Meshed numerical weather feature dimension reduction extraction method and system

Abstract

The present application belongs to the field of grid numerical weather feature extraction, and provides a grid numerical weather feature dimension reduction extraction method and system. Among them, the method comprises obtaining the corresponding relationship of load type and load power of a high-proportion new energy power system; according to the typical daily grid numerical weather measured data of the region to which different types of loads belong, a plurality of meteorological factors are obtained, and the correlation of the load power of different load types is calculated, the meteorological factors related to different load types are screened out, and an initial meteorological factor feature matrix is constructed; the initial meteorological factor feature matrix is reduced in dimension by using a principal component analysis method, and a reduced meteorological factor feature matrix is obtained; the reduced meteorological factor feature matrix is iteratively learned by using a preset deep learning model, so as to refine the grid meteorological feature scale, further extract the hidden features of different load types, and obtain a grid feature matrix representing different types of loads, which is used for load prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of grid numerical weather feature extraction, and particularly relates to a grid numerical weather feature dimension reduction extraction method and system. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] New energy power system refers to an electric power system formed by comprehensively utilizing renewable energy such as solar energy, wind energy, water energy, biomass energy and geothermal energy. The high-proportion new energy power system is to increase the utilization proportion of new energy to a high level. The load characteristics of the high-proportion new energy power system refer to the law that the active power and reactive power of the power load of the high-proportion new energy power system are changed with the change of the voltage at the load end and the system frequency. This characteristic is a core element in the operation of the high-proportion new energy power system, which deeply affects the overall performance of the power system. As the final consumer of electric energy, the change of the load characteristics is directly related to the analysis, design and control strategy of the high-proportion new energy power system.

[0004] However, in actual industrial production and life, the load types of the high-proportion new energy power system are various, different types of loads have different representative characteristics and power consumption behaviors, and the historical data of the load and the information content of the grid numerical weather forecast data are huge, so the difficulty and complexity of direct analysis are also great. SUMMARY

[0005] In order to solve the above problems, the present application provides a grid numerical weather feature dimension reduction extraction method and system, which can accurately and quickly extract weather features for load prediction.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] The first aspect of the present application provides a grid numerical weather feature dimension reduction extraction method.

[0008] A load feature extraction method based on grid numerical weather forecast feature dimension reduction, comprising:

[0009] Typical daily load curves of different types of loads in the high-proportion new energy power system are obtained to obtain the correspondence between the load types and the load power of the high-proportion new energy power system;

[0010] The multi-element meteorological factors are obtained according to the typical day grid numerical weather measured data of the region to which different types of loads belong, and then the correlation of the load power of different load types is calculated, the meteorological factors related to different load types are screened out, and an initial meteorological factor feature matrix is constructed.

[0011] The initial meteorological factor feature matrix is dimensionally reduced by using a principal component analysis method to obtain a dimensionally reduced meteorological factor feature matrix.

[0012] The dimensionally reduced meteorological factor feature matrix is iteratively learned by using a preset deep learning model to refine the grid meteorological feature scale, further extract the hidden features of different load types, and obtain a grid feature matrix representing different types of loads for load prediction.

[0013] The second aspect of the present application provides a grid numerical weather feature dimension reduction extraction system.

[0014] A grid numerical weather feature dimension reduction extraction system comprises:

[0015] A load type and power relationship determination module is configured to obtain typical daily load curves of different types of loads in a high-proportion new energy power system, and obtain the corresponding relationship between the load type and the load power of the high-proportion new energy power system.

[0016] A meteorological factor feature matrix construction module is configured to obtain multi-element meteorological factors according to the typical day grid numerical weather measured data of the region to which different types of loads belong, and then calculate the correlation of the load power of different load types, screen out meteorological factors related to different load types, and construct an initial meteorological factor feature matrix.

[0017] A meteorological factor feature matrix dimension reduction module is configured to dimensionally reduce the initial meteorological factor feature matrix by using a principal component analysis method to obtain a dimensionally reduced meteorological factor feature matrix.

[0018] A grid meteorological feature extraction module is configured to iteratively learn the dimensionally reduced meteorological factor feature matrix by using a preset deep learning model to refine the grid meteorological feature scale, further extract the hidden features of different load types, and obtain a grid feature matrix representing different types of loads for load prediction.

[0019] The third aspect of the present application provides a computer readable storage medium.

[0020] A computer readable storage medium has a computer program stored thereon, and the program is executed by a processor to implement the steps of the grid numerical weather feature dimension reduction extraction method described above.

[0021] The fourth aspect of the present application provides an electronic device.

[0022] An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the method for grid numerical weather feature dimensionality reduction when executing the program.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] (1) The present application obtains the correspondence between the load type and the load power of the high-proportion new energy power system by the typical daily load curve of different types of loads in the high-proportion new energy power system, can explore the similarities and differences of the power between the same type of loads in different seasons and the same type of loads in different categories in the same season, and mine the characteristics of the load power consumption behavior.

