Correlation-based photovoltaic output prediction method

Through the combination of the Bi-LSTM network, the distributed photovoltaic power generation prediction problem is solved, and the characterization and accurate prediction of the photovoltaic output correlation relationship is realized, which improves the grid stability.

CN119051003APending Publication Date: 2024-11-29HEBEI UNIV OF TECH
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202411154190.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

The power prediction of distributed photovoltaic power generation is difficult to accurately perform, especially due to the difficulty in collecting historical data and the lack of real-time monitoring data, which leads to the contradiction between the large power grid and the distributed power supply, affecting the stability of the power grid.

Method used

A correlation-based photovoltaic output prediction method is proposed. Through the concept of occlusion factor and the Copula function model, combined with the Bi-LSTM network, the deterministic and random parts of the photovoltaic output are separated, and the historical data of the reference photovoltaic power station is used to perform distributed photovoltaic power stations. Expect efforts to predict.

Benefits of technology

The photovoltaic output correlation relationship characterization based on a small amount of data is realized, the accuracy and reliability of distributed photovoltaic power generation prediction is improved, and the photovoltaic output characteristics and weather characteristics are fully considered.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119051003A_ABST
    Figure CN119051003A_ABST
Patent Text Reader

Abstract

The invention discloses a photovoltaic output prediction method based on correlation, and the method comprises the steps: carrying out the prediction of the output of a target date of a reference photovoltaic power station with sufficient historical photovoltaic power data and weather data through a conventional method, and calculating an unshielded factor of the reference photovoltaic power station on the basis of the historical photovoltaic power data and the weather data; establishing indexes according to the mean value and the mean absolute deviation of the non-shielding factors, dividing weather types, clustering weather, and further constructing a correlation model of the reference photovoltaic power station and the target photovoltaic power station under various weather conditions by using a Copula function; and according to output prediction data of a target date of the reference photovoltaic power station, calculating a corresponding non-shielding factor, and judging a weather type to which the reference photovoltaic power station belongs. And obtaining a non-shielding factor of the output of the target photovoltaic power station through the correlation model of the weather type, calculating according to a formula to obtain the actual output condition of the target photovoltaic power station, and finally realizing the output prediction of the target photovoltaic power station based on the output prediction data of the reference photovoltaic power station.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic power generation, and specifically relates to a photovoltaic output prediction method based on correlation, which accurately predicts photovoltaic power generation according to relevant historical data and NWP data. Background Art

[0002] Photovoltaic power generation is an important part of China's current energy structure. Accurate photovoltaic prediction is of great significance to the safe and stable operation of the power system. The power grid formulates dispatching plans and adjusts the grid frequency based on the photovoltaic prediction results to ensure grid stability and power supply quality.

[0003] Photovoltaic power generation can be divided into centralized photovoltaic and distributed photovoltaic. The characteristics of making full use of dispersed solar energy resources and consuming them locally have made people pay more and more attention to the development of distributed photovoltaic power generation. However, for large power grids, distributed photovoltaic is an uncontrollable source. The randomness of its output has caused the main contradiction between large power grids and distributed power sources. As its installed capacity increases year by year, it has become necessary to carry out research on the power prediction of distributed photovoltaic power generation. However, compared with traditional photovoltaic power stations, distributed photovoltaic power stations have problems such as difficulty in collecting historical data and lack of real-time monitoring data, which makes it difficult to predict the power of a single power station.

[0004] The main natural factors that affect photovoltaic power generation include irradiance, ambient temperature, humidity, etc., and irradiance is affected by season and weather. These factors are almost the same in the same region, which leads to a high correlation between photovoltaic outputs in the same region. Based on the above analysis, this paper proposes a photovoltaic output prediction method that takes correlation into account. Summary of the invention

[0005] In view of the deficiencies of the prior art, the present invention proposes a correlation photovoltaic output prediction method. Photovoltaic output can be regarded as a combination of a deterministic part and a stochastic part, wherein the deterministic part ignores the influence of uncertain factors such as shadows, cloud obstructions, weather changes, and temperature changes in the actual photovoltaic power generation process, and adopts the global solar radiation intensity model for modeling; the stochastic part focuses on the influence of the above uncertain factors, and proposes the concept of "photovoltaic output unobstructed factor" to describe the influence of uncertain factors on photovoltaic power generation output. Then, indicators are established based on the unobstructed factor, weather is clustered, and then a correlation model of photovoltaic output under various weather conditions is constructed using the Copula function; finally, distributed photovoltaic output prediction based on centralized photovoltaic output prediction data is realized.

[0006] To achieve the above object, the present invention adopts the following technical solution: a photovoltaic output prediction method based on correlation, characterized in that the method comprises the following steps:

[0007] Step 1: Use the prediction model to predict the photovoltaic output of the reference photovoltaic power station in the same area as the target photovoltaic power station on the target date, which specifically includes the following steps:

[0008] Step 1.1: Data preprocessing

[0009] The historical photovoltaic power data, temperature and irradiance data of the reference photovoltaic power station were collected and preprocessed. The data was cleaned by removing outliers and filling missing values ​​to obtain three kinds of raw data. Then, the three kinds of raw data were normalized by 0-1 standardization. Finally, the complete data were sampled at fixed time intervals, and invalid data were removed to obtain a data set including three kinds of time series data.

