A short-term photovoltaic power prediction method and system for real-time control

Abstract

The present invention belongs to the technical field of photoelectric prediction and evaluation, and provides a short-term photovoltaic power prediction method and system for real-time control, comprising: performing Euclidean distance calculation on a meteorological and power comprehensive set and a meteorological condition data set to be predicted, screening the meteorological and power comprehensive set based on the degree of data deviation, dividing the screened meteorological and power comprehensive set into a first analysis set and a second analysis set, adding limiting conditions according to the value distribution of the data and in combination with a chi-square test method, screening multiple LSTM models through verification tests of the first analysis set and the second analysis set, and calculating the verification performance value of the model according to the verification performance of the LSTM model after screening, combining the prediction results of the multiple LSTM models and the verification performance of the model, and comprehensively outputting predicted power generation data, thereby improving the comprehensiveness and comprehensiveness of the power generation prediction on the basis of accurate prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of photoelectric prediction, and in particular relates to a short-term photovoltaic power prediction method and system for real-time control. Background Art

[0002] Due to the intermittent and fluctuating nature of photovoltaic power generation, if a large number of photovoltaic power sources are connected to the grid without accurate power forecasting, this can lead to problems such as grid voltage fluctuations and frequency instability. By using short-term photovoltaic power forecasting, grid operators can take measures in advance, such as adjusting reactive power compensation equipment and optimizing grid operation modes, effectively mitigating the impact of photovoltaic power fluctuations on the grid and improving grid stability and reliability. Therefore, it is of great significance to study a short-term photovoltaic power forecasting method and system for real-time control. Existing technologies fail to fully consider the deviation between the meteorological and power integrated set and the dataset of the meteorological conditions to be predicted, resulting in a poor match between the dataset used for model training and the actual conditions to be predicted. Furthermore, existing technologies fail to fully consider the distribution of data values ​​and lack a scientific method for adding limiting conditions, which affects the performance and adaptability of model training and verification. More importantly, existing technologies may rely on a single model for power generation forecasting, which lacks comprehensiveness and is difficult to cope with complex and changing real-world situations.

[0003] To this end, the present invention provides a short-term photovoltaic power prediction method and system for real-time control. Summary of the Invention

[0004] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0005] The technical solution adopted by the present invention to solve its technical problem is:

[0006] In a first aspect, the present invention provides a short-term photovoltaic power prediction method for real-time control, comprising the following steps: S1: obtaining historical time-sharing meteorological condition data and corresponding power generation data of the photovoltaic power station through real-time monitoring of meteorological monitoring equipment and a photoelectric monitoring system within the photovoltaic power station, integrating the meteorological and power comprehensive sets of different historical time-sharing periods, calculating the Euclidean distance between the meteorological and power comprehensive sets of different historical time-sharing periods and the meteorological condition data set to be predicted, and screening a power prediction similarity set based on the calculation results;

[0007] S2: dividing the power prediction similarity set into a first analysis set and a second analysis set, adding a first limiting condition for the division between the first analysis set and the second analysis set based on the value distribution of the data in the divided first analysis set and the second analysis set; if the first condition is met, constructing a first data histogram and a second data histogram based on the first analysis set and the second analysis set, and adding a second limiting condition for the division between the first analysis set and the second analysis set in combination with a chi-square test method;

[0008] S3: When the first analysis set and the second analysis set meet the added restriction conditions, the LSTM model is trained and verified using the first analysis set and the second analysis set, and the LSTM model is screened according to the training and verification results to obtain a photovoltaic power prediction model;

[0009] S4: The meteorological condition dataset to be predicted is used as the input of multiple photovoltaic power prediction models. Through the prediction of multiple photovoltaic power prediction models and combined with the verification performance of the photovoltaic power prediction models, the photovoltaic power generation power is predicted comprehensively and accurately.