[0025] (2) The present application obtains the multivariate meteorological factors according to the grid numerical weather measured data of the typical day of the region to which the different types of loads belong, and then calculates the correlation of the load power of different load types, selects the meteorological factors related to different load types, realizes the preliminary screening of the multidimensional meteorological data, simplifies the homogeneous meteorological data, avoids data redundancy, eliminates low-correlation meteorological factors, and improves the reliability.

[0026] (3) The present application uses the principal component analysis method to reduce the dimension of the initial meteorological factor feature matrix to obtain the reduced meteorological factor feature matrix, realizes the dimensionality reduction processing of the multivariate meteorological data, represents the original high-dimensional information through the construction of a low-dimensional expression, reduces the data dimension while retaining the original information, and reduces the calculation complexity.

[0027] (4) The present application uses a preset deep learning model to iteratively learn the reduced meteorological factor feature matrix, further extracts the grid hidden features of the region to which the different load types belong, obtains the grid feature matrix representing the different types of loads, and uses it for load prediction; wherein the different load type features are mainly spatial features, including latitude and longitude information to provide a theoretical basis for load prediction. BRIEF DESCRIPTION OF DRAWINGS

[0028] The drawings accompanying the specification of the present application form a part thereof, serve to provide further understanding of the present application, and together with the exemplary embodiments of the present application and their description, explain the present application, and do not constitute an improper limitation of the present application.

[0029] Figure 1 The grid numerical weather feature dimensionality reduction method flowchart of the embodiment of the present application;

[0030] Figure 2 The grid numerical weather feature dimensionality reduction system structure schematic diagram of the embodiment of the present application;

[0031] Figure 3A flow chart of a multivariate weather data dimension reduction process based on a Pearson correlation coefficient and principal component analysis according to an embodiment of the present application.

[0032] Figure 4 An algorithm diagram for a Pearson correlation coefficient-based correlation analysis of multivariate load and load data according to an embodiment of the present application.

[0033] Figure 5 An algorithm diagram for principal component analysis-based extraction of principal components of a high-dimensional weather matrix and dimension reduction visualization according to an embodiment of the present application.

[0034] Figure 6 A residual neural network Res-Net residual module structure diagram according to an embodiment of the present application.

[0035] Figure 7 A load feature dimension reduction module example diagram based on a residual neural network Res-Net according to an embodiment of the present application. DETAILED DESCRIPTION

[0036] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0037] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application pertains.

[0038] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and furthermore, it should be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of a feature, step, operation, device, component, and / or combinations thereof.

[0039] Figure 1 A gridded numerical weather feature dimension reduction extraction method flow chart according to an embodiment of the present application. According to the present application, a gridded numerical weather feature dimension reduction extraction method is provided. Figure 1 According to the present application, a gridded numerical weather feature dimension reduction extraction method is provided.

[0040] S101: Obtain a typical daily load curve of different types of load based on gridded numerical weather feature dimension reduction extraction, to obtain a load type and load power correspondence relationship of a high-proportion new energy power system.

[0041] The load types in the high-proportion new energy power system can be classified according to voltage levels or load application fields. For example, according to the voltage level of the grid-connected distributed photovoltaic, the load can be classified into 220V voltage level, 380V voltage level and 10kV voltage level; according to the load application field, the load can be classified into scientific research industry, investment industry, manufacturing industry, education industry, processing industry and other different industry fields.

[0042] Among them, the specific date corresponding to the typical day of the peak load is determined, and the typical day load curve of different types of loads is extracted.

[0043] The typical day refers to a day in a season that has representative load characteristics. By drawing the 24-hour load curve of the typical day, the change rule and characteristics of the load in the season can be better understood. The typical day includes summer typical day, winter typical day, spring typical day and autumn typical day. Due to the large difference in climate conditions and people's activity habits in different seasons, the load characteristics also show obvious seasonal changes. The summer typical day is usually affected by high-temperature weather and air conditioning power consumption, and the load peak often occurs in the evening; while the winter typical day is affected by heating equipment power consumption, and the load peak may occur in the morning and evening peak period. The load characteristics in spring and autumn are relatively stable, but they are still affected by some special factors, such as the Spring Festival, National Day and other holidays, as well as irregular large-scale activity power consumption, etc.

[0044] S102: According to the typical day grid data of the region where the different types of loads belong to, the multi-element weather factor is obtained, and then the correlation between the load power of different load types is calculated, the weather factor related to different load types is selected, and the initial weather factor feature matrix is constructed, as shown in Figure 4 .

