A photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division

Abstract

The application discloses a photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division, and belongs to the technical field of photovoltaic cluster power short-term prediction. The method comprises the following steps: acquiring the historical power of all power stations in a photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the longitude and latitude data; dividing all the power stations into n sub-regions according to the longitude and latitude data; acquiring the generalized weather pattern of the cluster on the predicted day; training a graph convolutional neural network based on the D+1 day forecast irradiance of all the power stations, the D day historical power and the calculated Kendall coefficient; obtaining power prediction models under three generalized weather patterns respectively; inputting the feature matrix and the adjacency matrix of the predicted day into the power prediction model under the corresponding generalized weather pattern according to the generalized weather pattern of the cluster on the predicted day; and obtaining the prediction result. The application significantly improves the power prediction accuracy and the model generalization ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic cluster power short-term prediction, and particularly relates to a photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division. BACKGROUND

[0002] With the construction of sand desert photovoltaic large base in China, more and more photovoltaic power stations emerge in the form of clusters. Accurate short-term prediction of photovoltaic cluster power generation is helpful for the dispatching department to reasonably arrange power generation plan and improve new energy utilization rate.

[0003] At present, the prediction methods of photovoltaic cluster power mainly include accumulation method, statistical upscaling method and direct prediction method. The accumulation method obtains the total power of the cluster by summing up the prediction results of each power station, but if the accuracy of one power station is low due to lack of data or other problems, the accuracy of the total power of the cluster will be affected. The statistical upscaling method depends on the accuracy of the selected representative power station. The direct prediction method needs to construct a suitable prediction model to fully extract the characteristics of the cluster so as to reduce the prediction error. In recent years, due to the fact that the graph convolutional neural network can represent the photovoltaic cluster in the form of graph and fully extract the spatio-temporal correlation of each power station, it has been widely used.

[0004] However, for a photovoltaic cluster, it contains multiple power stations, and due to the long distance, each power station may be in a different weather pattern, so that the photovoltaic cluster does not have a unified weather pattern. However, in fact, the combination of different weather patterns in the cluster will still make the cluster show different output characteristics. In this case, using one model to fit multiple mapping relationships will still increase the fitting difficulty. At present, there is no generalized weather pattern for photovoltaic clusters, and no prediction model is fitted respectively to improve the prediction accuracy.

[0005] Overall, the existing research lacks exploration of the generalized weather pattern of the cluster, and does not explore the division method of the generalized weather pattern of the cluster, and does not construct a power prediction model respectively for different generalized weather patterns of the cluster. The present application proposes a photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division, which extends the weather pattern of a small area to the generalized weather pattern of the cluster scale, and then constructs a dynamic graph convolutional neural network prediction model respectively for different generalized weather patterns, extracts the dynamic spatio-temporal correlation characteristics of each power station, avoids using one prediction model to fit multiple different output mapping relationships, and thus improves the model accuracy.

[0006] In view of the above problems, a photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division is needed to solve the above problems existing in the traditional method and realize the short-term prediction of photovoltaic cluster power. SUMMARY

[0007] The application aims to provide a photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division, which significantly improves the power prediction accuracy and model generalization ability by dividing the photovoltaic cluster generalized weather pattern and modeling accordingly.

[0008] To achieve the above-mentioned purpose, the technical scheme adopted by the application is as follows:

[0009] A photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division, comprising:

[0010] Step 1: Obtain the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the latitude and longitude data, and divide the power data and the irradiance data into a training set and a test set;

[0011] Step 2: According to the latitude and longitude data of all power stations, use a K-means clustering model to divide them into n sub-regions;

[0012] Step 3: Obtain the weather pattern of each sub-region every day, and obtain the generalized weather pattern of the cluster on the predicted day based on the weather pattern of each sub-region every day;

[0013] Step 4: Obtain the D+1 day forecast irradiance and the D day historical power of all power stations, and calculate the Kendall coefficient every day, and train the graph convolutional neural network based on the D+1 day forecast irradiance, the D day historical power and the Kendall coefficient of all power stations, and obtain the power prediction model under three kinds of generalized weather patterns respectively;

[0014] Step 5: According to the generalized weather pattern of the cluster prediction day, input the feature matrix and the adjacency matrix of the predicted day into the power prediction model under the corresponding generalized weather pattern to obtain the prediction result.