[0010] Step 1.2: Divide the dataset

[0011] Use a fixed-size sliding window to slice the three time series data in the data set, and combine the three time series data slices at the same sampling time point as a training data; for the obtained training data, use no less than 60% as the training set, no less than 20% as the validation set, and the rest as the test set;

[0012] Step 1.3: Construction and training of prediction model

[0013] The prediction model is composed of an input layer, a Bi-LSTM network and an output layer; wherein the input layer receives input data; the Bi-LSTM network combines the bidirectional model and the gating mechanism of LSTM, and is composed of two independent LSTM networks; the forward LSTM network processes the data sequence from the first time step to the last time step, extracts the forward features, and the reverse LSTM network processes the data sequence from the last time step to the first time step, extracts the backward features; then the forward and backward features of a time step are concatenated, and the output vector formed is used as the final output of the time step; the output layer is a fully connected layer, which is used to convert the output of the Bi-LSTM network into a predicted value, that is, the predicted value of photovoltaic power;

[0014] The time series data slices of temperature and irradiance in a training data are used as the input of the prediction model, and the historical photovoltaic power is used as the reference value of the prediction model output; the network parameters of the prediction model are initialized by random assignment method, the learning rate is set to 0.005, the batch size is 64, and the maximum number of training rounds is 200; RMSE is selected as the loss function, and the Adam optimizer is used;

[0015] Input a batch of training data in the training set into the prediction model, calculate the RMSE value between the predicted value output by the prediction model for the batch of training data and the reference value, and use the Adam optimizer to reversely update the network parameters of the prediction model once according to the RMSE value to complete the training of one batch; use the network parameters when the last batch training is completed as the initial parameters for the next batch training, and repeat continuously until the last batch of training data in the training set is trained to complete a round of training; after each round of training, input the training data in the validation set into the prediction model in turn, and calculate the training loss value of the prediction model that has completed this round of training for the validation set; use the network parameters when the last round of training is completed as the initial parameters for the next round of training, and repeat continuously until the training loss value of the validation set does not decrease for several consecutive rounds or the training round reaches the maximum value, and the training of the prediction model is completed;

[0016] Input the training data in the test set into the trained prediction model in sequence, calculate the MAE and MAPE between the predicted value output by the prediction model and the reference value, and when MAE is less than 10 and MAPE is less than 5, the trained prediction model is tested as the optimal prediction model; otherwise, adjust the learning rate, batch size, and maximum number of training rounds, and repeat the training and testing process of the prediction model until the optimal prediction model is obtained;

[0017] Step 1.4: Reference PV power plant's PV output Pr on target date j Prediction

[0018] The temperature and irradiance data of the target date of the reference photovoltaic power station are first preprocessed according to the method in step 1.1, and then the sliding window of the same size in step 1.2 is used to perform data slicing processing in the order of sampling time points to obtain time series data slices of the two types of data, and the sampling time points in any two time series data slices are not repeated; the time series data slices of the same sampling time points of the above two types of data are sequentially input into the optimal prediction model in step 1.3 to obtain the predicted value output by the prediction model; the predicted value is denormalized to obtain the predicted value of the photovoltaic output at the corresponding sampling time point of the target date of the reference photovoltaic power station;

[0019] Step 2: Define the unshaded factor of photovoltaic output

[0020] The actual output P of the photovoltaic panel at sampling time t t It can be calculated using formula (2);

[0021]

[0022] Where: P stc is the output of the photovoltaic panel under standard test conditions; α Tis the power temperature coefficient of the photovoltaic panel; I t is the actual irradiance at sampling time t; T t is the actual temperature of the photovoltaic panel at sampling time t;

[0023] Transform equation (2) to separate the random factors in photovoltaic output and define the photovoltaic power station output P without considering the shading and temperature effects: tm , that is, no obstruction to output;

[0024]

[0025] Where: I tm It is the maximum value of irradiance that can be achieved without considering any shielding. This value is only related to the latitude and longitude, altitude and time of the area. Therefore, P tm Only the deterministic part of photovoltaic power generation output is included;

[0026] The unshaded factor of photovoltaic output is defined as:

[0027]

[0028] Where ε is the unshaded correction coefficient, which is used to correct the difference between the theoretical shading factor and the actual observed value. It is determined by the empirical estimation method. Based on the statistical analysis of a large amount of historical data, the ratio of the actual photovoltaic output to the theoretical output can be directly calculated to obtain the correction coefficient ε.

[0029] The unshading factor is used to characterize the unshading situation at each sampling time point, that is, the random part of the photovoltaic output;

[0030] From equation (2) to equation (4), we can know that:

[0031]

[0032] From formula (3) and formula (5), we can see that P tm The maximum value of photovoltaic output that can be achieved in theory is determined, while the actual photovoltaic output will be weakened by various random factors. t , describes the weakening effect of shadow, cloud cover, weather changes and temperature on photovoltaic output based on the theoretical maximum photovoltaic output. Its physical meaning is the ratio of the actual output of the photovoltaic power station to the deterministic output, so its value must not be greater than 1;

[0033] Step 3: Classify weather types based on historical data of reference PV power plants

[0034] Obtain the historical data of the reference photovoltaic power station in step 1, which includes the actual irradiance I at the sampling time point t t 、The power temperature coefficient of photovoltaic panels αT , the actual temperature T of the photovoltaic panel at sampling time t t , combined with the standard irradiance and standard temperature, according to formula (5), the unobstructed factor at each sampling time point is obtained; then the mean value of the unobstructed factor for each day is calculated and mean absolute error (MAE);

[0035]

[0036] m is the total number of sampling points per day, x j is the value of the unobstructed factor at sampling time point j on a certain day;

[0037] Use the mean of the unoccluded factors The mean absolute deviation (MAE) represents the degree of clearness of the day, and the weather is divided into three categories: sunny days with high clearness and no drastic fluctuations, cloudy days with low clearness and no drastic fluctuations, and bad weather with unstable clearness and drastic fluctuations.

