Distributed photovoltaic power station cluster power prediction model construction method and system, and distributed photovoltaic power station cluster power prediction method and system

Abstract

The invention provides a distributed photovoltaic power station cluster power prediction model construction and prediction method and system, and the method comprises the steps: obtaining historical operation data, selecting a plurality of representative power stations through employing a maximum correlation-minimum redundancy feature selection method, and solving the weight of each representative power station according to an eagle optimization algorithm; obtaining a power prediction sub-model of each representative power station based on a space-time diagram convolution model; performing weighted combination on the power prediction sub-models of the representative power stations to obtain a distributed photovoltaic power station cluster power prediction model; according to the method and the system, a plurality of representative power stations are selected through a maximum correlation-minimum redundancy feature selection method, the selected representative power stations are highly related to cluster power but have the lowest redundancy, weights of the representative power stations are determined through an eagle optimization algorithm, the representative power stations are predicted in combination with a space-time diagram convolution model, and the cluster power is predicted. The overall characteristics of the cluster and the spatial-temporal correlation of the power data of the photovoltaic power station can be better captured, so that the cluster power prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of cluster power prediction, and in particular relates to the construction of a distributed photovoltaic power station cluster power prediction model, a prediction method and a system. Background Art

[0002] The world is currently facing an energy crisis. The large amount of greenhouse gases and pollutants produced during the mining and use of traditional fossil fuel resources have led to global climate change. In this context, the development and utilization of clean renewable energy has become an urgent need. As a renewable energy source, solar energy has attracted much attention for its inexhaustible characteristics. In particular, photovoltaic power generation technology is considered to be one of the most promising new energy sources because it can directly convert solar energy into electrical energy with high efficiency and does not produce any pollutants. Photovoltaic power generation can not only effectively alleviate the problem of energy shortage and reduce carbon emissions, but also optimize the power system through distributed applications and promote the sustainability of energy utilization. Therefore, the widespread application and promotion of photovoltaic technology has become an important solution to the global energy crisis and environmental challenges. In recent years, driven by the goals of "carbon peak" and "carbon neutrality", the photovoltaic industry has developed rapidly. Due to the dispersed solar energy resources and low energy density, photovoltaic power generation is naturally suitable for distributed power generation mode. In addition, the distribution of wasteland resources and electricity load is extremely uneven, and distributed photovoltaic power generation will be the optimal form of large-scale photovoltaic application.

[0003] With the increase of distributed photovoltaic systems, traditional power dispatching and management methods are difficult to adapt to new demands, and the power system needs more accurate predictions to balance supply and demand. Distributed photovoltaic power generation has the characteristics of dispersion, intermittency and volatility. Its power generation is greatly affected by weather, seasons and day and night changes, which makes the power output of a single photovoltaic power station difficult to predict and manage. These characteristics increase the difficulty of grid dispatching and stable operation. The output of distributed photovoltaic clusters has better regularity than that of a single distributed power generation system, which greatly improves the power prediction accuracy of distributed photovoltaic cluster power generation and reduces the difficulty of power prediction. Therefore, cluster prediction is particularly important.

[0004] In cluster power prediction, the selection of representative power stations is crucial. If the selected representative power station is not typical and cannot represent the entire cluster, then even if the prediction accuracy of the representative power station is high, the prediction accuracy of the cluster cannot be guaranteed. Therefore, it is very important to select representative power stations that are highly correlated with cluster power to improve prediction accuracy. Current research mainly focuses on using time series-based methods for prediction. However, due to the spatial correlation of distributed photovoltaic systems, photovoltaic sites located in similar geographical and climatic environments usually show similar power changes at the same time point. This means that there is a certain correlation between the historical power of each site and the future power of the neighboring sites, that is, the spatiotemporal correlation. How to predict the power of a distributed photovoltaic power station cluster while considering the spatiotemporal correlation characteristics is an issue that needs further study. Summary of the invention

[0005] In order to overcome the above-mentioned deficiencies of the prior art, the present invention proposes a method for constructing a distributed photovoltaic power station cluster power prediction model, comprising:

[0006] Obtain the historical operation data of each power station in the distributed photovoltaic power station cluster;

[0007] Based on the historical operation data of each of the power stations, a maximum relevance-minimum redundancy feature selection method is used to select multiple representative power stations from each of the power stations, and the optimal weight of each of the representative power stations is obtained by solving the Egret Eagle optimization algorithm;

[0008] Based on the historical operation data of the plurality of representative power stations, the spatiotemporal graph convolution model is trained to obtain a power prediction sub-model of each representative power station;

[0009] Based on the optimal weight of each representative power station, the power prediction sub-models of the plurality of representative power stations are weighted and combined to obtain a constructed distributed photovoltaic power station cluster power prediction model.

[0010] Preferably, the method of selecting a plurality of representative power stations from each power station by using a maximum relevance-minimum redundancy feature selection method based on the historical operation data of each power station comprises:

[0011] Based on the historical power in the historical operation data of each of the power stations, a first comprehensive correlation coefficient between each of the power stations and the distributed photovoltaic power station cluster, and a second comprehensive correlation coefficient between each of the power stations are calculated; the first comprehensive correlation coefficient and the second comprehensive correlation coefficient are calculated according to the average of multiple sub-correlation coefficients;

[0012] Based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient, a maximum correlation-minimum redundancy feature selection method is adopted to select a plurality of representative power stations from each of the power stations.

[0013] Preferably, the method of selecting a plurality of representative power stations from each of the power stations using a maximum correlation-minimum redundancy feature selection method based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient includes:

[0014] Based on the first comprehensive correlation coefficients corresponding to the power stations, a power station with the largest first comprehensive correlation coefficient is selected as a representative power station, and power stations in the distributed photovoltaic power station cluster except the representative power station are used as remaining power stations;

[0015] Obtaining an average value of the second comprehensive correlation coefficients between each of the remaining power stations and the selected representative power station to determine the redundancy;

[0016] Based on each of the remaining power stations, the difference between the first comprehensive correlation coefficient corresponding to the remaining power stations and the redundancy is calculated; the remaining power station with the largest difference is selected as the representative power station, and the number of remaining power stations is updated, and the step of obtaining the average value of the second comprehensive correlation coefficient between each of the remaining power stations and the selected representative power station is returned to determine the new redundancy and the new representative power station, until the total capacity of all the selected representative power stations is not less than the set capacity threshold, the cycle is stopped, and the final multiple representative power stations are obtained.

[0017] Preferably, the plurality of sub-correlation coefficients include Pearson's correlation coefficient, Spearman's rank correlation coefficient and Kendall's rank correlation coefficient.

[0018] Preferably, the method of solving the optimal weight of each representative power station according to the Egret Eagle optimization algorithm includes:

[0019] Initializing the parameters of the egret optimization algorithm, wherein the number of egret groups in the parameters is the number of the plurality of representative power stations, each egret in the egret group corresponds to one of the representative power stations, and the position of each egret corresponds to the weight of one of the representative power stations;

[0020] Calculating the fitness value of each egret's position according to a preset fitness function, with the goal of finding a position with the minimum fitness value, and determining the best position for each egret;

[0021] Based on the egret optimization algorithm, the hunting behavior of the egret is simulated in the exploration phase, the position of the egret is preliminarily updated, and the optimal position of the egret is preliminarily updated according to the fitness function. In the development phase, the avoiding behavior of the egret is simulated, the position of the egret is updated again, and the optimal position of the egret is updated again according to the fitness function. The position and optimal position of the egret are updated repeatedly until the maximum number of iterations is reached. The optimal position of each egret when the maximum number of iterations is reached is used as the optimal weight of the corresponding representative power station.

[0022] Preferably, the step of calculating the fitness value of the position of each egret according to a preset fitness function comprises:

[0023] Based on the fitness function, the fitness value of the position of each egret is calculated according to the square difference between the weighted sum of the historical power of each representative power station and the historical total power of the distributed photovoltaic power station cluster.

[0024] Preferably, the spatiotemporal graph convolution model is trained based on the historical operation data of the plurality of representative power stations to obtain a power prediction sub-model of each representative power station, including:

[0025] Acquire historical power and geographical location information in historical operation data of each of the representative power stations, and determine static graph structure data of the plurality of representative power stations based on the geographical location information of the plurality of representative power stations;

[0026] Based on each of the representative power stations, the spatiotemporal graph convolution model is trained with the historical power of the representative power station and the static graph structure data as input to obtain a power prediction sub-model of the representative power station.