[0010] In a second aspect, the present invention provides a short-term photovoltaic power prediction system for real-time control, comprising the following modules: a data screening module: obtaining historical meteorological condition data and corresponding power generation data of the photovoltaic power station through real-time monitoring of meteorological monitoring equipment and a photoelectric monitoring system in the photovoltaic power station, integrating the meteorological and power comprehensive sets of different historical time periods, calculating the Euclidean distance between the meteorological and power comprehensive sets of different historical time periods and the meteorological condition data set to be predicted, and screening the power prediction similarity set based on the calculation results;

[0011] A partitioning and limiting module is configured to partition the power prediction similarity set into a first analysis set and a second analysis set, and to add a first limiting condition to the partition between the first analysis set and the second analysis set based on the value distribution of the data in the divided first analysis set and the second analysis set. If the first condition is met, a first data histogram and a second data histogram are constructed based on the first analysis set and the second analysis set, and a second limiting condition is added to the partition between the first analysis set and the second analysis set in combination with a chi-square test method.

[0012] Model testing module: When the first analysis set and the second analysis set meet the added constraints, the LSTM model is trained and verified using the first analysis set and the second analysis set. The LSTM model is screened based on the training and verification results to obtain a photovoltaic power prediction model.

[0013] Comprehensive prediction module: The meteorological condition dataset to be predicted is used as the input of multiple photovoltaic power prediction models. Through the prediction of multiple photovoltaic power prediction models and combined with the verification performance of the photovoltaic power prediction models, the photovoltaic power generation power is comprehensively and accurately predicted.

[0014] The beneficial effects of the present invention are as follows: by calculating the Euclidean distance between the meteorological and power comprehensive set and the meteorological condition data set to be predicted, the meteorological and power comprehensive set is screened based on the degree of data deviation, and the consistency of the data set used in subsequent model training is improved, which is conducive to improving the accuracy of subsequent power generation prediction. When the screened meteorological and power comprehensive set is divided into the first analysis set and the second analysis set, the limiting conditions are added according to the value distribution of the data and combined with the chi-square test method, so as to improve the balance of data between the first analysis set and the second analysis set, improve the performance and adaptability of subsequent model training and verification, and further improve the accuracy of power generation prediction. Most importantly, multiple LSTM models are screened through verification tests of the first analysis set and the second analysis set, and after screening, the verification performance value of the model is calculated according to the verification performance of the LSTM model. Finally, the prediction results of multiple LSTM models and the verification performance of the model are combined to output the predicted power generation data. Compared with the prediction of power generation data achieved by a single model, the comprehensiveness and comprehensiveness of power generation prediction are improved on the basis of accurate prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will be further described below with reference to the accompanying drawings.

[0016] Figure 1 is a flowchart of the steps of a short-term photovoltaic power prediction method for real-time control according to an embodiment of the present invention;

[0017] Figure 2 This is a flowchart of a short-term photovoltaic power prediction system for real-time control according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0019] Example 1

[0020] See also Figure 1 As shown, a short-term photovoltaic power prediction method for real-time control according to an embodiment of the present invention includes the following steps:

[0021] S1: Through real-time monitoring by meteorological monitoring equipment and photovoltaic monitoring systems within the photovoltaic power station, obtain the historical meteorological condition data and corresponding power generation data of the photovoltaic power station, and integrate them to obtain comprehensive sets of meteorological and power data at different historical times. Calculate the Euclidean distance between the comprehensive sets of meteorological and power data at different historical times and the meteorological condition data set to be predicted, and select similar sets for power prediction based on the calculated results;

[0022] Specifically, in S1, the method of obtaining the historical time-sharing meteorological condition data and the corresponding power generation data of the photovoltaic power station through real-time monitoring of the meteorological monitoring equipment and the photoelectric monitoring system in the photovoltaic power station is as follows:

[0023] Meteorological monitoring equipment: A dedicated meteorological monitoring station is installed in the photovoltaic power station to collect real-time meteorological data, including various meteorological parameters such as light intensity, temperature, humidity, wind speed, wind direction, and air pressure. For example, a light radiation sensor is used to measure light intensity, a temperature sensor is used to obtain ambient temperature, and an anemometer is used to measure wind speed. The equipment will record data at set time intervals (such as every minute, every 5 minutes, etc.) to form historical time-sharing meteorological data;

[0024] Photovoltaic power station monitoring system: Photovoltaic power stations are equipped with a monitoring system that can monitor and record the operating data of photovoltaic modules, inverters and other equipment in real time, including power generation data. The system can collect information such as the output power and total power generation of each inverter, and store it in chronological order to form historical power generation data, providing support for more refined data analysis;

[0025] The integration of meteorological and power comprehensive sets of different historical time periods in S1 is divided according to the time series of meteorological condition data and corresponding power generation data. Specifically, the meteorological condition data and power generation data of the same historical time period are integrated into a meteorological and power comprehensive set;