[0038] The average of the calculated unobstructed factors for each day and mean absolute deviation MAE are normalized respectively, and the weather characteristic vector of the hth day is T h , T h =(x h , m h ), where x h is the mean value of the unobstructed factor on the hth day The normalized value of m h is the normalized value of the mean absolute deviation MAE of the unobstructed factor on the hth day; according to the weather feature vector of each day, the K-means clustering algorithm is used to cluster with the Euclidean distance as the similarity index, and its objective function is shown in formula (8);

[0039]

[0040] Where: K is the number of cluster categories, K = 3, namely sunny, cloudy, and bad weather; N is the total number of days of historical data, N k is the number of samples in the k-th cluster; x k 、m k is the eigenvalue of the weather eigenvector of the cluster center of the kth class; x ik 、m ik is the eigenvalue of the weather feature vector of the i-th sample of the k-th category; clustering takes the minimum S as the iteration goal; the K clusters obtained by the last clustering are the K weather category data groups required, and the corresponding samples in each cluster are and MAE value to determine the weather category corresponding to the group;

[0041] Step 4: Obtain the same historical data of the target photovoltaic power station in the same time period as the reference photovoltaic power station in step 3, and calculate the unshading factor of each sampling time point of each day in the corresponding time period of the target photovoltaic power station; and divide the unshading factor of each sampling time point of each day in the corresponding time period of the target photovoltaic power station into K data sets according to the dates corresponding to the K weather category data groups of the reference photovoltaic power station divided in step 3 according to the same date structure; Under each weather category, use the Copula function to establish a correlation model between the unshading factor of each sampling time point of each day in the corresponding time of the target photovoltaic power station and the unshading factor of each sampling time point of each day in the corresponding time of the reference photovoltaic power station;

[0042] Step 5: Based on the PV output on the target date predicted by the prediction model in step 1, the unshading factor predicted on the target date of the reference PV power station is calculated using formula (4), and the weather category of the day is obtained based on the mean and mean absolute deviation of the unshading factor of the day;

[0043] Step 6: Based on the weather category of the target date of the reference photovoltaic power station obtained in step 5 and the unobstructed factor predicted on the target date of the reference photovoltaic power station, combined with the correlation model of the corresponding weather category in step 4, calculate the unobstructed factor of each sampling time point predicted on the target date of the target photovoltaic power station; then, according to formula (11), calculate the actual output of the target photovoltaic power station predicted on each sampling time point on the target date;

[0044]

[0045] In the above formula, is the actual output predicted at the sampling time point t of the target date for the target PV power station, is the unobstructed factor of the sampling time point t predicted on the target date of the target PV power station, It is the deterministic part of photovoltaic power generation output and is only related to the latitude and longitude, altitude and time of the target photovoltaic power station.

[0046] Compared with the existing methods, the beneficial effects of the present invention are as follows: the present invention designs a photovoltaic output prediction method that takes correlation into account. The present invention proposes the concept of an unshaded factor, uses historical output data and solar irradiance data to calculate the photovoltaic output under unshaded conditions, and separates the deterministic and random components of photovoltaic output; and establishes indicators based on the mean and mean absolute deviation of the unshaded factor, divides weather types, clusters weather, and then uses Copula functions to construct a correlation model for photovoltaic output under various weather conditions; finally, the BiLSTM model is used to predict the output of a reference photovoltaic power station, and finally the target power station output prediction based on the reference power station output prediction data is achieved. This method fully considers the photovoltaic output characteristics and the actual local weather characteristics, and can characterize the correlation relationship of photovoltaic output based on a small amount of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The figure is a flow chart of an embodiment of a photovoltaic output prediction method based on correlation according to the present invention. DETAILED DESCRIPTION

[0048] The present invention is further explained below in conjunction with the embodiments and drawings, but this is not intended to limit the scope of protection of the present application.

[0049] The present invention proposes a photovoltaic output prediction method based on correlation (see Figure 1 ), the method includes the following contents:

[0050] Step 1: Use the BiLSTM model to predict the photovoltaic output (i.e., photovoltaic power data) of the reference photovoltaic power station in the same region as the target photovoltaic power station on the target date, which specifically includes the following steps:

[0051] Step 1.1: Data preprocessing

[0052] The collected historical photovoltaic power data, temperature and irradiance data were preprocessed. The data were cleaned by removing outliers and filling missing values ​​to obtain three types of original data. Then, the three types of original data were normalized using 0-1 standardization. Finally, the complete data were sampled at fixed time intervals, and invalid data were eliminated to obtain a data set including three types of time series data.

[0053] Step 1.2: Divide the dataset

[0054] A fixed-size sliding window is used to slice the three time series data in the data set, and the three time series data slices at the same sampling time point are combined as a training data; for the several training data obtained, 70% of them are used as training sets, 20% as validation sets, and the remaining 10% as test sets.

