Distribution network distributed photovoltaic prediction system and method based on multi-data coupling

Abstract

The invention discloses a distribution network distributed photovoltaic prediction system and method based on multi-data coupling, and relates to the field of distributed photovoltaic prediction, and the method comprises the steps: constructing a distribution network clustering regionalization model based on edge nodes, and dividing the distribution network distributed photovoltaic; based on a seasonal SARIMA algorithm model, predicting a long-term baseline trend of distributed photovoltaic power generation power data characteristics of each edge node; constructing a short-term prediction model of the electric power data characteristics, and predicting the short-term electric power data characteristic data value of each edge node; constructing an electric power data feature comprehensive prediction model, and predicting an electric power data feature comprehensive value of each edge node; the method comprises the following steps: acquiring an electric power data characteristic comprehensive value of each edge node by utilizing an internet of things technology, and constructing an electric power data characteristic database; the method has the advantages that the distributed photovoltaic of the distribution network is effectively divided in detail, and the distributed photovoltaic prediction precision and the resource optimization rate are improved by using a long and short term prediction mode.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic (PV) forecasting, specifically to a distributed PV forecasting system and method for distribution networks based on multi-data coupling. Background Technology

[0002] As the penetration rate of distributed photovoltaic (PV) power in distribution networks continues to increase, its output is affected by multiple factors such as geographical environment, meteorological conditions (e.g., cloud cover, sunlight intensity), and installation attributes (e.g., tilt angle, orientation), exhibiting strong randomness, volatility, and intermittency. These characteristics pose unprecedented challenges to the power balance, power quality, protection configuration, and dispatch operation of distribution networks. Accurate distributed PV power forecasting is crucial for ensuring the safe and stable operation of distribution networks, optimizing dispatch resources, and improving absorption capacity. It is a core technological support for achieving safe, stable, and economical operation of distribution networks and is of great significance for absorbing green energy and enhancing the grid's regulation capabilities.

[0003] Existing distributed photovoltaic (PV) forecasting for distribution networks largely relies on a single data source (such as historical power data) or a single time-scale model, making it difficult to balance long-term trends with short-term fluctuations. Furthermore, the dispersed nature and significant individual differences of distributed PV units, coupled with a lack of targeted regionalization methods, limit forecast accuracy, making it difficult to meet the needs of refined distribution network management and resulting in low resource optimization rates. Summary of the Invention

[0004] To address the aforementioned technical problems, a distribution network distributed photovoltaic (PV) forecasting system and method based on multi-data coupling is provided. This technical solution solves the problems mentioned in the background technology, such as the difficulty in balancing long-term trends and short-term fluctuations. Furthermore, the dispersed distribution of PV units and their significant individual differences, coupled with the lack of targeted regionalization methods, limit forecasting accuracy and make it difficult to meet the needs of refined distribution network management and low resource optimization rates.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A distribution network distributed photovoltaic (PV) forecasting method based on multi-data coupling includes:

[0007] Based on the distribution of distributed photovoltaic power in the distribution network, edge nodes are set up, and a distribution network clustering regionalization model is constructed based on the edge nodes to divide the distributed photovoltaic power in the distribution network.

[0008] Based on each edge node, and using a seasonal SARIMA algorithm model, the long-term baseline trend of distributed photovoltaic power generation data characteristics of each edge node is predicted.

[0009] Based on cloud motion tracking, a short-term prediction model for power data characteristics is constructed for each edge node to predict the short-term power data characteristic values ​​of each edge node.

[0010] Based on the seasonal SARIMA algorithm model and the short-term prediction model of power data characteristics, a comprehensive prediction model of power data characteristics is constructed to predict the comprehensive value of power data characteristics of each edge node.

[0011] By utilizing Internet of Things (IoT) technology and based on a comprehensive prediction model of power data characteristics, the comprehensive value of power data characteristics of each edge node is obtained, and a power data characteristic database is constructed.

[0012] Preferably, the step of setting edge nodes based on the distribution of distributed photovoltaic power in the distribution network and constructing a distribution network clustering regionalization model based on the edge nodes to divide the distributed photovoltaic power in the distribution network specifically includes:

[0013] Based on the distribution of distributed photovoltaic power in the distribution network, each location where distributed photovoltaic power is installed independently is set as a distributed photovoltaic unit;

[0014] Obtain the geographic coordinates, historical power generation data, and installation attribute information of all distributed photovoltaic units;

[0015] Based on the geographical coordinates, historical power generation data, and installation attribute information of distributed photovoltaic units, multidimensional feature vectors related to their power generation characteristics are extracted.

[0016] Based on the Z-score standardization formula, the multidimensional feature vectors are normalized to eliminate the influence of data dimensions.

[0017] Based on the Euclidean distance formula and combined with the normalized multidimensional feature vector, the Euclidean distance value between any two units is calculated to quantify the similarity between the two units.

[0018] Based on the k-distance graph method, the neighborhood radius of the density clustering algorithm is determined. With minimum number of points threshold ;

[0019] Based on the Euclidean distance between any two units, obtain its Number of internal cells;

[0020] Among them, for distributed photovoltaic units ,That Defined as: all with The Euclidean distance is less than or equal to The set of units, i.e. , For the first Multidimensional feature vectors of distributed photovoltaic units For the first Multidimensional feature vectors of distributed photovoltaic units For the first The and the first Euclidean distance between the multidimensional feature vectors of a distributed photovoltaic unit;

[0021] judge If the number of internal cells is not less than the minimum number of points threshold, then determine the cell. If it is the core object, then the determination unit is used. These are non-core objects;

[0022] by The density reachability is used as the criterion for judging the current cluster expansion. All distributed photovoltaic units are traversed to form a virtual predictive clustering region.

