Flexible load high adjustable potential data extraction method based on improved full convolutional network model, storage medium and electronic equipment

Abstract

A flexible load high adjustable potential data extraction method based on an improved full convolutional network model comprises the following steps: firstly, collecting electrical load data of a user, pre-processing the electrical load data, taking the processed data and selected adjustment potential index characteristic parameters as input, and extracting the adjustment potential index characteristic parameters; the method comprises the following steps: extracting potential time sequence data of a user by establishing a full convolutional network model, and determining the number of filters and the size of a convolution kernel of an FCN model through a BKA improved by combining refraction reverse learning, a dynamic switching probability and a Levy flight strategy; and substituting the finally obtained number of filters and the convolution kernel size back to the FCN model to extract time sequence data of the potential of the user. A user load characteristic index value is calculated according to the selected adjustment potential index data to achieve extraction of load high adjustable potential data points, and key data points can be screened out through convolution and pooling of load data points in original data under load characteristic index mapping and processing of a multi-stage classifier.

Description

Technical Field

[0001] The present invention belongs to the technical field of data feature extraction, and specifically relates to a flexible load high-adjustable potential data extraction method, storage medium and electronic device based on an improved full convolutional network model. Background Art

[0002] There is great demand elasticity for user flexible loads. However, due to the wide variety of flexible load types and electrical equipment, the diverse and complex interaction modes and the huge amount of electricity consumption data, it is difficult to accurately characterize the true adjustment potential of user flexible loads. It is necessary to effectively extract data points with high adjustable potential for flexible loads.

[0003] The existing methods for extracting load high adjustable potential data are mainly divided into two types: data analysis method and cluster grouping analysis method. Data analysis method is highly dependent on data, cannot adapt well to future changes, and cannot handle nonlinear and complex relationships. Cluster grouping analysis method has the disadvantages of difficult parameter selection, sensitivity to noise and outliers in data, and poor result interpretation.

[0004] Fully Convolutional Networks (FCN) is formed by modifying the existing VGG classification network structure. It converts all the fully connected layers in VGG into convolutional layers. In order to solve the problem of image size reduction caused by convolution and pooling, the image size is restored by upsampling, and pixel-to-pixel and end-to-end learning is realized on the original pixel area method, and the refined result with the same size as the original image is output. The FCN network structure is mainly composed of convolutional layers, deconvolutional layers and multi-layer classifiers. The fully convolutional layers are some classic CNN networks. They extract features through multi-layer convolution and pooling, while the deconvolutional layers obtain semantic segmentation images with the same size as the original image by upsampling. Finally, all features can be classified and the results are output by multi-layer classifiers. However, the number of filters and convolution kernel size of the FCN network are basically obtained by random selection or empirical debugging. This parameter selection method has limitations and inappropriate parameters will lead to low prediction efficiency of the FCN model, easy to fall into local optimality, and low extraction accuracy.

[0005] BKA (Black-Winged Kite Algorithm, BKA) is a meta-heuristic optimization algorithm inspired by the migration and predation behavior of BK. It not only captures the flight and predation behavior of BK in nature, but also deeply simulates their high adaptability to environmental changes and target locations. In BKA, the Cauchy mutation strategy is introduced, which is a probability distribution strategy that can help the algorithm jump out of the local optimum and increase the probability of finding a better solution in the global search space. And a leadership strategy is integrated to imitate the leadership role of the leader in the BK population to ensure that the algorithm can effectively utilize the current best solution and guide the search direction. This method not only helps to improve the efficiency of the algorithm in utilizing the current search area, but also effectively balances the dynamic relationship between global search and local search, ensuring that potential competitive new areas are not ignored in the process of pursuing the optimal solution. However, due to the fixed p value of the algorithm, it also has certain defects in balancing global and local searches. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a load high adjustable potential data extraction method based on improved BKA optimized FCN, which is reasonably designed, accurate, reliable and easy to implement in engineering. The present invention introduces a refraction reverse learning mechanism to initialize the BK population, thereby increasing species diversity; introduces a dynamic switching probability to improve the constant value p, thereby achieving the effect of balancing the local and global exploration proportions of BK, and improving the ability of the algorithm to jump out of the local optimal value; adopts the Levy flight strategy to adjust the position of the leader and disturb it, thereby enhancing the local escape ability of the leader and improving the ability of the algorithm to obtain the global optimal solution. The improved BKA is used to solve the problem of low data extraction accuracy caused by the randomness of the FCN network model.

