Model updating method and device based on photovoltaic power generation prediction

Abstract

The present application discloses a model updating method and device based on photovoltaic power generation prediction, the method is: pre-processing the sampling data of the day to obtain the training data of the day; based on the first sample, the second sample and the third sample, combined with the eigenvalue calculation formula, determining the eigenvalue corresponding to the data of the day; based on the eigenvalue corresponding to the data of the specified date in the historical time, calculating the coefficient matrix of the regression model to obtain the updated target model. The method determines the eigenvalue of the training data obtained by pre-processing the sampling data of the day based on the eigenvalue calculation formula, and calculates the coefficient matrix using the eigenvalue corresponding to the data of the specified date to update the target model, effectively improving the efficiency of the modeling and prediction process of the target model, and at the same time, it can realize the incremental update and forgetting of the training data of the target model on the basis of non-repetitive and large-scale reading of historical time series data (i.e., each time series data contained in the sampling data of previous dates).

Description

Technical Field

[0001] The present application relates to the field of photovoltaic power generation, and in particular to a model updating method and device based on photovoltaic power generation power prediction. Background Art

[0002] In recent years, the rapid development of new energy power generation technology has led to a continuous increase in the total installed capacity of photovoltaic power generation equipment, and has also driven related research in the field of photovoltaic power generation prediction. Photovoltaic power generation is mainly affected by irradiance and equipment health status. Generally, historical irradiance and equipment power generation data are used to train models, and irradiance data is used for prediction. To improve the prediction accuracy, the model is usually built at the inverter level.

[0003] Depending on the model application scenario, historical data ranging from 3 to 180 days needs to be read. If the model training data is not updated or updated less frequently, the accuracy of using the old model to predict current data may decrease with the change of seasons and environment. If the model update frequency is high, the number of inverters in the photovoltaic power station ranges from dozens to thousands. If the historical data of recent days is read from each inverter every day for cleaning and modeling, it will occupy a lot of computing resources and consume a long computing time. In addition, the same historical data may be read and cleaned repeatedly for multiple days. Summary of the invention

[0004] The present application provides a model updating method and device based on photovoltaic power generation prediction, which aims to improve the modeling and prediction efficiency of the photovoltaic power generation model, while realizing incremental updating and forgetting of the training data of the target model.

[0005] In order to achieve the above objectives, this application provides the following technical solutions:

[0006] A model updating method based on photovoltaic power generation prediction includes:

[0007] Preprocessing the sampling data of the day to obtain the training data of the day; the training data includes a first sample, a second sample and a third sample corresponding to a plurality of sampling time points; the first sample represents the power generation power of the first device; the second sample represents the power generation power of the second device; the third sample represents the observation value of the preset indicator;

[0008] Based on the first sample, the second sample and the third sample, in combination with the characteristic value calculation formula, the characteristic value corresponding to the data of the day is determined; the characteristic value calculation formula is obtained by analyzing the target model using the least squares method; the target model includes a regression model that characterizes the relationship between the power generation of the first device, the power generation of the second device and the preset indicator at the same time point;

[0009] Based on the eigenvalues ​​corresponding to the specified date data within the historical time, the coefficient matrix of the regression model is calculated to obtain an updated target model; the updated target model is used to predict the power generation power of the first device.

[0010] Optionally, determining the characteristic value corresponding to the data of the day based on the first sample, the second sample, and the third sample in combination with the characteristic value calculation formula includes:

[0011] Based on the analysis of the target model, the characteristic value calculation formula is obtained;

[0012] Based on the first sample, the second sample and the third sample, combined with the characteristic value calculation formula, the characteristic value corresponding to the data of the day is calculated; the characteristic values ​​include n, C, D, E, F, G, H, L, and M; n is determined based on the total number of the sampling time points; C is determined based on the cumulative sum of the squares of the second samples corresponding to all sampling time points; D is determined based on the cumulative sum of the second samples corresponding to all sampling time points; E is determined based on the cumulative sum of the products of the first samples and the second samples corresponding to all sampling time points; F is determined based on the cumulative sum of the first samples corresponding to all sampling time points; G is determined based on the cumulative sum of the products of the second samples and the third samples corresponding to all sampling times; H is determined based on the cumulative sum of the squares of the third samples corresponding to all sampling times; L is determined based on the cumulative sum of the third samples corresponding to all sampling time points; M is determined based on the cumulative sum of the products of the first samples and the third samples corresponding to all sampling time points.

[0013] Optionally, if the second sample or the third sample participates in the calculation of the eigenvalue calculation formula, the eigenvalue includes n, C, D, E, and F; if the second sample and the third sample participate in the calculation of the eigenvalue calculation formula, the eigenvalue includes at least n, C, D, E, F, G, H, L, and M.

[0014] Optionally, preprocessing the sampling data of the day to obtain the training data of the day includes:

[0015] The sampling data of the day is obtained; the sampling data includes first time series data, second time series data and third time series data; the first time series data includes the power generation of the first device at multiple sampling time points; the second time series data includes the power generation of the second device at multiple sampling time points; the third time series data includes the observation values ​​of the preset indicators at multiple sampling time points; the first device includes photovoltaic power generation equipment whose equipment properties do not meet the requirements; the second device includes photovoltaic power generation equipment whose equipment properties meet the requirements; the preset indicators include irradiance;

[0016] Preprocessing the first time series data to obtain a plurality of first samples and corresponding timestamps;

[0017] Preprocessing the second time series data to obtain a plurality of second samples and corresponding timestamps;

[0018] Preprocessing the third time series data to obtain a plurality of third samples and corresponding timestamps;

[0019] According to the corresponding timestamps, the first sample, the second sample, and the third sample of the day are aligned to obtain the training data of the day.

[0020] Optionally, preprocessing the second time series data to obtain a plurality of second samples and corresponding timestamps includes:

[0021] Eliminate data in the second time series data whose data amount does not meet a preset range, so as to obtain second time series data after data amount determination;

[0022] Normalizing the second time series data after the data amount determination to obtain a plurality of candidate samples and corresponding timestamps;

[0023] Selecting a target sample whose value meets a preset range from the plurality of candidate samples;

[0024] Determine multiple target sample sets based on the timestamps corresponding to the target samples obtained by screening; the target sample sets include multiple target samples belonging to the same timestamp;

[0025] A plurality of second samples are determined based on the arithmetic mean of the data within a given quantile of the plurality of target sample sets.