[0023] After the traversal is complete, the units that are not assigned to any cluster are set as independent distributed photovoltaic prediction units.

[0024] The virtual prediction clustering region and the distributed photovoltaic prediction unit are respectively used as the edge nodes of the distributed photovoltaic distribution network, and a distribution network clustering regionalization model is constructed based on the edge nodes.

[0025] Preferably, the step of constructing a short-term prediction model of power data characteristics based on cloud motion tracking for each edge node, and predicting the short-term power data characteristic values ​​of each edge node specifically includes:

[0026] Deploy all-sky imagers or utilize satellite remote sensing data to acquire dynamic cloud image sequences of each edge node in real time;

[0027] Based on the dynamic cloud map sequence of each edge node, calculate the cloud's moving speed, direction, height, and optical thickness;

[0028] Based on the cloud's movement speed, direction, height, and optical thickness, and using an LSTM neural network model, the influence weights of the cloud on the electrical power data feature data are obtained.

[0029] Based on the influence weight of cloud layers on the characteristic data of power data, a short-term prediction model for the characteristic data of power data is constructed to predict the short-term characteristic data values ​​of power data for each edge node.

[0030] Preferably, the short-term prediction model based on the seasonal SARIMA algorithm model and power data characteristics, which constructs a comprehensive prediction model of power data characteristics to predict the comprehensive value of power data characteristics for each edge node, specifically includes:

[0031] Based on the seasonal SARIMA algorithm model, with a fixed long period Execute once; each time it is executed, the model generates a future preset range. Internal, in terms of time resolution A sequence of baseline values ​​for power generation data at intervals;

[0032] Based on the characteristic baseline value sequence of power generation data, a corresponding absolute timestamp is set for each value in the sequence. , ,in, This represents the start time of model execution. , , This represents the number of data points in the baseline value sequence;

[0033] Short-term prediction models based on power data characteristics, with fixed short periods The module executes in a rolling fashion, and each time it executes, it outputs a time period in the future. Internal, in terms of time resolution A sequence of correction coefficients for power generation data characteristics at intervals;

[0034] Based on the power generation data characteristic correction coefficient sequence, a corresponding absolute timestamp is set for each value in the sequence. , ,in, This is the absolute start time of this prediction execution. , , To correct the number of data points in the coefficient sequence;

[0035] Define the baseline value sequence of power generation data characteristics itself as the target time axis. This timeline contains all absolute time points in the baseline value sequence. and its corresponding baseline power ,in, ;

[0036] The absolute timestamp of the power generation data characteristic correction coefficient sequence Based on the nearest neighbor matching method, find the match with Find the closest correction factor and assign it to ;

[0037] Based on baseline power and correction factor A comprehensive prediction model for power data features is constructed to predict the comprehensive value of power data features for each edge node.

[0038] The expression for the comprehensive prediction model based on the characteristics of the power data is as follows:

[0039]

[0040] In the formula, This is a comprehensive value representing the characteristics of electrical power data.

[0041] Furthermore, this solution proposes a distribution network distributed photovoltaic (PV) forecasting system based on multi-data coupling to implement the aforementioned distribution network distributed PV forecasting method based on multi-data coupling, including:

[0042] The node partitioning module is used to set edge nodes according to the distribution of distributed photovoltaic power in the distribution network, and to construct a distribution network clustering regionalization model based on the edge nodes to partition the distributed photovoltaic power in the distribution network.

[0043] The photovoltaic prediction module is used to predict the long-term baseline trend of distributed photovoltaic power generation data characteristics of each edge node based on a seasonal SARIMA algorithm model; to construct a short-term prediction model of power data characteristics based on cloud motion tracking for each edge node, and to predict the short-term power data characteristic values ​​of each edge node; and to construct a comprehensive prediction model of power data characteristics based on the seasonal SARIMA algorithm model and the short-term prediction model of power data characteristics, and to predict the comprehensive value of power data characteristics of each edge node.

[0044] The database module is used to utilize Internet of Things (IoT) technology, based on a comprehensive prediction model of power data characteristics, to obtain the comprehensive value of power data characteristics of each edge node, and to construct a power data characteristic database.

[0045] Preferably, the photovoltaic prediction module includes:

[0046] The long-term trend unit is used to predict the long-term baseline trend of the distributed photovoltaic power generation data characteristics of each edge node based on the seasonal SARIMA algorithm model.

[0047] A short-term prediction unit is used to construct a short-term prediction model of the power data characteristics of each edge node based on cloud motion tracking, and to predict the short-term power data characteristic data values ​​of each edge node.

