Photovoltaic power short-term prediction method based on multi-meteorological-parameter directed weighted dynamic complex networked representation

Abstract

A photovoltaic power short-term prediction method based on multi-meteorological parameter directed weighted dynamic complex networked representation comprises the following steps: S10, collecting historical and real-time power data of each power generation unit of a target photovoltaic power station, and obtaining various meteorological element forecast data from various meteorological forecast sources; s20, regarding the power generation units 1, 2,..., N and the meteorological elements 1, 2,..., M as nodes, and connecting all N + M nodes in pairs to form a complex network; s30, determining the direction of an edge between any two nodes in the complex network by adopting a causal analysis method, determining the weight of the edge by adopting a mutual information analysis method, and forming a directed weighted complex network; s40, taking data of more than one year in history corresponding to each node, and dynamically updating the directed weighted complex network at different moments by adopting an attention mechanism model to form a directed weighted dynamic complex network; and S50, establishing and training a graph neural network prediction model based on historical directed weighted dynamic complex network data, and realizing photovoltaic power short-term prediction by inputting real-time data of each node. The technical problem to be solved by the invention is to provide a photovoltaic power short-term prediction method based on multi-meteorological-parameter directed weighted dynamic complex networked representation, and the problem of poor photovoltaic power prediction effect caused by low data quality of a single prediction source of an existing station is solved.

Description

Technical Field

[0001] The invention relates to the technical field of short-term prediction of photovoltaic power stations, and in particular to a photovoltaic power short-term prediction method based on directed weighted dynamic complex network representation of multiple meteorological parameters. Background Art

[0002] Photovoltaic power generation is a key way to utilize solar energy, contributing to energy and environmental protection, but its power volatility poses challenges to the stability and economic operation of the power grid. Accurate prediction of photovoltaic power in the short term is particularly critical, as it directly affects the real-time dispatching and energy allocation of the power grid, and is an important link in ensuring energy transformation and economical and efficient operation of the power system.

[0003] At present, the existing short-term prediction methods for photovoltaic power stations are mainly based on short-term prediction models of numerical weather forecasts, including traditional statistical methods and artificial intelligence algorithms, support vector machines and other machine learning methods. In general, a single photovoltaic power station is equipped with only one numerical weather forecast. In the case of missing forecast source data, its short-term prediction accuracy will be greatly reduced. In addition, the variability of weather conditions requires that forecast data must be multi-dimensional and updated at a high frequency. Data from a single source may not respond adequately when facing local meteorological events or sudden changes, and thus the reliability of the forecast will be greatly reduced. In addition, the prediction model structure of the traditional prediction method is simple, and a single deep learning network model is difficult to learn too much useful information from the drastic fluctuations. It ignores the complex interrelationships between multiple meteorological parameters, resulting in limited prediction accuracy and inability to effectively cope with the uncertainty brought about by climate change. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a photovoltaic power short-term prediction method based on directed weighted dynamic complex network characterization of multiple meteorological parameters, so as to solve the problem that the photovoltaic power prediction effect is poor due to the low quality of the single forecast source data of the existing site.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a photovoltaic power short-term prediction method based on directed weighted dynamic complex network representation of multiple meteorological parameters, comprising the following steps: S10, collecting historical and real-time power data of each power generation unit of the target photovoltaic power station, and various meteorological element forecast data obtained from various meteorological forecast sources; S20, treating power generation units 1, 2, ..., N and meteorological elements 1, 2, ..., M as nodes, and connecting all N+M nodes in pairs to form a complex network; S30, for the edge between any two nodes in the complex network, a causal analysis method is used to determine its direction, and a mutual information analysis method is used to determine its weight, so as to form a directed weighted complex network; S40, taking historical data corresponding to each node for more than one year, and using the attention mechanism model to dynamically update the directed weighted complex network at different times to form a directed weighted dynamic complex network; S50. Establish and train a graph neural network prediction model based on historical directed weighted dynamic complex network data, and realize short-term prediction of photovoltaic power by inputting real-time data of each node.