[0045] Among them, the multi-element weather factor includes but is not limited to temperature, pressure, irradiance, wind direction and speed, cloud cover and precipitation. Various types of weather data are usually obtained by numerical weather prediction (NWP), which is a huge spatial weather feature composed of multi-dimensional weather information. Among all the weather features, a certain large feature is often subdivided into more detailed categories, for example: for the wind speed weather feature, the wind speed at different positions from the ground (10 meters, 30 meters, 70 meters, etc.) can be obtained through numerical weather prediction. With the change of the distance to the ground, the wind direction also changes accordingly. Although the numbers are different, such as wind speed and wind direction, these features have the same characteristics. Therefore, if the principal component of all weather features is directly extracted, not only will it increase the complexity of the input data, but it will also interfere with the final result.

[0046] In the implementation process, the correlation between the multiple meteorological factors and the load power of different load types is represented by the Pearson correlation coefficient. According to the Pearson correlation coefficient, the meteorological factor indexes are preliminarily screened, the correlation coefficient of each meteorological factor and the load power is calculated, and the low-correlation meteorological factors are removed to avoid affecting the subsequent analysis results.

[0047] The positive and negative of the Pearson coefficient and the absolute value size can clearly reflect the correlation direction and strength between multiple variables. The Pearson correlation coefficient is used to preliminarily screen and remove all meteorological factors, and remove homogeneous redundant meteorological factors and low-correlation unnecessary meteorological factors.

[0048] The Pearson coefficient, or Pearson correlation coefficient, is a statistical measure of the degree of linear correlation between the continuous sequences of two variables. In this embodiment, the method is used to solve the dependence between the multiple meteorological factors and the load. The absolute value of Pearson is indeed proportional to the correlation strength between variables. When the value of the Pearson coefficient is positive, there is a positive correlation between the meteorological factors and the load. As the value of the meteorological factor increases, the load power increases accordingly. The greater the positive value of Pearson, the stronger the positive correlation between them. When the value of the Pearson coefficient is negative, it indicates that the meteorological factors and the load are negatively correlated. The smaller the negative value of Pearson, the more obvious the trend of reverse change, that is, the more significant the negative correlation.

[0049] Generally, the Pearson correlation coefficient obtained by calculation can be divided into the following five value ranges, and the correlation strengths represented by each type are as follows: the correlation coefficient is in the interval of 0.8-1.0, which is extremely strong correlation; the correlation coefficient is in the interval of 0.6-0.8, which is strong correlation; the correlation coefficient is in the interval of 0.4-0.6, which is moderate correlation; the correlation coefficient is in the interval of 0.2-0.4, which is weak correlation; and the correlation coefficient is in the interval of 0.0-0.2, which is extremely weak correlation or no correlation.

[0050] For sequences X and Y, there is a linear correlation between them, and the correlation coefficient is defined as follows.

[0051]

[0052] Where cov(X, Y) is the covariance, σ X , and σ Y are the standard deviations of the two sequences, and E(X) and E(Y) are the expectations of the two sequences.

[0053]

[0054]

[0055] As shown above, formula (2), (3), (4) are respectively the calculation formula of covariance, standard deviation, and expected value. After simplification and deformation, the correlation formula can be changed as follows:

[0056]

[0057] According to the above correlation formula, the Pearson coefficient formula is as follows:

[0058]

[0059] Wherein, x i , y i In this embodiment, the representative load power sequence and the corresponding meteorological factor, Indicates the sample mean of the two sequences.

[0060] The mathematical properties that Pearson coefficient satisfies are symmetry, shift invariance and scale invariance, which are explained as follows.

[0061] (1) Symmetry

[0062] ρ XY = ρ YX (7)

[0063] As shown above, the Pearson correlation coefficient does not affect the calculation result of the two variables x i , y i to be solved, that is, it has symmetry.

[0064] (2) Shift invariance

[0065] In the calculation of Pearson correlation coefficient, the numerator (covariance of both) and the denominator (variance of each) are both subjected to the operation of subtracting the mean (centering, or mean normalization) to eliminate the influence of the mean, so the change of the mean of the sequence X and Y will not affect the Pearson correlation coefficient between them.

[0066] 1) Covariance part: the sequence X and Y are respectively subtracted by the mean of each. This shows that no matter how the mean of the sequence X or Y changes, as long as the relative change relationship between the sequence X and Y remains unchanged, the value of the covariance will not change.

[0067] 2) Variance part: the variance of the sequence X and Y The dispersion degree between each of the sequences X and Y and their mean values is measured respectively. Since the mean value is also involved in the variance calculation, when the mean value changes, the variance will also change accordingly. However, it can be observed that the denominator of the Pearson correlation coefficient is the product of the two standard deviations. This indicates that when the mean values of the sequences X and Y change simultaneously, the change of the variance of each of them will offset each other, thereby keeping the value of the variance unchanged.

[0068] (3) Scale invariance

[0069] Scale invariance means that if we perform an arbitrary linear transformation (i.e., multiply by a constant or add a constant) on the two sequences X and Y, the Pearson correlation coefficient between them will not change.