[0026] Optionally, aligning the first sample, the second sample, and the third sample in any period according to the corresponding timestamps to obtain the training data of the day includes:

[0027] Aligning the first sample, the second sample, and the third sample of the day according to the corresponding timestamps to obtain a data point set of the day; the data point set includes data points corresponding to multiple timestamps; the data points are determined based on the first sample, the second sample, and the third sample belonging to the same timestamp;

[0028] Eliminate data points whose timestamps in the data point set are within a time period of abnormal device operation;

[0029] Using a density clustering algorithm, target data points meeting preset conditions are screened from the data point set and determined as the training data for the day.

[0030] A model updating device based on photovoltaic power generation prediction, comprising:

[0031] A data preprocessing unit, used to preprocess the sampling data of the day to obtain the training data of the day; the training data includes a first sample, a second sample and a third sample corresponding to a plurality of sampling time points; the first sample represents the power generation of the first device; the second sample represents the power generation of the second device; the third sample represents the observation value of the preset indicator;

[0032] an eigenvalue determination unit, configured to determine the eigenvalue corresponding to the data of the day based on the first sample, the second sample, and the third sample in combination with an eigenvalue calculation formula; the eigenvalue calculation formula is obtained by analyzing the target model using the least squares method; the target model includes a regression model that characterizes the relationship between the power generation of the first device, the power generation of the second device, and the preset indicator at the same time point;

[0033] A model updating unit is used to calculate the coefficient matrix of the regression model based on the eigenvalues ​​corresponding to the specified date data in the historical time to obtain an updated target model; the updated target model is used to predict the power generation power of the first device.

[0034] A storage medium includes a stored program, wherein the program is executed by a processor to execute the model updating method based on photovoltaic power generation prediction.

[0035] An electronic device, comprising: a processor, a memory and a bus; the processor and the memory are connected via the bus;

[0036] The memory is used to store programs, and the processor is used to run programs, wherein the program is executed by the processor to execute the model updating method based on photovoltaic power generation prediction.

[0037] The technical solution provided by the present application pre-processes the sampling data of the day to obtain the training data of the day. Based on the first sample, the second sample and the third sample, in combination with the eigenvalue calculation formula, the eigenvalue corresponding to the data of the day is determined. Based on the eigenvalue corresponding to the data of the specified date in the historical time, the coefficient matrix of the regression model is calculated to obtain the updated target model. Based on the eigenvalue calculation formula, the method determines the eigenvalue of the training data obtained by pre-processing the sampling data of the day, and calculates the coefficient matrix using the eigenvalue corresponding to the data of the specified date to update the target model, effectively improving the efficiency of the modeling and prediction process of the photovoltaic power generation model, and at the same time, it can realize the incremental update and forgetting of the training data of the photovoltaic power generation model on the basis of non-repetitive and large-scale reading of historical time series data (that is, each time series data contained in the sampling data of previous dates). BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 A schematic flow chart of a model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application;

[0040] Figure 2 A schematic flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application;

[0041] Figure 3 A schematic flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application;

[0042] Figure 4 A schematic flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application;

[0043] Figure 5 A schematic diagram of the architecture of a model updating device based on photovoltaic power generation prediction provided in an embodiment of the present application;

[0044] Figure 6 A schematic flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application;

[0045] Figure 7 A schematic flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0047] In this application, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of more restrictions, the elements defined by the sentence "comprise one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0048] like Figure 1 As shown, it is a flow chart of a model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application, which includes the following steps.

[0049] S101: Preprocess the sampling data of the day to obtain the training data of the day.

[0050] The training data includes a first sample, a second sample and a third sample corresponding to multiple sampling time points, the first sample represents the power generation power of the first device, the second sample represents the power generation power of the second device, and the third sample represents the observed value of the preset indicator.

[0051] In some examples, the first device and the second device are both photovoltaic power generation devices (eg, inverters) in the same photovoltaic station, and the power generation power of the photovoltaic power generation device can be understood as the active power of the inverter.

[0052] In some examples, the sampling time point of the day can be set by the technician according to the actual situation.

[0053] Optionally, preprocess the sampling data of the day to obtain the training data of the day. For details, see Figure 2 The steps are shown and the corresponding explanations.

[0054] S102: Based on the first sample, the second sample, and the third sample, in combination with the characteristic value calculation formula, determine the characteristic value corresponding to the data of the day.

[0055] The characteristic value calculation formula is obtained by analyzing the target model using the least square method, and the target model includes a regression model that characterizes the relationship between the power generation of the first device, the power generation of the second device and the preset index at the same time point. Generally speaking, the target model is used to serve as a photovoltaic power generation model.

[0056] It should be noted that the power generation of the equipment is theoretically linearly related to the solar irradiance at the same time (i.e. the preset index). However, in the actual production process, it is affected by various external factors and may be nonlinear to a certain extent. Considering the complexity of the photovoltaic power prediction model and the fact that the nonlinear relationship can generally be approximated as a linear relationship, the binary linear model is selected as the regression model.

[0057] The target model has an analytical solution. For the univariate linear regression model Y=XK, the corresponding coefficient matrix can be calculated using the least squares method. The specific calculation formula can be found in formula (1).

[0058] (1)

[0059] In formula (1), X represents the independent variable, Y represents the dependent variable, K represents the coefficient matrix, and X is an n×2-dimensional matrix, Y is an n-dimensional column vector, K is a 2-dimensional column vector, and n is a positive integer (which can be determined based on the total number of sampling time points). If the independent variables and dependent variables are sorted by time and then divided into multiple sub-matrices according to date, the univariate linear regression model can be transformed into formula (2).

[0060] (2)

[0061] In formula (2), Representative Day's The independent variable data, Representative Day's dependent variable data, and X and Y are decomposed and daily data is taken as a sub-matrix, as shown in formula (3).

[0062] (3)

[0063] Substitute formula (3) into formula (1) and expand it to obtain formula (4).

[0064] (4)

[0065] make , then we get formula (5).

[0066] (5)

[0067] Based on formula (5), we only need to calculate and save n in the daily training data. j , C j , D j 、E j 、F j, you can select daily training data to update the coefficient matrix of the target model.

