Photovoltaic generating capacity prediction method and system based on data sequence resolution compression scale optimization, and storable medium

Abstract

The invention provides a photovoltaic power generation capacity prediction method and system based on data sequence resolution compression scale optimization and a storable medium, and relates to the field of electric power. According to an input photovoltaic power generation capacity data sequence, data compression processing is carried out according to different time resolutions, the prediction accuracy is improved from the perspective of data processing, and the prediction efficiency is improved. And a single hidden layer structure is adopted, and a width learning system (BLS) which expands the number of nodes in width is used as a prediction tool, so that historical data can be well fitted, and the calculation efficiency of the model is ensured. When the method is used for predicting the electric quantity of the photovoltaic power station, the accuracy of a prediction result can be improved, more reliable information support is provided for planning, operation, stable analysis and control and transaction of a power system, and the method has important significance in improving the comprehensive energy efficiency of the system and promoting new energy consumption.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and in particular relates to a photovoltaic power generation prediction method, system and storable medium based on data sequence resolution compression scale optimization. Background Art

[0002] Accurate and reliable photovoltaic power generation forecasts provide information support for power system planning, operation, stability analysis and control, and trading. They are of great significance to improving the overall energy efficiency of the system and promoting the consumption of new energy. They are an important guarantee for the safe, economical and efficient operation of new energy power systems.

[0003] Most existing photovoltaic power point predictions are implemented using neural network deep learning algorithms, which predict by learning the laws displayed by a large amount of data. Therefore, the characteristics of the input data have a great impact on the performance of the prediction model, including the choice of time resolution. When describing the historical data of photovoltaic output for a specific time length, the higher the time resolution selected, the more data samples are obtained, which provides the model with the opportunity to access more information and patterns, so the fitting effect is better. However, for the prediction target of a specific time length, the increase in time resolution will inevitably lead to an increase in the prediction step. This will cause the original problem to become a multi-step prediction. The disadvantage is that the prediction error may accumulate with the increase in the number of steps, resulting in the continuous accumulation and amplification of the error, and this phenomenon is particularly obvious for predicting longer sequences. Therefore, for the point prediction target of photovoltaic output, an algorithm model based on data sequence resolution compression scale optimization is constructed, and the most appropriate time resolution of input data is selected, which can effectively improve the accuracy of point prediction. Summary of the invention

[0004] The purpose of the present invention is to provide a photovoltaic power generation prediction method, system and storage medium based on data sequence resolution compression scale optimization, according to the input photovoltaic power generation data sequence, data compression processing is performed according to different time resolutions, and the prediction accuracy is improved from the perspective of data processing, and a single hidden layer structure is used, and a width learning system (BLS) with a node number expanded in width is used as a prediction tool, which can better fit historical data and ensure the model calculation efficiency. Using this method to predict the power of photovoltaic power stations can improve the accuracy of the prediction results, provide more reliable information support for power system planning, operation, stability analysis and control, and trading, and is of great significance to improving the comprehensive energy efficiency of the system and promoting the consumption of new energy.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] The photovoltaic power generation prediction method based on data sequence resolution compression scale optimization includes:

[0007] Step 1: Collect the power generation data of the photovoltaic power station and the meteorological data that the power station can provide, determine the input factors, and construct the input data sequence;

[0008] Step 2: Remove outliers, classify by weather, classify by season, etc. for the input data sequence;

[0009] Step 3: compress the processed data according to different time resolutions, and perform data normalization to obtain power generation sequences with different time resolutions;

[0010] Step 4: Use data sequences with different time resolutions to train the BLS model, divide it into training set and test set, use the training set to determine the value of each parameter in the BLS model, and use the test set to output the prediction results;

[0011] Step 5: Calculate the accuracy of the prediction results at different time resolutions based on the point prediction evaluation index, select the time resolution with the highest accuracy as the optimal power generation prediction data series resolution compression scale for the photovoltaic power station, and use the BLS model obtained at this time resolution as the power generation prediction model.

[0012] Furthermore, in step 1, when determining the input factors, since the non-sunny irradiance can be calculated by the weather type coefficient, the prediction error can be within 5%, so the measured irradiance value is superimposed with a random error of ±5% as the prediction value and as one of the input factors of the model; since the time interval of the numerical weather forecast is long and there is a certain lag, the weather forecast data is not used as the input of the prediction model, and only the local measured meteorological data of the photovoltaic site at adjacent times is used as a reference; considering that there is a strong correlation between the historical data and the output at the current time in the short-term forecast and the ultra-short-term forecast, the historical output at several adjacent times is used as the input factor.