[0070] In the implementation process, for the two sequences X and Y, the Pearson correlation coefficient between them has been calculated as ρ XY , and now a linear transformation is performed on the sequences X and Y, and the following can be obtained:

[0071]

[0072] where a, b, c, and d are arbitrary constants, and in order to ensure that the sequence after the linear transformation is meaningful, the constants a and c are non-zero constants. After the linear transformation, the Pearson correlation coefficient between the sequences X and Y is calculated as follows:

[0073]

[0074] In summary, since the Pearson correlation coefficient is calculated by standardizing the variables, the linear transformation will not change the linear relationship between the two variables or their relative changes. Scale invariance makes the Pearson coefficient suitable for comparing data of different scales or units.

[0075] By calculating the Pearson coefficient between the load power and the multi-element weather, the preliminary screening of the multi-element weather data can be realized, the homogeneous data can be simplified and removed, and the data with little effect on the load power, such as air pressure, can be filtered, which optimizes the data for the next principal component analysis.

[0076] In the implementation process, the process of constructing the initial weather factor feature matrix includes:

[0077] Calculating the variance and standard deviation of the weather factors related to different load types selected, and performing data standardization processing on the weather factors of different dimensions and units;

[0078] According to the standardized weather factors, a covariance matrix is constructed and the eigenvalues and eigenvectors are calculated to obtain the initial weather factor feature matrix.

[0079] For example,Figure 4 As shown in FIG. 3, the algorithm diagram of the Pearson correlation coefficient for the correlation analysis of the multi-position load and the load data is shown. The Pearson correlation coefficient calculated can preliminarily determine that, for the meteorological homogeneous factors representing the wind speed, the absolute value of the correlation coefficient of the wind speed at 100 meters and the load is the largest, so the wind speed at 100 meters is selected as the preliminary meteorological influencing factor to enter the principal component analysis in the next step, and the wind speed at 10 meters, the wind speed at 30 meters and the wind speed at 170 meters are removed to avoid their entering the principal component analysis to affect the result of the principal component dimension reduction.

[0080] S103: Dimension reduction is performed on the initial meteorological factor feature matrix by using the principal component analysis method to obtain a dimension-reduced meteorological factor feature matrix.

[0081] In the specific implementation process, in the process of dimension reduction on the initial meteorological factor feature matrix by using the principal component analysis method, the finally selected principal components are determined according to the contribution rate of each meteorological factor and the total contribution rate, and the initial meteorological factor feature matrix is dimension-reduced and converted into a new relationship represented by the principal components, as shown in FIG. 4. Figure 5

[0082] The principal component analysis method is a commonly used data analysis method, mainly used for data dimension reduction and feature extraction. When analyzing the load power influencing factors, the meteorological data often contains multiple variables, such as temperature, pressure, irradiance, wind direction and speed, cloud cover, precipitation, etc. There is a correlation between these meteorological factors and the load in different degrees. The principal component analysis method can convert multiple related variables into a few uncorrelated principal components, thereby simplifying the data structure and making the analysis more intuitive and efficient.

[0083] The principle of the principal component analysis method is based on the correlation between variables, and the original variables are converted into new linear combination variables through orthogonal transformation, that is, the principal components are extracted. These new variables can retain as much original data information as possible while having better interpretability and independence.

[0084] Specifically, the principal component analysis idea is to find an orthogonal transformation matrix to convert the original data matrix into a new coordinate system, and the data variance on each coordinate axis (i.e., the principal component) of the new coordinate system decreases in turn, and the data on each coordinate axis is independent. In this way, PCA can extract the main features of the data while reducing the dimension of the data.

[0085] The implementation steps of the principal component analysis method are as follows:

[0086] (1) For multi-meteorological data, assuming that there are n samples and p features, x i,j ​Let Xij represent the jth feature of the ith sample, then the feature matrix X of the original sample can be represented as:

[0087]

[0088] (2) Calculate the mean of each feature and the standard deviation S j

[0089]

[0090] (3) Weather data standardization

[0091] Standardization is a common data preprocessing technique aimed at converting variables of different dimensions into numerical values with the same scale, so that they can be compared and analyzed on the same scale. In the original data of the multivariate weather data, different variables correspond to different dimensions. In order to eliminate the adverse effects of dimension and numerical difference on data analysis, the original matrix is usually standardized.

[0092]

[0093] where X ij represents the standardized data.

[0094] (4) Construct the covariance matrix

[0095]

[0096] The covariance matrix measures the correlation between variables in the data set. If the covariance between two variables is positive, they are positively correlated; if it is negative, they are negatively correlated; if it is 0, they are not correlated.

[0097] (5) Calculate the covariance matrix of the standardized sample

[0098]

[0099] (6) Calculate the eigenvalues and eigenvectors of the covariance matrix

[0100] Calculate the eigenvalues of the matrix covariance matrix R, arrange them in order of size, calculate the corresponding eigenvectors, and standardize them so that their total length is 1.