[0015] Further, in step 1, the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the latitude and longitude data are obtained, and the power data and the irradiance data are divided into a training set and a test set, specifically:

[0016] Obtain the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the latitude and longitude data;

[0017] Divide the historical power of all power stations, the total historical power of the cluster and the D+1 day forecast irradiance into a training set and a test set according to the ratio of 7:3.

[0018] Further, in step 2, according to the latitude and longitude data of all power stations, use a K-means clustering model to divide them into n sub-regions, specifically:

[0019] Input the longitude and latitude data of all power stations into the K-means clustering model, set the number of clusters, and obtain the category label to which each power station belongs;

[0020] Adjust the number of clusters, and calculate the error sum of squares of the clusters corresponding to each number of clusters;

[0021] Select the number of clusters with the smallest error sum of squares as the optimal number of clusters, and divide to obtain n sub-regions.

[0022] Further, in step 3, the weather pattern of each sub-region each day is obtained, and the generalized weather pattern of the cluster on the predicted day is obtained based on the weather pattern of each sub-region each day, specifically:

[0023] Calculate the feature quantity of the daily forecast irradiance of each power station, and normalize it to obtain the normalized feature quantity;

[0024] Fit the Gaussian mixture clustering model based on the normalized feature quantity of all power stations in each sub-region of the training set, set the number of clusters to 3, obtain the weather pattern recognition model of each sub-region, and obtain the weather pattern of each sub-region each day based on the weather pattern recognition model;

[0025] Calculate the feature quantity of the historical total power of the cluster, and obtain three generalized weather patterns of the cluster based on the feature quantity of the historical total power of the cluster.

[0026] Further, the feature quantity of the daily forecast irradiance includes the relative clearness index, the maximum daily irradiance, the daily irradiance variance, the number of daily intersection points, and the Shannon entropy.

[0027] Further, the feature quantity of the historical total power of the cluster includes the daily maximum value, the daily average value, and the number of daily intersection points.

[0028] Further, based on the feature quantity of the historical total power of the cluster, three generalized weather patterns of the cluster are obtained, specifically:

[0029] The daily maximum value, the daily average value, and the number of daily intersection points are subjected to maximum-minimum normalization to obtain the normalized daily maximum value, the normalized daily average value, and the normalized number of daily intersection points.

[0030] Fit a new Gaussian mixture clustering model based on the normalized daily maximum value, the normalized daily average value, and the normalized number of daily intersection points, set the number of clusters to 3, obtain three cluster centers, extract the feature vectors of the three cluster centers, which respectively represent three output characteristics of the cluster, wherein the three output characteristics correspond to three generalized weather patterns, calculate the feature vectors of the cluster centers, obtain the feature quantity of the cluster on the predicted day, and calculate the generalized weather pattern of the predicted day based on the feature quantity of the cluster on the predicted day.

[0031] Further, in step 4, the D+1 day predicted irradiance and the D day historical power of all power stations are obtained, and the Kendall coefficient of each day is calculated, and the graph convolutional neural network is trained based on the D+1 day predicted irradiance, the D day historical power and the Kendall coefficient of all power stations, and three power prediction models under different generalized weather modes are trained respectively, and specifically, the power prediction model under each generalized weather mode is obtained.

[0032] The D+1 day predicted irradiance and the D day historical power of all power stations are obtained to form a feature matrix.

[0033] The Kendall correlation coefficient of the D+1 day predicted irradiance and the D day historical power of all power stations is calculated to form an adjacency matrix.

[0034] The feature matrix and the adjacency matrix of each day under different generalized weather modes of the photovoltaic cluster are input into the graph convolutional neural network, the prediction model is trained on the training set, and the power prediction model under each generalized weather mode is trained respectively.

[0035] Further, in step 5, according to the generalized weather mode of the cluster prediction day, the feature matrix and the adjacency matrix of the predicted day are input into the power prediction model under the corresponding generalized weather mode to obtain the prediction result, and specifically, the prediction result of each power station is obtained.