[0007] The present invention solves the existing technical problems by adopting the following technical solutions: a method for extracting highly adjustable potential data of flexible loads based on an improved fully convolutional network model, comprising the following steps:

[0008] Step S1: determining a user load characteristic index that can measure the user's adjustable potential;

[0009] Step S2: using Newton interpolation method to fill the missing values ​​in the original user load data to obtain processed user load data;

[0010] Step S3: using the Pearson coefficient to select the user load characteristic index, select the user core load characteristic index, and calculate the index value of the selected user core load characteristic index corresponding to the processed user load data according to the user load characteristic index calculation formula;

[0011] Step S4: Divide the processed user load data into a training set and a test set;

[0012] Step S5: constructing the FCN network model;

[0013] Step S6: constructing a mean square error (MSE) function for the number of filters k and the convolution kernel size l of the FCN network model to evaluate whether the data points extracted by the FCN network model will have better performance in terms of high adjustable potential relative to the original user load data;

[0014] The mean square error objective function is:

[0015] Where: n is the number of load data; is the mean of all elements of the Pearson correlation coefficient matrix in the original data;

[0016] is the mean of all elements of the Pearson correlation coefficient matrix of the data points extracted by FCN;

[0017] Step S7: Solve the mean square error objective function based on the improved BKA optimization algorithm to determine the initial filter number k and convolution kernel size l of the FCN network model;

[0018] Step S8: Use the user load data and index values ​​in the test set as input, and the initial number of filters k and the convolution kernel size l as output to substitute back into the FCN network model for training to obtain the data points that best reflect the high load adjustability potential.

[0019] User load characteristic indicators include: all-day load rate, daily peak-to-valley difference rate, peak load rate, valley load rate, normal load rate, and daily minimum load rate;

[0020] User load characteristic index calculation formula:

[0021] The full-day load rate is: α1=P av / P max ; The daily peak-to-valley difference rate is expressed as: The peak load rate is: α3=P av.peak / P av ; The load rate during the valley period is: α4=P av.val / P av ; The average load rate is: α5=P av.sh / P av ; Daily minimum load rate is: α6 = P min / P av ;

[0022] Where: P av represents the daily average load, P max Indicates the maximum daily load, P min Indicates the minimum daily load, P av.peak represents the average load during the daily peak period, Pav.val represents the average load during the daily valley period, P av.sh Indicates the average load during the daily average period.

[0023] Step S2 is specifically as follows:

[0024] In view of the nonlinear and non-stationary characteristics of user power load data, the Newton interpolation method is used to establish the interpolation function using the complete points in the data, and then the filling value is calculated based on the horizontal coordinates of the missing items;

[0025] For a set of data: n0(x0,f(x0)),n1(x1,f(x1)),…,n k (x k ,f(x k ))Construct Newton interpolation polynomial: Where f(x0,x1,…,x k ) is the k-order difference quotient of function f(x), that is:

[0026]

[0027] From the above formula, we can conclude that each order difference quotient can be derived from the previous order difference quotient, and the value of the interpolation point can be calculated according to the interpolation polynomial.

[0028] Where: n i represents the i-th data point, x i is the coordinate of the ith data point x, f(x i ) is in x i The function value at , i is 0 ~ K, N(x) is the Newton interpolation polynomial we want to construct, which is used to approximate the original function f(x), and f(x) is the original unknown function.

[0029] The Pearson coefficient is used to select the core load characteristic indicators of users, which are as follows:

[0030] The Pearson coefficient is used to calculate the Pearson coefficients between various indicators and between various indicators and the original user load data, and the indicators with the Pearson coefficients between various indicators and the user load data greater than 0.5 are selected;

[0031] The Pearson correlation coefficient calculation formula is:

[0032]

[0033] In the formula, r represents the Pearson correlation coefficient, x i represents the i-th load data, represents the average value of all user load data, y i represents the load characteristic index value corresponding to the i-th load data, It represents the mean value of all load characteristic index values, i represents the i-th power point, and n represents the number of all power points.

[0034] Step S5 includes the following steps:

[0035] Step S5.1: Build the FCN network model architecture. The FCN network model architecture includes three layers: convolution layer, global pooling layer and multi-classification layer. The number of filters in the convolution layer is set to k, the convolution kernel size is set to l, and the learning step size is set to 1.

[0036] Step S5.2: The relationship between the input data and the output data in the convolutional layer is:

[0037] y ij =f ks ({x si+δi,sj+δj} 0≤δi,δj＜k )

[0038] Where: k is the size of the convolution kernel, s is the step size or sampling factor, f ks Determine the type of layer, x ij is the data vector written at position (i, j) in a particular layer, y ij is the data vector output at position (i, j) in a specific layer; x si+δi,sj+δj is the convolution kernel coverage area with (si,sj) as the center and size k×k in the input data matrix x; si is the row coordinate of the center of the current convolution kernel, sj is the column coordinate of the center of the current convolution kernel; the two variables δi and δj represent the offset of the convolution kernel in the row direction and column direction respectively;

[0039] Step S5.3: The transformation rules of the convolution kernel and step size are as follows:

[0040]