[0101] λ1≥λ2≥L≥λ p ≥0 (18)

[0102]

[0103] where λ1, λ2, L λ p are the eigenvalues of the P weather data, L1, L2, L Lp is a feature vector.

[0104] (7) Determine the principal component and calculate the contribution rate

[0105] F i = l i1 X1+l i2 X2+L+l ip X p ,(i=1,2,K,p) (20)

[0106]

[0107] Where, F i represents the i-th principal component, is the contribution rate of the i-th principal component, ψ i is the cumulative contribution rate of the first i principal components. Generally, the cumulative contribution rate of the selected principal components is more than 85%.

[0108] Finally, the conversion matrix is constructed to convert the p features of the original data to the new space represented by m features, as shown in the following formula.

[0109]

[0110] Through principal component analysis, the multivariate meteorological data is further reduced in dimension on the basis of eliminating some meteorological factors by Pearson correlation coefficient. The new principal component matrix representing the original features is reconstructed after dimension reduction, so as to realize the extraction and combination of meteorological factors having strong correlation with load output, and as a premise, the simplification of huge data is realized.

[0111] After screening the multivariate meteorological data by Pearson correlation coefficient and principal component analysis, the meteorological factors affecting the regional load are screened out, and the low-dimensional feature matrix is reconstructed. Next, the embodiment of the application will take the low-dimensional matrix reconstructed after principal component analysis as the basis, input the low-dimensional matrix into the residual network for further learning, and mine the hidden features of the load.

[0112] S104: Iterative learning of the dimension-reduced meteorological factor feature matrix by using a preset deep learning model, to refine the grid meteorological feature scale, further extract hidden features of different load types, and obtain a grid feature matrix representing different types of loads, for load prediction.

[0113] Wherein, the grid feature matrix includes load type, load power and corresponding grid weather space features. The grid weather space features include longitude and latitude and corresponding weather information.

[0114] In one or more embodiments, the preset deep learning model can adopt a Res-Net residual network.

[0115] The reduced dimension meteorological factor feature matrix is iteratively learned by using a Res-Net residual network.

[0116] Res-Net is a variant based on CNN, and its main feature is to introduce a residual block, so that the network training degree is deeper, and the gradient disappearance problem in deep network is solved. Res-Net realizes the optimization of network structure by constructing residual units and bottleneck residual units, each of which contains a plurality of residual blocks.

[0117] As shown in Figure 6 , the main change of Res-Net in structure is the introduction of residual blocks. Each residual block contains a plurality of convolutional layers and a "shortcut" route across several layers, so that information can be directly transmitted from a certain layer in front to a certain layer behind. This design enables the network to avoid the problem of gradient disappearance while deepening, and comprehensively improves the performance of the network.

[0118] The residual structure can be represented by the following formula:

[0119] x l+1 =x l +F(x l ,W l ) (24)

[0120] Through recursion, the feature expression of any deep layer unit L can be obtained:

[0121]

[0122] From the above formula, the feature x L of any deep layer unit L can be expressed as the feature x l of a shallow layer unit l plus a residual function, which shows that any deep layer unit L and shallow layer unit l have residual characteristics.

[0123] Similarly, for any deep layer unit L, its feature is:

[0124]

[0125] In the above formula, the feature x L of the deep layer unit is the sum of the outputs of all residual functions and x0, and x0 is the sum of the outputs of all residual functions and x0, which is a summation operation, which is better than a series of matrix vector products of the plain network, that is: , which greatly reduces the computational complexity.

[0126] For back propagation, assuming that the loss function is E, the following formula can be obtained according to the back propagation chain rule.

[0127]

[0128] Observing the above derivative formula, the derivative is divided into two parts:

[0129] (1) Transmission without weight layer:

[0130] (2) Transmission through weight layer:

[0131] The former can ensure that the signal directly returns to any shallow layer and does not appear the phenomenon of gradient disappearance.

[0132] In the Res-Net network, the residual learning process is different from the mapping process in the traditional CNN network. In the traditional CNN network, the mapping process of signal transmission is realized between two adjacent units. Different from this, the Res-Net network structure changes the increment F(x) on the input x of each layer, and completes the whole learning process through the jump connection.

[0133] The identity mapping process in the Res-Net residual network model is derived as follows.

[0134] Suppose there are two adjacent convolution units in the residual unit, and the convolution layer in the front position has a weight W n and a bias b n . After adding the weight, the output is B n . After the Relu activation function calculation, the expression is as follows.

[0135] B n =B n x+b n (28)

[0136] Similarly, suppose the convolution layer in the rear position has a weight W n+1 and a bias b n+1 . After adding the weight, the output is B n +1 . After the Relu activation function calculation, the expression is as follows.