[0036] According to the generalized weather mode of the cluster prediction day, the feature matrix and the adjacency matrix of the predicted day are input into the power prediction model under the corresponding generalized weather mode to obtain the prediction result of each power station.

[0037] The prediction results of each power station are added to obtain the predicted power of the cluster.

[0038] In summary, the present application has at least one of the following beneficial technical effects:

[0039] 1. The prediction concept is innovated, the existing research lacks the definition and division of the cluster scale composite weather characteristics, and the present application first extends the single weather mode of a small area to the generalized weather mode of the cluster scale. The mode is not only the generalization of the weather type, but also the generalization of the spatial distribution, which can more accurately represent the overall output characteristics of the cluster under different weather combinations, and provides a new theoretical perspective and heuristic guidance for photovoltaic cluster power prediction.

[0040] 2. The prediction accuracy of the photovoltaic cluster power is effectively improved. The traditional method uses a single model to fit the complex mapping relationship under all weather conditions, and the model fitting is difficult and low in precision. The present application divides different generalized weather modes, and constructs a special prediction model for each mode, realizes the "divide and conquer", avoids the mutual interference between different output characteristics, significantly reduces the prediction error, and performs excellently in the power dispatching scene with extremely high prediction accuracy requirement.

[0041] 3. The method fully exploits the spatio-temporal correlation characteristics of photovoltaic clusters, enhancing the robustness of the model. The dynamic graph convolutional neural network used in the present application can accurately capture the correlation between each power station under different weather conditions through a dynamic adjacency matrix. This method not only fully extracts the spatial structure features and time series features of the cluster, but also enables the model to adapt to the dynamic changes of meteorological conditions, having stronger generalization ability and stability compared to models that only consider time series or static space.

[0042] 4. The method solves the prediction problem caused by uneven weather in the cluster and has strong engineering practicability. The present application effectively aggregates local weather information into the output characteristics of the entire cluster through the technical path of "sub-region division-local weather recognition-contribution fusion", thereby accurately identifying the generalized weather pattern. This method overcomes the "averagism" error caused by treating the cluster as a whole weather unit in traditional methods, and has high application value in spatially distributed photovoltaic clusters, such as the Shaguo Desert Large Photovoltaic Base. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The figure is a schematic diagram of the method of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0045] As shown in Figure 1 The present application provides a photovoltaic cluster power short-term prediction method based on spatio-temporal generalized weather pattern division, comprising:

[0046] Step 1: Obtain the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the latitude and longitude data, and divide the power data and irradiance data into training set and test set;

[0047] Step 2: Divide all power stations into n sub-regions using the K-means clustering model according to the latitude and longitude data of all power stations;

[0048] Step 3: Obtain the weather pattern of each sub-region every day, and obtain the generalized weather pattern of the cluster on the day to be predicted based on the weather pattern of each sub-region every day;

[0049] Step 4: Obtain the D+1-day predicted irradiance of all power stations and the D-day historical power, and calculate the Kendall coefficient of each day. Train the graph convolutional neural network based on the D+1-day predicted irradiance of all power stations, the D-day historical power, and the Kendall coefficient, and obtain the power prediction model under three general weather patterns respectively;

[0050] Step 5: According to the general weather pattern of the predicted day, input the feature matrix and adjacency matrix of the predicted day into the power prediction model under the corresponding general weather pattern to obtain the prediction result.

[0051] In step 1, the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1-day predicted irradiance, and the latitude and longitude data are obtained, and the power data and irradiance data are divided into training set and test set, specifically:

[0052] Obtain the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1-day predicted irradiance, and the latitude and longitude data;

[0053] Divide the historical power of all power stations, the total historical power of the cluster, and the D+1-day predicted irradiance into training set and test set according to the ratio of 7:3.

[0054] In step 2, according to the latitude and longitude data of all power stations, use the K-means clustering model to divide them into n sub-regions, specifically:

[0055] Input the latitude and longitude data of all power stations into the K-means clustering model, set the cluster number n=3, and obtain the category label of each power station;

[0056] Adjust the cluster number n to an integer between 2 and 10, calculate the sum of squared errors of each cluster number, and the calculation formula is:

[0057] (1)

[0058] In the formula, is the sample point in the cluster , is the center of the th cluster, i is the cluster number of clustering; Select the cluster number with the smallest sum of squared errors as the best cluster number n, and divide it into n sub-regions.