[0041] In the above formula: f ks represents a convolution kernel with a size of k and a step size of s; g k′s′ represents a convolution kernel with a size of k′ and a stride of s′; Represents a convolution operation, which means performing a convolution operation on the input data; Indicates the use of convolution kernel f ks Perform a convolution operation on the input data, which will generate a new feature map, and then the convolution kernel g k′s′ Apply it to the feature map and perform another convolution operation; The equivalent convolution kernel size and stride after the composite of the two convolution operations are k′+(k-1)s′, ss′;

[0042] The upsampling convolution formula is:

[0043]

[0044] In the above formula: γ is the upsampling factor, y ij is the user electricity consumption data point with high adjustable potential at position (i, j) after data reconstruction; x represents the input data matrix; α and β are the indexes inside the convolution kernel, which take values ​​0 or 1 and are used to traverse the four pixels around the reference point; i / γ and j / γ are used to map the coordinates i, j of the output data back to the coordinates of the input data; {i / γ} and {i / j} represent the decimal parts of i / γ and i / j, respectively, indicating the offset near the reference point coordinates; [i / γ] and [j / γ] represent the integer parts of i / γ and j / γ, respectively, indicating the reference point coordinates in the input data; |1-α-{i / γ}| and |1-β-{i / j}| calculate the weights of the data points in the horizontal and vertical directions, respectively.

[0045] In step S7, the mean square error objective function is solved based on the improved BKA optimization algorithm to determine the initial number of filters k and the convolution kernel size l of the FCN network model, specifically:

[0046] Step S7.1: Initialize the position state of the BK individual to vector σ = (k, l), where vectors k and l correspond to the initial number of filters k and convolution kernel size l of the FCN network model respectively. The position of each BK individual combines the information of the initial number of filters k and convolution kernel size l parameters, and each BK individual represents a solution to the mean square error objective function; set the initial parameters of BKA: population size pop, dimension dim, maximum number of iterations T, lower limit of solution space l b 、Find the upper limit u of the solution space b , the parameter p used to control the balance between global and local search, the variable r = rand (0, 1), then the position of the randomly initialized pop BK in the optimization range is y i,j , each BK individual is a potential solution that satisfies the mean square error objective function;

[0047] Step S7.2: Use the refraction reverse learning strategy to generate a refraction reverse population

[0048] Step S7.3: Merge the initial population and the refraction reverse population generated by the refraction reverse learning strategy According to the mean square error objective function, the fitness value of the entire population is calculated in the initial stage of iteration, and the fitness values ​​are arranged in descending order. The first pop BK individuals with the best fitness value are selected as the initial population. In this initialization stage, the individual with the best fitness value in the initial population is selected as the leader. The larger the mean square error, the greater the fitness value of the BK individual.

[0049] Step S7.4: Since BK has two kinds of attack behaviors during flight, which are used for global exploration and local search processes respectively, and these two processes are controlled by constant p∈[0,1], the dynamic switching probability is used to optimize BKA to balance the proportion of local and global searches to achieve a better optimization strategy, and update the position and fitness value of the best BK individual in the current BK population under the current attack strategy;

[0050] Step S7.5: During the migration of the BK population, if the fitness value of the current population is less than the fitness value of the random population, the leader of the current population will be given leadership and join the migrating population; if the fitness value of the current population is greater than the fitness value of the random population, it will guide the population until it reaches the destination. The position of the leader in each population of the BK individuals changes according to this strategy. Since the range of the global optimal solution has been basically determined in the later search, the Levy flight strategy is used to adjust the leader position, disturb it, and enhance its local escape ability, which can improve the quality of the algorithm's later solutions and enable the algorithm to converge to the global optimum;

[0051] Step S7.6: Record the BK individual with the best fitness value in each iteration process, repeat the iteration process until the maximum number of iterations T is reached, and select the BK with the highest fitness value from the best BK leader individual in each iteration as the final result of the optimization.

[0052] In step S7.1, the position state of the BK individual is initialized as follows:

[0053] Initialize the population as follows:

[0054]

[0055] In the above formula: BK ij is the i-th BK in the j-th dimension;

[0056] Evenly distribute the position of each BK as follows:

[0057] X i =BK lb +rand(BK ub -BK lb )

[0058] Where: X i is the position of the i-th black-winged kite individual;

[0059] BK lb and BK ub are the lower and upper limits of the i-th black-winged kite individual in the j-th dimension;

[0060] rand is a random value between [0,1];

[0061] During the initialization process, BKA selects the individual with the best fitness value in the initial population as the leader X L , which is considered the best position for BK, and is selected as follows:

[0062] f best =min(f(X i ))

[0063] X L =X(find(f best = = f(X i )))

[0064] Where: f(X i ) represents the fitness value of the i-th black-winged kite individual;

[0065] f best Indicates the best fitness value;