[0027] Preferably, the determining of the static graph structure data of the plurality of representative power stations based on the geographical location information of the plurality of representative power stations comprises:

[0028] Based on the geographical location information of each of the representative power stations, construct an adjacency matrix of the plurality of representative power stations, and use the adjacency matrix as the static graph structure data;

[0029] The adjacency matrix is ​​expressed as:

[0030]

[0031] Wherein, A(i,j) is the adjacency matrix, D(i,j) is the geographical distance between representative power station i and representative power station j; S is the variance of the geographical distance between each representative power station; ε is the distance threshold; are the latitudes of power station i and power station j respectively; i , j are the longitudes of power station i and j respectively; R is the radius of the earth.

[0032] Preferably, the historical power of the representative power station and the static graph structure data are used as input to train the spatiotemporal graph convolution model to obtain the power prediction sub-model of the representative power station, including:

[0033] Acquire the historical power of the representative power station and the static graph structure data to construct a data set; divide the data set into a training set and a test set according to a set ratio;

[0034] Based on the training set, the historical power of the representative power station in the current historical training period and the space-time graph structure data are used as input, and the historical power of the representative power station in the next historical training period of the current historical training period is used as output, the space-time graph convolution model is trained, the optimizer is used to optimize and update the training parameters of the space-time graph convolution model, and the test set is used to evaluate the space-time graph convolution model during the training process until the loss function of the space-time graph convolution model converges, so as to obtain the trained space-time graph convolution model.

[0035] Based on the same inventive concept, the present invention also provides a distributed photovoltaic power station cluster power prediction model construction system, including:

[0036] The first data acquisition module is used to acquire the historical operation data of each power station in the distributed photovoltaic power station cluster;

[0037] A weight calculation module is used to select a plurality of representative power stations from the power stations based on the historical operation data of the power stations by using the maximum relevance-minimum redundancy feature selection method, and obtain the optimal weight of each representative power station by solving the optimization algorithm according to the Heron Eagle optimization algorithm;

[0038] A sub-model training module, used to train the spatiotemporal graph convolution model based on the historical operation data of the plurality of representative power stations, to obtain a power prediction sub-model of each of the representative power stations;

[0039] The combined model building module is used to weightedly combine the power prediction sub-models of the multiple representative power stations based on the optimal weight of each representative power station to obtain a constructed distributed photovoltaic power station cluster power prediction model.

[0040] Preferably, the weight calculation module includes:

[0041] A coefficient calculation unit, configured to calculate, based on the historical power in the historical operation data of each of the power stations, a first comprehensive correlation coefficient between each of the power stations and the distributed photovoltaic power station cluster, and a second comprehensive correlation coefficient between each of the power stations; the first comprehensive correlation coefficient and the second comprehensive correlation coefficient are calculated based on the average of a plurality of sub-correlation coefficients;

[0042] A power station selection unit is used to select a plurality of representative power stations from each of the power stations by using a maximum correlation-minimum redundancy feature selection method based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient.

[0043] Preferably, the power station selection unit is specifically used for:

[0044] Based on the first comprehensive correlation coefficients corresponding to the power stations, a power station with the largest first comprehensive correlation coefficient is selected as a representative power station, and power stations in the distributed photovoltaic power station cluster except the representative power station are used as remaining power stations;

[0045] Obtaining an average value of the second comprehensive correlation coefficients between each of the remaining power stations and the selected representative power station to determine the redundancy;

[0046] Based on each of the remaining power stations, the difference between the first comprehensive correlation coefficient corresponding to the remaining power stations and the redundancy is calculated; the remaining power station with the largest difference is selected as the representative power station, and the number of remaining power stations is updated, and the step of obtaining the average value of the second comprehensive correlation coefficient between each of the remaining power stations and the selected representative power station is returned to determine the new redundancy and the new representative power station, until the total capacity of all the selected representative power stations is not less than the set capacity threshold, the cycle is stopped, and the final multiple representative power stations are obtained.

[0047] Preferably, the plurality of sub-correlation coefficients include Pearson's correlation coefficient, Spearman's rank correlation coefficient and Kendall's rank correlation coefficient.

[0048] Preferably, the weight calculation module further includes:

[0049] An initialization unit, used for initializing the parameters of the egret optimization algorithm, wherein the number of egret groups in the parameters is the number of the plurality of representative power stations, each egret in the egret group corresponds to one of the representative power stations, and the position of each egret corresponds to the weight of one of the representative power stations;

[0050] A position determination unit, used for calculating the fitness value of the position of each egret according to a preset fitness function, with the goal of finding the position with the minimum fitness value, and determining the best position of each egret;

[0051] The solving unit is used for simulating the hunting behavior of the egret eagle based on the egret optimization algorithm in the exploration phase, preliminarily updating the position of the egret eagle, and preliminarily updating the optimal position of the egret eagle according to the fitness function; simulating the avoiding behavior of the egret eagle in the development phase, updating the position of the egret eagle again, and updating the optimal position of the egret eagle again according to the fitness function, and iteratively updating the position and optimal position of the egret eagle again until the maximum number of iterations is reached, and the optimal position of each egret eagle when the maximum number of iterations is reached is used as the optimal weight of the corresponding representative power station.

[0052] Preferably, the position determination unit is specifically used for:

[0053] Based on the fitness function, the fitness value of the position of each egret is calculated according to the square difference between the weighted sum of the historical power of each representative power station and the historical total power of the distributed photovoltaic power station cluster.

[0054] Preferably, the sub-model training module is specifically used for:

[0055] Acquire historical power and geographical location information in historical operation data of each of the representative power stations, and determine static graph structure data of the plurality of representative power stations based on the geographical location information of the plurality of representative power stations;

[0056] Based on each of the representative power stations, the spatiotemporal graph convolution model is trained with the historical power of the representative power station and the static graph structure data as input to obtain a power prediction sub-model of the representative power station.

[0057] Preferably, the sub-model training module is specifically used for:

[0058] Based on the geographical location information of each of the representative power stations, construct an adjacency matrix of the plurality of representative power stations, and use the adjacency matrix as the static graph structure data;

[0059] The adjacency matrix is ​​expressed as:

[0060]

[0061] Wherein, A(i,j) is the adjacency matrix, D(i,j) is the geographical distance between representative power station i and representative power station j; S is the variance of the geographical distance between each representative power station; ε is the distance threshold; are the latitudes of power station i and power station j respectively; i , j are the longitudes of power station i and j respectively; R is the radius of the earth.

[0062] Preferably, the sub-model training module is specifically used for:

[0063] Acquire the historical power of the representative power station and the static graph structure data to construct a data set; divide the data set into a training set and a test set according to a set ratio;

[0064] Based on the training set, the historical power of the representative power station in the current historical training period and the space-time graph structure data are used as input, and the historical power of the representative power station in the next historical training period of the current historical training period is used as output, the space-time graph convolution model is trained, the optimizer is used to optimize and update the training parameters of the space-time graph convolution model, and the test set is used to evaluate the space-time graph convolution model during the training process until the loss function of the space-time graph convolution model converges, so as to obtain the trained space-time graph convolution model.

[0065] Based on the same inventive concept, the present invention also provides a distributed photovoltaic power station cluster power prediction method, comprising:

[0066] Obtain the actual power of each representative power station in the distributed photovoltaic power station cluster during the set historical period;

[0067] Taking the actual power of each representative power station in a set historical period as input, a pre-built distributed photovoltaic power station cluster power prediction model is used to perform prediction to obtain the predicted power of the distributed photovoltaic power station cluster in the period to be predicted;

[0068] Wherein, the time interval between the set historical period and the period to be predicted does not exceed a preset time threshold;

[0069] The distributed photovoltaic power station cluster power prediction model is constructed based on the distributed photovoltaic power station cluster power prediction model construction method as described above.

[0070] Based on the same inventive concept, the present invention also provides a distributed photovoltaic power station cluster power prediction system, comprising:

[0071] The second data acquisition module is used to obtain the actual power of each representative power station in the distributed photovoltaic power station cluster during a set historical period;

[0072] A prediction module is used to use the actual power of each representative power station in a set historical period as input, and use a pre-built distributed photovoltaic power station cluster power prediction model to perform prediction, so as to obtain the predicted power of the distributed photovoltaic power station cluster in the period to be predicted;

[0073] Wherein, the time interval between the set historical period and the period to be predicted does not exceed a preset time threshold;

[0074] The distributed photovoltaic power station cluster power prediction model is constructed based on the distributed photovoltaic power station cluster power prediction model construction method as described above.

[0075] Based on the same inventive concept, the present invention also provides a computer device, including: one or more processors;

[0076] A memory for storing one or more programs;

[0077] When the one or more programs are executed by the one or more processors, a distributed photovoltaic power station cluster power prediction model construction method as described above, or a distributed photovoltaic power station cluster power prediction method as described above is implemented.

