unified modeling method for finite grid cells by regional gridding and the construction of grid equivalent stations. At the same time, it utilizes the spatiotemporal correlation of adjacent grids or nearby complete stations to reconstruct and correct historical power sequences, enabling previously unusable information from completely black / semi-complete stations to be included in the same prediction system, fundamentally improving the availability of station group data and the stability of model training.
[0024] This invention constructs a two-level similar day screening mechanism based on the enhanced power sequence, first globally and then grid-based, to ensure that the training samples are consistent with the weather background and local power output pattern of the target day. The similar days are used for sample weighting and pattern prior constraints, thereby reducing the averaging error caused by the mixture of different weather types and improving the prediction accuracy and robustness under complex meteorological conditions.
[0025] This invention integrates mesoscale numerical weather prediction with irradiance or clear sky index information from a few complete weather stations, and employs a 3D-CNN distributed photovoltaic (PV) prediction model based on similar daily samples to output a full-grid high-resolution power prediction. This invention can achieve full-domain prediction coverage of distributed PV in distributed scenarios with incomplete data, while maintaining accuracy and robustness. Attached Figure Description
[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0027] Figure 1This is a schematic diagram of a multi-level distributed photovoltaic power prediction method. Figure 2 A flowchart illustrating the grid division and positioning process; Figure 3 A flowchart illustrating the process of filtering similar dates; Figure 4 This is a schematic diagram of a three-dimensional convolutional neural network model framework. Detailed Implementation
[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0030] like Figure 1 As shown, this application provides a multi-level distributed photovoltaic power prediction method, including the following steps: S10. Determine the area scope based on the administrative boundary or dispatch / distribution network management boundary of the target area. Perform matrix-based grid subdivision of the target area under a unified coordinate system to generate grid cells with unique numbers. Adaptively set the spatial resolution of the grid cells. Determine the representative spatial location points and validity of each grid cell. Output the set of grid cells covering the target area and the center location point. Map all distributed photovoltaic sites in the target area to the corresponding grid cells. Construct grid equivalent sites in each grid cell and calculate the equivalent installed capacity of the grid equivalent sites. In step S10, the spatial resolution of the grid cell can be adaptively set according to the density of distributed photovoltaic sites, the scale of spatial changes in the meteorological field, and the scale of operational forecasts. Specifically, when the distributed photovoltaic sites are dense or the spatial gradient of the cloud cluster is large, a finer grid is used to improve the accuracy of spatial representation; when the distributed photovoltaic sites are sparse or the operational focus is more on the regional scale summary results, a coarser grid is used to balance computational complexity and generalization ability.
[0031] After gridding, a representative spatial location is determined for each grid cell. Preferably, the representative spatial location point is the center point of the grid cell. The grid center point is used as the spatial location point of the grid and serves as a unified reference location for meteorological driving quantity extraction, irradiance output, and power prediction output, ensuring the consistency of subsequent multi-source data fusion.
[0032] The validity determination of each grid cell is as follows: When the overlap area between a grid cell and the region boundary is greater than or equal to a set threshold, the grid cell is retained. When the overlap area between a grid cell and the region boundary is less than a set threshold, the grid cell is trimmed, merged, or removed.
[0033] For boundary grids or grid cells in irregular regions, the validity can be determined by the ratio of the overlapping area between the grid and the region boundary. Boundary grids can be trimmed or merged when necessary to ensure the complete coverage of the target region by the grid set and the computational stability.
[0034] Furthermore, when a distributed photovoltaic site is mapped to its corresponding grid cell, consistency verification needs to be performed by combining the distribution network topology information or distance threshold.
[0035] After completing the regional gridding, all distributed photovoltaic (PV) sites within the region are spatially assigned. Based on the site's latitude and longitude coordinates, each site is mapped to its corresponding grid cell, establishing a site-grid assignment relationship. For cases where coordinate discrepancies exist or sites are located near grid boundaries, consistency verification can be further performed using distribution network topology information or distance thresholds to reduce misclassification due to location errors. After site assignment is completed, sites are aggregated and managed using the grid as the basic organizational unit, thus achieving a structured transformation from a discrete distribution of multiple sites to a set of grid cells. This facilitates subsequent spatiotemporal correlation modeling, data augmentation, and unified prediction.