[0066] X represents the spatial position;

[0067] X L Indicates the position of the black-winged kite individual with the best fitness value (the leader in the population);

[0068] The mathematical model of BK attack behavior is as follows:

[0069]

[0070] In the above formula: and denotes the position of the i-th BK in the j-th dimension in the t-th and (t+1)-th iteration steps respectively; is a random number ranging from 0 to 1; T is the total number of iterations, t is the number of iterations completed so far; r is a random number between 0 and 1;

[0071] The mathematical model of BK migration behavior is as follows:

[0072]

[0073] m=2×sin(r+π / 2)

[0074] In the above formula: denotes the leading scorer of BK in the jth dimension at the tth iteration so far; and represents the position of the i-th BK in the j-th dimension in the t-th and (t+1)-th iteration steps respectively; F i represents the fitness value of the current position in the jth dimension obtained by any BK in t iterations; F rirepresents the fitness value of the random position of the jth dimension obtained from any BK in the tth iteration; C(0,1) represents the Cauchy mutation;

[0075] The one-dimensional Cauchy distribution is a continuous probability distribution with two parameters. The probability density function of the one-dimensional Cauchy distribution is:

[0076]

[0077] x: is a random variable, representing the value of calculating its probability density;

[0078] δ: is the scale parameter of the distribution, which controls the width of the distribution, that is, the degree of diffusion of the distribution;

[0079] μ: is the location parameter of the distribution, which determines the center position of the distribution, that is, the location of the peak in the Cauchy distribution;

[0080] When δ=1,μ=0, its probability density function becomes the standard form, and the following is the exact formula:

[0081]

[0082] In step S7.2, a refraction reverse learning strategy is adopted to generate a refraction reverse population Specifically:

[0083] According to the principle of refraction reverse learning, the reverse learning formula can be derived as follows: * =v+uy, Extending to the high-dimensional space of BKA, setting n1 = 1, we can get the following:

[0084]

[0085] In the above formula, y i,j is the position of the i-th BK in the population in the j-dimensional position (i=1,2,...,D; j=1,2,...,N), D is the population number, N is the dimension; for y i,j The refraction reverse position; y represents the position of the initialization BK individual, y * represents the refraction reverse position of y, u j represents the upper limit of the optimal range of the black-winged kite individual in the j-dimensional position, v j represents the lower limit of the optimal search range of the black-winged kite individual in the j-dimensional position, ε represents the scaling factor, and ω represents the refractive index.

[0086] Dynamic switching probability is used to optimize BKA, specifically:

[0087] The formula of dynamic switching probability p is: p=0.6-0.1×(MaxIter-t) / MaxIter, where: MaxIter is the maximum number of iterations, and t is the current number of iterations.

[0088] The leader position is adjusted using the Levy flight strategy, specifically:

[0089]

[0090] In the above formula: represents the position of the jth dimension of the ith individual at time t; C(0,1) represents a uniformly distributed random number from 0 to 1; represents the Levy distribution factor of the jth dimension at time t; L (λ) represents the scaling factor of the Levy distribution; F i represents the fitness value of the i-th individual; F ri represents the reference fitness value of the i-th individual; m is a constant;

[0091] L (λ) Random search path for Levy flight:

[0092] L(s,λ)=η~s -λ ,1<λ≤3

[0093] Among them, s is the random step length of Levy flight; η is a probability density function in power form with λ as the power; since the Levy distribution is very complex, the Mantegna algorithm is used to simulate it, and the step length calculation formula is as follows:

[0094]

[0095] In the formula, η and ν are two normal random variables with standard deviations, σ η and σ ν is the standard deviation of its corresponding normal distribution, The calculation formula is as follows:

[0096]

[0097] The present invention also provides a non-volatile computer storage medium, which stores computer executable instructions, and the computer executable instructions can execute the flexible load high-adjustable potential data extraction method based on the improved full convolutional network model as described in any one of claims 1-10.

[0098] The present invention also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the characteristic of the electronic device is that the instructions are executed by the at least one processor so that the at least one processor can execute the flexible load high-adjustable potential data extraction method based on the improved fully convolutional network model as described in any one of claims 1-10.

[0099] The deep learning method adopted in the present invention can solve the problems existing in the above methods, and classify the electricity consumption data of industrial users based on load characteristic index mapping, and construct an industrial electricity consumption data set with high adjustable potential. At the same time, in view of the low efficiency of the traditional method of traversing all hyperparameter combinations of the network using the grid search method, this paper adopts the intelligent optimization algorithm BKA to optimize the number of filters and convolution kernel size of the FCN network, which can improve efficiency and accuracy.