[0026] In S1, the Euclidean distance between the comprehensive set of meteorological and power data at different historical time periods and the meteorological condition dataset to be predicted is calculated separately, specifically including:

[0027] Euclidean distance calculates the straight-line distance between two points in multidimensional space. For a meteorological condition dataset, each meteorological characteristic (such as temperature, humidity, wind speed, and light intensity) can be considered a dimension in the multidimensional space. By calculating the Euclidean distance between the comprehensive meteorological and power dataset and the meteorological condition dataset to be predicted, the degree of deviation between the two can be quantified using a specific numerical value. For example, when the Euclidean distance is small, it indicates that the meteorological conditions to be predicted and some samples in the meteorological and power comprehensive set are close in position in the multidimensional meteorological feature space, that is, the meteorological conditions of the two are relatively similar; conversely, it indicates that the meteorological conditions are significantly different, which makes it possible to more intuitively understand the degree of difference between different meteorological condition combinations, and helps to quickly judge the similarity between the meteorological conditions to be predicted and historical samples. For the meteorological and power comprehensive sets and the meteorological condition data sets to be predicted at different historical time points, n dimensions are obtained according to the number of different meteorological condition data in the meteorological and power comprehensive sets, where each dimension corresponds to one meteorological condition data. The obtained meteorological and power comprehensive set is: (qx1, qx2, qx3...qxn), and the obtained meteorological condition data set to be predicted is (dqx1, dqx2, dqx3...dqxn); where qxn represents the meteorological condition data under the nth dimension, and dqxn represents the meteorological condition data to be predicted under the nth dimension;

[0028] For two n-dimensional data sets, the Euclidean distance dj between the meteorological and power comprehensive set and the meteorological condition data set to be predicted is calculated. The specific calculation formula is:

[0029]

[0030] Among them, qxi represents the i-th meteorological condition data in the meteorological and power comprehensive set, and dqxi represents the i-th meteorological condition data in the meteorological condition data set to be predicted;

[0031] The method of screening the power prediction similarity set in S1 specifically includes:

[0032] Meteorological data usually contains multiple types of features, such as numerical temperature, humidity, and quantified wind direction. The Euclidean distance method has good compatibility with these different types of numerical features. The Euclidean distance is used to calculate the deviation, and there is no need to use different distance measurement methods for different types of data. The Euclidean distance dj between the calculated meteorological and power comprehensive set and the meteorological condition dataset to be predicted is compared with the set Euclidean distance threshold djy. If dj < djy, it means that the similarity between the meteorological and power comprehensive set and the meteorological condition dataset to be predicted is high, and the meteorological and power comprehensive set is recorded as the power prediction similarity set. Otherwise, it means that the similarity between the meteorological and power comprehensive set and the meteorological condition dataset to be predicted is low, and no operation is performed.

[0033] It should be noted that the purpose of screening similar power prediction sets in S1 is that similar data samples can better reflect the actual operation of the photovoltaic system under the meteorological conditions to be predicted. By screening data samples similar to the meteorological conditions to be predicted for training and verification, the model can learn the changing patterns of photovoltaic power under specific meteorological conditions, thereby more accurately predicting future photovoltaic power output. In addition, the original meteorological and power comprehensive set may contain data under various meteorological conditions. Screening data samples with high similarity can eliminate irrelevant information, allowing the model to focus on learning features and patterns related to the scenario to be predicted, thereby improving the accuracy of the prediction.

[0034] S2: dividing the power prediction similarity set into a first analysis set and a second analysis set, adding a first limiting condition for the division of the first analysis set and the second analysis set based on the value distribution of the data in the divided first analysis set and the second analysis set; if the first condition is met, constructing a first data histogram and a second data histogram based on the first analysis set and the second analysis set, and adding a second limiting condition for the division of the first analysis set and the second analysis set in combination with a chi-square test method;

[0035] In S2, 70% of the power prediction similarity sets are divided into the first analysis set, and 30% of the power prediction similarity sets are divided into the second analysis set;

[0036] In S2, the first limiting condition is added to the partition between the first analysis set and the second analysis set based on the value distribution of the data in the partitioned first analysis set and the second analysis set. Specifically, the first limiting condition is added as follows:

[0037] For the same data in the first analysis set and the second analysis set, obtain the value range of the data in the first analysis set and record it as [cqmin-cqmax]; obtain the value range of the data in the second analysis set and record it as [yqmin-yqmax], where min is the minimum value and max is the maximum value;

[0038] If the distribution value ranges of the same data in the first analysis set and the second analysis set differ significantly, for example, the light intensity in the first analysis set is mainly between 0-1000 W / m², while in the second analysis set it is between 500-1500 W / m², then the model is likely to have not learned the pattern of light intensity values ​​in the second analysis set during training, resulting in poor performance on the second analysis set. Based on the difference in the distribution value ranges of the same data in the first and second analysis sets, a first limiting condition is added to the division between the first and second analysis sets, specifically including:

[0039] The distribution value range difference CY of the same data in the first analysis set and the second analysis set is calculated by performing difference processing on the corresponding value endpoints. The specific calculation formula is:

[0040]

[0041] Whether the value distribution of the same data in the first analysis set and the second analysis set is reasonable is judged by calculating the distribution value range difference CY. The distribution value range difference CY is calculated, and a distribution value range difference threshold CYy is set. If CY≤CYy, the value range distribution of the same data in the first analysis set and the second analysis set is considered reasonable. If CY>CYy, the value range distribution of the same data in the first analysis set and the second analysis set is considered unreasonable.

[0042] The distribution value range difference CY is for the same data in the first analysis set and the second analysis set. The first limiting condition added to the division of the first analysis set and the second analysis set is for the data as a whole in the first analysis set and the second analysis set. Therefore, it is necessary to calculate a value that can represent the difference in the distribution value range of the data as a whole in the first analysis set and the second analysis set, that is, the overall distribution difference value ZC. The first limiting condition is added based on the overall distribution difference value ZC. The method for obtaining the overall distribution difference value specifically includes:

[0043] The data in the first analysis set and the second analysis set are identified based on the comparison result of the distribution value range difference CY corresponding to the same data in the first analysis set and the second analysis set with the distribution value range difference threshold CYy. Specifically, if CY≤CYy, the same data in the first analysis set and the second analysis set are recorded as same-range data. Conversely, if CY>CYy, the same data in the first analysis set and the second analysis set are recorded as different-range data. The specific calculation method of the overall distribution difference value ZC is:

[0044] Count the number of data in different ranges and compare it with the number of data in the same range to get the proportion of data in different ranges.

[0045] Based on the non-same range data, the distribution value range difference CY corresponding to the non-same range data is obtained, and the deviation is calculated with the distribution value range difference threshold CYy. After the calculated deviation is averaged, the ratio is calculated with the distribution value range difference threshold CYy to obtain the distribution value range deviation degree of the non-same range data;

[0046] The overall distribution difference value ZC is calculated by summing the proportion of the number of data in different ranges and the degree of deviation of the distribution value range of data in different ranges;

[0047] The first limiting condition is added by the overall distribution difference value ZC, specifically: the overall distribution difference value ZC ≤ the overall distribution difference threshold;

[0048] The method of constructing the first data histogram and the second data histogram according to the first analysis set and the second analysis set in S2 specifically includes:

[0049] For the data with the same range in the first analysis set and the second analysis set, the value ranges of the data with the same range in the first analysis set and the second analysis set are integrated to obtain the target construction range, specifically:

[0050] Select the value range [cqmin-cqmax] of the data in the first analysis set and the value range [yqmin-yqmax] of the data in the second analysis set. Based on the comparison between the minimum value cqmin of the value range of the first analysis set and the minimum value yqmin of the value range in the second analysis set, select the smaller one as the minimum value of the target construction range. Similarly, based on the comparison between the maximum value cqmax of the value range of the first analysis set and the maximum value yqmax of the value range in the second analysis set, select the larger one as the maximum value of the target construction range.

[0051] Before constructing the first data histogram and the second data histogram, the target construction range is obtained. The purpose is to make the first data histogram and the second data histogram have the same horizontal distribution range (horizontal axis range), which is conducive to providing an accurate basis for the data support used in the subsequent chi-square test calculation;

[0052] The target construction range is taken as the X-axis, and the target construction range is divided into k sub-intervals of equal length; the frequency (number of samples) of the same range data in the first analysis set and in each sub-interval is taken as the Y-axis to construct the first data histogram. Similarly, the second data histogram is constructed. The difference between the first data histogram and the second data histogram is that the same range data corresponding to the first data histogram are in the first analysis set, and the same range data corresponding to the second data histogram are in the second analysis set.