[0068] In addition, the training data of different dates can be weighted, and the weights can be assigned in a decreasing manner as the date decreases, so as to achieve the purpose of high weights for recent training data and low weights for old training data. After the training data is weighted, X in formula (1) becomes AX, and Y becomes AY, where the weight matrix A is an n×n dimensional matrix as shown in formula (6), and the non-diagonal elements are all 0, and the diagonal elements are the weights of each data point. Since the weights of the data points on the same date are the same, the values ​​of the 1st to n1th elements are the same, and the values ​​of the n1+1st to n1+n2th elements are the same. Substituting into formula (1) can obtain formula (7).

[0069] (6)

[0070] (7)

[0071] After obtaining the analytical solution of the univariate linear regression model, it can be extended to the bivariate linear model. The bivariate linear model can still be expressed as Y = XK, where Y is an n-dimensional column vector, X is an n × 3-dimensional matrix, and K is a 3-dimensional column vector. The expansion is shown in formula (8).

[0072] (8)

[0073] In formula (8), Y i No change, X i It becomes as shown in formula (9).

[0074] (9)

[0075] Substituting formula (9) into the analytical solution, we get the coefficient matrix as shown in formula (10).

[0076] (10)

[0077] In formula (10), According to formula (10), the characteristic value of the training data is calculated for each first device (non-benchmark inverter), that is, the n of the training data of the day is calculated for each non-benchmark inverter. j , C j , D j 、E j 、F j , G j , H j , L j 、M jThere are 9 eigenvalues ​​in total. After saving these 9 eigenvalues ​​to the database, when training data for a specified date is needed, only these 9 eigenvalues ​​can be read. It is no longer necessary to read all the time series data (i.e., the first time series data, the second time series data, and the third time series data) for the specified date, nor is it necessary to repeat preprocessing such as data cleaning. For photovoltaic stations that only have daily irradiance data (the third time series data) or only have benchmark inverter data (i.e., the second time series data), it is only necessary to read n time series data for the specified date. j , C j , D j 、E j 、F j These 5 eigenvalues.

[0078] To this end, the eigenvalue calculation formula includes: as well as .

[0079] Optionally, the implementation process of determining the eigenvalue corresponding to the data of the day based on the first sample, the second sample and the third sample in combination with the eigenvalue calculation formula can be summarized as follows: based on the analysis of the target model, the eigenvalue calculation formula is obtained; based on the first sample, the second sample and the third sample, in combination with the eigenvalue calculation formula, the eigenvalue corresponding to the data of the day is calculated; the eigenvalues ​​include n, C, D, E, F, G, H, L, and M; n is determined based on the total number of sampling time points; C is determined based on the cumulative sum of the squares of the second samples corresponding to all sampling time points; D is determined based on the total number of sampling time points. The cumulative sum of the second samples corresponding to the sampling time points is determined; E is determined based on the cumulative sum of the products of the first samples and the second samples corresponding to all sampling time points; F is determined based on the cumulative sum of the first samples corresponding to all sampling time points; G is determined based on the cumulative sum of the products of the second samples and the third samples corresponding to all sampling times; H is determined based on the cumulative sum of the squares of the third samples corresponding to all sampling times; L is determined based on the cumulative sum of the third samples corresponding to all sampling time points; M is determined based on the cumulative sum of the products of the first samples and the third samples corresponding to all sampling time points.

[0080] In some examples, the number of the first sample, the second sample, and the third sample of the same photovoltaic power generation device is the same, and n can be regarded as the number of sampling time points. In formula (7), x represents the second sample, y represents the first sample, z represents the third sample, j represents the index of the date, and i represents the index of the sample.

[0081] Optionally, if the second sample or the third sample participates in the calculation of the eigenvalue calculation formula, the eigenvalue includes n, C, D, E, and F; if the second sample and the third sample participate in the calculation of the eigenvalue calculation formula, the eigenvalue includes at least n, C, D, E, F, G, H, L, and M.

[0082] It should be emphasized that the number of eigenvalues ​​is affected by the input time series data. If the input time series data includes both the active power of the benchmark inverter (i.e. the second sample) and the daily irradiance of the site (i.e. the third sample), there are 9 eigenvalues; if one of the above two types of time series data is empty, it becomes 5 eigenvalues. If other types of time series data such as temperature are added on the basis of the above two types of time series data, the number of eigenvalues ​​will increase further.

[0083] S103: Calculate the coefficient matrix of the regression model based on the eigenvalues ​​corresponding to the specified date data in the historical time to obtain an updated target model.

[0084] The updated target model is used to predict the power generation of the first device.

[0085] It should be noted that the eigenvalues ​​corresponding to the data of the specified date can be obtained from the database, and the eigenvalues ​​corresponding to the data of the specified date can be substituted into formula (10) to calculate the coefficient matrix K of the regression model. In addition, for each inverter (i.e., photovoltaic power generation equipment), there is a set of weight coefficients and eigenvalues ​​of historical training data, and a coefficient matrix is ​​calculated. At this time, it is required that the number of weight coefficients is the same as the number of days of historical data, and the date of the eigenvalue corresponds correctly to the weight value.

[0086] In some examples, since the eigenvalues ​​are calculated for daily training data, the eigenvalues ​​of the latest training data can be selected to calculate the coefficient matrix, so as to update the target model daily without significantly increasing the amount of calculation. In addition, according to the needs of the scenario, the coefficient matrix of the regression model can be calculated by selecting the eigenvalues ​​corresponding to the data on a specified date to achieve on-demand modeling instead of fixing the model coefficients in advance, thus making the photovoltaic power generation prediction model more flexible.

[0087] In a possible implementation, when using data from the same date in the past five years for modeling, the characteristic values ​​of these five days can be directly read; when using data from the same month in the past three years for modeling, the characteristic values ​​of all dates within these three months can be directly read. Since the present application saves characteristic values ​​instead of models, the date of the training data used for modeling can be flexibly selected. In addition, the time series data of large photovoltaic stations one or two years ago may be transferred to other databases, and modeling using the original time series data is more difficult. The amount of data required for modeling using characteristic values ​​is less and can be stored in the same database for a long time.

[0088] It should be noted that the hyperparameters used by the target model are not determined in daily calculations. The determination of hyperparameters consumes more computing resources and takes a long time to calculate. Once the hyperparameters are determined, they are rarely changed frequently. Therefore, when determining the hyperparameters, use offline modeling to read in multiple days of sampling data at one time. After the aforementioned data preprocessing steps, the data is divided into a training set and a test set. The eigenvalues ​​of the data in the training set are calculated on a daily basis, and then the coefficient matrix is ​​calculated based on the hyperparameters to obtain the regression model under this set of hyperparameters. The test set data is brought in for prediction and the prediction results are evaluated, and finally a set of appropriate hyperparameters is selected. The hyperparameters involved include, but are not limited to: the number of days of historical data used, and the weight coefficients assigned to the data of each day.