[0100] The present invention uses a refraction reverse learning mechanism to initialize the Black-Winged Kite (BK) population, increases the species diversity in the initial stage of the algorithm, and enhances the algorithm's global exploration ability in the early stage; at the same time, a dynamic switching probability is used to improve the constant value p, to achieve the effect of balancing the local and global exploration proportions of BK, and improve the algorithm's ability to jump out of the local optimal value; finally, the Levy flight strategy is used to adjust the position of the leader and disturb it, thereby enhancing the leader's local escape ability and improving the algorithm's ability to obtain the global optimal solution. By applying the improved BKA to extract load high adjustable potential data points, the problem of overly random parameter selection in the FCN network during data extraction is solved, and the efficiency and accuracy of the data extraction process are also improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 It is a flowchart of a method for extracting load high adjustable potential data based on improved BKA optimized FCN of the present invention;

[0102] Figure 2 Schematic diagram for reverse learning of refraction. DETAILED DESCRIPTION

[0103] The technical solution of the present invention will be clearly and completely described below in conjunction with embodiments.

[0104] The design idea of ​​the present invention is: the number of filters k and the size of the convolution kernel l of the model are optimized by fusing BKA with the FCN network, the mean square error function is used as the objective function of BKA, and BKA is applied to the FCN network to solve the problem of low data extraction accuracy caused by the randomness of the FCN network model. The iterative method is simple, the convergence speed is fast, and it is easy to implement in engineering. In BKA, each BK is a potential solution that satisfies the objective function. In addition, the BK population is initialized by introducing a refractive reverse learning mechanism, which increases the species diversity; the constant value p is improved by introducing a dynamic switching probability to achieve the effect of balancing the local and global exploration proportions of BK, and the ability of the algorithm to jump out of the local optimal value is improved; the Levy flight strategy is used to adjust the position of the leader and disturb it, which strengthens the local escape ability of the leader and improves the ability of the algorithm to obtain the global optimal solution.

[0105] Based on the above design concept, the present invention proposes a method for extracting load high adjustable potential data points based on the improved BKA optimized FCN network, such as Figure 1 As shown, the following steps are included:

[0106] Step S1: determining a user load characteristic index that can measure the user's adjustable potential;

[0107] The adjustable potential of users can be measured by the user load characteristic indicators, which include: all-day load rate, daily peak-to-valley difference rate, peak load rate, valley load rate, flat load rate, and daily minimum load rate;

[0108]

[0109] Where: P represents load, subscripts sum, av, max, and min represent sum, average, maximum, and minimum respectively, peak, val, and sh represent the power consumption characteristics of peak time, valley time, and flat time respectively, and α1 to α6 reflect the power consumption characteristics of users at different times of the day. Each indicator has different emphasis and physical meaning, and its importance is also different when used to reflect the power consumption characteristics of users;

[0110] Step S2: using Newton interpolation method to fill the missing values ​​in the original user load data to obtain processed user load data;

[0111] In practical applications of machine learning, deep learning and other algorithms, if there are missing items in the data set, the calculation will be interrupted; the missing values ​​in the user's power load data are processed, and the interpolation method is used to process time series data, which can better retain the time trend of the missing data. In view of the nonlinear and non-stationary characteristics of user load data, the Newton interpolation method is used to establish an interpolation function using the complete points in the data, and then the filling value is calculated based on the horizontal coordinate of the missing item;

[0112] For a set of data: n0(x0,f(x0)),n1(x1,f(x1)),…,n k (x k ,f(x k ))Construct Newton interpolation polynomial: Where f(x0,x1,…,x k ) is the k-order difference quotient of function f(x), that is:

[0113]

[0114] It can be seen from the above formula that each order difference quotient can be deduced from the previous order difference quotient, and the value of the interpolation point can be calculated according to the interpolation polynomial to obtain each order difference quotient.

[0115] Step S3: using the Pearson coefficient to select the user load characteristic index, select the user core load characteristic index, and calculate the index value of the selected user core load characteristic index corresponding to the processed user load data according to the user load characteristic index calculation formula;

[0116] The Pearson coefficient is used to calculate the Pearson coefficients between various indicators and between various indicators and the original user load data. The indicators with Pearson coefficients greater than 0.5 between various indicators and user load data are selected, and the values ​​of the selected indicators corresponding to the user load data are calculated and used as the input features of the FCN network.

[0117] Due to the huge amount of user load data and the large number of user characteristic indicators, some of these indicators may not reflect the user's adjustable potential. The Pearson correlation coefficient can accurately show the corresponding relationship between the evaluation indicators and the selection results for time series data. Therefore, the Pearson correlation coefficient is used to select indicators that can accurately reflect the user's adjustable potential.