[0013] Furthermore, step 2 specifically includes:

[0014] Step 2.1: Classify the data according to whether it is sunny or not. The judgment method is: set the maximum difference threshold to distinguish between fluctuation and non-fluctuation. It is considered that a scenario where the output difference between any two adjacent points in a day does not exceed 10% of the power station capacity is defined as a sunny day; otherwise, it is a non-sunny day.

[0015] Step 2.2: Divide the data by season. The evaluation method is: divide according to the date of the data. In my country, March to May is spring, June to August is summer, September to November is autumn, and December to February of the following year is winter.

[0016] Step 2.3: Remove the dates with 12 consecutive missing or abnormal data from the data set. For missing data and abnormal phenomena on other dates, linear interpolation is used to fill them in.

[0017] Furthermore, step 3 specifically includes:

[0018] Step 3.1: Change the time resolution of the input data and define the data compression process as a mapping process F: For any N 4-hour PV output data, it can be represented as a 5-minute PV output sequence D containing the N 4-hour output information. 48×N At the same time, there is a series of compressed photovoltaic output sequences D containing the same information C×N Correspondingly,

[0019] F:D 48×N →D C×N (1)

[0020] Where: C = 48 / k, k∈N + , C∈N + , k is defined as the compression resolution. Generally speaking, the selectable k∈K={1,2,3,6,12,24,48} ​​means that the historical photovoltaic power generation data is counted according to the time intervals of 5 minutes, 10 minutes, 15 minutes, 30 minutes, 1 hour, 2 hours, and 4 hours, respectively. When k is selected as 48, the 4-hour statistical data is used, and the model is a single-step prediction; when k is selected as 1, it means that data compression is not performed, and the results of the output 48 points need to be averaged as one photovoltaic power prediction data. At this time, the prediction step length is the longest, which is 48 steps.

[0021] Step 3.2: Perform the Min-Max Normalization method on the data, also known as "deviation standardization", which is to perform linear transformation on the original data and map the value range to [0,1] to standardize the data. The formula is as follows:

[0022] X N =(XX min ) / (X max -X min ) (2)

[0023] Where: X is the original data, Xmin and Xmax are the minimum and maximum values ​​of the column data respectively, and X N It is the new value after normalization. Normalization can transform features of different scales or units to the same scale, so that the model will not deviate due to the scale of the features during training; and maintain the relative size relationship of each feature in the original data.

[0024] Furthermore, step 4 specifically includes:

[0025] Step 4.1: Divide the training set and the test set, select the last 10 days of data in each season as test data, and the rest as training data.

[0026] Step 4.2: Extract features from the input to obtain a set of BLS model feature nodes. Since sparse feature learning has high computational efficiency and can effectively express the characteristics of the data set, thereby improving the efficiency of the prediction model, sparse learning based on the Least Absolute Shrinkage and Selection Operator (LASSO) is used to implement the construction of BLS feature nodes. Solve the optimization problem shown in formula (3):

[0027]

[0028] Where: X is the input data for feature extraction, and there are L moments in the training set (denoted as t1, t2, …, t L ), then the input matrix of BLS is X train =[x(t1),x(t2),…,x(t L )] T ; W F is the coefficient matrix of feature selection; Z is the randomly generated feature matrix. The first item is to solve the linear equation XW F -The least squares estimate of Z; the second term is the l1 norm, which is used to estimate W F The sparsity constraint is imposed; λ is a regularization parameter used to control the relative size between the two items. The alternating direction method of multipliers (ADMM) is used to solve it. By introducing the auxiliary variable V, equation (3) can be equivalently expressed as:

[0029]

[0030] Where: g(V)=λ||V||1. Consider the augmented Lagrangian function of this optimization problem:

[0031] L(W F ,V,Λ)=f(W F )+g(V)+Λ T (W F -V) (5)

[0032] Where: Λ is the Lagrange multiplier of the optimization problem. Based on ADMM, W can be updated alternately F , V and Λ to solve the equation. The iterative process is as follows:

[0033]

[0034] Where: I is the unit matrix; S is the soft threshold function, and its expression is:

[0035]