[0089] In some examples, combined Figure 2-Figure 4 The method shown in the embodiment of the present application, the model updating method based on photovoltaic power generation prediction, can be summarized as follows: Figure 6 For further steps, see Figure 7 The steps shown are specifically implemented as follows.

[0090] Step 1. Read the data required for calculation from the database: the input data (sampled data of the specified date) includes the time series data of the active power measurement points of all inverters under the photovoltaic station on that day (including the first time series data and the second time series data), the time series data of the irradiance measurement points on that day (i.e., the third time series data), the installed capacity of all inverters, the numbers of all benchmark inverters, all available data on abnormal working time periods of inverters, the relevant parameter values ​​(i.e., hyperparameters) required for calculation, the characteristic values ​​of historical training data (i.e., the characteristic values ​​of each non-benchmark inverter corresponding to the required date in the historical time), and the predicted data as the input of the target model.

[0091] Step 2: Preprocess the input data. The preprocessing process includes many detailed operations: first, delete the null values ​​in all time series data; then process the irradiance time series data, divide the irradiance time series value by 1000, obtain the normalized irradiance time series data, and delete the normalized irradiance time series data greater than 1.4 or less than 0; then process the active power time series data of each inverter, including benchmark inverters and non-benchmark inverters. For each inverter, screen its time series data and installed capacity data. If the number of its time series data is less than 60, the quality of the inverter's time series data on that day is low. Poor, should not be retained as training data; for the time series data of the inverter that meets the data volume requirements, divide it by its own installed capacity to obtain the normalized active power time series data, and delete the normalized time series data greater than 1 or less than 0; if the device still has available time series data, and the number of time series data is greater than 60, then find out whether the device has data for abnormal working time periods. If there are one or more abnormal working time periods, it is necessary to judge the time series data to see whether the time when the time series data is generated is within the abnormal working time period, and only retain the time series data within the normal working time period.

[0092] In a possible implementation, the preprocessing process can be summarized as: deleting null values ​​in all sampled data, scaling the irradiance time series data, normalizing the inverter active power time series data, and eliminating out-of-limit data, eliminating the time series data of measurement points with less valid data on the day, and eliminating the time series data of the equipment during abnormal working time periods.

[0093] Step 3. After completing the preprocessing of all time series data, construct the training data for the day: the training data is composed of the input data and output data of the model, the input data is the normalized irradiance time series data (i.e., multiple third samples) and the normalized active power time series data of the benchmark inverter (i.e., multiple candidate samples), and the output data is the normalized active power time series data of each non-benchmark inverter on the day (i.e., multiple first samples); according to the input benchmark inverter number, filter the normalized active power time series data of all benchmark inverters, and align these data according to the time when the time series data is generated; for each The normalized active power values ​​of multiple benchmark inverters at moments are obtained, the 10% and 90% quantiles are calculated, and the data between these two quantile values ​​are filtered out, the selected data are averaged, and the compressed normalized active power time series data of the benchmark inverter is obtained (i.e., multiple second samples); the compressed normalized active power time series data of the benchmark inverter and the normalized irradiance time series data are aligned according to the moment when the time series data is generated, and the null values ​​are eliminated to obtain the input data of the training data; the output data of the training data directly uses the normalized active power time series data of the non-benchmark inverter after data preprocessing.

[0094] Step 4: Calculate the training data feature values for each non-reference inverter: For each non-reference inverter, the input data of the training data is the same, only the output data is different. Align the input data and the output data according to the time when the time-series data is generated, and剔除 null values to obtain the training data of the inverter on that date. Use the density clustering algorithm to identify the severely outlying points in the training data. Each point has a three-dimensional coordinate composed of the normalized active power value of the compressed reference inverter, the normalized irradiance value, and the normalized active power value of the device. Each time-series data value has been normalized to the range of 0-1. The non-noise points selected by the clustering are used as the training data actually used by the model. If the number of the filtered training data is not less than 60, then according to the eigenvalue calculation formula shown in the following formula (11), calculate n j 、C j 、D j 、E j 、F j 、G j 、H j 、L j 、M j , and save n j 、C j 、D j 、E j 、F j 、G j 、H j 、L j 、M j .

[0095] (11)

[0096] In formula (11), j represents the specified date, x represents the normalized active power of the compressed reference inverter (i.e., the second sample), z represents the normalized daily irradiance (i.e., the third sample), y represents the normalized active power of the specified device (i.e., the first sample), and n j represents the data volume (i.e., the number of sampling times).

[0097] Step 5: Calculate the coefficient matrix of each non-reference inverter model according to the input historical training data feature values: If the model uses short-term training data for prediction, the training data of the recent 5 days can be selected. The data dates are Day1, Day2, Day3, Day4, and Day5 from far to near, with Day1 < Day5. At the same time, different weight coefficients are assigned to different dates: Day1: α1 = 0.05; Day2: α2 = 0.05; Day3: α3 = 0.1; Day4: α4 = 0.3; Day5: α5 = 0.4. Calculate the coefficient matrix K according to the following formula (12).

[0098] (12)

[0099] In formula (12), j represents the date and θ represents the total number of days.

[0100] Step 6. Bring the predicted data into the model for prediction and output the prediction results: For each non-benchmark inverter, the input data for prediction is the same, which are the normalized active power value and the normalized irradiance value of the benchmark inverter, but the coefficient matrix is ​​different, and the predicted normalized active power value is also different; the normalized active power prediction value of each device is multiplied by the installed capacity of each device to obtain the device active power prediction value (that is, predict the active power of the second device).

[0101] In general, the embodiment of the present application realizes a photovoltaic device power prediction method that can incrementally update and forget training data, and relates to the field of photovoltaic power prediction technology. The prediction of this method can be summarized as follows: obtaining the time series data, device information, parameter values, historical calculation results, etc. required for updating the optical power model; preprocessing the time series data and constructing the input and output of the training data; calculating the characteristic values ​​of the training data according to the derived formula; using the characteristic values ​​of multiple days to calculate the model coefficients according to the formula; and finally using the model and prediction data to predict the active power of each device. This prediction method realizes the incremental update and forgetting of the training data of the optical power model on the basis of non-repetitive and large-scale reading of historical time series data.