[0036] To adapt to the characteristics of distributed photovoltaic (PV) systems—multi-point access, small individual capacity, and varying data quality—this invention constructs an equivalent grid site within each grid cell. Multiple distributed PV sites within the same grid are considered as an equivalent power generation unit at a grid location point, representing the overall contribution of PV power output within that grid. The equivalent installed capacity is the sum of the installed capacities of all distributed PV sites within the grid cell, including statistics for sites with completely blacked-out information. This serves as the basis for power normalization, capacity constraints, and upscaling. For sites with completely blacked-out information or lacking available historical power, this invention still includes them in the grid affiliation and incorporates them into the equivalent capacity statistics, ensuring that the equivalent site fully reflects the actual installed capacity within the grid. This provides a constraint boundary for subsequent power reconstruction and gridded power prediction based on spatiotemporal correlation.
[0037] like Figure 2As shown, through the above-mentioned grid division and positioning processing, the target area is matrix-grid-divided according to a unified coordinate system, and equivalent stations are constructed within the grid as the smallest predictive modeling unit. This includes a unified attribution and equivalent installed capacity statistics mechanism for stations with complete, semi-complete, and completely black information, so that the grid layer can fully reflect the actual installed capacity. This invention transforms the complex distributed station group prediction problem into a unified modeling problem oriented towards grid equivalent stations, laying the foundation for subsequent multi-source data enhancement, grid-based prediction, and multi-level aggregation.
[0038] S20. To enable the information in the fully black grid and semi-complete grid to participate in subsequent similar day selection and model training, the historical power data of the target grid and adjacent grids / nearest neighbor complete stations are preprocessed. Preprocessing includes timestamp alignment, deduplication of duplicate records, and uniformity of sampling intervals. For data with inconsistent sampling intervals, linear interpolation, spline interpolation, or preserving previous values can be used for resampling to form a historical power sequence with a uniform time resolution. For missing data periods, missing markers are retained for subsequent reconstruction processing.
[0039] Reference objects are selected based on grid adjacency, spatial distance, distribution network topology and / or historical correlation, and a power sequence mapping relationship between the target grid and the reference objects is established. After time unification is completed, a set of reference objects is constructed for the target grid. Reference objects can be adjacent grids and / or nearby complete sites, and their selection criteria include at least one of grid adjacency, spatial distance, distribution network topology, and historical power correlation. Preferably, a candidate set is first formed based on a distance threshold or topological adjacency, and then the historical correlation between the target grid and the candidate objects within the overlapping effective time period is used to sort them. Several reference objects with high correlation are selected for subsequent mapping relationship establishment and missing power reconstruction.
[0040] Furthermore, establishing the power sequence mapping relationship includes: First, the historical power sequences of the target grid and reference objects are normalized according to the equivalent installed capacity. Then, the weights of the reference objects are determined based on the historical correlation within the overlapping effective time period. Finally, the power of the target grid in the missing time period is reconstructed by using correlation weighting, distance weighting, or a combination of both. After normalization, the sum of the weights of each reference object is 1.
[0041] To reduce the impact of differences in installed capacity between different grids or sites on the reconstruction results, the historical power sequences of the target grid and the reference object can first be normalized according to the equivalent installed capacity. Then, a power sequence mapping relationship between the target grid and the reference object can be established based on the overlapping effective time periods. The mapping relationship can be determined by scaling, correlation-weighted regression, local linear regression, or a combination thereof. Preferably, the weight of the reference object is positively correlated with its historical power and negatively correlated with spatial distance, and the weight is normalized. Based on the mapping relationship, the historical power of the target grid with completely black information is reconstructed for all time periods; for the semi-complete target grid, only the missing time periods are filled in while retaining the original effective observations; at the junction of the original observations and the reconstructed values, a smooth transition processing can be used to reduce splicing abrupt changes.
[0042] The power of missing time periods in a completely black or semi-complete target grid is reconstructed. After obtaining the reconstructed power sequence, outliers are identified and corrected. Furthermore, the identification of anomalies includes, but is not limited to, physical boundary detection based on power non-negativity constraints and capacity upper limit constraints, abrupt change detection based on ramp rate threshold, and anomaly detection based on statistical outlier criteria. Among them, the identified outliers can be corrected by neighborhood interpolation, reference object mapping value replacement or weighted fusion, and the corrected power sequence can be subject to continuity constraints to improve the rationality and stability of the time series shape.
[0043] The reconstructed power sequence is subjected to physical constraint consistency correction to generate a complete historical power sequence, which is used to support subsequent screening of similar days and data-driven model training.
[0044] Physical constraints are applied to the reconstructed and corrected target grid power sequence for consistency correction, ensuring it meets constraints of non-negative power, capacity upper limit, and continuous intraday variation, thus generating a complete historical power sequence. Simultaneously, data quality labels are generated for each time period, serving as one of the inputs for grid-level similar day secondary screening, quality control, and sample weight allocation in step S30. Through step S20, the usability of historical samples can be improved while preserving the grid spatial structure, providing a consistent and reliable data foundation for subsequent similar day screening mechanisms and gridded power prediction model training.