[0006] Preferably, the step S10 includes the following steps: S101, obtaining forecast data of various meteorological elements from various meteorological forecast sources, and obtaining real-time and historical photovoltaic power data of the site; S102, processing the collected data, using a data cleaning method to filter out abnormal values ​​in the data, and using a normalization preprocessing method to unify the data format and range of different data sources.

[0007] Preferably, in step S102, abnormal data are removed from the numerical weather forecast data and historical photovoltaic power series data of each source. If the forecast value of a certain day in a certain forecast source is missing for more than 4 hours, the data of this day is deleted, otherwise it is filled using linear interpolation; at the same time, for the case where the time scales in different weather forecast sources are different and unified, the spline interpolation method is used to fill, ensuring that the time scale of each meteorological parameter and measured power data of each source is 15 minutes; in order to eliminate the difference in the dimensions of each data, the forecast data and power data need to be normalized according to the following formula to keep the data between 0 and 1: ; in and are the original data and the normalized data, and are the maximum and minimum values ​​in the sample.

[0008] Preferably, in step S30, the direction of the edges between the nodes in the network is first determined through the complex network obtained in step S20, and a vector autoregression (VAR) model is constructed by using a Granger causality test and considering any two nodes in the network, a power generation unit i and a meteorological element j: ; in, and are the values ​​of nodes i and j at time t, is the lag order, , , is the coefficient, is the error term; Statistical tests show The coefficient is significant, indicating that right There is Granger causality, that is, point to There is one directed edge; In a weighted network, the weight reflects the strength of the connection between nodes. Therefore, mutual information analysis is used to determine the weight of the edge. For nodes i and j, the mutual information is calculated. : Calculation of the maximum mutual information coefficient, ; in, and represents two random variables, Representation variables and The joint distribution probability of , Respectively represent variables and probability.

[0009] The mutual information value is used as the weight of the edge, reflecting the degree of information exchange between nodes. The weight matrix is ​​constructed based on the mutual information results. , the weight matrix Elements in Determined according to the mutual information value, it represents the weight of the edge between nodes i and j.

[0010] Preferably, the step S40 includes the following steps: Take the historical data of more than one year corresponding to each node for training, establish a graph neural network prediction model with directed weighted complex network data as input and historical photovoltaic power data of the station as output, and perform denormalization on the output photovoltaic power prediction results to obtain the photovoltaic power prediction results of the target station; A graph neural network prediction model is established with directed weighted complex network structure data as input and photovoltaic power at future moments as output; the historical data series corresponding to each node of the target power station for more than one year are divided into training set and test set in a ratio of 7:3. In the training set, the historical 3-day time series data of each node in the directed weighted complex network obtained by S30 is used as input, and the photovoltaic power data of the target station in the next 7 days is used as output. The initialization parameters of the short-term photovoltaic power prediction model based on the neural network are set and trained. The prediction model has a total of four layers of networks, including a graph learning layer, two graph convolution layers, and an output layer. Finally, the output photovoltaic power prediction results are denormalized to obtain the final short-term prediction results of the photovoltaic power of the target station. The denormalization processing method is shown in the following formula: .

[0011] The present invention provides a photovoltaic power short-term prediction method based on directed weighted dynamic complex network characterization of multiple meteorological parameters, which has the following beneficial effects: (1) The photovoltaic power station short-term prediction method of the present invention can effectively reduce the impact on the short-term prediction results even when a certain numerical weather forecast source data is missing, and considers the impact of the accuracy of multiple weather forecast sources on the photovoltaic power prediction of the target site; (2) The present invention analyzes the correlation between various meteorological elements by introducing complex network theory and graph representation technology, effectively capturing the deep-level relationships in multi-source meteorological data, which is beneficial to the training of prediction models and improves prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a short-term prediction result diagram of the present invention. DETAILED DESCRIPTION