[0102] It is understandable that the conventional method is to update the model every six months or quarterly, for example, using the training data from March 1 to May 31 to update the model on June 1, and using the training data from June 1 to August 31 to update the model on September 1. In this method, the model updated on June 1 is used in August. The incremental update method can update the model daily, using the training data from May 9 to August 8 to update the model on August 9. The so-called forgetting can be understood as the data before May 9 is no longer used for model training, and the model is not affected by the data before May 9.

[0103] Compared with the prior art, the embodiment of the present application is based on the principle of least squares method, divides the training data by date, selects appropriate thresholds for data cleaning based on business principles, calculates and saves the eigenvalues ​​of the training data of each day and each device for calculating the coefficient matrix of the regression model, selects the eigenvalues ​​of the required number of days as needed during prediction, combines the weight coefficients of each day, calculates the coefficient matrix of the regression model, and finally realizes the prediction of the power generation power of the photovoltaic equipment.

[0104] The process shown in S101-S103 above determines the eigenvalues ​​of the training data obtained by preprocessing the sampling data of the day based on the eigenvalue calculation formula, and uses the eigenvalue calculation coefficient matrix corresponding to the data of the specified date to update the target model. In addition, the sampling data is only read and processed on the day of generation, which reduces the hardware resource occupancy rate of the program and effectively improves the efficiency of the modeling and prediction process of the target model (i.e., photovoltaic power generation model). At the same time, the incremental update and forgetting of the training data of the target model can be realized on the basis of non-repetitive and large-scale reading of historical time series data (i.e., each time series data contained in the sampling data of previous dates). Since the data eigenvalues ​​are calculated first and then the model is updated when the program is running every day, the latest data can be used to update the model every day, so that the model can track the latest status of the device. When the modeling scenario changes, the training data date can be selected on demand to calculate the new model at a low computational cost, which improves the flexibility of modeling.

[0105] like Figure 2 As shown, it is a flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application, including the following steps.

[0106] S201: Obtain sampling data for the day.

[0107] Among them, the sampling data includes first time series data, second time series data and third time series data. The first time series data includes the power generation of the first device at multiple sampling time points, the second time series data includes the power generation of the second device at multiple sampling time points, and the third time series data includes the observation values ​​of preset indicators at multiple sampling time points. The first device includes photovoltaic power generation equipment whose equipment properties do not meet the requirements, and the second device includes photovoltaic power generation equipment whose equipment properties meet the requirements. The preset indicators include irradiance.

[0108] In some examples, device attributes include but are not limited to failure rate, communication stability rate, power limit possibility, etc. In a possible implementation, the device attribute not meeting the requirements may be: the failure rate is higher than the first calibration value, and / or the communication stability rate is lower than the second calibration value, and / or the power limit possibility is higher than the third calibration value.

[0109] In some examples, the sampling data of a photovoltaic station on a given day may include: time series data of active power measurement points of all inverters on that day (including first time series data and second time series data), time series data of irradiance measurement points of the photovoltaic station on that day (i.e., third time series data), installed capacity of all inverters under the photovoltaic station, numbers of all benchmark inverters (benchmark inverters can be regarded as second devices) under the photovoltaic station, and known abnormal working time periods of inverters.

[0110] It is understandable that in the process of photovoltaic power generation prediction, the inverter's active power measurement point time series data for the day, the inverter's installed capacity, and the relevant parameters required for calculation (i.e., the hyperparameters required for target modeling, such as the number of historical data days, the weight of training data on different dates, etc.) must be provided. If any one of them is missing, the calculation cannot continue. Taking into account the different construction conditions of different photovoltaic sites, at least one of the photovoltaic site's daily irradiance measurement point time series data and the number of all benchmark inverters under the site should be provided. The benchmark inverter can specifically refer to an inverter with a low failure rate, no power restriction, and stable communication. The site with a benchmark inverter is generally based on a sub-array, and all inverters under one or more sub-arrays are benchmark inverters. The known inverter abnormal working time period can be omitted. This input is used in the preprocessing of the sampling data to help improve the quality of the preprocessing.

[0111] In some examples, after obtaining the sampled data, the sampled data needs to be preprocessed. The preprocessing process mainly includes data quality judgment, time series data normalization, elimination of values ​​that seriously deviate from physical laws, elimination of data within non-normal working time periods, elimination of seriously outlier data points, etc.

[0112] Generally speaking, data quality judgment is a judgment of the overall situation of the time series data of a specific device on that day, including inverters and irradiation equipment. When the amount of time series data is less than the preset threshold, the amount of training data is too small, which is very likely to lead to underfitting and is not suitable for updating the model. Time series data normalization is because the installed capacity of different inverters is different. This value is related to the installed capacity and number of strings under the inverter. The installed capacity represents the maximum active power that the inverter can output. Therefore, although the actual active power range output by different inverters is different, the value range of the ratio of actual active power to installed capacity is the same, both between 0 and 1. Values ​​less than 0 and greater than 1 are data that do not conform to physical laws. Therefore, after the inverter active power is normalized, data that is not in the range of 0~1 should be eliminated. Irradiance refers to the radiation energy per unit area per unit time on the solid earth surface after solar radiation is absorbed, scattered, reflected, etc. by the atmosphere. The maximum value is about 1373W / m 2Therefore, data not in the range of 0~1400 does not conform to physical laws and should be eliminated. In addition, in order to make the order of magnitude of the two types of time series data consistent, the irradiance time series data needs to be scaled so that its maximum value is on the order of 10E0. Therefore, the irradiance time series data value is divided by 1000 to make its value range between 0~1.4. Photovoltaic power generation equipment is affected by various internal and external factors, resulting in power generation lower than normal. Internal factors include inverter and corresponding equipment failures, and external factors include shading, power restrictions, and drastic fluctuations in irradiance. If the site can provide the time period when each inverter is affected by the above factors, the time series data can be eliminated according to the list of each device and time period, and only the time series data within the normal working time period of the equipment is retained. In addition, the abnormal working time period of different inverters in the same site may be 0 or more, and there may be partial overlap between different types of abnormal working time periods. They should be eliminated one by one during data cleaning. For sites that cannot provide the time period when each inverter is affected by the above factors, it is necessary to analyze the business meaning and characteristics of the data, design appropriate rules, and eliminate them.