[0048] The comprehensive prediction unit is used to construct a comprehensive prediction model of power data features based on a short-term prediction model using a seasonal SARIMA algorithm model and power data features, and to predict the comprehensive value of power data features of each edge node.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention provides a distribution network distributed photovoltaic (PV) forecasting system and method based on multi-data coupling. By constructing a distribution network clustering regionalization model, the system regionalizes the dispersed and highly individualized distributed PV systems, obtaining virtual forecasting clusters with similar characteristics and distributed PV forecasting units with unique properties. This effectively reduces the deployment of data acquisition devices and improves system computational efficiency. Secondly, by constructing a seasonal SARIMA algorithm model and a short-term prediction model of power data characteristics, the system obtains the long-term baseline trend and short-term changes of the distributed PV power generation data characteristics of each edge node. Then, through a comprehensive prediction model of power data characteristics, the system obtains the comprehensive value of the power data characteristics of each edge node. That is, based on the long-term baseline trend, the short-term prediction corrects the long-term baseline trend, effectively improving the accuracy and stability of distributed PV forecasting at each edge node. Finally, by constructing a power data characteristic database, the system weights and sums the comprehensive values ​​of the collected power data characteristics of each edge node to obtain the total comprehensive forecast value of distributed PV in the distribution network. The data is then categorized and stored according to edge node type, making the massive and dispersed PV data orderly and searchable, thereby effectively achieving accurate forecasting and resource optimization of distributed PV in the distribution network. Attached Figure Description

[0051] Figure 1 This is a flowchart of the distributed photovoltaic prediction method for distribution networks based on multi-data coupling according to the present invention.

[0052] Figure 2 The flowchart for dividing the distributed photovoltaic power grid into distribution networks using the regionalized clustering model of the present invention is shown below;

[0053] Figure 3 The flowchart of the present invention is shown for predicting the short-term power data feature values ​​of each edge node using a short-term prediction model for constructing power data features. Detailed Implementation

[0054] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0055] Reference Figure 1 As shown, a distribution network distributed photovoltaic forecasting method based on multi-data coupling includes:

[0056] Based on the distribution of distributed photovoltaic power in the distribution network, edge nodes are set up, and a distribution network clustering regionalization model is constructed based on the edge nodes to divide the distributed photovoltaic power in the distribution network.

[0057] Based on each edge node, and using a seasonal SARIMA algorithm model, the long-term baseline trend of distributed photovoltaic power generation data characteristics of each edge node is predicted.

[0058] Based on cloud motion tracking, a short-term prediction model for power data characteristics is constructed for each edge node to predict the short-term power data characteristic values ​​of each edge node.

[0059] Based on the seasonal SARIMA algorithm model and the short-term prediction model of power data characteristics, a comprehensive prediction model of power data characteristics is constructed to predict the comprehensive value of power data characteristics of each edge node.

[0060] By utilizing Internet of Things (IoT) technology and based on a comprehensive prediction model of power data characteristics, the comprehensive value of power data characteristics of each edge node is obtained, and a power data characteristic database is constructed.

[0061] This can be explained as follows: Distribution network distributed photovoltaic (PV) forecasting refers to the process of estimating and predicting the power generation of distributed PV systems in the distribution network in advance. Its core objective is to predict the capacity of distributed PV systems over a future period by analyzing historical data, meteorological conditions, and PV system characteristics. Therefore, this solution constructs a distribution network clustering regionalization model to regionalize the dispersed and highly individualized distributed PV systems, obtaining virtual prediction clusters with similar characteristics and distributed PV prediction units with unique properties. This effectively reduces the deployment of data acquisition devices and improves the system's computational efficiency. Secondly, a seasonal SARIMA algorithm model is constructed. A short-term prediction model based on power data characteristics is used to obtain the long-term baseline trend and short-term changes of distributed photovoltaic power generation data characteristics at each edge node. Then, a comprehensive prediction model based on power data characteristics is used to obtain the comprehensive value of power data characteristics at each edge node. That is, based on the long-term baseline trend, the long-term baseline trend is corrected by short-term prediction, thereby effectively improving the accuracy and stability of distributed photovoltaic prediction at each edge node. Finally, by constructing a power data characteristic database, the comprehensive value of power data characteristics collected from each edge node is weighted and summed to obtain the total comprehensive prediction value of distributed photovoltaic power generation in the distribution network, thereby effectively realizing accurate prediction and resource optimization of distributed photovoltaic power generation in the distribution network.

[0062] Reference Figure 2 As shown, the specific steps for constructing a distribution network clustering regionalization model to divide the distributed photovoltaic power grid include:

[0063] Based on the distribution of distributed photovoltaic power in the distribution network, each location where distributed photovoltaic power is installed independently is set as a distributed photovoltaic unit;

[0064] Obtain the geographic coordinates, historical power generation data, and installation attribute information of all distributed photovoltaic units;

[0065] Based on the geographical coordinates, historical power generation data, and installation attribute information of distributed photovoltaic units, multidimensional feature vectors related to their power generation characteristics are extracted.

[0066] Based on the Z-score standardization formula, the multidimensional feature vectors are normalized to eliminate the influence of data dimensions.

[0067] Based on the Euclidean distance formula and combined with the normalized multidimensional feature vector, the Euclidean distance value between any two units is calculated to quantify the similarity between the two units.