[0078] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed, it implements a distributed photovoltaic power station cluster power prediction model construction method as described above, or a distributed photovoltaic power station cluster power prediction method as described above.

[0079] Compared with the closest prior art, the present invention has the following beneficial effects:

[0080] The present invention provides a method and system for constructing a power prediction model for a distributed photovoltaic power station cluster, including obtaining historical operation data of each power station in a distributed photovoltaic power station cluster; based on the historical operation data of each of the power stations, a maximum relevance-minimum redundancy feature selection method is used to select multiple representative power stations from each of the power stations, and the optimal weight of each of the representative power stations is obtained by solving the Egret Eagle optimization algorithm; based on the historical operation data of the multiple representative power stations, a spatiotemporal graph convolution model is trained to obtain a power prediction sub-model of each of the representative power stations; based on the optimal weight of each of the representative power stations, the power prediction sub-models of the multiple representative power stations are weightedly combined to obtain a constructed distributed photovoltaic power station cluster power prediction model; the method and system are based on the maximum relevance-minimum redundancy feature selection method to select multiple representative power stations from each of the power stations, and the optimal weight of each of the representative power stations is obtained by solving the spatiotemporal graph convolution model to obtain a power prediction sub-model of each of the representative power stations; based on the optimal weight of each of the representative power stations, the power prediction sub-models of the multiple representative power stations are weightedly combined to obtain a constructed distributed photovoltaic power station cluster power prediction model; The redundant feature selection method selects multiple representative power stations, so that the selected multiple representative power stations are highly correlated with the cluster power but have the lowest redundancy between each other, thereby improving the effectiveness and information content of multiple representative power stations; the Egret Optimization Algorithm has powerful global search capabilities, dynamic adjustment capabilities and fast convergence characteristics. The weights of multiple representative power stations are determined by the Egret Optimization Algorithm, which can effectively avoid falling into local optimality and ensure that the global optimal solution is found; the spatiotemporal graph convolution model inputs the data of multiple selected representative power stations for prediction to remove redundant information while retaining key feature information, which can better capture the overall characteristics of the cluster and improve the prediction performance; the spatiotemporal graph convolution model can more effectively capture the spatiotemporal correlation of photovoltaic power station power data, thereby improving the prediction accuracy of cluster power.

[0081] The present invention also provides a distributed photovoltaic power station cluster power prediction method and system, including obtaining the actual power of each representative power station in the distributed photovoltaic power station cluster in a set historical time period; taking the actual power of each representative power station in the set historical time period as input, using a pre-constructed distributed photovoltaic power station cluster power prediction model for prediction, and obtaining the predicted power of the distributed photovoltaic power station cluster in the time period to be predicted; wherein the time interval between the set historical time period and the time period to be predicted does not exceed a preset time threshold; the method and system improve the effectiveness and information volume of the multiple representative power stations by selecting multiple representative power stations that are highly correlated with the cluster power but have the lowest redundancy with each other, and when predicting through the pre-constructed distributed photovoltaic power station cluster power prediction model, it can more effectively capture the spatiotemporal correlation of photovoltaic power station power data, thereby improving the cluster power prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A schematic diagram of a distributed photovoltaic power station cluster power prediction model construction method provided by the present invention;

[0083] Figure 2 A schematic diagram of a specific process of selecting multiple representative power stations provided by the present invention;

[0084] Figure 3 A schematic diagram of a specific process of using the Heron Eagle optimization algorithm to solve and obtain the optimal weight of each representative power station provided by the present invention;

[0085] Figure 4 A schematic diagram of the structure of the spatiotemporal graph convolution model provided by the present invention;

[0086] Figure 5 A schematic diagram of the structure of a distributed photovoltaic power station cluster power prediction model construction system provided by the present invention;

[0087] Figure 6 A schematic diagram of a distributed photovoltaic power station cluster power prediction method provided by the present invention Figure 1 ;

[0088] Figure 7 A schematic diagram of a distributed photovoltaic power station cluster power prediction method provided by the present invention Figure 2 ;

[0089] Figure 8 A schematic diagram of prediction results of a distributed photovoltaic power station cluster power prediction method provided by the present invention;

[0090] Fig. 9 A schematic diagram comparing prediction results of a distributed photovoltaic power station cluster power prediction method provided by the present invention;

[0091] Fig.10 A schematic diagram of the structure of a distributed photovoltaic power station cluster power prediction system provided by the present invention;

[0092] Fig.11 The present invention provides a schematic diagram of the structure of an electronic device. DETAILED DESCRIPTION

[0093] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0094] Embodiment 1:

[0095] The present invention provides a method for constructing a distributed photovoltaic power station cluster power prediction model. Figure 1 As shown, including:

[0096] S1. Obtain the historical operation data of each power station in the distributed photovoltaic power station cluster;

[0097] S2. Based on the historical operation data of each power station, a maximum relevance-minimum redundancy feature selection method is used to select multiple representative power stations from each power station, and the optimal weight of each representative power station is obtained according to the Heron Eagle optimization algorithm;

[0098] S3. Based on the historical operation data of multiple representative power stations, the spatiotemporal graph convolution model is trained to obtain a power prediction sub-model for each representative power station;

[0099] S4. Based on the optimal weight of each representative power station, the power prediction sub-models of multiple representative power stations are weighted and combined to obtain a constructed distributed photovoltaic power station cluster power prediction model.

[0100] With the rapid popularization of distributed photovoltaic power generation systems, accurate power forecasting has become a key requirement for ensuring grid stability and optimizing energy scheduling. Compared with the power forecast of a single photovoltaic power station, the prediction of a distributed photovoltaic cluster is more challenging. In order to improve the accuracy of distributed photovoltaic cluster power forecasting, the present invention selects multiple representative power stations through the maximum correlation-minimum redundancy feature selection method, so that the selected multiple representative power stations are highly correlated with the cluster power but have the lowest redundancy with each other, thereby improving the effectiveness and information volume of multiple representative power stations; the Egret Optimization Algorithm has powerful global search capabilities, dynamic adjustment capabilities and fast convergence characteristics. The weights of multiple representative power stations are determined by the Egret Optimization Algorithm, which can effectively avoid falling into the local optimum and ensure that the global optimal solution is found; the spatiotemporal graph convolution model inputs the data of multiple selected representative power stations for prediction, which can remove redundant information while retaining key feature information, and can better capture the overall characteristics of the cluster and improve the prediction performance; the spatiotemporal graph convolution model can more effectively capture the spatiotemporal correlation of photovoltaic power station power data, thereby improving the prediction accuracy of cluster power.

[0101] In this embodiment, the historical operation data includes historical power, and the sum of the historical power of each power station is the historical total power of the cluster. By selecting representative power stations based on the historical power and training the spatiotemporal graph convolution model, a distributed photovoltaic power station cluster power prediction model can be obtained in a more targeted manner, making the prediction results of the distributed photovoltaic power station cluster power prediction model more accurate.

[0102] Before using the historical power to select representative power stations, data preprocessing operations may also be performed on the historical power, including: preprocessing the historical power data of each power station, removing abnormal values, and filling in missing data.

[0103] When selecting representative power stations in a distributed photovoltaic power station cluster area, usually only a single correlation coefficient or indicator is used for selection, such as the Pearson correlation coefficient (Pearson correlation coefficient) or mutual information. However, this single indicator selection method has some significant disadvantages and limitations. First, a single indicator often fails to fully reflect the correlation and mutual influence between the various power stations. Secondly, it is easily disturbed by data noise and outliers, resulting in unstable and unreliable selection results. In addition, since only a single indicator is considered, the influence of other important factors may be ignored, resulting in a certain deviation in the selection results. Therefore, the present invention selects to use multiple sub-correlation coefficients, including but not limited to the following: Pearson correlation coefficient, Spearman rank correlation coefficient and Kendall rank correlation coefficient.

[0104] Among them, the Pearson correlation coefficient r focuses on the linear relationship between two variables and is suitable for detecting the linear correlation between variables. The calculation formula is:

[0105]

[0106] Where x p is the p-th observed value of the power of each power station, that is, the p-th observed value of the historical power of each power station, y p is the p-th observed value of the cluster power, that is, the p-th observed value of the historical total power of the distributed photovoltaic power station cluster; are the power of each power station and the mean power of the cluster respectively; n is the total number of sample observations.

[0107] The Spearman correlation coefficient ρ measures the monotonic relationship between two variables and is suitable for identifying correlations that remain monotonic even if they are nonlinear. The calculation formula is:

[0108]

[0109] Where d p For each pair of observations x p and p The rank difference, d p =R(x p )-R(y p ), where R(x p ) and R(y p ) are x p and p rank; n is the total number of sample observations.