[0113] S202: Preprocess the first time series data to obtain a plurality of first samples and corresponding timestamps.

[0114] The first sample may be regarded as the power generation power of the first device (eg, a non-benchmark inverter).

[0115] In some examples, the timestamp corresponding to the first sample can be understood as the sampling time point corresponding to the power generation of the first device.

[0116] It is understandable that the preprocessing of the first time series data may include data quality judgment, time series data normalization, elimination of values ​​that seriously deviate from physical laws, elimination of data within non-normal working time periods, elimination of seriously outlier data points, and the like.

[0117] S203: Preprocess the second time series data to obtain a plurality of second samples and corresponding timestamps.

[0118] The second sample may be regarded as the power generation power of the second device (eg, a benchmark inverter).

[0119] In some examples, the timestamp corresponding to the second sample can be understood as the sampling time point corresponding to the power generation of the second device.

[0120] It is understandable that the preprocessing of the second time series data may include data quality judgment, time series data normalization, elimination of values ​​that seriously deviate from physical laws, elimination of data within non-normal working time periods, elimination of seriously outlier data points, and the like.

[0121] Optionally, the second time series data is preprocessed to obtain a plurality of second samples and corresponding timestamps. For details, see Figure 3 The steps are shown and the corresponding explanations.

[0122] S204: Preprocess the third time series data to obtain a plurality of third samples and corresponding timestamps.

[0123] Among them, the third sample can be regarded as the observed value of the preset indicator (such as irradiance).

[0124] In some examples, the timestamp corresponding to the third sample can be understood as the sampling time point corresponding to the observation value of the preset indicator.

[0125] It is understandable that the preprocessing of the third time series data may include data quality judgment, time series data normalization, elimination of values ​​that seriously deviate from physical laws, elimination of data within non-normal working time periods, elimination of seriously outlier data points, and the like.

[0126] S205: Align the first sample, the second sample, and the third sample of the day according to the corresponding timestamps to obtain the training data of the day.

[0127] Among them, due to the existence of accidental factors such as environment and human factors, there are still seriously outlier data points in the training data. The number of such data points is small and the distance between each other is far. These abnormal data can be identified through density clustering calculation to improve the quality of training data.

[0128] Optionally, the first sample, the second sample, and the third sample of the day are aligned according to the corresponding timestamps to obtain the training data of the day. For details, see Figure 4 The steps are shown and the corresponding explanations.

[0129] The process shown in S201-S205 above pre-processes the sampling data of the day to obtain the training data of the day, thereby effectively ensuring the data quality of the first sample, the second sample and the third sample in the training data.

[0130] like Figure 3 As shown, it is a flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application, including the following steps.

[0131] S301: Eliminate data whose data volume does not meet a preset range in the second time series data to obtain second time series data after data volume determination.

[0132] Among them, eliminating the data whose data volume does not meet the preset range in the second time series data is essentially to determine the data volume of the second time series data.

[0133] S302: Normalize the second time series data after the data amount is determined to obtain multiple candidate samples and corresponding timestamps.

[0134] The candidate samples can be understood as the normalized power generation of the second device (benchmark inverter).

[0135] In some examples, the timestamp corresponding to the candidate sample can be understood as the sampling time point corresponding to the normalized power generation power of the second device.

[0136] S303: Filtering target samples whose values ​​fall within a preset range from multiple candidate samples.

[0137] Among them, from multiple candidate samples, target samples whose values ​​meet the preset range are screened to eliminate abnormal data.

[0138] S304: Determine multiple target sample sets based on the timestamps corresponding to the target samples obtained by screening.

[0139] The target sample set includes multiple target samples belonging to the same timestamp.

[0140] S305: Determine a plurality of second samples based on the arithmetic mean of the data within a given quantile of the plurality of target sample sets.

[0141] For each target sample set, the arithmetic mean of data within a given quantile of multiple target samples in the target sample set is calculated, and the second sample is determined based on the arithmetic mean.

[0142] It should be noted that since there is more than one second device (such as a benchmark inverter), it is necessary to compress the normalized active power time series data (i.e., multiple candidate samples) of all benchmark inverters to obtain one-dimensional data that changes over time (i.e., multiple target samples). The power generation performance of different benchmark inverters is slightly different. Although the change trend of the active power of each inverter over time is consistent, the normalized power values ​​are different. The compressed data of the benchmark inverter (i.e., multiple second samples) ignores the differences in the power generation capacity of each device and can be used as a benchmark for the power generation capacity of the site at each time of the day. During compression, the normalized active power of all benchmark inverters is aligned in time, and the data within the 10% to 90% quantile (i.e., within the given quantile) of all benchmark inverter time series data at each time is filtered, and the data within this interval is averaged as the compressed normalized power value of the benchmark inverter at that time (i.e., multiple second samples).

[0143] The process shown in S301-S305 above can be performed by screening a target sample from a plurality of candidate samples and determining the second sample based on the arithmetic mean of the target sample set.

[0144] like Figure 4 As shown, it is a flow chart of another model updating method based on photovoltaic power generation prediction provided in an embodiment of the present application, including the following steps.

[0145] S401: Align the first sample, the second sample, and the third sample of the day according to the corresponding timestamps to obtain a data point set of the day.

[0146] The data point set includes data points corresponding to multiple timestamps, and the data points are determined based on a first sample, a second sample, and a third sample belonging to the same timestamp.

[0147] S402: Eliminate data points in the data point set whose timestamps are within a time period of abnormal device operation.

[0148] Among them, eliminating data points in the data point set whose timestamps are within the abnormal working period of the equipment can eliminate invalid data points in the data point set and improve the reliability of the data point set.

[0149] S403: Using a density clustering algorithm, target data points that meet preset conditions are selected from the data point set to be determined as training data for the day.

[0150] Among them, the density clustering algorithm is used to filter out seriously outlier data points in the data point set to select target data points that meet the preset conditions.

[0151] In some examples, each data point has a three-dimensional coordinate point consisting of a compressed normalized active power value of a benchmark inverter (i.e., the second sample), a normalized irradiance value (i.e., the third sample), and a normalized active power value of a non-benchmark inverter (i.e., the first sample). Since the first sample, the second sample, and the third sample have been normalized (e.g., normalized to a range of 0-1), eps in the hyperparameters of the density clustering algorithm can be set to 0.03 and min_samples can be set to 20, and the non-noise points (i.e., target data points) selected by clustering are used as the training data finally used.