[0036] The feature extraction of input factors only needs to be iteratively calculated according to formula (6). The input matrix X is composed of the input factor vectors required for photovoltaic power generation prediction at each moment. Suppose there are L moments (denoted as t1, t2, …, t L ), then X=[x(t1),x(t2),…,x(t L )] T X obtains a series of feature nodes through the above feature extraction method, where the i-th group of feature maps is obtained by the following formula:

[0037] F i =φ(XW F,i +β F,i ) i=1,2,L,N g (8)

[0038] Where: The weight matrix W associated with the feature node F,i and the bias term β F,i are all randomly generated, and the value range of each element is [-1,1]; the mapping function φ is indirectly given by formula (6).

[0039] Step 4.3, solve the BLS model parameter W, assuming that BLS includes N g Group feature maps (each group contains N f feature nodes) and N e Enhanced nodes, and the grid search method is used to determine its parameters N g , N f and N e The combination of feature maps obtained in step 4.2 is recorded as the matrix F = [F1, F2, ..., F Ng ], this N g *N f The feature nodes generate the enhanced nodes of BLS according to formula (9):

[0040] E=ζ(FW E +β E ) (9)

[0041] Where: The weight matrix W associated with the enhanced node E and the bias term β E It is also randomly generated, and the value range of each element is [-1,1]; ζ(·) is the activation function, and the hyperbolic tangent function is selected here, and its expression is:

[0042]

[0043] Then the mathematical model of BLS can be expressed as:

[0044]

[0045] Where: W represents the weight connecting from feature nodes and enhancement nodes to the output layer; is the output of the neural network, which is the predicted value in the photovoltaic prediction model. In order to simplify the expression, the matrix corresponding to the BLS middle layer (including feature nodes and enhancement nodes) is recorded as H = [F, E].

[0046] During the training process, the randomly generated weights W F,i With W E , bias β F,i With β E All of them no longer change, so we only need to solve the output weight W. Therefore, the training of BLS is equivalent to solving the least squares solution of the linear system Y = HW:

[0047]

[0048] Where: is the Moore-Penrose generalized inverse matrix of matrix H, also known as the pseudoinverse matrix; Y is the target output matrix of BLS, which is given by the training set. Since the prediction target set in this paper is 4 hours in the future, multi-step prediction is required, so the output data corresponding to each moment is a vector, that is, Y = [y(t1), y(t2),…, y(t L )] T ,in

[0049]

[0050] Where: y(t) is the photovoltaic power generation at time t; is the sampling interval of photovoltaic power generation; C is the number of steps for ultra-short-term prediction.

[0051] Furthermore, in step 5, the normalized root mean square error (nRMSE) is used to evaluate and analyze the prediction error, and its expression is:

[0052]

[0053] Where: N s is the total number of samples; P cap is the installed capacity of photovoltaic power generation system; i is the measured value; is the predicted value. In order to reflect the impact of the prediction step on the prediction error, this paper calculates nRMSE for different prediction time scales.

[0054] The solution of the present invention also includes a photovoltaic power generation prediction system based on data sequence resolution compression scale optimization, the system comprising:

[0055] The acquisition module is used to collect the power generation data of the photovoltaic power station and the meteorological data that the power station can provide, determine the input factors, and construct the input data sequence;

[0056] The data preprocessing module removes outliers, classifies by weather, and classifies by season for the input data sequence;

[0057] The time resolution compression processing module compresses the processed data according to different time resolutions and performs data normalization operations to obtain power generation sequences with different time resolutions;

[0058] Model training module, using data sequences with different time resolutions to train the BLS model, dividing it into training set and test set, using the training set to determine the value of each parameter in the BLS model, and using the test set to output the prediction results;

[0059] The optimal compression scale module is screened, and the accuracy of the prediction results at different time resolutions is calculated according to the point prediction evaluation index. The time resolution with the highest accuracy is selected as the optimal power generation prediction data series resolution compression scale of the photovoltaic power station, and the BLS model obtained at this time resolution is used as the power generation prediction model.