[0059] In step 3, obtain the weather pattern of each sub-region every day, and obtain the general weather pattern of the cluster on the predicted day based on the weather pattern of each sub-region every day, specifically:

[0060]

[0061] ​Step 301: Calculate the feature quantity of the daily forecast irradiance of each power station, and normalize it to obtain the normalized feature quantity;

[0062] Step 302: Fit the Gaussian mixture clustering model based on the normalized feature quantity of all power stations in each sub-region of the training set, set the cluster number to 3, obtain the weather pattern recognition model of each sub-region, and obtain the weather pattern of each sub-region each day based on the weather pattern recognition model;

[0063] Step 303: Calculate the feature quantity of the historical total power of the cluster, and obtain three generalized weather patterns of the cluster based on the feature quantity of the historical total power of the cluster.

[0064] In step 301, the feature quantity of the daily forecast irradiance of each power station is calculated, and normalized to obtain the normalized feature quantity, which is specifically:

[0065] The feature quantity of the daily forecast irradiance of each power station is calculated, which is:

[0066] 1. Relative clearness index:

[0067] (2)

[0068] In the formula, represents the discrete sampling value of the predicted irradiance, m =1, 2,..., M ; M represents the number of sampling points in 24 hours in the power station ， represents the maximum sampling value of the total irradiance in a day in the power station;

[0069] 2. Daily maximum irradiance:

[0070] (3)

[0071] In the formula, represents the discrete sampling value of the corrected irradiance, m=1, 2,..., M; M represents the number of sampling points in 24 hours in the power station;

[0072] 3. Daily irradiance variance:

[0073] (4)

[0074] 4. Number of daily intersection points:

[0075] (5)

[0076] (6)

[0077] In the formula, denotes the selected period, denotes the exponential smoothing value of the previous time period, denotes the original irradiance sequence;

[0078] 5. Shannon entropy:

[0079] (6)

[0080] wherein, denotes the predicted irradiance at each frequency;

[0081] The maximum and minimum values of the daily maximum irradiance, the daily irradiance variance, the number of daily cross points, and the Shannon entropy are normalized, and the normalization formula is:

[0082] (7)

[0083] wherein, is the original data, is the minimum value in the original data, is the maximum value in the original data.

[0084] In step 302, the Gaussian mixture clustering model is fitted based on the normalized feature quantities of all power stations in each sub-region of the training set, the number of clusters is set to 3, the weather pattern recognition model of each sub-region is obtained, and the weather pattern of each sub-region each day is obtained based on the weather pattern recognition model, which is specifically:

[0085] The Gaussian mixture clustering model is fitted based on the normalized feature quantities of all power stations in each sub-region of the training set, the number of clusters is set to 3, the weather pattern recognition model of each sub-region is obtained, and the weather pattern of each sub-region each day is obtained based on the weather pattern recognition model, wherein the weather pattern includes sunny, cloudy, or rainy days.

[0086] In step 303, the feature quantity of the historical total power of the cluster is calculated, and the three generalized weather patterns of the cluster are obtained based on the feature quantity of the historical total power of the cluster, which is specifically:

[0087] The feature quantity of the historical total power of the cluster on the training set is calculated, which is specifically:

[0088] 1. Daily maximum value:

[0089] (8)

[0090] wherein, p is the historical total power of the cluster;

[0091] 2. Daily average value:

[0092] (9)

[0093] 3. Number of intraday crossover points:

[0094] (10)

[0095] The obtained daily power characteristics of the cluster are normalized by performing maximum and minimum value normalization to obtain the normalized intraday maximum value. Normalized intraday average Normalized intraday crossover points ;

[0096] Normalized inner maximum value Normalized intraday average Normalized intraday crossover points The data is input into a new Gaussian mixture clustering model, with the number of clusters set to 3, resulting in three cluster centers. Their feature vectors are then extracted. , , These represent the three output characteristics of the cluster;

[0097] Since the power output characteristics of a power cluster essentially depend on the combination of weather patterns from its individual power stations, and this combination can be called the generalized weather pattern of the cluster, the three power output characteristics of the cluster can represent three generalized weather patterns. This allows us to obtain the generalized weather pattern of the cluster on the training set. , , It is viewed as a three-dimensional weather score encompassing a generalized sunny pattern, a generalized cloudy pattern, and a generalized rainy pattern.