[0137] B n+1 =B n+1 x+B n+1 (29)

[0138] Let F(x) represent the output B n+1 after adding the weight, F(x) = Z m+1 = W m+1 a m +b m+1 , and combining equation (28) and equation (29), the expression of F(x) is as follows.

[0139] F(x) = W m+1 g(W m x+b m )+b m+1 (30)

[0140] The F(x) and x are processed through the activation function Relu in the residual network, and the final expression is as follows.

[0141] g(F(x)+x) = g(W m+1 g(W m x+b m )+b m+1 +x) (31)

[0142] The above flow illustrates in detail how to construct the output feature of the residual learning unit with the input and output features of the convolution layer. The analysis result shows that in the Res-Net residual network, the construction of the output feature must be activated through the ReLU function, and finally completed. This step ensures that the network can learn more complex and effective feature representation, thereby improving the performance of the entire network.

[0143] Wherein, as shown in Figure 7 , the process of further extracting the grid-based hidden features of the region belonging to different load types by using the Res-Net residual network to iteratively learn the dimension-reduced meteorological factor feature matrix is as follows:

[0144] (1) The dimension-reduced data is input into the Res-Net residual neural network model for deep learning as input;

[0145] (2) Considering the regional similarity of meteorology, the input meteorological data is learned in the Res-Net residual neural network using the convolution layer and the residual module, and the existing 3km*3km grid meteorological data grid is upscaled;

[0146] (3) The meteorological features in the divided region are extracted (such as the highest temperature, average temperature, etc. of each grid point in the region) through the global average pooling layer in the Res-Net network, forming a meteorological element value matrix with representative meaning, and completing the extraction of load features.

[0147] The multi-dimensional meteorological array after calculating the Pearson correlation coefficient and principal component analysis is input into the residual neural network in the embodiment, and after learning and processing through the above convolution layer and residual module, a meteorological matrix representing load features will be obtained.

[0148] Subsequently, feature dimension reduction is performed through the global pooling layer of the Res-Net, and the meteorological matrix is further subjected to feature screening, so as to obtain a low-dimensional and high-representative load feature matrix. The global average pooling layer is generally used as the last part of the Res-Net network, and its main function is to compress the spatial dimension of the last convolution feature map into a 1x1 grid, so that each feature map corresponds to a global average feature value. The global average pooling layer has two main advantages:

[0149] (1) Reducing the number of parameters: since the global average pooling layer compresses the spatial dimension of the feature map, the number of parameters in the subsequent fully connected layer (if any) is greatly reduced, which helps to prevent overfitting and accelerate the training and inference process.

[0150] (2) Improving the robustness of the model to spatial transformation: the global average pooling layer considers all spatial information of the feature map, so it is not sensitive to small changes in spatial position, and the robustness of the model is enhanced.

[0151] The specific implementation of the global average pooling is to perform an average operation on each channel (or feature map) of the feature map, that is, to calculate the average value of all pixel values, and the pooling result is a numerical value, which represents the global information of the feature map. This process is performed for all feature maps, so the output dimension is the same as the number of channels of the feature map.

[0152] After data processing by the global average pooling layer, the features are reduced in dimension to obtain representative load features. Based on the processing of the algorithm of the present application, a load feature extraction and dimension reduction method is obtained, which provides an algorithm basis for subsequent load prediction work.

[0153] As can be known from the above introduction, compared with the traditional CNN model, the Res-Net residual network introduces a residual module, which overcomes the problems of gradient disappearance and gradient explosion, solves the adverse effects brought by the increase of the depth of the CNN, and can better serve the extraction of load features.

[0154] The Res-Net residual network learning further extracts the hidden features of the regional load, constructs a new feature matrix according to the original multi-dimensional meteorological data, obtains the grid-based low-dimensional meteorological features representing a single load, facilitates the subsequent data as input to participate in the operation and analysis of the load prediction model, makes the grid-based power characteristics as a mapping of the single-valued power characteristics, and lays a foundation for the extraction of load features.

[0155] Figure 2 is a schematic structural diagram of a grid-based numerical weather feature dimension reduction extraction system in an embodiment of the present application, which corresponds to the grid-based numerical weather feature dimension reduction extraction method of Figure 1 , and the grid-based numerical weather feature dimension reduction extraction method of Figure 2As shown, the grid-based numerical weather feature dimensionality reduction extraction system in the embodiment can include:

[0156] The load type and power relationship determination module 201 is configured to obtain typical daily load curves of different types of loads in the high-proportion new energy power system, and obtain the corresponding relationship between the load type and the load power of the high-proportion new energy power system.

[0157] The meteorological factor feature matrix construction module 202 is configured to obtain multivariate meteorological factors according to the typical daily grid-based numerical weather measured data of the region where the different types of loads belong, and further calculate the correlation between the load power of different load types, filter out the meteorological factors related to different load types, and construct an initial meteorological factor feature matrix.