[0053] In S2, the chi-square test method is combined to add a second limiting condition for the division of the first analysis set and the second analysis set, specifically including:

[0054] Extract the frequencies corresponding to the sub-interval ranges in the first data histogram and the second data histogram, and calculate the expected frequency Ei corresponding to each sub-interval range. Specifically, the expected frequency Ei calculation formula is:

[0055]

[0056] Among them, Ri represents the sum of the i-th row, Ci represents the sum of the i-th column, and N is the total number of samples;

[0057] The calculation of the expected frequency Ei is illustrated as shown in Table 1. Assume that the frequencies of the data in the same range in the first analysis set A in the five intervals [0,20), [20,40), [40,60), [60,80), and [80,100] are 15, 20, 30, 25, and 10 respectively; and in the second analysis set B in the five intervals [0,20), [20,40), [40,60), [60,80), and [80,100] are 12, 22, 28, 26, and 12 respectively;

[0058] Table 1: Frequency corresponding to subinterval range;

[0059] The total number of samples N is 15+20+30+25+10+12+22+28+26+12=190;

[0060] For the interval [0,20), the expected frequency ;

[0061] After the expected frequency calculation is completed, the chi-square test method is used to calculate the chi-square statistic KZ of the first analysis set and the second analysis set. The specific calculation formula is:

[0062]

[0063] Among them, xi 2 represents the chi-square statistic calculated for the i-th data in the same range; k represents the number of sub-intervals; Qi is the frequency of the i-th sub-interval, and Ei is the expected frequency corresponding to the i-th sub-interval;

[0064] Adding a second limiting condition through the chi-square statistic KZ of the first analysis set and the second analysis set, specifically: chi-square statistic KZ ≤ chi-square statistic threshold;

[0065] S3, when the first analysis set and the second analysis set meet the added restriction conditions, the LSTM model (recurrent neural network machine learning model) is trained and verified using the first analysis set and the second analysis set, and the LSTM model is screened according to the training and verification results to obtain a photovoltaic power prediction model;

[0066] Among them, LSTM models include but are not limited to supervised learning models, unsupervised learning models, semi-supervised learning models, and reinforcement learning models;

[0067] In S3, the first analysis set and the second analysis set are used to train and verify the LSTM model. The LSTM model is screened according to the training and verification results to obtain a photovoltaic power prediction model. Specifically, the method includes:

[0068] Use the first analysis set to train the LSTM model, obtain the trained LSTM model, and use the second analysis set to verify the LSTM model. Specifically:

[0069] The meteorological condition data included in the second analysis set is used as input, and the trained LSTM model outputs the power generation data. The power generation data output by the LSTM model is analyzed for deviation from the power generation data corresponding to the input meteorological condition data to obtain an output deviation value sc;

[0070] For example, the power generation data corresponding to the meteorological condition data is recorded as (fd1, fd2, fd3...fdz), where z represents the data volume of the power generation data; the power generation data output by the LSTM model is recorded as (sfd1, sfd2, sfd3...sfdz), where z represents the data volume of the power generation data;

[0071] The specific calculation formula for the output deviation value sc is:

[0072]

[0073] Among them, sfdi represents the i-th power generation data in the power generation data output by the LSTM model, and fdi represents the i-th power generation data in the power generation data corresponding to the meteorological condition data;

[0074] Compare the output deviation value sc with the output deviation threshold, and determine whether the input meteorological condition data has passed verification based on the comparison result. Specifically:

[0075] If the output deviation value sc ≤ the output deviation threshold, the input meteorological condition data is considered to have passed the verification;

[0076] If the output deviation value sc> the output deviation threshold, it is considered that the input meteorological condition data verification fails;

[0077] Counting the number of meteorological condition data that have passed the verification, and performing a ratio process on the number of meteorological condition data included in the second analysis set to obtain a verification pass value;

[0078] It should be noted that the validation pass value reflects the ratio of the amount of data in the second analysis set that the LSTM model passed validation to the total amount of data in the second analysis set. By understanding the passing degree of the LSTM model in the second analysis set validation, the LSTM model can be screened, which is beneficial to the applicability of subsequent photovoltaic power generation model applications, thereby improving the accuracy of photovoltaic power generation prediction.