[0098] The following is a summary of the overall steps involved in identifying generalized weather models, which include:

[0099] S1: Input the irradiance characteristics of each power station in each region of the cluster on the test set into the Gaussian mixture clustering model fitted for each region in step 302 to obtain the weather pattern of each sub-region on the predicted day.

[0100] S2: Calculate the contribution of each sub-region to three metrics on the training set: daily maximum, daily average, and number of daily crossover points.

[0101] 1. Peak contribution:

[0102] (11)

[0103] In the formula, Indicates the first z Peak contribution of each sub-region Indicates the first z Power sequence of each sub-region Represents the power sequence of the cluster;

[0104] 2. Mean contribution:

[0105] (12)

[0106] In the formula, For the first z The mean contribution of each sub-region For the first z Sub-region t The daily average of total power. For the cluster t The daily average of total power. The number of days in the training set;

[0107] 3. Contribution to volatility:

[0108] (13)

[0109] In the formula, For the first z The contribution of fluctuations in each sub-region For the first z Sub-region t The number of cross points of the total daily power. Cluster No. t The number of crossover points in the total daily power. This represents the total number of sub-regions.

[0110] S3: Combine the weather patterns of the sub-regions on the predicted day to select the weather scores for each power characteristic, multiply them by the contribution of each sub-region, and calculate the three power characteristics of the cluster on the predicted day. If the weather pattern of a certain sub-region is a sunny pattern, then its weather score is... If it is a cloudy mode, its weather score is: If it is a cloudy / rainy pattern, its weather score is: The three power characteristics of the cluster on the predicted date are calculated as follows:

[0111] 1. Daily maximum value:

[0112] (14)

[0113] In the formula, For the first z Each sub-region is in weather pattern w The highest weather score within a given time period. w Sunny, cloudy, or rainy days are suitable. For the first z Peak contribution of each sub-region This represents the total number of sub-regions.

[0114] 2、Daily average value:

[0115] (15)

[0116] wherein, is the weather score of the mth sub-region on the day when the weather pattern is z , and the weather pattern can be taken as sunny, cloudy, and rainy, w is the average contribution degree of the mth sub-region; w z wherein,

[0117] 3、Daily intersection point number:

[0118] (16)

[0119] wherein, is the weather score of the mth sub-region on the day when the weather pattern is z , and the weather pattern can be taken as sunny, cloudy, and rainy, w is the fluctuation contribution degree of the mth sub-region; w z

[0120] S4: input the three power characteristic quantities of the obtained cluster into the clustering model fitted in step 303 to obtain the generalized weather pattern of the cluster on the predicted day.

[0121] In step 4, the D+1 day predicted irradiance and the D day historical power of all power stations are obtained, and the Kendall coefficient of each day is calculated. Based on the D+1 day predicted irradiance, the D day historical power, and the Kendall coefficient of all power stations, the graph convolutional neural network is trained to obtain power prediction models under three generalized weather patterns, specifically:

[0122] The D+1 day predicted irradiance and the D day historical power of all power stations are obtained to form a feature matrix, which is:

[0123] (17)

[0124] wherein, is the D+1 predicted irradiance vector of the mth power station, is the historical previous day power vector of the mth power station, is the number of power stations, and the feature matrix has a size of ;

[0125] The Kendall correlation coefficient of the D+1 day predicted irradiance and the D day historical power of all power stations is calculated to form an adjacency matrix, which is:​​​​​

[0126] (18)

[0127] In the formula, For the first m The power station and the first i Kendall correlation coefficient of D+1 predicted irradiance among the power plants For the first i The power station and the first m Kendall correlation coefficient and critical matrix of historical power output of individual power stations on the previous day. Size is ;

[0128] The daily feature matrix and adjacency matrix of the photovoltaic cluster under different generalized weather patterns are input into a graph convolutional neural network, and a prediction model is trained on the training set to obtain power prediction models under the three generalized weather patterns.