[0060] The present invention also includes a storable medium, including: a processor and a memory coupled to the processor, the memory storing a computer program, and the computer program, when executed by the processor, implements the steps of the photovoltaic power generation prediction method based on data sequence resolution compression scale optimization.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] The present invention considers processing photovoltaic output historical data with different time resolutions as the input data of the model, compares the changes in the errors of the prediction results to the fixed prediction target under different time resolutions, optimizes the prediction model according to the problem of the non-linear relationship between the prediction error and the compression scale, proposes a photovoltaic power generation prediction method and system based on data sequence resolution compression scale optimization, and uses the BLS model to reflect higher prediction accuracy and computational efficiency in photovoltaic power point prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1The present invention is a flow chart of a photovoltaic power generation prediction method based on data sequence resolution compression scale optimization.

[0064] Figure 2 2 is a network structure diagram of a breadth learning system (BLS) model in an embodiment. DETAILED DESCRIPTION

[0065] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0066] The following detailed description is an exemplary description, which is intended to provide further detailed description of the present invention. Unless otherwise specified, all technical terms used in the present invention have the same meaning as those generally understood by those skilled in the art to which the present application belongs. The terms used in the present invention are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention.

[0067] See also Figure 1 As shown,

[0068] The present invention provides a photovoltaic power generation prediction method based on data sequence resolution compression scale optimization, comprising:

[0069] Step 1: Collect the power generation data of the photovoltaic power station and the meteorological data that the power station can provide, determine the input factors, and construct the input data sequence;

[0070] Among them, when determining the input factors, since the non-sunny irradiance can be calculated by the weather type coefficient, the prediction error can be within 5%, so the measured irradiance value is superimposed with a random error of ±5% as the prediction value and as one of the input factors of the model; since the time interval of numerical weather forecast is long and there is a certain lag, the weather forecast data is not used as the input of the prediction model, and only the local measured meteorological data of the photovoltaic site at adjacent times is used as a reference; considering that there is a strong correlation between the historical data and the output at the current time in short-term and ultra-short-term forecasts, the historical output at several adjacent times is used as the input factor.

[0071] Step 2: Remove outliers, classify by weather, classify by season, etc. for the input data sequence;

[0072] Step 2.1: Classify the data according to whether it is sunny or not. The judgment method is: set the maximum difference threshold to distinguish between fluctuation and non-fluctuation. It is considered that the scenario where the output difference between any two adjacent points in a day does not exceed 10% of the power station capacity is defined as a sunny day; otherwise, it is a non-sunny day.

[0073] Step 2.2: Divide the data by season. The evaluation method is: divide the data according to the date. In my country, March to May is spring, June to August is summer, September to November is autumn, and December to February of the following year is winter.

[0074] Step 2.3: Remove the dates with 12 consecutive missing or abnormal data from the data set. For missing data and abnormal phenomena on other dates, linear interpolation is used to fill them in.

[0075] Step 3: compress the processed data according to different time resolutions, and perform data normalization to obtain power generation sequences with different time resolutions;

[0076] Step 3.1: Change the time resolution of the input data and define the data compression process as a mapping process F: For any N 4-hour PV output data, it can be represented as a 5-minute PV output sequence D containing the N 4-hour output information. 48×N At the same time, there is a series of compressed photovoltaic output sequences D containing the same information C×N Correspondingly,

[0077] F:D 48×N →D C×N (15)

[0078] Where: C = 48 / k, k∈N + , C∈N + , k is defined as the compression resolution. Generally speaking, the selectable k∈K={1,2,3,6,12,24,48} ​​means that the historical photovoltaic power generation data is counted according to the time intervals of 5 minutes, 10 minutes, 15 minutes, 30 minutes, 1 hour, 2 hours, and 4 hours, respectively. When k is selected as 48, the 4-hour statistical data is used, and the model is a single-step prediction; when k is selected as 1, it means that data compression is not performed, and the results of the output 48 points need to be averaged as one photovoltaic power prediction data. At this time, the prediction step length is the longest, which is 48 steps.

[0079] Step 3.2: Perform the Min-Max Normalization method on the data, also known as "deviation standardization", which is to perform linear transformation on the original data and map the value range to [0,1] to standardize the data. The formula is as follows:

[0080] X N =(XX min ) / (X max -X min ) (16)

[0081] Where: X is the original data, Xmin and Xmax are the minimum and maximum values ​​of the column data respectively, and X N It is the new value after normalization. Normalization can transform features of different scales or units to the same scale, so that the model will not deviate due to the scale of the features during training; and maintain the relative size relationship of each feature in the original data.