[0068] Based on the k-distance graph method, the neighborhood radius of the density clustering algorithm is determined. With minimum number of points threshold ;

[0069] Based on the Euclidean distance between any two units, obtain its Number of internal cells;

[0070] Among them, for distributed photovoltaic units ,That Defined as: all with The Euclidean distance is less than or equal to The set of units, i.e. , For the first Multidimensional feature vectors of distributed photovoltaic units For the first Multidimensional feature vectors of distributed photovoltaic units For the first The and the first Euclidean distance between the multidimensional feature vectors of a distributed photovoltaic unit;

[0071] judge If the number of internal cells is not less than the minimum number of points threshold, then determine the cell. If it is the core object, then the determination unit is used. These are non-core objects;

[0072] by The density reachability is used as the criterion for judging the current cluster expansion. All distributed photovoltaic units are traversed to form a virtual predictive clustering region.

[0073] After the traversal is complete, the units that are not assigned to any cluster are set as independent distributed photovoltaic prediction units.

[0074] The virtual prediction clustering region and the distributed photovoltaic prediction unit are respectively used as the edge nodes of the distributed photovoltaic distribution network, and a distribution network clustering regionalization model is constructed based on the edge nodes.

[0075] This can be explained by the fact that distributed photovoltaic (PV) systems, typically installed on residential rooftops and in industrial and commercial buildings, are small-scale and dispersedly connected to the distribution network. Their PV capacity is more susceptible to local weather conditions (such as building shading and localized cloud cover), exhibiting greater volatility and randomness, making prediction more difficult. However, meteorological elements such as cloud cover, sunlight, and temperature are spatially continuous and do not affect only one rooftop while skipping adjacent ones. Furthermore, deploying numerous meteorological sensors in densely packed, geographically similar units within the same area is redundant and wasteful of resources. Therefore, this solution constructs a regionalized distribution network clustering model to regionalize the dispersed and highly individualized distributed PV systems, obtaining virtual prediction clusters with similar characteristics and clusters with unique distributed PV systems. This measurement unit effectively reduces the deployment of data acquisition devices and improves the system's computational efficiency. Geographic coordinate features include at least longitude, latitude, and altitude; historical power generation data features include at least average power generation and average power output; and installation attribute features include at least installed capacity, roof tilt angle, and orientation angle. Since historical power generation data is a continuous value that changes over time, directly calculating the Euclidean distance using the power value of a single time point will lead to deviations due to randomness (such as instantaneous weather fluctuations). Therefore, by setting a time window (such as 1 day, 1 week, or 1 month) and calculating the average power within the window, short-term fluctuations can be smoothed, reflecting the unit's stable power generation level during that period, making it a statistically significant stable numerical feature. The neighborhood radius... With minimum number of points threshold The steps to determine this are as follows: Setting The initial value is ,in, Based on experience, the dimension of the multidimensional feature vector is set to increase by 1, or to 4 or 5, and each unit is calculated to its nth dimension. The Euclidean distances of the nearest neighbor cells are calculated, and the distance values ​​corresponding to all cells are sorted in descending order to generate a k-distance sequence. The k-distance sequence graph is plotted, and the distance value corresponding to the inflection point where the curve in the graph changes from steep to gentle is determined as the optimal neighborhood radius. Specifically, traversing all distributed photovoltaic units to form a virtual predictive clustering region includes: starting from any core object, and... All units within the current cluster are included in the current cluster. Then, the newly added units in the neighborhood are traversed. If a new core object is found, the units in its neighborhood that have not yet belonged to any cluster are also included in the current cluster. This process is repeated until no new units can be added to the cluster, and finally a virtual predicted clustering region is formed.

[0076] The prediction of the long-term baseline trend of distributed photovoltaic power generation data characteristics of each edge node based on the seasonal SARIMA algorithm model specifically includes:

[0077] For each edge node, the historical total power generation data and meteorological data of the same sequence are aggregated to construct the edge node dataset;

[0078] The data in the edge node dataset is filtered, normalized, outlier removed, and interpolated.

[0079] The processed edge node dataset is divided into a training set and a validation set.

[0080] Based on the seasonal SARIMA algorithm model, seasonal difference processing is performed on the centralized data of the processed edge node dataset to obtain the baseline prediction value of power generation at each future time.

[0081] The trained seasonal SARIMA algorithm model is used to backtest the training set and make predictions on the data in the validation set.

[0082] Based on the root mean square error, the error index between the predicted results and the true values ​​is calculated to verify the accuracy and generalization ability of the seasonal SARIMA algorithm model.

[0083] It can be explained that the characteristic sequence of photovoltaic power generation data contains two main components: one is the long-term trend and seasonality determined by the Earth's revolution, rotation and local climate patterns; the other is the random noise and short-term fluctuations caused by short-term weather fluctuations. The seasonal SARIMA algorithm model, as a commonly used technique for long-term trends and seasonality, is good at accurately extracting and predicting these long-term and stable regular components from historical data due to its seasonal difference capability. It can effectively exclude most of the long-term trend power baseline prediction values ​​caused by short-term weather disturbances. This provides a stable benchmark framework for the subsequent introduction of cloud tracking models that focus on handling short-term fluctuations, enabling the two models to have a clear division of labor and efficient coupling.

[0084] Reference Figure 3 As shown, the short-term prediction model for constructing power data features specifically predicts the short-term power data feature values ​​of each edge node, including:

[0085] Deploy all-sky imagers or utilize satellite remote sensing data to acquire dynamic cloud image sequences of each edge node in real time;

[0086] Based on the dynamic cloud map sequence of each edge node, calculate the cloud's moving speed, direction, height, and optical thickness;

[0087] Based on the cloud's movement speed, direction, height, and optical thickness, and using an LSTM neural network model, the influence weights of the cloud on the electrical power data feature data are obtained.