[0118] The Pearson correlation coefficient is:

[0119]

[0120] In the formula, r represents the Pearson correlation coefficient, x i represents the i-th load data, represents the average value of all user load data, y i represents the load characteristic index value corresponding to the i-th load data, Represents the mean of all load characteristic index values; the range of the correlation coefficient is [-1,1]. When r is positive, it indicates a positive correlation between the features, that is, the vectors have the same direction; when r is negative, it indicates a negative correlation between the features, that is, the vectors have opposite directions; when r is 0, there is no correlation at all. Therefore, the distance between r and 0 can be used to map the correlation between each load characteristic index and the user load data, so as to select the most appropriate time series evaluation index.

[0121] Step S4: Divide the pre-processed user load data into a training set and a test set, and input the user load data in the test set and the corresponding selected index values ​​as features into the FCN network respectively;

[0122] Step S5: construct an FCN network model and set basic parameters of the FCN network model;

[0123] Step S6: for the number of filters k and the convolution kernel size l of the FCN network model, a mean square error (MSE) function is constructed to evaluate whether the data points extracted by the full convolutional network (FCN) will have better performance in terms of high adjustable potential relative to the original power load data. This function will quantify the accuracy of the FCN model in identifying and extracting power characteristic data points with high adjustable potential;

[0124] The mean square error objective function is:

[0125] Where: n is the number of load data; is the mean of all elements of the Pearson correlation coefficient matrix in the original data;

[0126] is the mean of all elements of the Pearson correlation coefficient matrix of the data points extracted by FCN;

[0127] Step S7: Solve the mean square error objective function based on the improved BK optimization algorithm to determine the initial filter number k and convolution kernel size l of the FCN network model;

[0128] Step S8: The user load data and index values ​​in the test set are used as input, and the initial number of filters k and the convolution kernel size l are used as output to be substituted back into the FCN network for training to obtain the data points that best reflect the high potential for load adjustment.

[0129] The fully convolutional network (FCN) learns the input time series characteristics by establishing convolutional layers and pooling layers, and upsamples the data characteristics of the last convolutional layer through the deconvolution layer, so that the data sequence after multiple convolutions and pooling is restored to the state before the input, and then all the characteristics can be classified. Finally, the results are output through a multi-layer classifier to obtain highly adjustable potential data points, thereby converting the input time series characteristic data set into a user-adjustable potential data set;

[0130] Construct an FCN network model and set the basic parameters of the FCN network model. The specific implementation method of this step includes the following steps:

[0131] Step S5.1: Build the FCN network model architecture. The FCN model is a network consisting of a 3-layer convolution structure, a global layer, and a multi-classification layer. The number of filters is set to k, the convolution kernel size is set to l, and the learning step size is set to 1. Both k and l are obtained by BKA optimization.

[0132] Step S5.2: The relationship between the input data and the output data in the convolutional layer is:

[0133] y ij =f ks ({x si+δi,sj+δj} 0≤δi,δj＜k )

[0134] Where: k is the size of the convolution kernel, s is the step size or sampling factor, f ks Determine the type of layer, x ij is the data vector written at position (i, j) in a particular layer, y ij is the data vector output at position (i, j) in a specific layer;

[0135] Step S5.3: The transformation rules of the convolution kernel and step size are as follows:

[0136]

[0137] The upsampling convolution formula is:

[0138]

[0139] In the above formula: γ is the upsampling factor, y ij It is the user electricity consumption data point with high adjustable potential after data reconstruction.

[0140] Based on the improved BKA optimization algorithm, the mean square error objective function is solved to determine the initial number of filters k and the convolution kernel size l of the FCN network model. The specific implementation method of this step includes the following steps:

[0141] Step S7.1: Initialize the position state of the BK individual to vector σ = (k, l), where vectors k and l correspond to the initial number of filters k and convolution kernel size l of the FCN network model respectively. The position of each BK individual combines the information of the initial number of filters k and convolution kernel size l parameters, and each BK individual represents a solution to the mean square error objective function; set the initial parameters of BKA: population size pop, dimension dim, maximum number of iterations T, lower limit of solution space l b 、Find the upper limit u of the solution space b , the parameter p used to control the balance between global and local search, the variable r = rand (0, 1), then the position of the randomly initialized pop BK in the optimization range is y i,j , each BK individual is a potential solution that satisfies the mean square error objective function;

[0142] Step S7.2: Use the refraction reverse learning strategy to generate a refraction reverse population

[0143] Step S7.3: Merge the initial population and the population generated by the refraction reverse learning strategy, calculate the fitness value of the entire population in the initial stage of iteration according to the mean square error objective function, and sort them in descending order, select the first pop BK individuals with the best fitness value as the initial population, and select the individual with the best fitness value in the initial population as the leader in this initialization stage. The larger the mean square error, the greater the fitness value of the BK individual;