[0082] Step 4: Use data sequences with different time resolutions to train the BLS model, divide it into training set and test set, use the training set to determine the value of each parameter in the BLS model, and use the test set to output the prediction results;

[0083] Step 4.1: Divide the training set and the test set, select the last 10 days of data in each season as test data, and the rest as training data.

[0084] Step 4.2: Extract features from the input to obtain a set of BLS model feature nodes. Since sparse feature learning has high computational efficiency and can effectively express the characteristics of the data set, thereby improving the efficiency of the prediction model, sparse learning based on the Least Absolute Shrinkage and Selection Operator (LASSO) is used to implement the construction of BLS feature nodes. Solve the optimization problem shown in formula (3):

[0085]

[0086] Where: X is the input data for feature extraction, and there are L moments in the training set (denoted as t1, t2, …, t L ), then the input matrix of BLS is X train =[x(t1),x(t2),…,x(t L )] T ; W F is the coefficient matrix of feature selection; Z is the randomly generated feature matrix. The first item is to solve the linear equation XW F -The least squares estimate of Z; the second term is the l1 norm, which is used to estimate W F The sparsity constraint is imposed; λ is a regularization parameter used to control the relative size between the two items. The alternating direction method of multipliers (ADMM) is used to solve it. By introducing the auxiliary variable V, equation (3) can be equivalently expressed as:

[0087]

[0088] Where: g(V)=λ||V||1. Consider the augmented Lagrangian function of this optimization problem:

[0089] L(W F ,V,Λ)=f(W F )+g(V)+Λ T (W F -V) (19)

[0090] Where: Λ is the Lagrange multiplier of the optimization problem. Based on ADMM, W can be updated alternately F , V and Λ to solve the equation. The iterative process is as follows:

[0091]

[0092] Where: I is the unit matrix; S is the soft threshold function, and its expression is:

[0093]

[0094] The feature extraction of input factors only needs to be iteratively calculated according to formula (6). The input matrix X is composed of the input factor vectors required for photovoltaic power generation prediction at each moment. Suppose there are L moments (denoted as t1, t2, …, t L ), then X=[x(t1),x(t2),…,x(t L )] T X obtains a series of feature nodes through the above feature extraction method, where the i-th group of feature maps is obtained by the following formula:

[0095] F i =φ(XW F,i +β F,i ) i=1,2,L,N g (twenty two)

[0096] Where: The weight matrix W associated with the feature node F,i and the bias term β F,i are all randomly generated, and the value range of each element is [-1,1]; the mapping function φ is indirectly given by formula (6).

[0097] Step 4.3, solve the BLS model parameter W, assuming that BLS includes N g Group feature maps (each group contains N f feature nodes) and N e Enhanced nodes, and the grid search method is used to determine its parameters N g , N f and N e The combination of feature maps obtained in step 4.2 is recorded as the matrix F = [F1, F2, ..., F Ng ], this Ng *N f The feature nodes generate the enhanced nodes of BLS according to formula (9):

[0098] E=ζ(FW E +β E ) (twenty three)

[0099] Where: The weight matrix W associated with the enhanced node E and the bias term β E It is also randomly generated, and the value range of each element is [-1,1]; ζ(·) is the activation function, and the hyperbolic tangent function is selected here, and its expression is:

[0100]

[0101] Then the mathematical model of BLS can be expressed as:

[0102]

[0103] Where: W represents the weight connecting from feature nodes and enhancement nodes to the output layer; is the output of the neural network, which is the predicted value in the photovoltaic prediction model. In order to simplify the expression, the matrix corresponding to the BLS middle layer (including feature nodes and enhancement nodes) is recorded as H = [F, E].

[0104] During the training process, the randomly generated weights W F,i With W E , bias β F,i With β E All of them no longer change, so we only need to solve the output weight W. Therefore, the training of BLS is equivalent to solving the least squares solution of the linear system Y = HW:

[0105]

[0106] Where: is the Moore-Penrose generalized inverse matrix of matrix H, also known as the pseudoinverse matrix; Y is the target output matrix of BLS, which is given by the training set. Since the prediction target set in this paper is 4 hours in the future, multi-step prediction is required, so the output data corresponding to each moment is a vector, that is, Y = [y(t1), y(t2),…, y(t L )] T ,in

[0107]

[0108] Where: y(t) is the photovoltaic power generation at time t; is the sampling interval of photovoltaic power generation; C is the number of steps for ultra-short-term prediction.