[0079] comparing the verification pass value with the verification pass threshold;

[0080] If the verification pass value is greater than or equal to the verification pass threshold, the LSTM model is considered to have passed the verification of the second analysis set, and the LSTM model is marked as a photovoltaic power prediction model;

[0081] If the verification pass value is less than the verification pass threshold, the LSTM model is considered to have failed the verification of the second analysis set, and no operation is performed;

[0082] S4, using the meteorological condition dataset to be predicted as input to multiple photovoltaic power prediction models, and comprehensively and accurately predicting photovoltaic power generation through the predictions of multiple photovoltaic power prediction models and combining the verification performance of the photovoltaic power prediction models;

[0083] Specifically, the meteorological condition dataset to be predicted is used as input, and multiple predicted power generation data are obtained through the output of multiple photovoltaic power prediction models. These data are then integrated to obtain a predicted power generation data group (yfd1, yfd2, yfd3, ..., yfdb), where yfdb represents the predicted power generation data corresponding to the b-th photovoltaic power prediction model, and b represents the number of photovoltaic power prediction models.

[0084] The verification pass values ​​corresponding to each photovoltaic power prediction model are integrated into a verification pass data group (yt1, yt2, yt3...ytb), where ytb represents the verification pass value corresponding to the b-th photovoltaic power prediction model, and b represents the number of photovoltaic power prediction models;

[0085] In order to comprehensively and accurately predict photovoltaic power generation, the verification performance coefficient YBXi of the photovoltaic power prediction model is calculated based on the performance of the photovoltaic power prediction model. The specific calculation formula is:

[0086]

[0087] Where YBXi represents the verification performance coefficient of the i-th photovoltaic power prediction model, and yti represents the verification pass value of the i-th photovoltaic power prediction model;

[0088] The final photovoltaic power generation power prediction result ZYF is:

[0089]

[0090] Among them, yfdi represents the predicted power generation data corresponding to the i-th photovoltaic power prediction model;

[0091] The technical solution of the above embodiment is: by calculating the Euclidean distance between the meteorological and power comprehensive set and the meteorological condition data set to be predicted, the meteorological and power comprehensive set is screened based on the degree of data deviation, thereby improving the consistency of the data set used in subsequent model training, thereby facilitating the improvement of the accuracy of subsequent power generation prediction. When dividing the screened meteorological and power comprehensive set into a first analysis set and a second analysis set, limiting conditions are added based on the data value distribution and combined with the chi-square test method to improve the balance of data between the first analysis set and the second analysis set, thereby improving the performance and adaptability of subsequent model training and verification, thereby further improving the accuracy of power generation prediction. Most importantly, multiple LSTM models are screened through verification tests of the first analysis set and the second analysis set, and after screening, the verification performance value of the model is calculated based on the verification performance of the LSTM model. Finally, the prediction results of multiple LSTM models and the verification performance of the model are combined to output the predicted power generation data. Compared with the prediction of power generation data achieved by a single model, the comprehensiveness and comprehensiveness of power generation prediction are improved on the basis of accurate prediction.

[0092] Example 2

[0093] like Figure 2 As shown, a short-term photovoltaic power prediction system for real-time control according to an embodiment of the present invention includes the following modules:

[0094] Data screening module: Through real-time monitoring of the meteorological monitoring equipment and photovoltaic monitoring system within the photovoltaic power station, the historical meteorological condition data and corresponding power generation data of the photovoltaic power station are obtained. The meteorological and power comprehensive sets of different historical time periods are integrated to obtain the Euclidean distance between the meteorological and power comprehensive sets of different historical time periods and the meteorological condition data set to be predicted. The similar power prediction sets are screened based on the calculated results.

[0095] A partitioning and limiting module is configured to partition the power prediction similarity set into a first analysis set and a second analysis set, and to add a first limiting condition to the partition between the first analysis set and the second analysis set based on the value distribution of the data in the divided first analysis set and the second analysis set. If the first condition is met, a first data histogram and a second data histogram are constructed based on the first analysis set and the second analysis set, and a second limiting condition is added to the partition between the first analysis set and the second analysis set in combination with a chi-square test method.