[0088] Based on the influence weight of cloud layers on the characteristic data of power data, a short-term prediction model for the characteristic data of power data is constructed to predict the short-term characteristic data values ​​of power data for each edge node.

[0089] This can be explained by the fact that the power data characteristics output of distributed photovoltaic (PV) systems depend on solar radiation intensity, and clouds are the most direct and drastic factor affecting short-term fluctuations in solar radiation intensity. Rapid cloud movement causes sudden increases and decreases in surface solar radiation intensity, and the optical thickness and height of clouds determine the degree of attenuation of solar radiation. In contrast, meteorological factors such as temperature and humidity have a slow and indirect impact on PV power. Most importantly, cloud movement is predictable. Therefore, by using all-sky imagers or satellite remote sensing data, the position and movement trends of clouds can be observed in advance. Based on these movement patterns, it is possible to predict when clouds will cover each edge node, thereby obtaining the influence weight of clouds on the power data characteristics. This facilitates the construction of a short-term prediction model for power data characteristics, predicting the short-term power data characteristics values ​​of each edge node.

[0090] The calculation of cloud movement speed, direction, height, and optical thickness specifically includes:

[0091] The movement speed and direction of clouds are determined using a cross-correlation algorithm, with the specific formula as follows:

[0092]

[0093] In the formula, This represents the normalized cross-correlation values ​​between two consecutive cloud images. This represents the number of pixels the cloud layer has shifted along the horizontal x-axis and vertical y-axis. For the pixel coordinates of the cloud map, , The values ​​represent the grayscale values ​​of the cloud image at different coordinates at different times. , These represent the average values ​​of the pixels within the corresponding window, where the cloud's movement speed is: The direction of cloud movement is: , It is a two-parameter arctangent function. The time interval between two consecutive cloud images;

[0094] Normalized cross-correlation values ​​in the cross-correlation algorithm traverse all possible Find the two cloud images and The displacement with the highest similarity (i.e., the largest function value) is then used to find the displacement with the highest similarity. To calculate the speed and direction of cloud movement;

[0095] The height of clouds is determined using a stereoscopic method, with the specific formula as follows:

[0096]

[0097] In the formula, The height of the clouds. The baseline distance between all-sky imagers. and It is the angle of the cloud center relative to the zenith direction of each all-sky imager;

[0098] The optical thickness of clouds is determined using the Lambert-Beer law, with the specific formula as follows:

[0099]

[0100] In the formula, The optical thickness of the cloud layer. For air quality, This is the actual value measured by the light intensity sensor. The intensity of sunlight on a clear, cloudless day;

[0101] The weighting of the influence of cloud layer on the feature data of electrical power data specifically includes:

[0102] The cloud's moving speed, direction, height, optical thickness, plane coordinates of each edge node, and plane coordinates of the cloud layer are used as inputs.

[0103] The weight of the influence of cloud layer on the characteristic data of power data is used as the output quantity;

[0104] Based on historical data, a training sample set for the LSTM neural network model is constructed, and outliers are removed, interpolated, and normalized in the training sample set. The training sample set consists of a training set and a validation set.

[0105] To minimize the error between the predicted and measured power values, the loss function of the LSTM neural network model is constructed using the mean square error.

[0106] Based on the cloud's movement speed and direction, the planar coordinates of each edge node, and the cloud's planar coordinates, a physical association between the edge nodes and the cloud is constructed.

[0107] In this process, after constructing the physical association between edge nodes and clouds, the association results are used as derived features, which, together with the original input, constitute the input feature vector of the LSTM neural network model for model training and inference. By introducing physical association features, the model can more accurately capture the spatial interaction between clouds and edge nodes, thereby improving the prediction accuracy of the influencing weights.

[0108] The LSTM neural network model is trained based on the training set, and the influence weights of cloud layer on the feature data of electric power are obtained according to the real-time monitored input.

[0109] The physical association between the edge nodes and the cloud layer specifically includes:

[0110]

[0111] In the formula, This represents the effective distance the cloud layer moves along the direction pointing towards the edge node within the prediction time. To predict the duration of the time window, This is the azimuth angle of the edge node relative to the current position of the cloud layer;

[0112]

[0113] In the formula, The horizontal distance between the edge node and the cloud layer. The plane coordinates of the clouds, These are the planar coordinates of the edge nodes;

[0114] The short-term prediction model expression for the power data characteristics is as follows:

[0115]

[0116] In the formula, This represents a short-term forecast of the electrical power at the edge nodes. The weighting of cloud cover influence on the characteristic data of electrical power data. This represents the electrical power value under clear, cloudless conditions.

[0117] The short-term power generation prediction sequence obtained by the short-term prediction model based on the characteristics of power data can be further derived into short-term prediction values ​​of other power data characteristics, such as the average power generation, cumulative power generation, and power fluctuation rate within the time period, according to their physical meaning.