[0152] The process shown in S401-S403 above can utilize a density clustering algorithm to improve the quality of the training data of the day.

[0153] like Figure 5 As shown, it is a schematic diagram of the architecture of a model updating device based on photovoltaic power generation prediction provided in an embodiment of the present application, including the units shown below.

[0154] The data preprocessing unit 100 is used to preprocess the sampling data of the day to obtain the training data of the day; the training data includes a first sample, a second sample and a third sample corresponding to multiple sampling time points; the first sample represents the power generation power of the first device; the second sample represents the power generation power of the second device; the third sample represents the observation value of the preset indicator.

[0155] Optionally, the data preprocessing unit 100 is specifically used to: obtain sampling data of the day; the sampling data includes first time series data, second time series data and third time series data; the first time series data includes the power generation of the first device at multiple sampling time points; the second time series data includes the power generation of the second device at multiple sampling time points; the third time series data includes the observation values ​​of preset indicators at multiple sampling time points; the first device includes photovoltaic power generation equipment whose equipment properties do not meet the requirements; the second device includes photovoltaic power generation equipment whose equipment properties meet the requirements; the preset indicators include irradiance; preprocess the first time series data to obtain multiple first samples and corresponding timestamps; preprocess the second time series data to obtain multiple second samples and corresponding timestamps; preprocess the third time series data to obtain multiple third samples and corresponding timestamps; according to the corresponding timestamps, align the first sample, the second sample and the third sample of the day to obtain the training data of the day.

[0156] Optionally, the data preprocessing unit 100 is specifically used to: eliminate data in the second time series data whose data volume does not meet the preset range to obtain second time series data after data volume judgment; normalize the second time series data after data volume judgment to obtain multiple candidate samples and corresponding timestamps; from the multiple candidate samples, filter target samples whose values ​​meet the preset range; determine multiple target sample sets based on the timestamps corresponding to the filtered target samples; the target sample set includes multiple target samples belonging to the same timestamp; determine multiple second samples based on the arithmetic mean of the data within a given quantile of the multiple target sample sets.

[0157] Optionally, the data preprocessing unit 100 is specifically used to: align the first sample, the second sample and the third sample of the day according to the corresponding timestamps to obtain the data point set of the day; the data point set includes data points corresponding to multiple timestamps; the data points are determined based on the first sample, the second sample and the third sample belonging to the same timestamp; eliminate the data points in the data point set whose timestamps are within the abnormal working time period of the equipment; use a density clustering algorithm to filter the target data points that meet the preset conditions from the data point set and determine them as the training data for the day.

[0158] The eigenvalue determination unit 200 is used to determine the eigenvalue corresponding to the data of the day based on the first sample, the second sample and the third sample in combination with the eigenvalue calculation formula; the eigenvalue calculation formula is obtained by analyzing the target model using the least squares method; the target model includes a regression model that characterizes the relationship between the power generation power of the first device, the power generation power of the second device and the preset indicator at the same time point.

[0159] Optionally, the eigenvalue determination unit 200 is specifically used to: obtain the eigenvalue calculation formula based on the analysis of the target model; calculate the eigenvalue corresponding to the data of the day based on the first sample, the second sample and the third sample in combination with the eigenvalue calculation formula; the eigenvalues ​​include n, C, D, E, F, G, H, L, and M; n is determined based on the total number of sampling time points; C is determined based on the cumulative sum of the squares of the second samples corresponding to all sampling time points; D is determined based on the cumulative sum of the second samples corresponding to all sampling time points; E is determined based on the cumulative sum of the products of the first samples and the second samples corresponding to all sampling time points; F is determined based on the cumulative sum of the first samples corresponding to all sampling time points; G is determined based on the cumulative sum of the products of the second samples and the third samples corresponding to all sampling times; H is determined based on the cumulative sum of the squares of the third samples corresponding to all sampling times; L is determined based on the cumulative sum of the third samples corresponding to all sampling time points; M is determined based on the cumulative sum of the products of the first samples and the third samples corresponding to all sampling time points.

[0160] Optionally, the eigenvalue determination unit 200 is specifically used to: if the second sample or the third sample participates in the calculation of the eigenvalue calculation formula, the eigenvalue includes n, C, D, E, F; if the second sample and the third sample participate in the calculation of the eigenvalue calculation formula, the eigenvalue includes at least n, C, D, E, F, G, H, L, M.

[0161] The model updating unit 300 is used to calculate the coefficient matrix of the regression model based on the eigenvalues ​​corresponding to the specified date data in the historical time to obtain an updated target model; the updated target model is used to predict the power generation of the first device.

[0162] Each of the above-mentioned units determines the eigenvalues ​​of the training data obtained by preprocessing the sampling data of the day based on the eigenvalue calculation formula, and uses the eigenvalue calculation coefficient matrix corresponding to the data of the specified date to update the target model, thereby effectively improving the efficiency of the modeling and prediction process of the photovoltaic power generation model. At the same time, it can realize the incremental update and forgetting of the training data of the photovoltaic power generation model on the basis of non-repetitive and large-scale reading of historical time series data (that is, each time series data contained in the sampling data of previous dates).

[0163] The present application also provides a computer-readable storage medium, which includes a stored program, wherein the program executes the model updating method based on photovoltaic power generation power prediction provided by the present application.

[0164] The present application also provides an electronic device, including: a processor, a memory and a bus. The processor and the memory are connected via a bus, the memory is used to store programs, and the processor is used to run the programs, wherein the model update method based on photovoltaic power generation prediction provided by the present application is executed when the programs are run.

[0165] In addition, the functions described above in the embodiments of the present application may be at least partially performed by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0166] Although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present application. Certain features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination.

[0167] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in this application (but not limited to) by each other to form a technical solution.

Claims

1. A model updating method based on photovoltaic power generation prediction, characterized in that: include: Preprocessing the sampling data of the day to obtain the training data of the day; The training data includes a first sample, a second sample and a third sample corresponding to a plurality of sampling time points; the first sample represents the power generation of the first device; the second sample represents the power generation of the second device; the third sample represents the observed value of the preset indicator; Based on the first sample, the second sample, and the third sample, in combination with an eigenvalue calculation formula, determining an eigenvalue corresponding to the data of the day; The characteristic value calculation formula is obtained by analyzing the target model using the least square method; the target model includes a regression model that characterizes the relationship between the power generation of the first device, the power generation of the second device and the preset index at the same time point; Based on the eigenvalues ​​corresponding to the specified date data within the historical time, the coefficient matrix of the regression model is calculated to obtain an updated target model; the updated target model is used to predict the power generation power of the first device.