[0055] Step 1.3: Construction and training of prediction model

[0056] The prediction model consists of an input layer, a Bi-LSTM network (bidirectional long short-term memory neural network, which is a prior art) and an output layer. The input layer receives input data; the Bi-LSTM network combines the bidirectional model and the gating mechanism of LSTM, and is composed of two independent LSTM networks. The forward LSTM network processes the data sequence from the first time step to the last time step, extracts the forward features, and the reverse LSTM network processes the data sequence from the last time step to the first time step, extracts the backward features; then the forward and backward features of a time step are spliced, and the output vector formed is used as the final output of the time step. The output layer is a fully connected layer, which is used to convert the output of the Bi-LSTM network into a predicted value, that is, the predicted value of photovoltaic power.

[0057] The time series data slices of temperature and irradiance in a training data are used as the input of the prediction model, and the historical photovoltaic power is used as the reference value of the prediction model output; the network parameters of the prediction model are initialized by random assignment method, the learning rate is set to 0.005, the batch size is 64, and the maximum number of training rounds is 200; RMSE (root mean square error) is selected as the loss function, and the Adam optimizer is used.

[0058] A batch of training data in the training set is input into the prediction model, and the RMSE value between the predicted value output by the prediction model for the batch of training data and the reference value is calculated. According to the RMSE value, the network parameters of the prediction model are reversely updated once using the Adam optimizer to complete the training of one batch; the network parameters at the completion of the last batch training are used as the initial parameters for the next batch training, and the process is repeated continuously until the last batch of training data in the training set is trained to complete a round of training; after each round of training, the training data in the validation set is input into the prediction model in turn, and the training loss value of the prediction model for the validation set after the round of training is calculated; the network parameters at the completion of the last round of training are used as the initial parameters for the next round of training, and the process is repeated continuously until the training loss value of the validation set does not decrease for several consecutive rounds (such as 10 rounds) or the training round reaches the maximum value, and the training of the prediction model is completed.

[0059] The training data in the test set are input into the trained prediction model in turn, and the MAE (mean absolute error) and MAPE (mean absolute percentage error) between the predicted value output by the prediction model and the reference value are calculated. When MAE is less than 10 and MAPE is less than 5, the trained prediction model is tested as the optimal prediction model; otherwise, the learning rate, batch size, and maximum number of training rounds are adjusted, and the training and testing process of the prediction model is repeated until the optimal prediction model is obtained.

[0060] Step 1.4: Reference PV power station's PV output Pr on target date j Prediction

[0061] The temperature and irradiance data of the target date of the reference photovoltaic power station are first preprocessed according to the method in step 1.1, and then the data are sliced ​​in the order of sampling time points using the sliding window of the same size in step 1.2 to obtain time series data slices of the two types of data, and the sampling time points in any two time series data slices are not repeated; the time series data slices of the same sampling time points of the above two types of data are sequentially input into the optimal prediction model in step 1.3 to obtain the predicted values ​​output by the prediction model; the predicted values ​​are denormalized to obtain the predicted values ​​of the photovoltaic output at the corresponding sampling time points of the target date of the reference photovoltaic power station.

[0062] The method parameters of the denormalization process are:

[0063] P rj =Z rj ×σ P +μ P (1)

[0064] Among them, P rj is the predicted value of the photovoltaic output at sampling time point j on the target date of the reference photovoltaic power station, Z rj is the predicted value at sampling time point j output by the prediction model, σ P is the standard deviation of the original data of historical PV power data in step 1.1, μ P is the mean value of the original data of historical PV power data in step 1.1.

[0065] Step 2: Define the photovoltaic output unobstructed factor

[0066] Photovoltaic power generation is based on the photovoltaic effect, using photovoltaic modules to convert solar energy into electrical energy. The actual output P of the photovoltaic panel at the sampling time t t It can be calculated using formula (2).

[0067]

[0068] Where: P stc Under standard test conditions (corresponding to irradiance I stc =1000W / m 2 , temperature T stc =25℃) output of photovoltaic panels; α T is the power temperature coefficient of the photovoltaic panel; I t is the actual irradiance at sampling time t; T t is the actual temperature of the photovoltaic panel at the sampling time point t.

[0069] It can be seen from formula (2) that the actual irradiance I t The actual temperature T t . Actual irradiance I t Affected by many external factors such as the position of the sun, shadows, temperature, cloud cover, weather changes, etc., the temperature T t The influence of the temperature coefficient α of the photovoltaic panel T Therefore, we transform formula (2) to separate the random factors in the photovoltaic output and define the photovoltaic power station output P without considering the shading and temperature effects. tm , that is, unobstructed output. Unobstructed output P tm The calculation of can better characterize the deterministic part of photovoltaic output.

[0070]

[0071] Where: I tm It is the maximum value of irradiance that can be achieved without considering any shielding. This value is only related to the latitude and longitude, altitude and time of the area. tm It only includes the deterministic part of photovoltaic power generation output.

[0072] The photovoltaic output unobstructed factor is defined as:

[0073]

[0074] Where ε is the unshaded correction coefficient, which is used to correct the difference between the theoretical shading factor and the actual observed value. It is determined by the empirical estimation method. Based on the statistical analysis of a large amount of historical data, the ratio of the actual photovoltaic output to the theoretical output can be directly calculated to obtain the correction coefficient ε.

[0075] The unobstructed factor is used to characterize the unobstructed situation at each sampling time point, that is, the random part of the photovoltaic output. Accordingly, under the condition of knowing the unobstructed output and the actual output, the actual photovoltaic output at each sampling time point can also be restored.

[0076] From equation (2) to equation (4), we can know that:

[0077]

[0078] From formula (3) and formula (5), we can see that P tm The maximum value of photovoltaic output that can be achieved in theory is determined, while the actual photovoltaic output will be weakened by various random factors. The photovoltaic output has no shading factor η tIt describes the weakening effect of shadow, cloud cover, weather changes and temperature on photovoltaic output based on the theoretical maximum photovoltaic output. Its physical meaning is the ratio of the actual output of the photovoltaic power station to the deterministic output, so its value must not be greater than 1.