[0158] The meteorological factor feature matrix dimensionality reduction module 203 is configured to use a principal component analysis method to reduce the dimensionality of the initial meteorological factor feature matrix, and obtain a reduced meteorological factor feature matrix.

[0159] The grid-based meteorological feature extraction module 204 is configured to use a preset deep learning model to iteratively learn the reduced meteorological factor feature matrix to refine the grid-based meteorological feature scale, further extract the hidden features of different load types, and obtain a grid-based feature matrix representing different types of loads for load prediction.

[0160] The correlation between the multivariate meteorological factors and the load power of different load types is represented by a Pearson correlation coefficient, or the reduced meteorological factor feature matrix is iteratively learned by using a Res-Net residual network.

[0161] Figure 2 The specific implementation process of the modules 201-204 in the grid-based numerical weather feature dimensionality reduction extraction system is the same as the specific implementation process of the steps S101-S104 in the method 100, and will not be described in detail here. Figure 1

[0162] In particular, according to the embodiments of the present application, the processes described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments of the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program comprising program code for performing the steps of the method shown above. In such embodiments, the computer program can be downloaded and installed from a network by a communication part, and / or installed from a detachable medium. When the computer program is executed by a central processing unit, various functions defined in the device of the present application are performed. Figure 1

[0163] wherein, Figure 1 ​​The computer program instructions of the illustrated method can also be loaded into a computer readable storage medium such that the computer program instructions cause a computer or other programmable data processing apparatus to operate in a specific manner, such that a product of manufacture including an instruction means which implements the function specified in the flowchart or flowcharts and / or blocks is tangibly embodied in the computer readable storage medium. Figure 1 The computer program instructions of the illustrated method can also be loaded into a computer readable storage medium such that the computer program instructions cause a computer or other programmable data processing apparatus to operate in a specific manner, such that a product of manufacture including an instruction means which implements the function specified in the flowchart or flowcharts and / or blocks is tangibly embodied in the computer readable storage medium. Figure 1 The computer program instructions of the illustrated method can also be loaded into a computer readable storage medium such that the computer program instructions cause a computer or other programmable data processing apparatus to operate in a specific manner, such that a product of manufacture including an instruction means which implements the function specified in the flowchart or flowcharts and / or blocks is tangibly embodied in the computer readable storage medium.

[0164] It is understood by those skilled in the art that all or part of the flowcharts in the above-mentioned embodiment methods can be implemented by a computer program instructing relevant hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the flowcharts of the above-mentioned embodiment methods. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0165] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for grid-based numerical weather feature dimensionality reduction extraction, characterized in that, The method comprises the following steps: Obtain the typical daily load curve of different types of loads in a high-proportion new energy power system, and obtain the correspondence between the load type and the load power of the high-proportion new energy power system; The correlation between the multiple meteorological factors and the load power of different load types is represented by the Pearson correlation coefficient; The Pearson correlation coefficient is used to preliminarily screen and eliminate all meteorological factors, eliminate homogeneous redundant meteorological factors and unnecessary meteorological factors with low correlation, and avoid affecting the results of principal component dimension reduction; According to the typical daily grid numerical weather observation data of the region where the different types of loads belong to, the multiple meteorological factors are obtained, and the correlation between the multiple meteorological factors and the load power of different load types is calculated, the meteorological factors related to different load types are screened out, and an initial meteorological factor feature matrix is constructed; The initial meteorological factor feature matrix is reduced by using the principal component analysis method, and a reduced meteorological factor feature matrix is obtained, so as to realize the dimension reduction processing of the multiple meteorological data, represent the original high-dimensional information through the construction of a low-dimensional expression, and reduce the data dimension while retaining the original information; In the process of reducing the initial meteorological factor feature matrix by using the principal component analysis method, the final selected principal components are determined according to the contribution rate and total contribution rate of each meteorological factor, and the initial meteorological factor feature matrix is reduced and converted into a new relationship represented by the principal components; The contribution rate is calculated as follows: wherein, is the contribution rate of the first principal component, is the cumulative contribution rate of the first principal components, and the cumulative contribution rate of the selected principal components is more than 85%. The reduced meteorological factor feature matrix is iteratively learned by using a preset deep learning model, so as to refine the grid meteorological feature scale, further extract the hidden features of different load types, obtain a grid feature matrix representing different types of loads, and use the grid feature matrix for load prediction; wherein the features of different types of loads are spatial features, which include longitude and latitude information; the grid feature matrix includes load type, load power and corresponding grid weather spatial features; the grid weather spatial features include longitude and latitude and corresponding weather information; The reduced meteorological factor feature matrix is iteratively learned by using a Res-Net residual network, and the grid hidden features of the region where different load types belong to are extracted, and the process is as follows: The reduced data is input into the Res-Net residual neural network model for deep learning; Considering the regional similarity of meteorology, the convolution layer and the residual module are used in the Res-Net residual neural network to learn the input meteorological data and upgrade the existing grid meteorological data grid; The meteorological features in the divided region are extracted by the global average pooling layer in the Res-Net network to form a meteorological element numerical matrix with representative significance, and the load features are extracted; The multi-dimensional meteorological array after the calculation of the Pearson correlation coefficient and the principal component analysis is input into the residual neural network, and the meteorological matrix representing the load features is obtained after the learning and processing of the above convolution layer and residual module. The feature dimension is reduced through the global pooling layer of the Res-Net, the meteorological matrix is further subjected to feature screening, a low-dimensional and high-representative load feature matrix is obtained, and the global average pooling layer is used as the last part of the Res-Net network, so that the spatial dimension of the last convolution feature map is compressed into a 1x1 grid, and each feature map corresponds to a global average feature value.