2. The method according to claim 1, characterized in that Based on the first sample, the second sample, and the third sample, in combination with the characteristic value calculation formula, the characteristic value corresponding to the data of the day is determined, including: Based on the analysis of the target model, the characteristic value calculation formula is obtained; Based on the first sample, the second sample and the third sample, combined with the characteristic value calculation formula, the characteristic value corresponding to the data of the day is calculated; the characteristic values ​​include n, C, D, E, F, G, H, L, and M; n is determined based on the total number of the sampling time points; C is determined based on the cumulative sum of the squares of the second samples corresponding to all sampling time points; D is determined based on the cumulative sum of the second samples corresponding to all sampling time points; E is determined based on the cumulative sum of the products of the first samples and the second samples corresponding to all sampling time points; F is determined based on the cumulative sum of the first samples corresponding to all sampling time points; G is determined based on the cumulative sum of the products of the second samples and the third samples corresponding to all sampling times; H is determined based on the cumulative sum of the squares of the third samples corresponding to all sampling times; L is determined based on the cumulative sum of the third samples corresponding to all sampling time points; M is determined based on the cumulative sum of the products of the first samples and the third samples corresponding to all sampling time points.

3. The method according to claim 2, characterized in that If the second sample or the third sample participates in the calculation of the eigenvalue calculation formula, the eigenvalue includes n, C, D, E, and F. If the second sample and the third sample participate in the calculation of the eigenvalue calculation formula, the eigenvalue includes at least n, C, D, E, F, G, H, L, and M.

4. The method according to claim 1, characterized in that: Preprocessing the sampling data of the day to obtain the training data of the day includes: The sampling data of the day is obtained; the sampling data includes first time series data, second time series data and third time series data; the first time series data includes the power generation of the first device at multiple sampling time points; the second time series data includes the power generation of the second device at multiple sampling time points; the third time series data includes the observation values ​​of the preset indicators at multiple sampling time points; the first device includes photovoltaic power generation equipment whose equipment properties do not meet the requirements; the second device includes photovoltaic power generation equipment whose equipment properties meet the requirements; the preset indicators include irradiance; Preprocessing the first time series data to obtain a plurality of first samples and corresponding timestamps; Preprocessing the second time series data to obtain a plurality of second samples and corresponding timestamps; Preprocessing the third time series data to obtain a plurality of third samples and corresponding timestamps; According to the corresponding timestamps, the first sample, the second sample, and the third sample of the day are aligned to obtain the training data of the day.

5. The method according to claim 4, characterized in that Preprocessing the second time series data to obtain a plurality of second samples and corresponding timestamps includes: Eliminate data in the second time series data whose data amount does not meet a preset range, so as to obtain second time series data after data amount determination; Normalizing the second time series data after the data amount determination to obtain a plurality of candidate samples and corresponding timestamps; Selecting a target sample whose value meets a preset range from the plurality of candidate samples; Determine multiple target sample sets based on the timestamps corresponding to the target samples obtained by screening; the target sample sets include multiple target samples belonging to the same timestamp; A plurality of second samples are determined based on the arithmetic mean of the data within a given quantile of the plurality of target sample sets.

6. The method according to claim 4, characterized in that According to the corresponding timestamps, the first sample, the second sample, and the third sample in any period are aligned to obtain the training data of the day, including: Aligning the first sample, the second sample, and the third sample of the day according to the corresponding timestamps to obtain a data point set of the day; the data point set includes data points corresponding to multiple timestamps; the data points are determined based on the first sample, the second sample, and the third sample belonging to the same timestamp; Eliminate data points whose timestamps in the data point set are within a time period of abnormal device operation; Using a density clustering algorithm, target data points meeting preset conditions are screened from the data point set and determined as the training data for the day.

7. A model updating device based on photovoltaic power generation prediction, characterized in that: include: A data preprocessing unit, used to preprocess the sampling data of the day to obtain the training data of the day; The training data includes a first sample, a second sample and a third sample corresponding to a plurality of sampling time points; the first sample represents the power generation of the first device; the second sample represents the power generation of the second device; the third sample represents the observed value of the preset indicator; a characteristic value determining unit, configured to determine a characteristic value corresponding to the data of the current day based on the first sample, the second sample, and the third sample in combination with a characteristic value calculation formula; The characteristic value calculation formula is obtained by analyzing the target model using the least square method; the target model includes a regression model that characterizes the relationship between the power generation of the first device, the power generation of the second device and the preset index at the same time point; A model updating unit is used to calculate the coefficient matrix of the regression model based on the eigenvalues ​​corresponding to the specified date data in the historical time to obtain an updated target model; the updated target model is used to predict the power generation power of the first device.

8. The device according to claim 7, characterized in that The characteristic value determination unit is specifically used for: Based on the analysis of the target model, the characteristic value calculation formula is obtained; Based on the first sample, the second sample and the third sample, combined with the characteristic value calculation formula, the characteristic value corresponding to the data of the day is calculated; the characteristic values ​​include n, C, D, E, F, G, H, L, and M; n is determined based on the total number of the sampling time points; C is determined based on the cumulative sum of the squares of the second samples corresponding to all sampling time points; D is determined based on the cumulative sum of the second samples corresponding to all sampling time points; E is determined based on the cumulative sum of the products of the first samples and the second samples corresponding to all sampling time points; F is determined based on the cumulative sum of the first samples corresponding to all sampling time points; G is determined based on the cumulative sum of the products of the second samples and the third samples corresponding to all sampling times; H is determined based on the cumulative sum of the squares of the third samples corresponding to all sampling times; L is determined based on the cumulative sum of the third samples corresponding to all sampling time points; M is determined based on the cumulative sum of the products of the first samples and the third samples corresponding to all sampling time points.

9. A storage medium, characterized in that: The storage medium includes a stored program, wherein the program, when executed by a processor, executes the model updating method based on photovoltaic power generation prediction according to any one of claims 1 to 6.

10. An electronic device, characterized in that: include: processor, memory, and bus; The processor is connected to the memory via the bus; The memory is used to store programs, and the processor is used to run programs, wherein the program, when run by the processor, executes the model updating method based on photovoltaic power generation prediction according to any one of claims 1 to 6.

Images