[0096] Model testing module: When the first analysis set and the second analysis set meet the added constraints, the LSTM model is trained and verified using the first analysis set and the second analysis set. The LSTM model is screened based on the training and verification results to obtain a photovoltaic power prediction model.

[0097] Comprehensive prediction module: The meteorological condition dataset to be predicted is used as the input of multiple photovoltaic power prediction models. Through the prediction of multiple photovoltaic power prediction models and combined with the verification performance of the photovoltaic power prediction models, the photovoltaic power generation power is comprehensively and accurately predicted.

[0098] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A short-term photovoltaic power prediction method for real-time control, characterized by: include: The historical meteorological condition data of the photovoltaic power station and the corresponding power generation data are integrated to obtain a comprehensive set of meteorological and power data; The deviation analysis is carried out with the meteorological condition data set to be predicted, and a similar power forecast set is obtained by screening; The power prediction similarity set includes a first analysis set and a second analysis set; Using the overall distribution difference value as a first limiting condition for dividing the first analysis set into the second analysis set, wherein the overall distribution difference value is calculated by the data distribution difference between the first analysis set and the second analysis set; Using a chi-square statistic as a second limiting condition for dividing the first analysis set and the second analysis set, wherein the chi-square statistic is calculated by performing a chi-square test on a data histogram constructed from the first analysis set and the second analysis set; Using the first analysis set and the second analysis set to train and verify multiple LSTM models, the multiple LSTM models are screened to obtain a photovoltaic power prediction model; The meteorological condition dataset to be predicted is used as the input of multiple photovoltaic power prediction models. The photovoltaic power generation power of the photovoltaic power station is predicted through the prediction of multiple photovoltaic power prediction models and combined with the training performance of the photovoltaic power prediction models.

2. A short-term photovoltaic power prediction method for real-time control according to claim 1, characterized in that: The power prediction similarity set is obtained in the following manner: The historical meteorological condition data and power generation data of photovoltaic power stations are obtained, and the meteorological and power comprehensive sets of different historical time periods are integrated. The Euclidean distance between the meteorological and power comprehensive sets and the meteorological condition data set to be predicted is calculated respectively, and the power prediction similarity set is selected based on the calculation results.

3. The method for short-term photovoltaic power prediction for real-time control according to claim 1, characterized in that: The first limiting condition is added as follows: Obtaining the same data in the first analysis set and the second analysis set, and extracting the value ranges of the data in the first analysis set and the second analysis set respectively; Perform difference processing on the distribution value ranges of the same data in the first analysis set and the second analysis set on the corresponding value endpoints, and calculate the difference in the distribution value ranges; Identify the same data in the first analysis set and the second analysis set based on the difference in distribution value ranges to obtain data with different ranges; Based on the data in different ranges, the overall distribution difference value is calculated; The overall distribution difference value ≤ the overall distribution difference threshold is used as the first limiting condition.

4. The method for short-term photovoltaic power prediction for real-time control according to claim 3, characterized in that: The method for identifying the same data in the first analysis set and the second analysis set is: when the distribution value range difference is greater than the distribution value range difference threshold, the corresponding same data in the first analysis set and the second analysis set are recorded as non-same range data; conversely, the corresponding same data in the first analysis set and the second analysis set are recorded as same range data.

5. The method for short-term photovoltaic power prediction for real-time control according to claim 3, characterized in that: The overall distribution difference value is calculated as follows: Count the number of data in different ranges and compare it with the number of data in the same range to get the proportion of data in different ranges. Based on the non-same range data, the distribution value range difference corresponding to the non-same range data is obtained, and the deviation is calculated with the distribution value range difference threshold. After the calculated deviation is averaged, the ratio is calculated with the distribution value range difference threshold to obtain the distribution value range deviation degree of the non-same range data; The overall distribution difference value is calculated by summing the proportion of the number of data in different ranges and the degree of deviation of the distribution value range of data in different ranges.

6. The method for short-term photovoltaic power prediction for real-time control according to claim 4, characterized in that: The data histogram includes a first data histogram and a second data histogram; The method of constructing the first data histogram and the second data histogram specifically includes: Obtain the data with the same range in the first analysis set and the second analysis set, integrate the value ranges of the data with the same range in the first analysis set and the second analysis set to obtain the target construction range, specifically: The minimum value of the value range of the first analysis set and the minimum value of the value range of the second analysis set are selected, whichever is smaller as the minimum value of the target construction range; the maximum value of the value range of the first analysis set and the maximum value of the value range of the second analysis set are selected, whichever is larger as the maximum value of the target construction range; The target construction range is taken as the X-axis, and the target construction range is divided into k sub-intervals of equal length; the frequency of the same range data in the first analysis set and in each sub-interval is taken as the Y-axis to construct the first data histogram. Similarly, the second data histogram is constructed.