[0109] Step 5: Calculate the accuracy of the prediction results at different time resolutions based on the point prediction evaluation index, select the time resolution with the highest accuracy as the optimal power generation prediction data series resolution compression scale for the photovoltaic power station, and use the BLS model obtained at this time resolution as the power generation prediction model.

[0110] Among them, the normalized Root-Mean-Square Error (nRMSE) is used to evaluate and analyze the prediction error, and its expression is:

[0111]

[0112] Where: N s is the total number of samples; P cap is the installed capacity of photovoltaic power generation system; i is the measured value; is the predicted value. In order to reflect the impact of the prediction step on the prediction error, this paper calculates nRMSE for different prediction time scales.

[0113] Example 1

[0114] Taking the photovoltaic output and meteorological data of a photovoltaic power station in Yulara, Australia in the summer of 2017 and 2018 (January, February and December) as an example, the power generation in the next 4 hours is predicted;

[0115] Collect the power generation data of the photovoltaic power station and the meteorological data (including solar irradiance, temperature, and wind speed measurement data) that the power station can provide, determine the input factors, and construct the input data sequence; the photovoltaic information, data format, and prediction model input and output selection of the photovoltaic power station are shown in Table 1:

[0116] Table 1 Basic information of Yulara PV dataset in summer

[0117]

[0118]

[0119] The steps of the power system key node identification method are as follows:

[0120] According to step 2.1, classify the data according to whether it is sunny or not;

[0121] Divide the data by season according to step 2.2;

[0122] Handle missing and outliers data according to step 2.3;

[0123] Compress data according to different time resolutions according to formula (1) in step 3.1;

[0124] Normalize the data according to formula (2) in step 3.2;

[0125] Divide the training set and test set according to step 4.1;

[0126] Iteratively calculate the characteristic parameter W of the input data according to formula (6) in step 4.2 F ;

[0127] Solve the weight W from the node layer to the output layer in the BLS model according to formula (12) in step 4.3 and obtain the prediction result y(t) according to formula (13);

[0128] The prediction error is evaluated and analyzed according to formula (14) in step 5, and the results are as follows:

[0129]

[0130] It can be seen that there is no obvious linear relationship between the prediction error nRMSE and the compression scale, so it is difficult to find a universal compression scale k. According to the prediction errors recorded in Table 1, the error is the smallest when 10 minutes is selected as the data resolution compression scale on a sunny day, and the error is the smallest when 4 hours is selected as the data resolution compression scale on a sunny day. Compared with not performing data resolution compression, the prediction accuracy is improved. Therefore, the method proposed in the present invention can improve the prediction accuracy to a certain extent in the problem of compression scale selection.

[0131] Example 2

[0132] The present invention also provides a photovoltaic power generation prediction system based on data sequence resolution compression scale optimization, comprising:

[0133] The acquisition module is used to collect the power generation data of the photovoltaic power station and the meteorological data that the power station can provide, determine the input factors, and construct the input data sequence;

[0134] The data preprocessing module removes outliers, classifies by weather, and classifies by season for the input data sequence;

[0135] The time resolution compression processing module compresses the processed data according to different time resolutions and performs data normalization operations to obtain power generation sequences with different time resolutions;

[0136] Model training module, using data sequences with different time resolutions to train the BLS model, dividing it into training set and test set, using the training set to determine the value of each parameter in the BLS model, and using the test set to output the prediction results;

[0137] The optimal compression scale module is screened, and the accuracy of the prediction results at different time resolutions is calculated according to the point prediction evaluation index. The time resolution with the highest accuracy is selected as the optimal power generation prediction data series resolution compression scale of the photovoltaic power station, and the BLS model obtained at this time resolution is used as the power generation prediction model.

[0138] Example 3

[0139] The present invention also provides a photovoltaic power generation prediction system based on data sequence resolution compression scale optimization, comprising: a processor and a memory coupled to the processor, the memory storing a computer program, and the computer program, when executed by the processor, implements the steps of the photovoltaic power generation prediction method based on data sequence resolution compression scale optimization.