[0118] The short-term prediction model based on the seasonal SARIMA algorithm model and power data characteristics constructs a comprehensive prediction model of power data characteristics to predict the comprehensive value of power data characteristics for each edge node. Specifically, this includes:

[0119] Based on the seasonal SARIMA algorithm model, with a fixed long period Execute once; each time it is executed, the model generates a future preset range. Internal, in terms of time resolution A sequence of baseline values ​​for power generation data at intervals;

[0120] Based on the characteristic baseline value sequence of power generation data, a corresponding absolute timestamp is set for each value in the sequence. , ,in, This represents the start time of model execution. , , This represents the number of data points in the baseline value sequence;

[0121] Short-term prediction models based on power data characteristics, with fixed short periods The module executes in a rolling fashion, and each time it executes, it outputs a time period in the future. Internal, in terms of time resolution A sequence of correction coefficients for power generation data characteristics at intervals;

[0122] Based on the power generation data characteristic correction coefficient sequence, a corresponding absolute timestamp is set for each value in the sequence. , ,in, This is the absolute start time of this prediction execution. , , To correct the number of data points in the coefficient sequence;

[0123] Define the baseline value sequence of power generation data characteristics itself as the target time axis. This timeline contains all absolute time points in the baseline value sequence. and its corresponding baseline power ,in, ;

[0124] The absolute timestamp of the power generation data characteristic correction coefficient sequence Based on the nearest neighbor matching method, find the match with Find the closest correction factor and assign it to ;

[0125] Based on baseline power and correction factor A comprehensive prediction model for power data features is constructed to predict the comprehensive value of power data features for each edge node.

[0126] The expression for the comprehensive prediction model based on the characteristics of the power data is as follows:

[0127]

[0128] In the formula, This is a comprehensive value representing the characteristics of electrical power data.

[0129] This can be explained by the fact that the seasonal SARIMA algorithm model excels at capturing long-term, stable trends and seasonality, while the short-term prediction model based on power data characteristics can accurately depict short-term fluctuations. Therefore, it's necessary to improve the accuracy and stability of distributed photovoltaic (PV) forecasting in the distribution network by providing trend anchors through a long-term baseline and dynamically adjusting short-term correction coefficients. Specifically, the seasonal SARIMA algorithm model predicts long-term baseline trends, while the short-term prediction model based on power data characteristics has different start times (e.g., SARIMA executes daily, while the short-term model executes on a rolling basis every minute). Aligning with absolute timestamps ensures that each long-term baseline value matches the most suitable correction coefficient, avoiding correction deviations caused by time misalignment. Furthermore, the execution cycles of the seasonal SARIMA algorithm model and the short-term prediction model based on power data characteristics can be adjusted independently in different scenarios. For example, due to the large computational load, the seasonal SARIMA algorithm model has a cycle of 1 hour, while the short-term prediction model based on power data characteristics has a cycle of 5 minutes due to high real-time requirements. Absolute time alignment does not limit the model's execution frequency; dynamic fusion is achieved simply by matching timestamps, adapting to the characteristics of photovoltaic power that are slow to change in the long term and fast to change in the short term.

[0130] The process of utilizing Internet of Things (IoT) technology, based on a comprehensive prediction model of power data characteristics, to obtain the comprehensive value of power data characteristics of each edge node, and constructing a power data characteristic database specifically includes:

[0131] Using IoT technology, a sensing network covering all edge nodes of the distribution network is built to obtain the data collected by each edge node, the output data of the short-term prediction model of seasonal SARIMA algorithm model and power data characteristics, and the comprehensive value of power data characteristics.

[0132] Construct an electric power data feature database, and classify, store, and record data transmitted in real time via IoT technology according to edge node type;

[0133] Extract the comprehensive value of power data features from each edge node, and obtain the total comprehensive prediction value of distributed photovoltaic power in the distribution network by weighted summation;

[0134] The weight of each edge node is determined by the historical prediction error of each edge node.

[0135] This can be explained by the fact that by utilizing IoT technology to build a sensing network covering all edge nodes, data can be collected and transmitted in real time and comprehensively. This includes raw data collected from edge nodes, long-term baseline data from seasonal SARIMA algorithm models, corrected data from short-term prediction models of power data characteristics, and the final comprehensive value of power data characteristics. By aggregating the data from each edge node, the total comprehensive prediction value of distributed photovoltaic power in the distribution network is obtained through weighted summation. Furthermore, by classifying and storing data according to edge node type, the massive and dispersed photovoltaic data becomes ordered and searchable. Whether for subsequent model training, error analysis, or distribution network scheduling decisions, the required data can be quickly retrieved from the database, improving data utilization efficiency. The weight of each edge node is determined by the historical prediction error of each edge node in the following specific way:

[0136] For each edge node, select a historical period within one month and obtain the root mean square error between its predicted value and the actual value;

[0137] Based on the root mean square error of each edge node, obtain the reciprocal of its root mean square error;

[0138] Sum the reciprocals of the root mean square errors of all edge nodes, and obtain the weight of each edge node using a normalization formula.

[0139] Furthermore, based on the same inventive concept as the aforementioned distribution network distributed photovoltaic forecasting method based on multi-data coupling, this solution proposes a distribution network distributed photovoltaic forecasting system based on multi-data coupling, comprising:

[0140] The node partitioning module is used to set edge nodes according to the distribution of distributed photovoltaic power in the distribution network, and to construct a distribution network clustering regionalization model based on the edge nodes to partition the distributed photovoltaic power in the distribution network.

[0141] The photovoltaic prediction module is used to predict the long-term baseline trend of distributed photovoltaic power generation data characteristics of each edge node based on a seasonal SARIMA algorithm model; to construct a short-term prediction model of power data characteristics based on cloud motion tracking for each edge node, and to predict the short-term power data characteristic values ​​of each edge node; and to construct a comprehensive prediction model of power data characteristics based on the seasonal SARIMA algorithm model and the short-term prediction model of power data characteristics, and to predict the comprehensive value of power data characteristics of each edge node.