[0110] Kendall correlation coefficient τ measures the consistency of the rank between two variables and is suitable for detecting the relative order relationship between pairs of observations. The calculation formula is:

[0111]

[0112] Where C is the order consistent logarithm, that is, the observation value pair (x p ,y p ) and (x q ,y q ) satisfies (x p -x q )(y p -y q )>0; D is the order consistent logarithm, that is, the observed value pair (x p ,y p ) and (x q ,y q ) satisfies (x p -x q )(y p -y q )<0; n is the total number of sample observations, x q is the qth observed value of the power of each power station, that is, the pth observed value of the historical power of each power station, y q is the qth observation of the cluster power.

[0113] The comprehensive correlation coefficient Z obtained by adding and averaging these three seed correlation coefficients is expressed as:

[0114]

[0115] In the above S2, the comprehensive correlation coefficient is used as an indicator for selecting representative power stations. The comprehensive correlation coefficient has a more comprehensive evaluation capability and can simultaneously consider different types of correlations such as linear correlation, monotonic correlation and disordered correlation. By comprehensively considering the advantages of different types of correlation coefficients, the possible deviations and deficiencies of a single indicator are reduced.

[0116] Specifically, in this embodiment, the comprehensive correlation coefficient is divided into a first comprehensive correlation coefficient between each power station and the distributed photovoltaic power station cluster, and a second comprehensive correlation coefficient between each power station. The calculation process of the first comprehensive correlation coefficient and the second comprehensive correlation coefficient refers to the calculation process of the above-mentioned comprehensive correlation coefficient.

[0117] In this embodiment, selecting multiple representative power stations in S2 may include:

[0118] Based on the historical power in the historical operation data of each power station, a first comprehensive correlation coefficient between each power station and the distributed photovoltaic power station cluster, and a second comprehensive correlation coefficient between each power station are calculated; the first comprehensive correlation coefficient and the second comprehensive correlation coefficient are calculated according to the average of multiple sub-correlation coefficients;

[0119] Based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient, a maximum correlation-minimum redundancy feature selection method is adopted to select multiple representative power stations from each power station.

[0120] The Maximum Relevance Minimum Redundancy (mRMR) feature selection method aims to select features that are highly correlated with the target variable but have the lowest redundancy with each other. Specifically, this selection method improves the effectiveness and information content of the feature set by optimizing the correlation between the features and the target variable and the redundancy between the features. Therefore, using the mRMR feature selection method to select representative power stations can ensure that the power of the selected representative power stations is highly correlated with the total output power of the cluster, and the information redundancy between them is minimal.

[0121] In this embodiment, in S2, based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient, selecting multiple representative power stations from each power station using the maximum correlation-minimum redundancy feature selection method may include:

[0122] Based on the first comprehensive correlation coefficients corresponding to the power stations, the power station with the largest first comprehensive correlation coefficient is selected as the representative power station, and the power stations in the distributed photovoltaic power station cluster except the representative power station are taken as the remaining power stations;

[0123] Obtaining an average value of the second comprehensive correlation coefficient between each remaining power station and the selected representative power station to determine the redundancy;

[0124] Based on each remaining power station, the difference between the first comprehensive correlation coefficient and the redundancy corresponding to the remaining power station is calculated; the remaining power station with the largest difference is selected as the representative power station, and the number of remaining power stations is updated, and the process returns to the above step of obtaining the average value of the second comprehensive correlation coefficient between each remaining power station and the selected representative power station, and determines the new redundancy and the new representative power station, until the total capacity of all the selected representative power stations is not less than the set capacity threshold, and the cycle is stopped to obtain the final multiple representative power stations.

[0125] For example, the mRMR feature selection method is used to select representative power stations such as Figure 2 First, the power data of all power stations, that is, the historical power, is obtained, and then the comprehensive correlation coefficient between the power of the M power stations and the total power of the cluster is calculated. Here, the comprehensive correlation coefficient is the first comprehensive correlation coefficient mentioned above. The power station with the largest comprehensive correlation coefficient is selected as the representative power station to obtain a representative power station;

[0126] Next, the comprehensive correlation coefficients between all the remaining M-1 power stations and the representative power station selected last time are calculated. Here, the comprehensive correlation coefficient is the second comprehensive correlation coefficient mentioned above. The average value of each second comprehensive correlation coefficient is the redundancy. The difference between the first comprehensive correlation coefficient and the redundancy between the remaining M-1 power stations and the cluster power is calculated. The power station with the largest calculated difference is selected as the representative power station to obtain the second representative power station. Then, the step of obtaining the second representative power station is repeated in the M-2 power stations.

[0127] The prediction error will decrease as the number of representative power plants increases, but when the sum of the standard capacities of all representative power plants approaches 50% of the total capacity of the cluster, the prediction error will reach a reference value and will no longer change significantly. Therefore, it is necessary to determine whether the sum of the capacities of all selected power plants reaches or exceeds 50% of the total capacity, that is, to set a capacity threshold. If it reaches or exceeds 50%, these power plants are output and the process ends; if it does not reach 50%, the steps of calculating the comprehensive correlation and selecting power plants are repeated until the sum of the capacities of the selected representative power plants reaches or exceeds 50% of the total capacity of the cluster.

[0128] Secretary bird optimization algorithm (SBOA) is a novel meta-heuristic algorithm inspired by the hunting and avoidance behavior of secretary bird. SBOA performs well in optimization problems due to its powerful global search ability, dynamic adjustment ability and fast convergence characteristics. By simulating the hunting and escape strategies of secretary bird, SBOA can effectively avoid falling into local optimality and ensure the global optimal solution. In addition, the SBOA algorithm has a simple structure, is easy to implement, and is suitable for various complex nonlinear and multi-peak optimization problems. These advantages enable SBOA to provide efficient and accurate solutions in practical applications in S2.

[0129] In this embodiment, in S2, the optimal weight of each representative power station is obtained according to the Heron Eagle optimization algorithm, including:

[0130] Initialize the parameters of the egret optimization algorithm, where the number of egret groups in the parameters is the number of multiple representative power stations, each egret in the egret group corresponds to a representative power station, and the position of each egret corresponds to the weight of a representative power station;

[0131] The fitness value of each egret's position is calculated according to a preset fitness function, with the goal of finding the position with the minimum fitness value, and determining the best position for each egret;

[0132] Based on the egret optimization algorithm, the hunting behavior of the egret is simulated in the exploration stage, the position of the egret is preliminarily updated, and the optimal position of the egret is preliminarily updated according to the fitness function. In the development stage, the avoiding behavior of the egret is simulated, the position of the egret is updated again, and the optimal position of the egret is updated again according to the fitness function. The position and optimal position of the egret are updated repeatedly until the maximum number of iterations is reached. The optimal position of each egret when the maximum number of iterations is reached is used as the optimal weight of the corresponding representative power station.

[0133] In this embodiment, the fitness value of each egret's position is calculated according to a preset fitness function, including:

[0134] Based on the fitness function, the fitness value of each egret's position is calculated according to the square difference between the weighted sum of the historical power of each representative power station and the historical total power of the distributed photovoltaic power station cluster.

[0135] Specifically, Figure 3 As shown in FIG. 1 , the specific process of using SBOA to calculate the optimal weight of the representative power station is as follows:

[0136] First, a fitness function is defined as the square error, which is the square difference between the sum of the historical power of each representative power station calculated according to the weight and the total power of the cluster, that is, the historical total power of the distributed PV power station cluster, to evaluate the prediction error of different weight combinations.

[0137] Next, the algorithm parameters are initialized and the initial population position, that is, the initial weight of each representative power station, is generated. The fitness value of each egret's position is calculated according to the preset fitness function, with the goal of finding the position with the minimum fitness value to determine the best position for each egret.

[0138] Then, in each iteration, according to the search strategy of SBOA, in the exploration phase, the hunting behavior of the egret eagle is simulated, the solution space is extensively searched and the position is updated according to the formula, and then the development phase is entered, the avoidance behavior is simulated, the promising area is finely searched and the position is further updated according to the formula, the new fitness value is calculated, and the current optimal weight combination is recorded and updated. After multiple iterations, the current iteration number t reaches the maximum iteration number T, and the algorithm finally outputs the optimal representative power station weight combination.

[0139] Specifically, in the exploration stage, when t<1 / 3T, the behavior of the egret eagle in searching for prey during hunting is simulated, when 1 / 3T≤t≤2 / 3T, the behavior of the egret eagle in consuming prey during hunting is simulated, and when t>2 / 3T, the behavior of the egret eagle in attacking prey during hunting is simulated.