[0013] See also Figure 1 , Figure 1 The present invention provides a photovoltaic power short-term prediction method based on a directed weighted dynamic complex network representation of multiple meteorological parameters, such as Figure 1 As shown, the following steps are included: Step S10, collecting historical and real-time power data of each power generation unit of the target photovoltaic power station, and various meteorological element forecast data obtained from various meteorological forecast sources; Step S20, treating the power generation units 1, 2, ..., N and the meteorological elements 1, 2, ..., M as nodes, and connecting all N+M nodes in pairs to form a complex network; Step S30, for the edge between any two nodes in the complex network, a causal analysis method is used to determine its direction, and a mutual information analysis method is used to determine its weight, so as to form a directed weighted complex network; Step S40, taking historical data of more than one year corresponding to each node, combining it with a directed weighted complex network, establishing and training a graph neural network prediction model, and realizing short-term photovoltaic power prediction by inputting real-time data of each node.

[0014] The step of collecting the historical and real-time power data of the target photovoltaic power station and the meteorological data obtained from a variety of meteorological forecast sources and preprocessing the data in step S10 includes: Step S101, obtain the multi-source numerical weather forecast meteorological parameter data predicted one day before the forecast day, and obtain the real-time and historical photovoltaic power data of the site. To ensure the modeling effect, at least two or more forecast sources are required to have a year of site historical forecast data and corresponding time historical power data.

[0015] Step S102, process the collected data, use data cleaning methods to filter out abnormal values ​​in the data, and use normalization preprocessing methods to unify the data format and range of different data sources. Abnormal data is eliminated for the numerical weather forecast data and historical photovoltaic power series data of each source. If the forecast value of a certain day in a forecast source is missing for more than 4 hours, the data of this day is deleted, otherwise it is filled using linear interpolation; at the same time, for the case where the time scales in different weather forecast sources are different and unified, the spline interpolation method is used to fill in, ensuring that the time scale of each meteorological parameter and measured power data of each source is 15 minutes; in order to eliminate the differences in the dimensions of each data, the forecast data and power data need to be normalized according to the following formula to keep the data between 0-1: ; in and are the original data and the normalized data, and are the maximum and minimum values ​​in the sample.

[0016] In the step S20, the power generation units 1, 2, ..., N and the meteorological elements 1, 2, ..., M are all regarded as nodes, and all N+M nodes are connected in pairs to form a complex network, including: Step S201, according to the number N of power generation units of the target site and the type M of meteorological elements in each meteorological forecast source of the target site as nodes, all N+M nodes are connected in pairs to form a complex network.

[0017] A complex network is a mathematical model used to describe and analyze the interaction and connection patterns between elements in a complex system. In many natural and artificial systems, the interactions between elements form a network structure, which is called a complex network due to the complexity of its structure and dynamic characteristics. To build a complex network, various parameters must first be converted into a network structure. In this scenario, each power generation unit and each meteorological parameter can be regarded as a node in the network, and these nodes are connected in pairs to form the basic structure of the network. Each node is connected to all other nodes, forming a fully connected network.

[0018] The adjacency matrix is ​​the core of this structure, which determines the connection relationship between nodes through the correlation between each meteorological parameter data. Elements in It can be defined as: If there is a connection between node i and node j, then The value of is set to 1 if there is no direct connection between node i and node j. At this time, set the element in the adjacency matrix is 0.

[0019] According to the number of power generation units N of the target site and the type of meteorological elements M in each meteorological forecast source of the target site as nodes, all N+M nodes are connected in pairs to form an adjacency matrix of (N+M)×(N+M) as the complex network result. Since all nodes are connected to each other at this time, the adjacency matrix will be a matrix with all elements equal to 1 except the diagonal.

[0020] In step S30, for the edge between any two nodes in the complex network, the causal analysis method is used to determine its direction, and the mutual information analysis method is used to determine its weight, so as to form a directed weighted complex network, including: Step S301, defining the direction of the edge between any two nodes in the complex network by a causal analysis method, and assigning weights to the connections according to a mutual information analysis method, thereby constructing a directed weighted complex network that reflects the interrelated influence effects between meteorological parameters.