[0079] Step 3: Classify weather types based on historical data of reference PV power plants

[0080] Obtain the historical data of the reference photovoltaic power station in step 1, which includes the actual irradiance I at the sampling time point t t 、The power temperature coefficient of photovoltaic panels α T , the actual temperature T of the photovoltaic panel at sampling time t t , combined with the standard irradiance and standard temperature, according to formula (5), the photovoltaic output unshaded factor at each sampling time point is obtained. Then the average value of the photovoltaic output unshaded factor for each day is calculated. and mean absolute error (MAE).

[0081]

[0082] m is the total number of sampling points per day, x j is the value of the unshaded factor of the photovoltaic output at sampling time point j on a certain day.

[0083] The average value of the photovoltaic output without shading factor The MAE represents the degree of clearness of the day and the severity of weather fluctuations of the day. The weather is roughly divided into three categories: sunny days with high clearness and no drastic fluctuations, cloudy days with low clearness and no drastic fluctuations, and severe weather with unstable clearness and drastic fluctuations. On sunny days, the average output at each moment is close to the photovoltaic output under standard conditions, and the average absolute deviation is the smallest; on cloudy days, the average output at each moment is small, and the average absolute deviation is relatively small; on severe weather, the average output at each moment is small, and the average absolute deviation is large.

[0084] The average value of the unshaded factor of the photovoltaic output calculated for each day and mean absolute deviation MAE are normalized respectively, and the weather characteristic vector of the hth day is T h , T h =(x h , m h ), where x h is the average value of the unshaded factor of the photovoltaic output on the hth day The normalized value of m h is the normalized value of the mean absolute deviation MAE of the unshaded factor of the photovoltaic output on the hth day. According to the weather feature vector of each day, the K-means clustering algorithm is used to cluster with the Euclidean distance as the similarity index, and its objective function is shown in formula (8).

[0085]

[0086] Where: K is the number of cluster categories. In this embodiment, K is set to 3, i.e., sunny days, cloudy days, and bad weather. N is the total number of days in historical data. k is the number of samples in the k-th cluster; x k 、m k is the eigenvalue of the weather eigenvector of the cluster center of the kth class; x ik 、m ik is the eigenvalue of the weather feature vector of the i-th sample of the k-th class; clustering takes the minimum S as the iterative goal.

[0087] The specific operation process of the K-means clustering algorithm is as follows:

[0088] Step 1: Randomly select K samples from N weather feature vectors as the initial cluster centers;

[0089] Step 2: Calculate the Euclidean distance d between each sample and all K initial cluster centers k , Assign each sample to the initial cluster center closest to it to complete a clustering;

[0090] Step 3: Each initial cluster center and the samples assigned to them constitute a cluster. For each cluster center, calculate the average value of the samples assigned to it, and use the average value as the first cluster center of each cluster;

[0091] Step 4: Repeat steps 2 and 3, calculate the Euclidean distance between each sample and the updated K cluster centers, and reassign each sample to the cluster center closest to it. Then update the cluster centers to the average of the samples assigned to them. Repeat steps 2 and 3 until the Euclidean distance between the previous and next changes of each cluster center is no more than 10. -6 , clustering is over. The K clusters obtained by the last clustering are the K weather category data groups required. And MAE value, determine the weather category corresponding to the group.

[0092] Step 4: Obtain the same historical data of the target photovoltaic power station as the reference photovoltaic power station in step 3 in the same time period, and calculate the photovoltaic output unobstructed factor of the target photovoltaic power station at each sampling time point every day (the target photovoltaic power station and the reference photovoltaic power station are in the same area and are geographically close, and the meteorological data of the target photovoltaic power station is the meteorological data of the reference photovoltaic power station). And according to the dates corresponding to the K weather category data groups of the reference photovoltaic power station divided in step 3, the photovoltaic output unobstructed factor of the target photovoltaic power station at each sampling time point every day is divided into K data sets according to the same date structure; under each weather category, use the Copula function to establish a correlation model between the photovoltaic output unobstructed factor of the target photovoltaic power station at each sampling time point every day and the photovoltaic output unobstructed factor of the reference photovoltaic power station at each sampling time point every day.

[0093] The specific implementation of establishing a correlation model under a weather category is as follows:

[0094] Step 1: Use the kernel density estimation (KDE) method to estimate the marginal distribution function of the target unobstructed factor dataset X and the reference unobstructed factor dataset Y of the same weather category. Through the kernel density estimation method, the marginal distribution function F of each sample can be obtained. X (x p ) and F Y (y p ), used as the input of the Copula function. A sample of the target unobstructed factor dataset X is the unobstructed factor time series data x of one day. p , where the unobstructed factor at sampling time point j is expressed as A sample of the reference unobstructed factor dataset Y is the unobstructed factor time series data y for one day. p , where the unobstructed factor at sampling time point j is expressed as

[0095] Step 2: According to the distribution characteristics of the unobstructed factor, select a suitable Copula function type. Here, FrankCopula is selected, and its function expression is as shown in formula (9).

[0096]

[0097] C F (u, v, λ) is the Copula function, u = F X (x p ), v = F Y (y p ), λ is the parameter of the Copula function, λ∈(-∞,+∞).