[0140] Spatio-Temporal Graph Convolutional Networks (STGCN) can effectively capture local and global features in spatiotemporal data, and has strong spatiotemporal relationship modeling capabilities and the flexibility to adapt to graph structured data. The power generation data of photovoltaic power stations has obvious spatiotemporal correlations, and traditional prediction methods are difficult to effectively capture this complex spatiotemporal relationship. In the above S3, STGCN can improve the accuracy and reliability of prediction through its powerful spatiotemporal feature extraction capabilities.

[0141] In S3, the spatiotemporal graph convolution model is trained based on the historical operation data of multiple representative power plants to obtain the power prediction sub-model of each representative power plant, which may include:

[0142] Obtaining historical power and geographic location information from historical operation data of each representative power station, and determining static graph structure data of the plurality of representative power stations based on the geographic location information of the plurality of representative power stations;

[0143] Based on each representative power station, the spatiotemporal graph convolution model is trained with the historical power and static graph structure data of the representative power station as input to obtain the power prediction sub-model of the representative power station.

[0144] Specifically, STGCN takes the historical power data of the power station and the static graph structure data composed of the adjacency matrix constructed using the geographical location information of the representative power station as input, and finally outputs the predicted power of each representative power station after spatiotemporal convolution operation. The historical power data and the static graph structure data constitute the graph structure data, which includes two parts: nodes and edges: one node corresponds to a representative power station, and each node represents the historical power data of a specific representative power station; the edges between nodes use the adjacency matrix constructed using the geographical location information of the representative power station to represent the degree of association between the representative power stations.

[0145] In this embodiment, based on the geographical location information of the multiple representative power stations, the static graph structure data of the multiple representative power stations is determined, including:

[0146] Based on the geographical location information of each representative power station, an adjacency matrix of multiple representative power stations is constructed, and the adjacency matrix is ​​used as static graph structure data;

[0147] The adjacency matrix is ​​represented as:

[0148]

[0149] Where A(i,j) is the adjacency matrix, D(i,j) is the geographical distance between representative power station i and representative power station j; S is the variance of the geographical distance between each representative power station; ε is the distance threshold; are the latitudes of power stations i and j, respectively, in radians; i , j are the longitudes of representative power stations i and j respectively, in radians; R is the radius of the earth, usually 6371 km.

[0150] In this embodiment, Figure 4 As shown in the figure, the STGCN model includes two spatiotemporal convolution modules, each of which contains two temporal convolution blocks and spatial convolution blocks. The temporal convolution block first performs a convolution operation, regards the time dimension as the width of the image, regards the number of nodes as the height of the image, and uses the convolution kernel to slide on the time dimension to extract the temporal features; secondly, a gating mechanism similar to the gated linear unit (GatedLinearUnit, GLU) is used to adjust the influence of different convolution results; finally, the gating results are integrated with the convolution results to extract the temporal features. The spatial convolution block uses the adjacency matrix to perform a weighted summation of the temporal features of each node, fuses the features of each node with the features of its neighboring nodes, and effectively captures the spatial correlation between nodes.

[0151] In this embodiment, the historical power and static graph structure data representing the power station are used as input to train the spatiotemporal graph convolution model to obtain a power prediction sub-model representing the power station, including:

[0152] Obtain historical power and static graph structure data representing the power station and construct a data set; divide the data set into a training set and a test set according to a set ratio;

[0153] Based on the training set, the historical power and spatiotemporal graph structure data representing the power station in the current historical training period are used as input, and the historical power of the power station in the next historical training period of the current historical training period is used as output. The spatiotemporal graph convolution model is trained, and the optimizer is used to optimize and update the training parameters of the spatiotemporal graph convolution model. The test set is used to evaluate the spatiotemporal graph convolution model during the training process until the loss function of the spatiotemporal graph convolution model converges to obtain a trained spatiotemporal graph convolution model.

[0154] In summary, in order to improve the accuracy of distributed photovoltaic cluster power prediction, the present invention considers the spatiotemporal correlation between distributed photovoltaic sites, and proposes a method for distributed photovoltaic cluster power prediction based on representative power station selection and spatiotemporal graph convolution network. Specifically, the present invention selects representative power stations based on comprehensive correlation coefficients: the comprehensive correlation coefficient is used as an indicator for selecting representative power stations. The comprehensive correlation coefficient has a more comprehensive evaluation capability, and can simultaneously consider different types of correlations such as linear correlation, monotonic correlation, and disordered correlation. By comprehensively considering the advantages of different types of correlation coefficients, the possible deviations and deficiencies of a single indicator are reduced; at the same time, the spatiotemporal graph convolution model is used to predict the power of distributed photovoltaic clusters: due to the spatial correlation of distributed photovoltaic systems, photovoltaic power stations located in similar geographical and climatic environments usually show similar power changes at the same time. This means that there is a certain correlation between the historical power of each site and the future power of the neighboring site, that is, the spatiotemporal correlation. Using the spatiotemporal graph convolution model, considering the evolution of nodes in the time series and the spatial relationship between nodes, it is possible to capture the temporal and spatial feature information at the same time, and effectively integrate them together for learning and prediction, thereby improving the prediction accuracy.

[0155] Embodiment 2:

[0156] Based on the same inventive concept, the present invention also provides a distributed photovoltaic power station cluster power prediction model construction system, such as Figure 5 As shown, including:

[0157] The first data acquisition module is used to acquire the historical operation data of each power station in the distributed photovoltaic power station cluster;

[0158] The weight calculation module is used to select multiple representative power stations from each power station based on the historical operation data of each power station, using the maximum relevance-minimum redundancy feature selection method, and obtain the optimal weight of each representative power station according to the Heron Eagle optimization algorithm;

[0159] A sub-model training module is used to train the spatiotemporal graph convolution model based on the historical operation data of multiple representative power stations to obtain a power prediction sub-model for each representative power station;

[0160] The combined model building module is used to weightedly combine the power prediction sub-models of multiple representative power stations based on the optimal weight of each representative power station to obtain a constructed distributed photovoltaic power station cluster power prediction model.

[0161] In this embodiment, the weight calculation module includes:

[0162] A coefficient calculation unit, used to calculate a first comprehensive correlation coefficient between each power station and a distributed photovoltaic power station cluster, and a second comprehensive correlation coefficient between each power station based on historical power in historical operation data of each power station; the first comprehensive correlation coefficient and the second comprehensive correlation coefficient are calculated according to the average of multiple sub-correlation coefficients;

[0163] The power station selection unit is used to select a plurality of representative power stations from each power station by using a maximum correlation-minimum redundancy feature selection method based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient.

[0164] In this embodiment, the power station selection unit is specifically used for:

[0165] Based on the first comprehensive correlation coefficients corresponding to the power stations, the power station with the largest first comprehensive correlation coefficient is selected as the representative power station, and the power stations in the distributed photovoltaic power station cluster except the representative power station are taken as the remaining power stations;

[0166] Obtaining an average value of the second comprehensive correlation coefficient between each remaining power station and the selected representative power station to determine the redundancy;

[0167] Based on each remaining power station, the difference between the first comprehensive correlation coefficient and the redundancy corresponding to the remaining power station is calculated; the remaining power station with the largest difference is selected as the representative power station, and the number of remaining power stations is updated, and the process returns to the above step of obtaining the average value of the second comprehensive correlation coefficient between each remaining power station and the selected representative power station, and determines the new redundancy and the new representative power station, until the total capacity of all the selected representative power stations is not less than the set capacity threshold, and the cycle is stopped to obtain the final multiple representative power stations.

[0168] In this embodiment, the plurality of sub-correlation coefficients include a Pearson correlation coefficient, a Spearman rank correlation coefficient, and a Kendall rank correlation coefficient.

[0169] In this embodiment, the weight calculation module also includes:

[0170] An initialization unit, used for initializing the parameters of the egret optimization algorithm, wherein the number of egret groups in the parameters is the number of multiple representative power stations, each egret in the egret group corresponds to a representative power station, and the position of each egret corresponds to the weight of a representative power station;

[0171] A position determination unit is used to calculate the fitness value of the position of each egret according to a preset fitness function, with the goal of finding the position with the minimum fitness value, and determining the best position of each egret;

[0172] The solving unit is used for simulating the hunting behavior of the egret eagle based on the egret optimization algorithm in the exploration phase, preliminarily updating the position of the egret eagle, and preliminarily updating the optimal position of the egret eagle according to the fitness function; simulating the avoiding behavior of the egret eagle in the development phase, updating the position of the egret eagle again, and updating the optimal position of the egret eagle again according to the fitness function, and iteratively updating the position and optimal position of the egret eagle again until the maximum number of iterations is reached, and the optimal position of each egret eagle when the maximum number of iterations is reached is used as the optimal weight of the corresponding representative power station.