2. The gridded numerical weather feature dimensionality reduction extraction method of claim 1, wherein, The process of constructing the initial meteorological factor feature matrix includes: calculating the variance and standard deviation of the meteorological factors related to different load types, and normalizing the data of the meteorological factors with different dimensions and units; constructing a covariance matrix according to the normalized meteorological factors, and calculating the eigenvalues and eigenvectors to obtain the initial meteorological factor feature matrix.

3. The gridded numerical weather feature dimensionality reduction extraction method of claim 1, wherein, The multiple meteorological factors include temperature, pressure, irradiance, wind direction and speed, cloud cover and precipitation.

4. A gridded numerical weather feature dimensionality reduction extraction system, characterized in that, It includes: a load type and power relationship determination module, which is used to obtain the typical daily load curve of different types of loads in a power system containing a high proportion of new energy power, and obtain the corresponding relationship between the load type and the load power of the power system containing a high proportion of new energy power; a meteorological factor feature matrix construction module, which is used to obtain multiple meteorological factors according to the typical daily grid numerical weather observation data of the region to which different types of loads belong, and then calculate the correlation between the load power of different load types, screen out meteorological factors related to different load types, and construct an initial meteorological factor feature matrix; the correlation between the multiple meteorological factors and the load power of different load types is represented by the Pearson correlation coefficient; all meteorological factors are preliminarily screened and removed by using the Pearson correlation coefficient, and homogeneous redundant meteorological factors and unnecessary meteorological factors with low correlation are removed to avoid affecting the results of principal component dimension reduction; a meteorological factor feature matrix dimension reduction module, which is used to reduce the dimension of the initial meteorological factor feature matrix by using a principal component analysis method to obtain a reduced meteorological factor feature matrix, so as to realize the dimension reduction processing of multiple meteorological data, represent the original high-dimensional information through a low-dimensional expression, and reduce the data dimension while retaining the original information; the principal component analysis method includes determining principal components and calculating a contribution rate, and the calculation of the contribution rate is specifically: wherein, is the contribution rate of the first principal component, is the cumulative contribution rate of the first principal components, and the cumulative contribution rate of the selected principal components is more than 85%. a grid meteorological feature extraction module, which is used to iteratively learn the reduced meteorological factor feature matrix by using a preset deep learning model to refine the grid meteorological feature scale, further extract the implicit features of different load types, and obtain a grid feature matrix representing different types of loads, which is used for load prediction; wherein the features of different types of loads are spatial features, which include longitude and latitude information; the grid feature matrix includes load type, load power and corresponding grid weather spatial features; the grid weather spatial features include longitude and latitude and corresponding weather information; the reduced meteorological factor feature matrix is iteratively learned by using a Res-Net residual network; the process of iteratively learning the reduced meteorological factor feature matrix by using a Res-Net residual network to extract the grid implicit features of the region to which different load types belong is as follows: the reduced data is input into the Res-Net residual neural network model for deep learning. Considering the regional similarity of meteorological data, the input meteorological data is learned by using convolutional layers and residual modules in the Res-Net residual neural network, and the existing grid meteorological data grid is upscaled; The meteorological feature value matrix with representation significance is formed by extracting the meteorological characteristics in the divided region through the global average pooling layer in the Res-Net network, and the load feature is extracted; After calculating the Pearson correlation coefficient and the principal component analysis of the multi-dimensional meteorological array, the meteorological array is input into the residual neural network, and the meteorological matrix representing the load feature is obtained through the learning and processing of the convolutional layers and residual modules. The feature dimension is reduced by the global pooling layer of the Res-Net, and the meteorological matrix is further feature screened to obtain a low-dimensional and high-representative load feature matrix. The global average pooling layer is the last part of the Res-Net network, and the spatial dimension of the last convolution feature map is compressed into a 1x1 grid, and each feature map corresponds to a global average feature value.

5. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the grid numerical weather feature dimension reduction extraction method of any one of claims 1-3.

6. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the grid numerical weather feature dimension reduction extraction method of any one of claims 1-3.

Images