[0140] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0141] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0142] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0143] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A photovoltaic power generation prediction method based on data sequence resolution compression scale optimization, characterized in that: include: Step 1: Collect the power generation data of the photovoltaic power station and the meteorological data that the power station can provide, determine the input factors, and construct the input data sequence; Step 2: Remove outliers, classify by weather, and classify by season for the input data sequence; Step 3: compress the processed data according to different time resolutions, and perform data normalization to obtain power generation sequences with different time resolutions; Step 4: Use data sequences with different time resolutions to train the BLS model, divide it into training set and test set, use the training set to determine the value of each parameter in the BLS model, and use the test set to output the prediction results; Step 5: Calculate the accuracy of the prediction results at different time resolutions based on the point prediction evaluation index, select the time resolution with the highest accuracy as the optimal power generation prediction data series resolution compression scale for the photovoltaic power station, and use the BLS model obtained at this time resolution as the power generation prediction model.

2. The photovoltaic power generation prediction method based on data sequence resolution compression scale optimization according to claim 1 is characterized in that: In step 1, when determining the input factors, the measured irradiance value is superimposed with a random error of ±5% as a predicted value and used as one of the input factors of the model; the local measured meteorological data of the photovoltaic site at adjacent times is used as a reference; and the historical output at several adjacent times is used as an input factor.

3. The photovoltaic power generation prediction method based on data sequence resolution compression scale optimization according to claim 2 is characterized in that: The step 2 specifically includes: Step 2.1: Classify the data according to whether it is sunny or not. The judgment method is: set the maximum difference threshold to distinguish between fluctuation and non-fluctuation. It is considered that a scenario where the output difference between any two adjacent points in a day does not exceed 10% of the power station capacity is defined as a sunny day; otherwise, it is a non-sunny day. Step 2.2: Divide the data by season. The judging method is: according to the date of the data, in my country, March to May is spring, June to August is summer, September to November is autumn, and December to the following February is winter; Step 2.3: Remove the dates with 12 consecutive missing or abnormal data from the data set, and use linear interpolation to fill in the missing data and abnormal phenomena in other dates.

4. The photovoltaic power generation prediction method based on data sequence resolution compression scale optimization according to claim 3 is characterized in that: The step 3 specifically includes: Step 3.1: Change the time resolution of the input data and define the data compression process as a mapping process F: For any N 4-hour PV output data, it is represented as a 5-minute PV output sequence D containing the N 4-hour output information. 48×N At the same time, there is a series of compressed photovoltaic output sequences D containing the same information C×N Correspondingly, F:D 48×N →D C×N (1) Where: C = 48 / k, k∈N + , C∈N + , k is defined as compression resolution, k∈K={1,2,3,6,12,24,48}, which means that the historical photovoltaic power generation data is counted according to the time intervals of 5 minutes, 10 minutes, 15 minutes, 30 minutes, 1 hour, 2 hours, and 4 hours, respectively. When k is selected as 48, the 4-hour statistics are used, and the model is a single-step prediction; when k is selected as 1, it means that data compression is not performed, and the results of the output 48 points need to be averaged as one photovoltaic power prediction data. At this time, the prediction step length is the longest, which is 48 steps; Step 3.2: Perform the Min-Max Normalization method on the data, also known as "deviation standardization", which performs linear transformation on the original data and maps the value range to [0,1] to standardize the data. The formula is as follows: X N =(X-X min ) / (X max -X min ) (2) Where: X is the original data, Xmin and Xmax are the minimum and maximum values ​​of the column data respectively, and X N It is a new value after normalization. Normalization converts features of different scales or units to the same scale, so that the model will not deviate due to the scale of the features during training; and maintain the relative size relationship of each feature in the original data.