[0173] In this embodiment, the location determination unit is specifically used for:

[0174] Based on the fitness function, the fitness value of each egret's position is calculated according to the square difference between the weighted sum of the historical power of each representative power station and the historical total power of the distributed photovoltaic power station cluster.

[0175] In this embodiment, the sub-model training module is specifically used for:

[0176] Obtaining historical power and geographic location information from historical operation data of each representative power station, and determining static graph structure data of the plurality of representative power stations based on the geographic location information of the plurality of representative power stations;

[0177] Based on each representative power station, the spatiotemporal graph convolution model is trained with the historical power and static graph structure data of the representative power station as input to obtain the power prediction sub-model of the representative power station.

[0178] In this embodiment, the sub-model training module is specifically used for:

[0179] Based on the geographical location information of each representative power station, an adjacency matrix of multiple representative power stations is constructed, and the adjacency matrix is ​​used as static graph structure data;

[0180] The adjacency matrix is ​​represented as:

[0181]

[0182] Where A(i,j) is the adjacency matrix, D(i,j) is the geographical distance between representative power station i and representative power station j; S is the variance of the geographical distance between each representative power station; ε is the distance threshold; are the latitudes of power station i and power station j respectively; i , j are the longitudes of power station i and j respectively; R is the radius of the earth.

[0183] In this embodiment, the sub-model training module is specifically used for:

[0184] Obtain historical power and static graph structure data representing the power station and construct a data set; divide the data set into a training set and a test set according to a set ratio;

[0185] Based on the training set, the historical power and spatiotemporal graph structure data representing the power station in the current historical training period are used as input, and the historical power of the power station in the next historical training period of the current historical training period is used as output. The spatiotemporal graph convolution model is trained, and the optimizer is used to optimize and update the training parameters of the spatiotemporal graph convolution model. The test set is used to evaluate the spatiotemporal graph convolution model during the training process until the loss function of the spatiotemporal graph convolution model converges to obtain a trained spatiotemporal graph convolution model.

[0186] Embodiment 3:

[0187] Based on the same inventive concept, the present invention also provides a distributed photovoltaic power station cluster power prediction method, such as Figure 6 As shown, including:

[0188] A1. Obtain the actual power of each representative power station in the distributed photovoltaic power station cluster during the set historical period;

[0189] A2. Taking the actual power of each representative power station in the set historical period as input, a pre-built distributed photovoltaic power station cluster power prediction model is used to perform prediction, and the predicted power of the distributed photovoltaic power station cluster in the period to be predicted is obtained;

[0190] The time interval between the historical period and the period to be predicted is set to not exceed a preset time threshold;

[0191] The distributed photovoltaic power station cluster power prediction model is constructed based on the distributed photovoltaic power station cluster power prediction model construction method of the previous embodiment.

[0192] Embodiment 4:

[0193] Based on the same inventive concept, the present invention also provides a distributed photovoltaic power station cluster power prediction method, such as Figure 7-9 As shown, the overall process is as follows Figure 7 As shown in the figure, it is divided into three parts, namely, the selection of representative power stations based on the maximum relevance-minimum redundancy feature selection method, the weight calculation of representative power stations based on the Heron eagle optimization algorithm, and the cluster power prediction based on the spatiotemporal graph convolution model. The calculation steps of cluster power prediction are as follows:

[0194] (1) Data preprocessing: The historical power data of each power station is preprocessed to remove outliers and fill in missing data.

[0195] (2) Representative power station selection: Based on the preprocessed historical power data, the comprehensive correlation coefficient between the power of each power station and the cluster power is calculated, and the maximum correlation-minimum redundancy feature selection method is used to select representative power stations that are highly correlated with the total cluster output power and have the least redundant information between them.

[0196] (3) Weight calculation: The selected historical power data of the representative power station and the historical power data of the cluster are input into the Egret Eagle optimization algorithm to calculate the optimal weight value of the representative power station.

[0197] (4) Calculate the adjacency matrix: The adjacency matrix is ​​calculated based on the geographical location information representing the power station.

[0198] (5) Power prediction of representative power stations: The spatiotemporal graph convolution model is used to input the historical power data of the representative power stations and the corresponding adjacency matrix to obtain the predicted power of each representative power station. In this embodiment, since the adjacency matrix is ​​integrated into the model and does not change for a cluster, there is no need to input the adjacency matrix again when predicting the actual cluster power.

[0199] (6) Cluster power calculation: The predicted cluster power is the sum of the predicted power of each representative power station multiplied by the corresponding weight, that is, the predicted power of the distributed photovoltaic power station cluster during the predicted period.

[0200] The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0201] A distributed photovoltaic power station cluster in Hebei Province was adopted, and ten distributed photovoltaic power stations with good data quality were selected as representative power stations. The time period was 324 days from January 1, 2021 to December 31, 2021, with a time resolution of 15 minutes.

[0202] Specifically, the comprehensive correlation coefficient is used as an indicator for the cluster, and the representative power stations are selected based on the maximum correlation-minimum redundancy feature selection method. The representative power stations are power station 10, power station 7, power station 1, power station 8, and power station 5. The weights of each representative power station are calculated using SBOA, which are 1.4873, 2.121, 1.1639, 1.3339, and 2.4452, respectively. The final cluster prediction power calculation formula is:

[0203] P C =1.4873×P 10 +2.121×P7+1.1639×P1+1.3339×P8+2.4452×P5;

[0204] Where P C is the predicted cluster power; P 10 , P7, P1, P8, and P5 are the predicted powers of power station 10, power station 7, power station 1, power station 8, and power station 5 respectively. The prediction results are as follows: Figure 8 As shown, the calculation errors MAE and RMSE are 1.47% and 2.14% respectively.

[0205] Mutual Information (MI) was used as a single indicator to compare with the comprehensive correlation coefficient. The representative power stations were selected based on the mRMR feature selection method. The representative power stations were Power Station 6, Power Station 8, Power Station 7, Power Station 9, and Power Station 5. The weights of the representative power stations were obtained by SBOA as 2.4376, 1.8956, 1.5068, 2.4855, and 1.8273, respectively. The final cluster prediction power calculation formula was:

[0206] P C =2.4376×P6+1.8956×P8+1.5068×P7+12.4855×P9+1.8273×P5;

[0207] Where P C is the predicted cluster power; P6, P8, P7, P9, and P5 are the predicted powers of power station 6, power station 8, power station 7, power station 9, and power station 5 respectively.

[0208] When using the STGCN model to predict the power of a photovoltaic power station cluster, there are two methods for calculating the cluster power: one is to input the historical power data of all power stations in the cluster and the adjacency matrix constructed by the geographical location information of the corresponding power stations into the STGCN model, which will output the predicted power of all power stations, and then substitute the predicted power of the representative power station into the cluster prediction power formula constructed above to obtain the predicted cluster power; the other is to input the adjacency matrix constructed by the historical power data of the representative power station and the geographical location information of the representative power station into the STGCN model, output the predicted power of all representative power stations, and then substitute it into the cluster power prediction formula to obtain the predicted cluster power. The first method needs to process the data of all power stations, and the dimensions of the input data and the adjacency matrix are high, resulting in high model calculation complexity and long training and inference time. The second method only processes the data of the representative power station, with a smaller amount of data, low calculation complexity, and faster model training and inference process.

[0209] Therefore, there are six methods for cluster power prediction, as shown in Table 1 below:

[0210] Table 1 Method description

[0211]

[0212] The prediction results obtained by the six methods are shown in Table 2 below. Fig. 9 As shown:

[0213] Table 2 Cluster prediction results

[0214]

[0215] Comparing the prediction results of methods 1, 2, 5 and methods 3, 4, and 6, whether the STGCN model or the BiLSTM model is used for prediction, the method using the comprehensive correlation coefficient as the feature selection index is better than the method using mutual information as the feature selection index. Comparing method 1 and method 2, the MAE (mean absolute error) and RMSE (root mean square error) of the prediction of 10 power stations input into the STGCN model are 1.68% and 2.32% respectively, and the MAE and RMSE of the prediction of the representative power station input are 1.47% and 2.14% respectively. The results show that the method of predicting the cluster power by inputting the representative power station data is better. Comparing method 2 and method 5, method 2 uses the STGCN model for prediction, and method 5 uses the BiLSTM model for prediction. The MAE and RMSE of method 2 are 1.47% and 2.14% respectively, while the MAE and RMSE of method 5 are 2.16% and 4.41% respectively. The results show that the prediction error of cluster power predicted by the STGCN model is smaller.