7. The method for short-term photovoltaic power prediction for real-time control according to claim 1, characterized in that: The second limiting condition is added as follows: Extract the frequencies corresponding to the sub-interval ranges in the first data histogram and the second data histogram, and calculate the expected frequencies corresponding to the sub-interval ranges using the expected frequency calculation formula; After the expected frequency calculation is completed, the chi-square test method is used to calculate the chi-square statistic KZ of the first analysis set and the second analysis set. The specific calculation formula is: ； Among them, xi 2 represents the chi-square statistic calculated for the i-th data in the same range; k represents the number of sub-intervals; Qi is the frequency of the i-th sub-interval, and Ei is the expected frequency corresponding to the i-th sub-interval; A second limiting condition is added by the chi-square statistic KZ of the first analysis set and the second analysis set: chi-square statistic KZ≤chi-square statistic threshold.

8. The method for short-term photovoltaic power prediction for real-time control according to claim 1, characterized in that: The photovoltaic power prediction model is constructed as follows: The meteorological condition data included in the second analysis set is used as input, and the trained LSTM model outputs the power generation data. The power generation data output by the LSTM model is analyzed for deviation from the power generation data corresponding to the input meteorological condition data to obtain an output deviation value. If the output deviation value is less than or equal to the output deviation threshold, the input meteorological condition data is considered to have passed the verification; Counting the number of meteorological condition data that have passed the verification, and performing a ratio process on the number of meteorological condition data included in the second analysis set to obtain a verification pass value; If the verification pass value ≥ the verification pass threshold, the LSTM model is considered to have passed the verification of the second analysis set, and the LSTM model is marked as a photovoltaic power prediction model.

9. The method for short-term photovoltaic power prediction for real-time control according to claim 1, characterized in that: The prediction of photovoltaic power generation power of the photovoltaic power station specifically includes: Taking the meteorological condition dataset to be predicted as input, multiple photovoltaic power prediction models are used to output multiple predicted power generation data, which are then integrated to obtain a predicted power generation data group. Integrate the verification pass values ​​corresponding to each photovoltaic power prediction model into a verification pass data group; According to the verification performance of the photovoltaic power prediction model, the verification performance coefficient YBXi of the photovoltaic power prediction model is calculated. The specific calculation formula is: ； Where YBXi represents the verification performance coefficient of the i-th photovoltaic power prediction model, yti represents the verification pass value of the i-th photovoltaic power prediction model, and b represents the number of photovoltaic power prediction models; The final photovoltaic power generation power prediction result ZYF is: ； Where yfdi represents the predicted power generation data corresponding to the i-th photovoltaic power prediction model, and b represents the number of photovoltaic power prediction models.

10. A short-term photovoltaic power prediction system for real-time control, characterized by: include: Data screening module: This module integrates the historical meteorological condition data and the corresponding power generation data of the photovoltaic power station to obtain a comprehensive set of meteorological and power data; The deviation analysis is carried out with the meteorological condition data set to be predicted, and a similar power forecast set is obtained by screening; A partitioning and limiting module: uses an overall distribution difference value as a first limiting condition for partitioning the first analysis set and the second analysis set, wherein the overall distribution difference value is calculated by the data distribution difference between the first analysis set and the second analysis set; and uses a chi-square statistic as a second limiting condition for partitioning the first analysis set and the second analysis set, wherein the chi-square statistic is calculated by performing a chi-square test on the data histogram constructed from the first analysis set and the second analysis set; Model testing module: Use the first analysis set and the second analysis set to train and verify the LSTM model, screen the LSTM model, and obtain a photovoltaic power prediction model; Comprehensive prediction module: The meteorological condition dataset to be predicted is used as the input of multiple photovoltaic power prediction models. The photovoltaic power generation power of the photovoltaic power station is predicted through the prediction of multiple photovoltaic power prediction models and the training performance of the photovoltaic power prediction models.

Images