[0021] As an advanced network model, directed weighted complex network introduces the concepts of directionality and weight on the basis of ordinary complex network. It reflects the asymmetry and multi-level interaction complexity within the system through the directed connection between nodes and the interaction relationship of different strengths. The use of directed weighted network can better reflect the actual situation of the interrelated influence effects between meteorological parameters.

[0022] Through the complex network obtained in S201, we first determine the direction of the edges between the nodes in the network, and use the Granger causality test to consider any two nodes in the network, such as power generation unit i and meteorological element j, to construct a vector autoregression (VAR) model: ; in, and are the values ​​of nodes i and j at time t, is the lag order, , , is the coefficient, is the error term.

[0023] If the statistical test shows The coefficient is significant, indicating that right There is Granger causality, that is, point to There is one directed edge.

[0024] In a weighted network, the weight reflects the strength of the connection between nodes. Therefore, mutual information analysis is used to determine the weight of the edge. For nodes i and j, calculate the mutual information : Calculation of the maximum mutual information coefficient,.

[0025] ; in, and represents two random variables, Representation variables and The joint distribution probability of , Respectively represent variables and probability.

[0026] The mutual information value is used as the weight of the edge, reflecting the degree of information exchange between nodes. The weight matrix is ​​constructed from the mutual information results. The weight matrix Elements in Determined according to the mutual information value, it represents the weight of the edge between nodes i and j.

[0027] In step S40, historical data corresponding to each node for more than one year is taken, combined with a directed weighted complex network, a graph neural network prediction model is established and trained, and the steps of realizing short-term photovoltaic power prediction by inputting real-time data of each node include: Step S401, take the historical data of more than one year corresponding to each node for training, establish a graph neural network prediction model with directed weighted complex network data as input and historical photovoltaic power data of the station as output, and perform denormalization on the output photovoltaic power prediction result to obtain the photovoltaic power prediction result of the target station.

[0028] Establish a graph neural network prediction model with directed weighted complex network structure data as input and photovoltaic power at future moments as output. Take the historical data series of more than one year corresponding to each node of the target power station and divide it into training set and test set in a ratio of 7:3. In the training set, the historical 3-day time series data of each node in the directed weighted complex network obtained in S301 is used as input, and the photovoltaic power data of the target station in the next 7 days is used as output. Set the initialization parameters of the short-term photovoltaic power prediction model based on the neural network and train it. The prediction model has a total of four layers of networks, including a graph learning layer, two graph convolution layers, and an output layer. Finally, the output photovoltaic power prediction results are denormalized to obtain the final short-term prediction results of the photovoltaic power of the target station. The denormalization processing method is shown in the following formula: .

[0029] To illustrate the effectiveness of the method of the present invention, the method of the present invention was used to predict the photovoltaic power of a photovoltaic power station in Inner Mongolia from September 1, 2023 to September 8, 2023. The comparison results between the predicted value and the actual value are as follows: Figure 2 As shown. Figure 2 It can be seen that, through actual testing, the short-term photovoltaic power prediction method of the present invention has an ideal effect and a high prediction accuracy.

[0030] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A photovoltaic power short-term prediction method based on directed weighted dynamic complex network representation of multiple meteorological parameters, characterized in that: The following steps are involved: S10, collecting historical and real-time power data of each power generation unit of the target photovoltaic power station, and various meteorological element forecast data obtained from various meteorological forecast sources; S20, treating power generation units 1, 2, ..., N and meteorological elements 1, 2, ..., M as nodes, and connecting all N+M nodes in pairs to form a complex network; S30, for the edge between any two nodes in the complex network, a causal analysis method is used to determine its direction, and a mutual information analysis method is used to determine its weight, so as to form a directed weighted complex network; S40, taking the historical data corresponding to each node for more than one year, and using the attention mechanism model to dynamically update the directed weighted complex network at different times to form a directed weighted dynamic complex network; S50. Establish and train a graph neural network prediction model based on historical directed weighted dynamic complex network data, and realize short-term prediction of photovoltaic power by inputting real-time data of each node.