[0098] Step 3: Estimate the parameters of the Copula function by using the maximum likelihood estimation (MLE) method and substitute them into formula (9) to obtain the time series data x without occlusion factor: p ,y p The correlation between the marginal distribution functions describes C[F X (x p ), F Y (y p )].

[0099] Step 4: Combine the marginal distribution function F X (x p ) and F Y (y p ), and the estimated Copula function, to construct the overall correlation model. The formula is as follows:

[0100] H(x p ,y p )=C[F X (x p ),F Y (y p )] (10)

[0101] Among them, H(x p ,y p ) is x p ,y p The joint distribution function of is, and C is the Copula function.

[0102] Step 5: Based on the PV output on the target date predicted by the BiLSTM model in step 1, the unobstructed factor of the PV output predicted on the target date of the reference power station is calculated using formula (4), and the weather category of the day is obtained based on the mean and mean absolute deviation of the PV output unobstructed factor on a day.

[0103] There are two reference methods to obtain the weather category of the day. The first method is to find the corresponding sample in which group the mean and mean absolute deviation of the photovoltaic output unobstructed factor of the day fall. and MAE value range, find the cluster and get the corresponding weather category; if the above method cannot be directly obtained, then the Euclidean distance between the mean and mean absolute deviation of the unobstructed factor of the photovoltaic output on that day and the mean and mean absolute deviation of the unobstructed factor of the photovoltaic output corresponding to the cluster center of each cluster is used, and the weather category corresponding to the cluster with the smallest distance is the weather category corresponding to the target date of the reference photovoltaic power station.

[0104] Step 6: According to the weather category of the reference PV power station target date obtained in step 5, substitute the PV output unobstructed factor predicted on the reference PV power station target date into the Copula correlation model of the corresponding weather category in step 4:

[0105] Use the marginal distribution function of the kernel density estimate calculated in step 4 to calculate the unobstructed factor time series data x of the target PV power station on the target date. t The mapping value F on its marginal distribution X (x t ), the unobstructed factor mapping value F of the target PV power station X (x t ) is substituted into the correlation model under the corresponding weather conditions obtained in step 4. Through the correlation model, combined with the marginal distribution function F of the reference photovoltaic power station, Y (y t ), calculate the unobstructed factor of the photovoltaic output at each sampling time point of the target date forecast of the target photovoltaic power station.

[0106] Then, according to formula (11), the actual output of the target photovoltaic power station predicted at each sampling time point on the target date is calculated.

[0107]

[0108] In the above formula, is the actual output predicted at the sampling time point t of the target date for the target PV power station, is the photovoltaic output unobstructed factor at sampling time point t predicted on the target date of the target photovoltaic power station, It is the deterministic part of photovoltaic power generation output and is only related to the latitude and longitude, altitude and time of the target photovoltaic power station.

[0109] In the technical solution of the present invention, the reference photovoltaic power station is a photovoltaic power station with sufficient historical data and weather data, and the target photovoltaic power station is a photovoltaic power station with historical power data. The reference photovoltaic power station and the target photovoltaic power station are located in the same area, that is, the two equal.