[0142] The database module is used to utilize Internet of Things (IoT) technology, based on a comprehensive prediction model of power data characteristics, to obtain the comprehensive value of power data characteristics of each edge node, and to construct a power data characteristic database.

[0143] The photovoltaic prediction module includes:

[0144] The long-term trend unit is used to predict the long-term baseline trend of the distributed photovoltaic power generation data characteristics of each edge node based on the seasonal SARIMA algorithm model.

[0145] A short-term prediction unit is used to construct a short-term prediction model of the power data characteristics of each edge node based on cloud motion tracking, and to predict the short-term power data characteristic data values ​​of each edge node.

[0146] The comprehensive prediction unit is used to construct a comprehensive prediction model of power data features based on a short-term prediction model using a seasonal SARIMA algorithm model and power data features, and to predict the comprehensive value of power data features of each edge node.

[0147] In summary, the advantages of this invention are: it effectively divides the distributed photovoltaic power grid into detailed categories and improves the accuracy of distributed photovoltaic forecasting and resource optimization rate by using long-term and short-term forecasting methods.

[0148] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A multi-data coupling-based distribution network distributed photovoltaic prediction method, characterized in that, The application relates to a power grid distributed photovoltaic power prediction method and device. According to the distribution of the power grid distributed photovoltaic power, edge nodes are arranged, and a power grid clustering regionalization model is constructed based on the edge nodes to divide the power grid distributed photovoltaic power. According to each edge node, a long-term baseline trend of the distributed photovoltaic power generation power data characteristics of each edge node is predicted based on a seasonal SARIMA algorithm model. According to each edge node, a short-term prediction model of the electric power data characteristics is constructed based on cloud layer movement tracking to predict the short-term electric power data characteristic data value of each edge node. Based on the seasonal SARIMA algorithm model and the short-term prediction model of the electric power data characteristics, a comprehensive prediction model of the electric power data characteristics is constructed to predict the comprehensive value of the electric power data characteristics of each edge node. Based on the comprehensive prediction model of the electric power data characteristics, the comprehensive value of the electric power data characteristics of each edge node is obtained by using the Internet of Things technology, and an electric power data characteristic database is constructed.

2. The multi-data coupling-based distribution network distributed photovoltaic prediction method according to claim 1, characterized in that, According to the distribution of the power grid distributed photovoltaic power, edge nodes are arranged, and a power grid clustering regionalization model is constructed based on the edge nodes to divide the power grid distributed photovoltaic power. According to the distribution of the power grid distributed photovoltaic power, each position where a distributed photovoltaic power is independently installed is set as a distributed photovoltaic unit. The geographic coordinates, historical power generation data and installation attribute information of all the distributed photovoltaic units are obtained. Based on the geographic coordinates, historical power generation data and installation attribute information of the distributed photovoltaic units, a multi-dimensional feature vector related to the power generation characteristics is extracted. Based on the Z-score standardization formula, the multi-dimensional feature vector is normalized to eliminate the influence of the data dimension. Based on the Euclidean distance formula, the Euclidean distance value of any two units is calculated by combining the normalized multi-dimensional feature vector to quantify the similarity between the two units. Determining a neighborhood radius for a density clustering algorithm based on a k-distance graph With a minimum point threshold ; Based on the Euclidean distance value of any two cells, the number of inner cells is obtained ​ wherein, for a distributed photovoltaic unit , is defined as the set of units whose Euclidean distance from is less than or equal to , i.e. , is the th distributed photovoltaic unit multi-dimensional feature vector, is the th distributed photovoltaic unit multi-dimensional feature vector, is the Euclidean distance between the th and the th distributed photovoltaic unit multi-dimensional feature vector. determining whether the number of inner cells is not less than a minimum point threshold, and if so, the determining unit determines that the object is a core object, and if not, the determining unit determines that the object is a non-core object; With The density accessibility is taken as the judgment basis of current cluster expansion, all distributed photovoltaic units are traversed, and a virtual prediction clustering area is formed. After the traversal is completed, the units that are not divided into any cluster are set as independent distributed photovoltaic prediction units. The virtual prediction clustering region and the distributed photovoltaic prediction unit are respectively set as the edge nodes of the power grid distributed photovoltaic power, and a power grid clustering regionalization model is constructed based on the edge nodes.