5. The photovoltaic power generation prediction method based on data sequence resolution compression scale optimization according to claim 4 is characterized in that: The step 4 specifically includes: Step 4.1, divide the training set and the test set, select the last 10 days of data in each season as test data, and the rest as training data; Step 4.2: Extract features from the input to obtain a set of BLS model feature nodes. Use sparse learning based on the Least Absolute Shrinkage and Selection Operator (LASSO) to construct the BLS feature nodes and solve the optimization problem shown in formula (3): Where: X is the input data for feature extraction, and there are L moments in the training set (denoted as t1, t2, …, t L ), then the input matrix of BLS is X train =[x(t1),x(t2),…,x(t L )] T ; W F is the coefficient matrix for feature selection; Z is a randomly generated feature matrix, and the first item is to solve the linear equation XW F -The least squares estimate of Z; the second term is the l1 norm, which is used to estimate W F Sparsity restriction is imposed; λ is a regularization parameter used to control the relative size between the two items. The alternating direction method of multipliers (ADMM) is used to solve it. By introducing the auxiliary variable V, formula (3) is equivalently expressed as: Where: g(V)=λ||V||1, considering the augmented Lagrangian function of the optimization problem: L(W F ,V,Λ)=f(W F )+g(V)+Λ T (W F -V) (5) Where: Λ is the Lagrange multiplier of the optimization problem, based on ADMM, and W is updated alternately F , V and Λ to solve the equation, the iterative process is as follows: Where: I is the unit matrix; S is the soft threshold function, and its expression is: The feature extraction of input factors is iteratively calculated according to formula (6); the input matrix X is composed of the input factor vectors required for photovoltaic power generation prediction at each moment. Assume that there are L moments (denoted as t1, t2, …, t L ), then X=[x(t1),x(t2),…,x(t L )] T , X obtains a series of feature nodes through the above feature extraction method, where the i-th group of feature maps is obtained by the following formula: F i =φ(XW F,i +b F,i ) i=1,2,L,N g (8) Where: The weight matrix W associated with the feature node F,i and the bias term β F,i are all randomly generated, and the value range of each element is [-1,1]; the mapping function φ is indirectly given by formula (6); Step 4.3, solve the BLS model parameter W, assuming that BLS includes N g Group feature maps, each group contains N f feature nodes and N e Enhanced nodes, and the grid search method is used to determine its parameters N g , N f and N e The optimal value of , the combination of feature maps obtained in step 4.2 is recorded as a matrix This g *N f The feature nodes generate the enhanced nodes of BLS according to formula (9): E=ζ(FW E +b E ) (9) Where: The weight matrix W associated with the enhanced node E and the bias term β E Randomly generated, the value range of each element is [-1,1]; ζ(·) is the activation function, and the hyperbolic tangent function is selected, and its expression is: Then the mathematical model of BLS is expressed as: Where: W represents the weight connecting from feature nodes and enhancement nodes to the output layer; is the output of the neural network, which is the predicted value in the photovoltaic prediction model. The matrix corresponding to the BLS middle layer (including feature nodes and enhancement nodes) is recorded as H = [F, E]; During the training process, the randomly generated weights W F,i With W E , bias β F,i With β E The training of BLS is equivalent to solving the least squares solution of the linear system Y = HW: Where: is the Moore-Penrose generalized inverse matrix of the matrix H, also called the pseudoinverse matrix; Y is the target output matrix of BLS, which is given by the training set. The output data corresponding to each moment is a vector, that is, Y = [y(t1), y(t2),…, y(t L )] T ,in Where: y(t) is the photovoltaic power generation at time t; is the sampling interval of photovoltaic power generation; C is the number of steps for ultra-short-term prediction.

6. The photovoltaic power generation prediction method based on data sequence resolution compression scale optimization according to claim 5 is characterized in that: In step 5, the normalized root mean square error (nRMSE) is used to evaluate and analyze the prediction error, and its expression is: Where: N s is the total number of samples; P cap is the installed capacity of photovoltaic power generation system; i is the measured value; For the predicted value, nRMSE is calculated for different prediction time scales.

7. A system using the photovoltaic power generation prediction method based on data sequence resolution compression scale optimization as described in any one of claims 1 to 6, characterized in that: include: The acquisition module is used to collect the power generation data of the photovoltaic power station and the meteorological data that the power station can provide, determine the input factors, and construct the input data sequence; The data preprocessing module removes outliers, classifies by weather, and classifies by season for the input data sequence; The time resolution compression processing module compresses the processed data according to different time resolutions and performs data normalization operations to obtain power generation sequences with different time resolutions; Model training module, using data sequences with different time resolutions to train the BLS model, dividing it into training set and test set, using the training set to determine the value of each parameter in the BLS model, and using the test set to output the prediction results; The optimal compression scale module is screened, and the accuracy of the prediction results at different time resolutions is calculated according to the point prediction evaluation index. The time resolution with the highest accuracy is selected as the optimal power generation prediction data series resolution compression scale of the photovoltaic power station, and the BLS model obtained at this time resolution is used as the power generation prediction model.

8. A computer storable medium, characterized in that: include: A processor and a memory coupled to the processor, the memory storing a computer program, wherein the computer program, when executed by the processor, implements the steps of a photovoltaic power generation prediction method based on data sequence resolution compression scale optimization according to any one of claims 1 to 6.

Images