[0216] Comparing method 1 and method 2, the MAE and RMSE of the STGCN model using the data of 10 power stations for prediction are 1.68% and 2.32% respectively, and the MAE and RMSE of the representative power station data for prediction are 1.47% and 2.14% respectively. The results show that the method of predicting cluster power by using the data of the representative power station is better. This difference shows that although using the data of 10 power stations can capture more spatial information, it may introduce more noise and redundant information, affecting the accuracy of the prediction. The representative power station selected based on the mRMR principle using the comprehensive correlation coefficient can effectively remove redundant information while retaining key feature information, which can better capture the overall characteristics of the cluster and improve the prediction performance while reducing the computational complexity.

[0217] Comparing Method 2 and Method 5, Method 2 uses the STGCN model for prediction, and Method 5 uses the BiLSTM model for prediction. The MAE and RMSE of Method 2 are 1.47% and 2.14% respectively, while the MAE and RMSE of Method 5 are 2.16% and 4.41% respectively. The results show that the prediction error of cluster power predicted by the STGCN model is smaller, which is the same as the result shown in Cluster 1 above. This significant difference mainly stems from the difference in the ability of the two methods to process spatiotemporal data. The STGCN model combines temporal convolution and spatial convolution, which can effectively capture the spatiotemporal correlation in photovoltaic power station data. Temporal convolution can extract the temporal characteristics of the historical power data of the power station and capture the pattern that changes over time; spatial convolution captures the spatial relationship between power stations through the adjacency matrix, reflecting the relationship between different power stations. Therefore, the STGCN model can more accurately model the complex spatiotemporal dependencies of photovoltaic power stations, thereby improving the prediction accuracy. In contrast, the BiLSTM model is more suitable for time series data. It can capture long-term dependencies in the time dimension through forward and backward processing. However, it is insufficient in capturing the spatial correlation between power stations and lacks the ability to model the geographical locations of power stations and their relationships. Therefore, the BiLSTM model cannot fully utilize the spatial correlation between power stations, and the prediction error is larger than that of the STGCN model.

[0218] Finally, the following conclusions were drawn through analysis:

[0219] (1) Using the comprehensive correlation coefficient as the selection index of the representative power station can effectively reduce the limitations of a single index and the impact of noise on the selection results, and the stability and reliability of feature selection are stronger;

[0220] (2) The STGCN model input selects representative power station data for prediction, which can remove redundant information while retaining key feature information, better capture the overall characteristics of the cluster, and improve prediction performance;

[0221] (3) The STGCN model can more effectively capture the spatiotemporal correlation of PV power plant power data, thereby improving the accuracy of cluster power prediction.

[0222] Embodiment 5:

[0223] Based on the same inventive concept, the present invention also provides a distributed photovoltaic power station cluster power prediction system, such as Fig.10 As shown, including:

[0224] The second data acquisition module is used to obtain the actual power of each representative power station in the distributed photovoltaic power station cluster during a set historical period;

[0225] The prediction module is used to use the actual power of each representative power station in the set historical period as input, and use the pre-built distributed photovoltaic power station cluster power prediction model to perform prediction, so as to obtain the predicted power of the distributed photovoltaic power station cluster in the period to be predicted;

[0226] The time interval between the historical period and the period to be predicted is set to not exceed a preset time threshold;

[0227] The distributed photovoltaic power station cluster power prediction model is constructed based on the distributed photovoltaic power station cluster power prediction model construction method of the previous embodiment.

[0228] Embodiment 6:

[0229] like Fig.11 As shown, the present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.

[0230] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a storage medium to implement the corresponding method flow or corresponding function, so as to implement a distributed photovoltaic power station cluster power prediction model construction method in the above embodiment, or a distributed photovoltaic power station cluster power prediction method in the above embodiment. Steps.

[0231] Embodiment 7:

[0232] Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It can be understood that the storage medium here can include both built-in storage media in electronic devices and, of course, extended storage media supported by electronic devices. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor loads and executes one or more instructions stored in the storage medium, which can implement a distributed photovoltaic power station cluster power prediction model construction method in the above embodiment, or a distributed photovoltaic power station cluster power prediction method in the above embodiment. Steps.

[0233] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0234] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0235] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0236] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0237] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims of the present invention.

Claims

1. A method for constructing a distributed photovoltaic power station cluster power prediction model, characterized in that: include: Obtain the historical operation data of each power station in the distributed photovoltaic power station cluster; Based on the historical operation data of each of the power stations, a maximum relevance-minimum redundancy feature selection method is used to select multiple representative power stations from each of the power stations, and the optimal weight of each of the representative power stations is obtained by solving the Egret Eagle optimization algorithm; Based on the historical operation data of the plurality of representative power stations, the spatiotemporal graph convolution model is trained to obtain a power prediction sub-model of each representative power station; Based on the optimal weight of each representative power station, the power prediction sub-models of the plurality of representative power stations are weighted and combined to obtain a constructed distributed photovoltaic power station cluster power prediction model.

2. The method according to claim 1, characterized in that The method of selecting a plurality of representative power stations from each power station by using a maximum relevance-minimum redundancy feature selection method based on the historical operation data of each power station includes: Based on the historical power in the historical operation data of each of the power stations, a first comprehensive correlation coefficient between each of the power stations and the distributed photovoltaic power station cluster, and a second comprehensive correlation coefficient between each of the power stations are calculated; the first comprehensive correlation coefficient and the second comprehensive correlation coefficient are calculated according to the average of multiple sub-correlation coefficients; Based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient, a maximum correlation-minimum redundancy feature selection method is adopted to select a plurality of representative power stations from each of the power stations.

3. The method according to claim 2, characterized in that The method of selecting a plurality of representative power stations from each of the power stations using a maximum correlation-minimum redundancy feature selection method based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient includes: Based on the first comprehensive correlation coefficients corresponding to the power stations, a power station with the largest first comprehensive correlation coefficient is selected as a representative power station, and power stations in the distributed photovoltaic power station cluster except the representative power station are used as remaining power stations; Obtaining an average value of the second comprehensive correlation coefficient between each of the remaining power stations and the selected representative power station to determine the redundancy; Based on each of the remaining power stations, the difference between the first comprehensive correlation coefficient corresponding to the remaining power stations and the redundancy is calculated; the remaining power station with the largest difference is selected as the representative power station, and the number of remaining power stations is updated, and the step of obtaining the average value of the second comprehensive correlation coefficient between each of the remaining power stations and the selected representative power station is returned to determine the new redundancy and the new representative power station, until the total capacity of all the selected representative power stations is not less than the set capacity threshold, the cycle is stopped, and the final multiple representative power stations are obtained.

4. The method according to claim 2, characterized in that The plurality of sub-correlation coefficients include a Pearson correlation coefficient, a Spearman rank correlation coefficient, and a Kendall rank correlation coefficient.

5. The method according to claim 1, characterized in that The method of obtaining the optimal weight of each representative power station according to the Egret Eagle optimization algorithm includes: Initializing the parameters of the egret optimization algorithm, wherein the number of egret groups in the parameters is the number of the plurality of representative power stations, each egret in the egret group corresponds to one of the representative power stations, and the position of each egret corresponds to the weight of one of the representative power stations; Calculating the fitness value of each egret's position according to a preset fitness function, with the goal of finding a position with the minimum fitness value, and determining the best position for each egret; Based on the egret optimization algorithm, the hunting behavior of the egret is simulated in the exploration phase, the position of the egret is preliminarily updated, and the optimal position of the egret is preliminarily updated according to the fitness function. In the development phase, the avoiding behavior of the egret is simulated, the position of the egret is updated again, and the optimal position of the egret is updated again according to the fitness function. The position and optimal position of the egret are updated repeatedly until the maximum number of iterations is reached. The optimal position of each egret when the maximum number of iterations is reached is used as the optimal weight of the corresponding representative power station.

6. The method according to claim 5, characterized in that The step of calculating the fitness value of the position of each egret eagle according to a preset fitness function includes: Based on the fitness function, the fitness value of the position of each egret is calculated according to the square difference between the weighted sum of the historical power of each representative power station and the historical total power of the distributed photovoltaic power station cluster.

7. The method according to claim 1, characterized in that The method of training the spatiotemporal graph convolution model based on the historical operation data of the plurality of representative power stations to obtain a power prediction sub-model of each representative power station includes: Acquire historical power and geographical location information in historical operation data of each of the representative power stations, and determine static graph structure data of the plurality of representative power stations based on the geographical location information of the plurality of representative power stations; Based on each of the representative power stations, the spatiotemporal graph convolution model is trained with the historical power of the representative power station and the static graph structure data as input to obtain a power prediction sub-model of the representative power station.