[0045] S30. First, determine the global similar day candidate set at the regional scale. Then, for each grid cell, based on the obtained complete historical power sequence, perform a second screening of the global similar day candidate set at the grid scale to form a grid-level similar day set. Perform quality control and weight allocation on the grid-level similar day set to obtain weather type-specific similar day sets for each grid cell. The process of selecting a global similar day candidate set at the regional scale is as follows: Using the overall weather background of the target area as a constraint, and utilizing numerical weather forecast data for the target day and meteorological observations or irradiance data from a few complete stations in the region, a global feature vector characterizing the regional-scale irradiance conditions is constructed; this vector is used to depict the three weather types of the target day—sunny, cloudy, and rainy—and their evolution over time. By combining the historical sample database, a global feature vector of the same dimension is constructed for each historical day. The global similarity between the historical day and the target day is calculated by using sample entropy. The top few historical days with the highest similarity are selected as the global similar day candidate set.
[0046] Furthermore, the calculation process in determining the global similarity candidate set is as follows: Let the global feature time series of the target date and a certain historical date be as follows: Let the embedding dimension be... Generally take The threshold is Then the vector distance is defined as: Calculate length and length Match ratio: The final calculation result of the sample entropy is as follows: Therefore, sample entropy The smaller the value, the more similar the overall evolution of the historical date and the target date. The smallest front One historical date is used as a candidate set of globally similar dates.
[0047] In step S30, for each grid cell, a secondary grid-level screening is performed at the grid scale to form a grid-level similarity set. The specific process is as follows: Using the grid center point as a reference position, the historical power curve or normalized power curve corresponding to the grid and its local meteorological / irradiance characteristics are extracted from the candidate set to construct a local feature vector to describe the power output pattern of the grid. The local feature vector preferably includes intraday key morphological indicators of the power curve, such as peak power and peak time, rising / falling slope, midday depression degree, fluctuation intensity, number and amplitude of climbing events, etc., and is combined with the irradiance sequence morphological characteristics of the grid as constraints.
[0048] Calculate the local similarity between the target day and each historical day within the global similar day candidate set, and select the historical days with the highest similarity to form the grid-level similar day set for this grid.
[0049] To improve the adaptability and stability of the gridded power prediction model under different weather conditions, a similar day screening mechanism is introduced. The meteorological-power evolution characteristics of the target day are mapped to the historical sample database. The set of historical dates that are closest to the overall weather pattern and local grid power output pattern of the target day are selected first, and used as the basis for model training weighting, feature construction and curve correction.
[0050] In the above scheme, quality control specifically involves removing historical days with obvious power outages, system failures, or data anomalies; weight allocation is determined based on local similarity, with historical days having higher weights for higher similarity; and weather types are categorized into three types: sunny, cloudy, and rainy.
[0051] Based on a two-level similarity date screening mechanism, a priori construction mechanism for sample organization and morphology is adopted, employing a two-level similarity date screening strategy of first global screening and then grid screening. Figure 3 As shown, a candidate sample set consistent with the weather background of the target day is first screened at the regional scale. Then, a secondary screening is conducted at the grid scale based on power patterns and local meteorological characteristics to form a comprehensive set of similar days for different weather types. Different sets of similar days can be formed for different grids within the region to reflect local differences. To enhance robustness, a consistency check and quality control mechanism is introduced for the grid-level similar day set. Historical days with obvious power outages, system failures, or data anomalies are removed, and similar days are weighted according to similarity. Finally, similar day sets for three weather types—sunny, cloudy, and rainy—are obtained for each grid. This ensures that the selected samples not only meet the consistency of the large-scale meteorological background of the region but also reflect the spatial non-uniformity of local cloud cover changes and distributed photovoltaic power output differences.
[0052] S40. Model Prediction: Using the daily-grid field as the sample unit, samples are extracted from the set of similar days of different weather types to construct a 3D-CNN input tensor containing grid spatial dimension, time dimension and variable dimension; the power signal obtained by mapping a few complete stations is used as the supervision signal, the loss is calculated by using a masking mechanism, the similarity score of similar days is converted into sample weights and introduced into the loss function, the 3D-CNN model is trained, and the grid-level photovoltaic power prediction result is output by using the trained model.
[0053] In this scheme, the variable dimension includes numerical weather prediction elements such as cloud cover, shortwave radiation correlation, temperature, humidity and wind, and their derived characteristics; the time dimension includes the forecast lead time sequence and the historical evolution sequence; and the grid spatial dimension is a row × column matrix of grid cells.