[0110] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A photovoltaic output prediction method based on correlation, characterized in that: The method comprises the following steps: Step 1: Use the prediction model to predict the photovoltaic output of the reference photovoltaic power station in the same area as the target photovoltaic power station on the target date, which specifically includes the following steps: Step 1.1: Data preprocessing The historical photovoltaic power data, temperature and irradiance data of the reference photovoltaic power station were collected and preprocessed. The data was cleaned by removing outliers and filling missing values ​​to obtain three kinds of raw data. Then, the three kinds of raw data were normalized by 0-1 standardization. Finally, the complete data were sampled at fixed time intervals, and invalid data were removed to obtain a data set including three kinds of time series data. Step 1.2: Divide the dataset Use a fixed-size sliding window to slice the three time series data in the data set, and combine the three time series data slices at the same sampling time point as a training data; for the obtained training data, use no less than 60% as the training set, no less than 20% as the validation set, and the rest as the test set; Step 1.3: Construction and training of prediction model The prediction model is composed of an input layer, a Bi-LSTM network and an output layer; wherein the input layer receives input data; the Bi-LSTM network combines the bidirectional model and the gating mechanism of LSTM, and is composed of two independent LSTM networks; the forward LSTM network processes the data sequence from the first time step to the last time step, extracts the forward features, and the reverse LSTM network processes the data sequence from the last time step to the first time step, extracts the backward features; then the forward and backward features of a time step are concatenated, and the output vector formed is used as the final output of the time step; the output layer is a fully connected layer, which is used to convert the output of the Bi-LSTM network into a predicted value, that is, the predicted value of photovoltaic power; The time series data slices of temperature and irradiance in a training data are used as the input of the prediction model, and the historical photovoltaic power is used as the reference value of the prediction model output; the network parameters of the prediction model are initialized by random assignment method, the learning rate is set to 0.005, the batch size is 64, and the maximum number of training rounds is 200; RMSE is selected as the loss function, and the Adam optimizer is used; Input a batch of training data in the training set into the prediction model, calculate the RMSE value between the predicted value output by the prediction model for the batch of training data and the reference value, and use the Adam optimizer to reversely update the network parameters of the prediction model once according to the RMSE value to complete the training of one batch; use the network parameters when the last batch training is completed as the initial parameters for the next batch training, and repeat continuously until the last batch of training data in the training set is trained to complete a round of training; after each round of training, input the training data in the validation set into the prediction model in turn, and calculate the training loss value of the prediction model that has completed this round of training for the validation set; use the network parameters when the last round of training is completed as the initial parameters for the next round of training, and repeat continuously until the training loss value of the validation set does not decrease for several consecutive rounds or the training round reaches the maximum value, and the training of the prediction model is completed; Input the training data in the test set into the trained prediction model in sequence, calculate the MAE and MAPE between the predicted value output by the prediction model and the reference value, and when MAE is less than 10 and MAPE is less than 5, the trained prediction model is tested as the optimal prediction model; otherwise, adjust the learning rate, batch size, and maximum number of training rounds, and repeat the training and testing process of the prediction model until the optimal prediction model is obtained; Step 1.4: Reference PV power station's PV output Pr on target date j Prediction The temperature and irradiance data of the target date of the reference photovoltaic power station are first preprocessed according to the method in step 1.1, and then the sliding window of the same size in step 1.2 is used to perform data slicing processing in the order of sampling time points to obtain time series data slices of the two types of data, and the sampling time points in any two time series data slices are not repeated; the time series data slices of the same sampling time points of the above two types of data are sequentially input into the optimal prediction model in step 1.3 to obtain the predicted value output by the prediction model; the predicted value is denormalized to obtain the predicted value of the photovoltaic output at the corresponding sampling time point of the target date of the reference photovoltaic power station; Step 2: Define the unshaded factor of photovoltaic output The actual output P of the photovoltaic panel at sampling time t t It can be calculated using formula (2); Where: P stc is the output of the photovoltaic panel under standard test conditions; α T is the power temperature coefficient of the photovoltaic panel; I t is the actual irradiance at sampling time t; T t is the actual temperature of the photovoltaic panel at sampling time t; Transform equation (2) to separate the random factors in photovoltaic output and define the photovoltaic power station output P without considering the shading and temperature effects: tm , that is, no obstruction to output; Where: I tm It is the maximum value of irradiance that can be achieved without considering any shielding. This value is only related to the latitude and longitude, altitude and time of the area. Therefore, P tm Only the deterministic part of photovoltaic power generation output is included; The unshaded factor of photovoltaic output is defined as: Where ε is the unshaded correction coefficient, which is used to correct the difference between the theoretical shading factor and the actual observed value. It is determined by the empirical estimation method. Based on the statistical analysis of a large amount of historical data, the ratio of the actual photovoltaic output to the theoretical output is directly calculated to obtain the correction coefficient ε. The unshading factor is used to characterize the unshading situation at each sampling time point, that is, the random part of the photovoltaic output; From equation (2) to equation (4), we can know that: From formula (3) and formula (5), we can see that P tm The maximum value of photovoltaic output that can be achieved in theory is determined, while the actual photovoltaic output will be weakened by various random factors. t , describes the weakening effect of shadow, cloud cover, weather changes and temperature on photovoltaic output based on the theoretical maximum photovoltaic output. Its physical meaning is the ratio of the actual output of the photovoltaic power station to the deterministic output, so its value must not be greater than 1; Step 3: Classify weather types based on historical data of reference PV power plants Obtain the historical data of the reference photovoltaic power station in step 1, which includes the actual irradiance I at the sampling time point t t 、The power temperature coefficient of photovoltaic panels α T , the actual temperature T of the photovoltaic panel at sampling time t t , combined with the standard irradiance and standard temperature, according to formula (5), the unobstructed factor at each sampling time point is obtained; then the mean X and mean absolute deviation MAE of the unobstructed factor of each day are calculated; m is the total number of sampling points per day, x j is the value of the unobstructed factor at sampling time point j on a certain day; The mean value X of the unobstructed factor is used to represent the degree of clearness of the day, and the mean absolute deviation MAE is used to represent the severity of weather fluctuations on that day. The weather is divided into three categories: sunny days with high clearness and no drastic fluctuations, cloudy days with low clearness and no drastic fluctuations, and bad weather with unstable clearness and drastic fluctuations. The average of the calculated unobstructed factors for each day and mean absolute deviation MAE are normalized respectively, and the weather characteristic vector of the hth day is T h , T h =(x h , m h ), where x h is the mean value of the unobstructed factor on the hth day The normalized value of m h is the normalized value of the mean absolute deviation MAE of the unobstructed factor on the hth day; according to the weather feature vector of each day, the K-means clustering algorithm is used to cluster with the Euclidean distance as the similarity index, and its objective function is shown in formula (8); Where: K is the number of cluster categories, K = 3, namely sunny, cloudy, and bad weather; N is the total number of days of historical data, N k is the number of samples in the k-th cluster; x k 、m k is the eigenvalue of the weather eigenvector of the cluster center of the kth class; x ik 、m ik is the eigenvalue of the weather feature vector of the i-th sample of the k-th class; clustering takes the minimum S as the iteration goal; the K clusters obtained by the last clustering are the required K weather category data groups, and the weather category corresponding to the cluster is determined according to the X and MAE values ​​corresponding to the samples in each cluster; Step 4: Obtain the same historical data of the target photovoltaic power station in the same time period as the reference photovoltaic power station in step 3, and calculate the unshading factor of each sampling time point of each day in the corresponding time period of the target photovoltaic power station; and divide the unshading factor of each sampling time point of each day in the corresponding time period of the target photovoltaic power station into K data sets according to the dates corresponding to the K weather category data groups of the reference photovoltaic power station divided in step 3 according to the same date structure; Under each weather category, use the Copula function to establish a correlation model between the unshading factor of each sampling time point of each day in the corresponding time of the target photovoltaic power station and the unshading factor of each sampling time point of each day in the corresponding time of the reference photovoltaic power station; Step 5: Based on the PV output on the target date predicted by the prediction model in step 1, the unshading factor predicted on the target date of the reference PV power station is calculated using formula (4), and the weather category of the day is obtained based on the mean and mean absolute deviation of the unshading factor of the day; Step 6: Based on the weather category of the target date of the reference photovoltaic power station obtained in step 5 and the unobstructed factor predicted on the target date of the reference photovoltaic power station, combined with the correlation model of the corresponding weather category in step 4, calculate the unobstructed factor of each sampling time point predicted on the target date of the target photovoltaic power station; then, according to formula (11), calculate the actual output of the target photovoltaic power station predicted on each sampling time point on the target date; In the above formula, is the actual output predicted at the sampling time point t of the target date for the target PV power station, is the unobstructed factor of the sampling time point t predicted on the target date of the target PV power station, It is the deterministic part of photovoltaic power generation output and is only related to the latitude and longitude, altitude and time of the target photovoltaic power station.