2. According to claim 1, a photovoltaic power short-term prediction method based on directed weighted dynamic complex network representation of multiple meteorological parameters is characterized by: The step S10 comprises the following steps: S101, obtaining forecast data of various meteorological elements from various meteorological forecast sources, and obtaining real-time and historical photovoltaic power data of the site; S102, processing the collected data, using a data cleaning method to filter out abnormal values ​​in the data, and using a normalization preprocessing method to unify the data format and range of different data sources.

3. According to claim 2, a photovoltaic power short-term prediction method based on directed weighted dynamic complex network representation of multiple meteorological parameters is characterized by: In step S102, the numerical weather forecast data and historical photovoltaic power series data of each source are subjected to abnormal data elimination. If the forecast value of a certain day in a certain forecast source is missing for more than 4 hours, the data of this day is deleted, otherwise it is filled using linear interpolation. At the same time, for the case where the time scales in different weather forecast sources are different, the spline interpolation method is used to fill in, ensuring that the time scale of each meteorological parameter and measured power data of each source is 15 minutes. In order to eliminate the difference in the dimensions of each data, the forecast data and power data need to be normalized according to the following formula to keep the data between 0 and 1: ； in and are the original data and the normalized data, and are the maximum and minimum values ​​in the sample.

4. According to claim 1, a photovoltaic power short-term prediction method based on directed weighted dynamic complex network characterization of multiple meteorological parameters is characterized by: In step S30, the direction of the edges between the nodes in the network is first determined through the complex network obtained in step S20, and a vector autoregression model is constructed by using a Granger causality test and considering any two nodes in the network, power generation unit i and meteorological element j: ； in, and are the values ​​of nodes i and j at time t, is the lag order, , , is the coefficient, is the error term; Statistical tests show The coefficient is significant, indicating that right There is Granger causality, that is, point to There is one directed edge; In a weighted network, the weight reflects the strength of the connection between nodes. Therefore, mutual information analysis is used to determine the weight of the edge. For nodes i and j, the mutual information is calculated. : Calculation of the maximum mutual information coefficient, ； in, and represents two random variables, Representation variables and The joint distribution probability of , Respectively represent variables and probability. The mutual information value is used as the weight of the edge, reflecting the degree of information exchange between nodes. The weight matrix is ​​constructed based on the mutual information results. , the weight matrix Elements in Determined according to the mutual information value, it represents the weight of the edge between nodes i and j.

5. The photovoltaic power short-term prediction method based on directed weighted dynamic complex network representation of multiple meteorological parameters according to claim 1 is characterized in that: The step S40 comprises the following steps: Take the historical data of more than one year corresponding to each node for training, establish a graph neural network prediction model with directed weighted complex network data as input and historical photovoltaic power data of the station as output, and perform denormalization on the output photovoltaic power prediction results to obtain the photovoltaic power prediction results of the target station; A graph neural network prediction model is established with directed weighted complex network structure data as input and photovoltaic power at future moments as output; the historical data series corresponding to each node of the target power station for more than one year are divided into training set and test set in a ratio of 7:

3. In the training set, the historical 3-day time series data of each node in the directed weighted complex network obtained by S30 is used as input, and the photovoltaic power data of the target station in the next 7 days is used as output. The initialization parameters of the short-term photovoltaic power prediction model based on the neural network are set and trained. The prediction model has a total of four layers of networks, including a graph learning layer, two graph convolution layers, and an output layer. Finally, the output photovoltaic power prediction results are denormalized to obtain the final short-term prediction results of the photovoltaic power of the target station. The denormalization processing method is shown in the following formula: 。

Images