[0054] In this scheme, the masking mechanism specifically calculates the loss only for grid locations covered by supervised signals; the introduction of sample weights makes high-similarity samples contribute more to the model parameter update than low-similarity samples, forming a training strategy dominated by similarity days and supported by expanded samples.
[0055] This step uses, for example Figure 4 The model shown is based on the irradiance prediction (3D-CNN) method trained on similar day sample sets. It integrates mesoscale numerical weather prediction and irradiance observations from a few complete stations to construct a three-dimensional input tensor of space-time-variable. It adopts mask supervision and similar day hierarchical sampling / weighted training strategies to achieve gridded power prediction.
[0056] This invention proposes a gridded multi-level power prediction method and consistency correction mechanism for distributed photovoltaic (PV) stations, aiming to achieve predictive capabilities that are fully coverable, scalable, and implementable in the context of a large number of distributed PV stations, spatial dispersion, and inconsistent data quality.
[0057] First, this invention transforms the prediction problem of a large number of discrete stations into a unified modeling method for finite grid cells by regional gridding and grid equivalent station construction. At the same time, it utilizes the spatiotemporal correlation of adjacent grids or nearby complete stations to reconstruct and correct historical power sequences, enabling previously unusable information from completely black / semi-complete stations to be included in the same prediction system, fundamentally improving the availability of station group data and the stability of model training.
[0058] Next, this invention constructs a two-level similar day screening mechanism based on the enhanced power sequence, first global and then grid, so that the training samples are consistent with the weather background and local power output pattern of the target day. The similar days are used for sample weighting and pattern prior constraints, thereby reducing the averaging error caused by the mixture of different weather types and improving the prediction accuracy and robustness under complex meteorological conditions.
[0059] Finally, to address the problem of insufficient meteorological input caused by the limited resolution of numerical weather prediction and the sparseness of observation stations, this invention integrates mesoscale numerical weather prediction with irradiance or clear sky index information from a few complete stations, and employs a 3D-CNN distributed photovoltaic prediction model based on similar daily sample organization to output a full-grid high-resolution power prediction. This invention can achieve full-domain prediction coverage from distributed photovoltaics in distributed scenarios with incomplete data, while taking into account accuracy and robustness, and has significant engineering application value and promotion significance.
[0060] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.
Claims
1. A multi-level distributed photovoltaic power prediction method, characterized in that, Includes the following steps: S10. Determine the area range based on the administrative boundary or dispatch / distribution network management boundary of the target area, perform matrix-based grid subdivision of the target area under a unified coordinate system, generate grid cells with unique numbers, and adaptively set the spatial resolution of the grid cells. Determine the representative spatial location point and validity of each grid cell, and output the set of grid cells covering the target area and the center location point; map all distributed photovoltaic sites in the target area to the corresponding grid cells, construct grid equivalent sites in each grid cell, and calculate the equivalent installed capacity of the grid equivalent sites; S20. Utilize the spatiotemporal correlation of adjacent grids or nearby complete stations to reconstruct the historical power curves of the target grid that is completely black or semi-complete in the correction information, and generate a complete historical power sequence. S30. First, determine the global similar day candidate set at the regional scale. Then, for each grid cell, based on the obtained complete historical power sequence, perform a second screening of the global similar day candidate set at the grid scale to form a grid-level similar day set. Perform quality control and weight allocation on the grid-level similar day set to obtain weather type-specific similar day sets for each grid cell. S40. Using the day-grid field as the sample unit, samples are extracted from the set of similar days of different weather types to construct a 3D-CNN input tensor containing grid spatial dimension, time dimension and variable dimension; the power signal obtained by mapping a few complete stations is used as the supervision signal, the loss is calculated by using a masking mechanism, the similarity score of similar days is converted into sample weights and introduced into the loss function to train the model, and the grid-level photovoltaic power prediction result is output by using the trained model.
2. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S10, the spatial resolution of the grid cell can be adaptively set according to the density of distributed photovoltaic sites, the spatial variation scale of the meteorological field, and the operational forecast scale, specifically as follows: A finer grid is used when distributed photovoltaic sites are dense or the spatial gradient of cloud clusters is large; a coarser grid is used when distributed photovoltaic sites are sparse or the business focuses more on regional scale aggregation results.
3. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S10, the representative spatial positioning point is the center point of the grid cell.
4. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S10, the validity determination of the mesh cell is specifically as follows: When the overlap area between a grid cell and the region boundary is greater than or equal to a set threshold, the grid cell is retained. When the overlap area between a grid cell and the region boundary is less than a set threshold, the grid cell is trimmed, merged, or removed.
5. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S10, when the distributed photovoltaic site is mapped to the corresponding grid cell, a consistency check is performed by combining the distribution network topology information or distance threshold.
6. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S10, the equivalent installed capacity is the sum of the installed capacity of each distributed photovoltaic site within the grid unit, and includes the installed capacity statistics of sites with completely blacked-out information.
7. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S20, the historical power data of the target grid and adjacent grids / nearby complete stations are preprocessed. The preprocessing includes timestamp alignment, deduplication of duplicate records, and uniformity of sampling intervals. Reference objects are selected based on grid adjacency, spatial distance, distribution network topology and / or historical correlation, and a power sequence mapping relationship between the target grid and the reference objects is established. The power of missing time periods in the completely black or semi-complete target grid is reconstructed, and outliers are identified and corrected. Physical constraint consistency correction is performed on the reconstructed power sequence to generate a complete historical power sequence.
8. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, Establishing the power sequence mapping relationship includes: First, the historical power sequences of the target grid and reference objects are normalized according to the equivalent installed capacity. Then, the weights of the reference objects are determined based on the historical correlation within the overlapping effective time period. Finally, the power of the target grid in the missing time period is reconstructed by using correlation weighting, distance weighting, or a combination of both. After normalization, the sum of the weights of each reference object is 1.
9. The multi-level distributed photovoltaic power prediction method according to claim 7, characterized in that, The identification and correction of anomalies include at least: physical boundary detection based on power non-negativity constraints and capacity upper limit constraints, abrupt change detection based on ramp rate threshold, and anomaly detection based on statistical outlier criteria. Among them, the identified anomalies are corrected by neighborhood interpolation, reference object mapping value replacement or weighted fusion, and the corrected power sequence is subject to continuity constraints.
10. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S30, a global similarity candidate set is selected at the regional scale, specifically as follows: Using the overall weather background of the target area as a constraint, and utilizing numerical weather forecast data for the target day and meteorological observations or irradiance data from a few complete stations in the area, a global feature vector characterizing the regional-scale irradiance conditions is constructed. By combining the historical sample database, a global feature vector of the same dimension is constructed for each historical day, and the global similarity between the historical day and the target day is calculated by the sample entropy. The top few historical days with the highest similarity are selected as the global similar day candidate set.
11. The multi-level distributed photovoltaic power prediction method according to claim 8, characterized in that, The calculation process in determining the global similarity candidate set is as follows: Let the global feature time series of the target date and a certain historical date be as follows: Let the embedding dimension be The threshold is Then the vector distance is defined as: Calculate length and length Match ratio: The final calculation result of the sample entropy is as follows: Therefore, sample entropy The smaller the value, the more similar the overall evolution of the historical date and the target date. The smallest front One historical date is used as a candidate set of globally similar dates.
12. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S30, a secondary grid-level screening is performed at the grid scale to form a grid-level similar day set. The specific process is as follows: Using the grid center point as a reference position, the historical power curve or normalized power curve corresponding to the grid and its local meteorological / irradiance characteristics are extracted from the global similar day candidate set to construct a local feature vector to describe the grid power output pattern. Calculate the local similarity between the target date and each historical date, and select the historical dates with the highest similarity to form a grid-level similar date set for the grid.
13. The multi-level distributed photovoltaic power prediction method according to claim 10, characterized in that, The local feature vector preferably includes intraday key morphological indicators of the power curve, such as peak power and peak time, rising / falling slope, midday dip degree, fluctuation intensity, number and amplitude of climbing events, etc., while combining the irradiance sequence morphological characteristics of the grid as constraints.
14. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S30, the quality control specifically involves removing historical days with obvious power outages, system failures, or data anomalies; the weight allocation is determined based on local similarity, with historical days having higher weights for higher similarity; and the weather type classification includes three categories: sunny, cloudy, and rainy.
15. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S40, the variable dimensions include numerical weather prediction elements such as cloud cover, shortwave radiation correlation, temperature, humidity and wind, and their derived features. The time dimension includes the forecast lead time sequence and the historical evolution sequence; The grid space dimension is a row × column matrix of grid cells.
16. The multi-level distributed photovoltaic power prediction method according to claim 1, characterized in that, In step S40, the masking mechanism specifically calculates the loss only for grid locations covered by the supervised signal; The introduction of sample weights makes high-similarity samples contribute more to the model parameter update than low-similarity samples, forming a training strategy dominated by similarity days and supported by expanded samples.