2. A photovoltaic output prediction method based on correlation according to claim 1, characterized in that: In step 1, the method for denormalizing the predicted value is: P rj =Z rj ×σ P +μ P (1) Among them, P rj is the predicted value of the photovoltaic output at sampling time point j on the target date of the reference photovoltaic power station, Z rj is the predicted value at sampling time point j output by the prediction model, σ P is the standard deviation of the original data of historical PV power data in step 1.1, μ P is the mean value of the original data of historical PV power data in step 1.

1.

3. The photovoltaic output prediction method based on correlation according to claim 1, characterized in that: In step 3, the specific operation process of the K-means clustering algorithm is as follows: Step 1: Randomly select K samples from N weather feature vectors as the initial cluster centers; Step 2: Calculate the Euclidean distance d between each sample and all K initial cluster centers k , Assign each sample to the initial cluster center closest to it to complete a clustering; Step 3: Each initial cluster center and the samples assigned to them constitute a group; for each cluster center, calculate the average value of the samples assigned to it, and use the average value as the first cluster center of each group; Step 4: Repeat steps 2 and 3, calculate the Euclidean distance between each sample and the updated K cluster centers, and reallocate each sample to the cluster center closest to it; then update the cluster centers to the average value of the samples assigned to them; repeat steps 2 and 3 until the Euclidean distance between the previous and next changes of each cluster center is no more than 10 -6 , clustering ends; the K clusters obtained by the last clustering are the required K weather category data groups. According to the X and MAE values ​​corresponding to the samples in each cluster, the weather category corresponding to the cluster is determined.

4. The photovoltaic output prediction method based on correlation according to claim 1, characterized in that: In step 4, the specific implementation method of establishing the correlation model between the target unobstructed factor and the reference unobstructed factor under a weather category is as follows: Step 1: Use the kernel density estimation method to estimate the marginal distribution function of the target unobstructed factor dataset X and the reference unobstructed factor dataset Y of the same weather category; obtain the marginal distribution function F of each sample through the kernel density estimation method. X (x p ) and F Y (y p ), used as the input of the Copula function; a sample of the target unobstructed factor dataset X is the unobstructed factor time series data x of one day p , where the unobstructed factor at sampling time point j is expressed as A sample of the reference unobstructed factor dataset Y is the unobstructed factor time series data y for one day. p , where the unobstructed factor at sampling time point j is expressed as Step 2: According to the distribution characteristics of the unobstructed factor, select a suitable Copula function type. Here, FrankCopula is selected, and its function expression is as shown in formula (9); C F (u, v, λ) is the Copula function, u = F X (x p ), v = F Y (y p ), λ is the parameter of the Copula function, λ∈(-∞,+∞); Step 3: Estimate the parameters of the Copula function by the maximum likelihood estimation method and substitute them into formula (9) to obtain the time series data x without occlusion factor: p ,y p The correlation between the marginal distribution functions describes C[F X (x p ), F Y (y p )]; Step 4: Combine the marginal distribution function F X (x p ) and F Y (y p ), and the estimated Copula function, to construct the overall correlation model; the formula is as follows: H(x p ,y p )=C[F X (x p ),F Y (y p )] (10) Among them, H(x p ,y p ) is x p ,y p The joint distribution function of , C is the Copula function.

5. The photovoltaic output prediction method based on correlation according to claim 1, characterized in that: In step 5, the method for obtaining the weather category to which the day belongs is to first find the group class according to the X and MAE value range corresponding to the samples in which the mean and mean absolute deviation of the unobstructed factor of the day fall, and obtain the corresponding weather category; if the above method cannot be obtained directly, the Euclidean distance between the mean and mean absolute deviation of the unobstructed factor of the day and the mean and mean absolute deviation of the unobstructed factor corresponding to the cluster center of each group class is used, and the weather category corresponding to the group class with the smallest distance is the weather category corresponding to the target date of the reference PV power station.

Citation Information

Cited By

  • Photovoltaic simulation output generation method and system, storage medium and electronic equipment

    CN119944669A

  • A method, system, storage medium, and electronic device for generating photovoltaic simulated output

    CN119944669B

  • Photovoltaic output coefficient prediction method and related device

    CN120914761A

  • Medium and long term photovoltaic output compensation calibration and prediction method based on error analysis

    CN121529842A

  • A Medium- and Long-Term Photovoltaic Output Compensation Calibration and Prediction Method Based on Error Analysis

    CN121529842B