[0129] In step 5, based on the generalized weather model of the predicted day, the feature matrix and adjacency matrix of the predicted day are input into the power prediction model under the corresponding generalized weather model to obtain the prediction result, specifically:

[0130] Based on the generalized weather model of the cluster prediction day, the feature matrix and adjacency matrix of the predicted day are input into the power prediction model of the corresponding generalized weather model to obtain the prediction results of each power station.

[0131] The predicted power of the cluster is obtained by summing the prediction results of each power station, specifically:

[0132] (19)

[0133] In the formula, For the forecast results of each power station, For the predicted power of the cluster, m This represents the total number of power stations.

[0134] This invention provides an embodiment in which data from a photovoltaic cluster is used to verify the method proposed in this invention, and a comparative experiment is set up to verify the effectiveness of the method. The comparison conditions are shown in Table 1.

[0135] Table 1 Comparison Condition Settings Table

[0136]

[0137] The following four indicators will be used for evaluation:

[0138] 1) Root mean square error: used to measure the deviation between the observed value and the true value.

[0139] (20)

[0140] wherein, is the actual total power (unit: MW), is the predicted total power (unit: MW), is the total output of the cluster (unit: MW), and the total output of the 38 photovoltaic power stations of the present application is 2150 MW, is the number of predicted values in a single prediction.

[0141] 2) Mean absolute error: can better reflect the actual situation of the error of the predicted value.

[0142] (21)

[0143] wherein, is the actual total power (unit: MW), is the predicted total power (unit: MW), is the total output of the cluster (unit: MW), is the number of predicted values in a single prediction.

[0144] 3) Determination coefficient: used to evaluate the degree of coincidence of the predicted value and the actual value.

[0145] (22)

[0146] wherein, is the actual total power (unit: MW), is the predicted total power (unit: MW), is the average value of the actual value (unit: MW), S SE is the sum of squared residuals, S ST is the total sum of squares.

[0147] 4) Correlation coefficient: used to evaluate the linear correlation of the predicted value and the true value.

[0148] (23)

[0149] wherein, is the actual total power (unit: MW), is the predicted total power (unit: MW), is the average value of the actual value (unit: MW), is the average value of the predicted value (unit: MW).

[0150] The evaluation indexes of each method calculated are shown in Table 2:

[0151] Table 2 Evaluation index table of each method

[0152]

[0153] Embodiments of the present application can be provided as a method, a system, or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and the like) embodying computer-readable program code.

[0154] The present application is described in reference to the flowchart illustrations and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0155] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0157] The contents of the present application which are not described in detail in the specification belong to the prior art known to those skilled in the art. It is hereby indicated that the above description is helpful for those skilled in the art to understand the present application, but does not limit the protection scope of the present application. Any equivalent replacement, modification, improvement and deletion of the above description without departing from the essential content of the present application falls within the protection scope of the present application.

Claims

1. A photovoltaic cluster power short-term prediction method based on spatiotemporal generalized weather pattern division, characterized in that, Comprise: Step 1: obtain the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the latitude and longitude data, and divide the power data and the irradiance data into a training set and a test set; Step 2: according to the latitude and longitude data of all power stations, use the K-means clustering model to divide into n sub-regions; Step 3: obtain the weather pattern of each sub-region every day, and obtain the generalized weather pattern of the cluster on the predicted day based on the weather pattern of each sub-region every day; Step 4: obtain the D+1 day forecast irradiance of all power stations and the D day historical power, and calculate the Kendall coefficient every day, train the graph convolutional neural network based on the D+1 day forecast irradiance of all power stations, the D day historical power and the Kendall coefficient, and obtain the power prediction model under three kinds of generalized weather patterns respectively; Step 5: according to the generalized weather pattern of the cluster prediction day, input the feature matrix and the adjacency matrix of the predicted day into the corresponding power prediction model under the generalized weather pattern to obtain the prediction result. 2.The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 1, characterized in that, In step 1, the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the latitude and longitude data are obtained, and the power data and the irradiance data are divided into a training set and a test set, specifically: Obtain the historical power of all power stations in the photovoltaic cluster, the total historical power of the cluster, the D+1 day forecast irradiance and the latitude and longitude data; Divide the historical power of all power stations, the total historical power of the cluster and the D+1 day forecast irradiance into a training set and a test set according to the ratio of 7:

3. 3.The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 2, characterized in that, In step 2, according to the latitude and longitude data of all power stations, use the K-means clustering model to divide into n sub-regions, specifically: Input the latitude and longitude data of all power stations into the K-means clustering model, set the cluster number, and obtain the category label of each power station; Adjust the cluster number, calculate the error sum of squares of each cluster number corresponding to the cluster; Select the cluster number with the smallest error sum of squares as the best cluster number, and divide to obtain n sub-regions. 4.The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 3, characterized in that, In step 3, the weather pattern of each sub-region every day is obtained, and the generalized weather pattern of the cluster on the predicted day is obtained based on the weather pattern of each sub-region every day, specifically: Calculate the feature quantity of the daily forecast irradiance of each power station, and normalize it to obtain the normalized feature quantity; Based on the normalized feature quantity of all power stations in each sub-region of the training set, fit the Gaussian mixture clustering model, set the cluster number to 3, obtain the weather pattern recognition model of each sub-region, and obtain the weather pattern of each sub-region every day based on the weather pattern recognition model; Calculate the feature quantity of the historical total power of the cluster, and obtain the three generalized weather patterns of the cluster based on the feature quantity of the historical total power of the cluster.

5. The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 4, characterized in that, The feature quantity of the daily forecast irradiance includes relative clearness index, daily maximum irradiance, daily irradiance variance, daily cross point number and Shannon entropy. 6.The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 5, characterized in that, The feature quantity of the historical total power of the cluster includes daily maximum value, daily average value and daily cross point number.

7. The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 6, characterized in that, Based on the feature quantity of the historical total power of the cluster, the three generalized weather patterns of the cluster are obtained, specifically: The daily maximum value, the daily average value and the daily cross point number are subjected to maximum minimum normalization processing to obtain a normalized daily maximum value, a normalized daily average value and a normalized daily cross point number; The normalized daily maximum value, the normalized daily average value and the normalized daily cross point number are used to fit a new Gaussian mixture clustering model, the number of clusters is set to 3, three cluster centers are obtained, feature vectors of the three cluster centers are extracted, and the three cluster centers represent three output characteristics of the cluster, wherein the three output characteristics correspond to three generalized weather patterns, the feature vectors of the cluster centers are calculated to obtain characteristic quantities of the cluster on the predicted day, and the generalized weather pattern of the predicted day is calculated based on the characteristic quantities of the cluster on the predicted day. 8.The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 7, characterized in that, In step 4, the D+1 day forecast irradiance of all power stations and the D day historical power are obtained, and the Kendall coefficient of each day is calculated, and the graph convolutional neural network is trained based on the D+1 day forecast irradiance of all power stations, the D day historical power and the Kendall coefficient, and the power prediction model under three generalized weather patterns is trained respectively, specifically: The D+1 day forecast irradiance of all power stations and the D day historical power are obtained to form a feature matrix; The Kendall correlation coefficient of the D+1 day forecast irradiance of all power stations and the D day historical power is calculated to form an adjacency matrix; The daily feature matrix and the adjacency matrix of the photovoltaic cluster under different generalized weather patterns are input into the graph convolutional neural network, the prediction model is trained on the training set, and the power prediction model under three generalized weather patterns is trained respectively. 9.The photovoltaic cluster power short-term prediction method based on space-time generalized weather pattern division according to claim 8, characterized in that, In step 5, according to the generalized weather pattern of the cluster prediction day, the feature matrix and the adjacency matrix of the predicted day are input into the power prediction model under the corresponding generalized weather pattern to obtain the prediction result, specifically: According to the generalized weather pattern of the cluster prediction day, the feature matrix and the adjacency matrix of the predicted day are input into the power prediction model under the corresponding generalized weather pattern to obtain the prediction result of each power station; The prediction results of each power station are added to obtain the predicted power of the cluster.

Images