8. The method according to claim 7, characterized in that The determining of the static graph structure data of the plurality of representative power stations based on the geographical location information of the plurality of representative power stations comprises: Based on the geographical location information of each of the representative power stations, construct an adjacency matrix of the plurality of representative power stations, and use the adjacency matrix as the static graph structure data; The adjacency matrix is ​​expressed as: Wherein, A(i,j) is the adjacency matrix, D(i,j) is the geographical distance between representative power station i and representative power station j; S is the variance of the geographical distance between each representative power station; ε is the distance threshold; are the latitudes of power station i and power station j respectively; i , j are the longitudes of power station i and j respectively; R is the radius of the earth.

9. The method according to claim 7, characterized in that The historical power of the representative power station and the static graph structure data are used as input to train the spatiotemporal graph convolution model to obtain the power prediction sub-model of the representative power station, including: Acquire the historical power of the representative power station and the static graph structure data to construct a data set; divide the data set into a training set and a test set according to a set ratio; Based on the training set, the historical power of the representative power station in the current historical training period and the space-time graph structure data are used as input, and the historical power of the representative power station in the next historical training period of the current historical training period is used as output, the space-time graph convolution model is trained, the optimizer is used to optimize and update the training parameters of the space-time graph convolution model, and the test set is used to evaluate the space-time graph convolution model during the training process until the loss function of the space-time graph convolution model converges, so as to obtain the trained space-time graph convolution model.

10. A distributed photovoltaic power station cluster power prediction model construction system, characterized in that: include: The first data acquisition module is used to acquire the historical operation data of each power station in the distributed photovoltaic power station cluster; A weight calculation module is used to select a plurality of representative power stations from the power stations based on the historical operation data of the power stations by using the maximum relevance-minimum redundancy feature selection method, and obtain the optimal weight of each representative power station by solving the optimization algorithm according to the Heron Eagle optimization algorithm; A sub-model training module, used to train the spatiotemporal graph convolution model based on the historical operation data of the plurality of representative power stations, to obtain a power prediction sub-model of each of the representative power stations; The combined model building module is used to weightedly combine the power prediction sub-models of the multiple representative power stations based on the optimal weight of each representative power station to obtain a constructed distributed photovoltaic power station cluster power prediction model.

11. The system according to claim 10, characterized in that The weight calculation module includes: A coefficient calculation unit, configured to calculate, based on the historical power in the historical operation data of each of the power stations, a first comprehensive correlation coefficient between each of the power stations and the distributed photovoltaic power station cluster, and a second comprehensive correlation coefficient between each of the power stations; the first comprehensive correlation coefficient and the second comprehensive correlation coefficient are calculated based on the average of a plurality of sub-correlation coefficients; A power station selection unit is used to select a plurality of representative power stations from each of the power stations by using a maximum correlation-minimum redundancy feature selection method based on the first comprehensive correlation coefficient and the second comprehensive correlation coefficient.

12. The system according to claim 11, characterized in that The power station selection unit is specifically used for: Based on the first comprehensive correlation coefficients corresponding to the power stations, a power station with the largest first comprehensive correlation coefficient is selected as a representative power station, and power stations in the distributed photovoltaic power station cluster except the representative power station are used as remaining power stations; Obtaining an average value of the second comprehensive correlation coefficient between each of the remaining power stations and the selected representative power station to determine the redundancy; Based on each of the remaining power stations, calculating a difference between a first comprehensive correlation coefficient corresponding to the remaining power station and the redundancy; The remaining power station with the largest difference is selected as the representative power station, and the number of the remaining power stations is updated, and the process returns to the above step of obtaining the average value of the second comprehensive correlation coefficient between each of the remaining power stations and the selected representative power station, and a new redundancy and a new representative power station are determined until the total capacity of all the selected representative power stations is not less than the set capacity threshold, and the cycle is stopped to obtain the final multiple representative power stations.

13. The system according to claim 11, characterized in that The plurality of sub-correlation coefficients include a Pearson correlation coefficient, a Spearman rank correlation coefficient, and a Kendall rank correlation coefficient.

14. The system according to claim 10, characterized in that The weight calculation module also includes: An initialization unit, used for initializing the parameters of the egret optimization algorithm, wherein the number of egret groups in the parameters is the number of the plurality of representative power stations, each egret in the egret group corresponds to one of the representative power stations, and the position of each egret corresponds to the weight of one of the representative power stations; A position determination unit, used for calculating the fitness value of the position of each egret according to a preset fitness function, with the goal of finding the position with the minimum fitness value, and determining the best position of each egret; The solving unit is used for simulating the hunting behavior of the egret eagle based on the egret optimization algorithm in the exploration phase, preliminarily updating the position of the egret eagle, and preliminarily updating the optimal position of the egret eagle according to the fitness function; simulating the avoiding behavior of the egret eagle in the development phase, updating the position of the egret eagle again, and updating the optimal position of the egret eagle again according to the fitness function, and iteratively updating the position and optimal position of the egret eagle again until the maximum number of iterations is reached, and the optimal position of each egret eagle when the maximum number of iterations is reached is used as the optimal weight of the corresponding representative power station.

15. The system of claim 14, wherein: The position determination unit is specifically used for: Based on the fitness function, the fitness value of the position of each egret is calculated according to the square difference between the weighted sum of the historical power of each representative power station and the historical total power of the distributed photovoltaic power station cluster.

16. The system of claim 10, wherein: The sub-model training module is specifically used for: Acquire historical power and geographical location information in historical operation data of each of the representative power stations, and determine static graph structure data of the plurality of representative power stations based on the geographical location information of the plurality of representative power stations; Based on each of the representative power stations, the spatiotemporal graph convolution model is trained with the historical power of the representative power station and the static graph structure data as input to obtain a power prediction sub-model of the representative power station.

17. The system of claim 16, wherein: The sub-model training module is specifically used for: Based on the geographical location information of each of the representative power stations, construct an adjacency matrix of the plurality of representative power stations, and use the adjacency matrix as the static graph structure data; The adjacency matrix is ​​expressed as: Wherein, A(i,j) is the adjacency matrix, D(i,j) is the geographical distance between representative power station i and representative power station j; S is the variance of the geographical distance between each representative power station; ε is the distance threshold; are the latitudes of power station i and power station j respectively; i , j are the longitudes of power station i and j respectively; R is the radius of the earth.

18. The system of claim 16, wherein: The sub-model training module is specifically used for: Acquire the historical power of the representative power station and the static graph structure data to construct a data set; divide the data set into a training set and a test set according to a set ratio; Based on the training set, the historical power of the representative power station in the current historical training period and the space-time graph structure data are used as input, and the historical power of the representative power station in the next historical training period of the current historical training period is used as output, the space-time graph convolution model is trained, the optimizer is used to optimize and update the training parameters of the space-time graph convolution model, and the test set is used to evaluate the space-time graph convolution model during the training process until the loss function of the space-time graph convolution model converges, so as to obtain the trained space-time graph convolution model.

19. A distributed photovoltaic power station cluster power prediction method, characterized in that: include: Obtain the actual power of each representative power station in the distributed photovoltaic power station cluster during the set historical period; Taking the actual power of each representative power station in a set historical period as input, a pre-built distributed photovoltaic power station cluster power prediction model is used to perform prediction to obtain the predicted power of the distributed photovoltaic power station cluster in the period to be predicted; Wherein, the time interval between the set historical period and the period to be predicted does not exceed a preset time threshold; The distributed photovoltaic power station cluster power prediction model is constructed based on the distributed photovoltaic power station cluster power prediction model construction method according to any one of claims 1 to 9.

20. A distributed photovoltaic power station cluster power prediction system, characterized in that: include: The second data acquisition module is used to obtain the actual power of each representative power station in the distributed photovoltaic power station cluster during a set historical period; A prediction module is used to use the actual power of each representative power station in a set historical period as input, and use a pre-built distributed photovoltaic power station cluster power prediction model to perform prediction, so as to obtain the predicted power of the distributed photovoltaic power station cluster in the period to be predicted; Wherein, the time interval between the set historical period and the period to be predicted does not exceed a preset time threshold; The distributed photovoltaic power station cluster power prediction model is constructed based on the distributed photovoltaic power station cluster power prediction model construction method according to any one of claims 1 to 9.

21. An electronic device, characterized in that: include: at least one processor and memory; The memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a distributed photovoltaic power station cluster power prediction model construction method as described in any one of claims 1 to 9, or a distributed photovoltaic power station cluster power prediction method as described in claim 19 is implemented.

22. A readable storage medium, characterized in that: An execution program is stored thereon, and when the execution program is executed, a distributed photovoltaic power station cluster power prediction model construction method as described in any one of claims 1 to 9, or a distributed photovoltaic power station cluster power prediction method as described in claim 19 is implemented.

Images