3. The multi-data coupling-based distribution network distributed photovoltaic prediction method according to claim 2, characterized in that, For each edge node, the historical total power generation data and the meteorological data at the same time sequence of each edge node are summarized to construct an edge node data set. The data in the edge node data set is filtered, normalized, abnormal value removed and interpolated. The data in the processed edge node data set is divided into a training set and a verification set. Based on the seasonal SARIMA algorithm model, the seasonal difference processing is performed on the data in the processed edge node data set to obtain the baseline prediction value of the power generation power at each future time. The trained seasonal SARIMA algorithm model is used to backtest the training set, and the data in the verification set is predicted. Based on the root mean square error, the error index between the prediction result and the true value is calculated to verify the accuracy and generalization ability of the seasonal SARIMA algorithm model. ​ 4. The multi-data coupling-based distribution network distributed photovoltaic prediction method according to claim 3, characterized in that, The short-term prediction model of the electric power data characteristics is constructed according to the cloud layer movement tracking of each edge node, and the short-term electric power data characteristic data value of each edge node is predicted, and the short-term prediction model of the electric power data characteristics specifically includes: A full-sky imager or satellite remote sensing data is deployed to obtain a dynamic cloud sequence of each edge node in real time; The moving speed, direction, height and optical thickness of the cloud layer are calculated according to the dynamic cloud sequence of each edge node; The influence weight of the cloud layer on the electric power data characteristics is obtained based on the LSTM neural network model according to the moving speed, direction, height and optical thickness of the cloud layer; The short-term prediction model of the electric power data characteristics is constructed based on the influence weight of the cloud layer on the electric power data characteristics, and the short-term electric power data characteristic data value of each edge node is predicted.

5. The multi-data coupling-based distribution network distributed photovoltaic prediction method according to claim 4, characterized in that, The comprehensive prediction model of the electric power data characteristics is constructed based on the seasonal SARIMA algorithm model and the short-term prediction model of the electric power data characteristics, and the comprehensive value of the electric power data characteristics of each edge node is predicted, and the comprehensive prediction model of the electric power data characteristics specifically includes: based on a seasonal SARIMA algorithm model, with a fixed long period is executed once, and each time the model generates a future pre-set range of time resolution of the generated power data feature baseline value sequence; setting a corresponding absolute time stamp for each value in the sequence of baseline values of the data characteristic of the generated power , wherein, is a start time of the execution of the model, , , is a number of data points in the sequence of baseline values. Short-term prediction model based on electrical power data features, to fix short periods is executed, the module outputs a future time period with a time resolution a sequence of correction factors for the interval of the generated power data features; According to the sequence of correction coefficients of the power generation data characteristics, an absolute time stamp corresponding to each value in the sequence is set , wherein, is the starting absolute time of the current prediction execution, , , is the number of data points in the sequence of correction coefficients; The sequence of baseline values of the power generation data feature is defined as a target time axis itself The time axis contains all absolute time points in the sequence of baseline values And their corresponding baseline power Wherein ; The absolute timestamp of the power generation data characteristic correction coefficient sequence Based on the nearest neighbor matching method, find the match with Find the closest correction factor and assign it to ; According to the baseline power and the correction coefficient , a comprehensive prediction model of the electric power data feature is constructed to predict the comprehensive value of the electric power data feature of each edge node. The expression of the comprehensive prediction model of the electric power data characteristics is: ； In the formula, is an electric power data feature synthesis value.

6. The multi-data coupling-based distribution network distributed photovoltaic prediction method according to claim 5, characterized in that, The comprehensive value of the electric power data characteristics of each edge node is obtained based on the comprehensive prediction model of the electric power data characteristics by using the Internet of Things technology, and the electric power data characteristic database is constructed, and the comprehensive prediction model of the electric power data characteristics specifically includes: The sensing network covering each edge node of the distribution network is built by using the Internet of Things technology, the collection data of each edge node, the output data of the seasonal SARIMA algorithm model and the short-term prediction model of the electric power data characteristics, and the comprehensive value of the electric power data characteristics are obtained; The electric power data characteristic database is constructed, and the data transmitted in real time through the Internet of Things technology is stored and recorded according to the type of the edge node; The comprehensive prediction value of the distributed photovoltaic of the distribution network is obtained by using the comprehensive value of the electric power data characteristics of each edge node through weighted summation; Wherein, the weight of each edge node is determined by the historical prediction error of each edge node.

7. A multi-data coupling-based distribution network distributed photovoltaic prediction system, characterized in that, The method for realizing the multi-data coupling-based distributed photovoltaic prediction of the distribution network according to any one of claims 1-6 includes: The node division module is used for setting the edge node according to the distribution of the distributed photovoltaic of the distribution network, and constructing the distribution network clustering regionalization model based on the edge node to divide the distributed photovoltaic of the distribution network; The photovoltaic prediction module is used for predicting the long-term baseline trend of the distributed photovoltaic power data characteristics of each edge node based on the seasonal SARIMA algorithm model according to each edge node, predicting the short-term electric power data characteristic data value of each edge node based on the cloud layer movement tracking, constructing the short-term prediction model of the electric power data characteristics, and predicting the comprehensive value of the electric power data characteristics of each edge node based on the seasonal SARIMA algorithm model and the short-term prediction model of the electric power data characteristics; The database module is used for obtaining the comprehensive value of the electric power data characteristics of each edge node based on the comprehensive prediction model of the electric power data characteristics by using the Internet of Things technology, and constructing the electric power data characteristic database.

8. The multi-data coupling based network distribution type photovoltaic prediction system according to claim 7, characterized in that, The photovoltaic prediction module includes: a long-term trend unit configured to predict a long-term baseline trend of the distributed photovoltaic power data feature of each edge node based on a seasonal SARIMA algorithm model according to each edge node; a short-term prediction unit configured to predict a short-term power data feature data value of each edge node based on a cloud layer movement tracking to construct a short-term prediction model of the power data feature according to each edge node; a comprehensive prediction unit configured to predict a comprehensive value of the power data feature of each edge node based on the seasonal SARIMA algorithm model and the short-term prediction model of the power data feature to construct a comprehensive prediction model of the power data feature.

Images