[0144] Step S7.4: During the flight, BK will adjust the angles of its wings and tail according to the wind speed, hover quietly to observe its prey, and then quickly dive to attack. This strategy includes different attack behaviors for global exploration and search. BK either hovers in the air, spreads its wings and maintains balance, or rushes towards its prey at an extremely fast speed. The two attack behaviors of BK are controlled by the parameter p. In the basic BK optimization algorithm, the local search and global search processes are controlled by the constant p∈[0,1]. However, a reasonable search process should conduct a more intense global search in the early stage of the algorithm to quickly locate the range of the global optimal solution in the search space. In the later stage of exploration, the local exploration ability should be enhanced to improve the optimization accuracy of the algorithm. Therefore, the dynamic switching probability is used to optimize BKA to balance the proportion of local and global searches to achieve a better optimization strategy. Update the position and fitness value of the best BK individual in the current BK population under the current attack strategy;

[0145] Step S7.5: During the migration of the BK population, if the fitness value of the current population is less than the fitness value of the random population, the leader of the current population will be given leadership and join the migrating population; if the fitness value of the current population is greater than the fitness value of the random population, it will guide the population until it reaches the destination. The position of the leader in each population of the BK individuals changes according to this strategy. Since the range of the global optimal solution has been basically determined in the later search, the Levy flight strategy is used to adjust the leader position, disturb it, and enhance its local escape ability, which can improve the quality of the algorithm's later solutions and enable the algorithm to converge to the global optimum;

[0146] The migration of BK populations is to adapt to seasonal changes in order to obtain better living conditions and resources. Migration is usually led by leaders, whose navigation skills are crucial to the success of the team. If the fitness value of the current population is less than that of the random population, the leader of the current population will give leadership and join the migrating population, indicating that it is not suitable to lead the population forward. On the contrary, if the fitness value of the current population is greater than that of the random population, it will guide the population until it reaches its destination. This strategy can dynamically select excellent leaders to ensure successful migration. The position of the leading bird in each population in BK changes according to this strategy, and the position of the leading bird in the current BK population is updated under the current migration strategy;

[0147] Since the flight trajectory of Levy flight is a random walk, it is composed of small step length (short distance) jumps gathered together and occasionally large step length (long distance) jumps, which alternate with each other. If this strategy is integrated into the BK optimization algorithm, in the later search, since the range of the global optimal solution has been basically determined, the small step length can improve the quality of the algorithm's later solutions and enable the algorithm to converge to the global optimum. However, if the position of each BK is adjusted in each iteration cycle, a lot of time consumption will be added. Therefore, this paper only adjusts the position of the leader, perturbs it, and enhances the local escape ability.

[0148] Step S7.6: Record the BK individual with the best fitness value in each iteration, repeat the iteration process until the maximum number of iterations T is reached, and select the BK with the highest fitness value from the best BK leader individual in each iteration as the final result of the optimization;

[0149] Initialize the position state of the BK individual. The specific implementation method of this step includes the following steps:

[0150] Initialize the population as follows:

[0151]

[0152] In the above formula: BK ij is the i-th knowledge BK of the j-th dimension

[0153] Evenly distribute the position of each BK as follows:

[0154] X i =BK lb +rand(BK ub -BK lb )

[0155] During the initialization process, BKA selects the individual with the best fitness value in the initial population as the leader X L , which is considered the best position for BK, and is selected as follows:

[0156] f best =min(f(X i )

[0157] X L =X(find(f best = = f(X i )))

[0158] The mathematical model of BK attack behavior is as follows:

[0159]

[0160] In the above formula: and denotes the position of the i-th BK in the j-th dimension in the t-th and (t+1)-th iteration steps respectively; is a random number ranging from 0 to 1; T is the total number of iterations, and t is the number of iterations completed so far;

[0161] The mathematical model of BK migration behavior is as follows:

[0162]

[0163] m=2×sin(r+π / 2)

[0164] In the above formula: denotes the leading scorer of BK in the jth dimension at the tth iteration so far; and represents the position of the i-th BK in the j-th dimension in the t-th and (t+1)-th iteration steps respectively; F i represents the fitness value of the current position in the jth dimension obtained by any BK in t iterations; F ri represents the fitness value of the random position of the jth dimension obtained from any BK in the t iteration; C(0,1) represents the Cauchy mutation; m is used to measure the fitness of the black-winged kite individual during the migration process The size of the position change is a constant;

[0165] The one-dimensional Cauchy distribution is a continuous probability distribution with two parameters. The following equation describes the probability density function of the one-dimensional Cauchy distribution:

[0166]

[0167] When δ=1,μ=0, its probability density function becomes the standard form, and the following is the exact formula:

[0168]

[0169] Adopt the refraction reverse learning strategy to generate the refraction reverse population The specific implementation method of this step includes the following steps:

[0170] The principle of refraction reverse learning is as follows Figure 2 As shown in the figure, the optimal solution range on the x-axis is [l,μ], the y-axis is the normal line, α and β are the angle of incidence and angle of refraction respectively, h and h * are the lengths corresponding to the incident and refracted light rays, respectively, and O is the midpoint of the optimal range [l,μ]. According to the mathematical midline geometric relationship, we get the following:

[0171]

[0172] From the refractive index n = sinα / sinβ, we can get:

[0173] Let k = h / h * , from the above formula we can get: When n=1 and k=1, the formula can be converted into a reverse learning formula: * =l+μ-x;

[0174] Will Extending to the high-dimensional space of BKA, let n = 1, we can get the following:

[0175]

[0176] In the above formula, x i,j is the position of the i-th sparrow in the population in the j-dimensional space (i = 1, 2, ..., D; j = 1, 2, ..., N), where D is the population size and N is the dimension; For x i,j The refraction reverse position; l j , μ j are the minimum and maximum values ​​of the j-th dimension of the search space respectively;

[0177] The BKA is optimized by using the dynamic switching probability. The specific implementation method of this step includes the following steps:

[0178] The formula of dynamic switching probability p is: p=0.6-0.1×(MaxIter-t) / MaxIter, where: MaxIter is the maximum number of iterations, and t is the current number of iterations.

[0179] The Lévy flight strategy is used to update the leader position of the BK population in migration behavior. The specific implementation method of this step includes the following steps:

[0180]

[0181] In the above formula, L (λ) Random search path for Levy flight:

[0182] L(s,λ)=η~s -λ ,1<λ≤3

[0183] Where s is the random step length of the Levy flight, and the continuous random jumps / steps of a cuckoo follow the power law step length distribution, as shown below:

[0184]

[0185] Both η and ν in the formula obey random normal distribution, and the calculation formula is as follows:

[0186]

Claims

1. A method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model, characterized in that: The following steps are involved: Step S1: determining a user load characteristic index that can measure the user's adjustable potential; Step S2: using Newton interpolation method to fill the missing values ​​in the original user load data to obtain processed user load data; Step S3: using the Pearson coefficient to select the user load characteristic index, select the user core load characteristic index, and calculate the index value of the selected user core load characteristic index corresponding to the processed user load data according to the user load characteristic index calculation formula; Step S4: Divide the processed user load data into a training set and a test set; Step S5: constructing the FCN network model; Step S6: constructing a mean square error (MSE) function for the number of filters k and the convolution kernel size l of the FCN network model to evaluate whether the data points extracted by the FCN network model will have better performance in terms of high adjustable potential relative to the original user load data; The mean square error objective function is: Where: n is the number of load data; is the mean of all elements of the Pearson correlation coefficient matrix in the original data; is the mean of all elements of the Pearson correlation coefficient matrix of the data points extracted by FCN; Step S7: Solve the mean square error objective function based on the improved BKA optimization algorithm to determine the initial filter number k and convolution kernel size l of the FCN network model; Step S8: Use the user load data and index values ​​in the test set as input, and the initial number of filters k and the convolution kernel size l as output to return to the FCN network model for training to obtain the data points that best reflect the high load adjustability potential.

2. According to claim 1, a method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model, characterized in that: User load characteristic indicators include: all-day load rate, daily peak-to-valley difference rate, peak load rate, valley load rate, normal load rate, and daily minimum load rate; User load characteristic index calculation formula: The full-day load rate is: α1=P av / P max ; The daily peak-to-valley difference rate is expressed as: The peak load rate is: α3=P av.peak / P av ; The load rate during the valley period is: α4=P av.val / P av ; The average load rate is: α5=P av.sh / P av ; Daily minimum load rate is: α6 = P min / P av ; Where: P av represents the daily average load, P max Indicates the maximum daily load, P min Indicates the minimum daily load, P av.peak represents the average load during the daily peak period, P av.val represents the average load during the daily valley period, P av.sh Indicates the average load during the daily average period.

3. According to claim 1, a method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model, characterized in that: Step S2 is specifically as follows: In view of the nonlinear and non-stationary characteristics of user power load data, the Newton interpolation method is used to establish the interpolation function using the complete points in the data, and then the filling value is calculated based on the horizontal coordinates of the missing items; For a set of data: n0(x0,f(x0)),n1(x1,f(x1)),…,n k (x k ,f(x k ))Construct Newton interpolation polynomial: Where f(x0,x1,…,x k ) is the k-order difference quotient of function f(x), that is: From the above formula, we can conclude that each order difference quotient can be derived from the previous order difference quotient, and the value of the interpolation point can be calculated according to the interpolation polynomial. Where: n i represents the i-th data point, x i is the coordinate of the ith data point x, f(x i ) is in x i The function value at , i is 0 ~ K, N(x) is the Newton interpolation polynomial we want to construct, which is used to approximate the original function f(x), and f(x) is the original unknown function.

4. According to claim 1, a method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model is characterized in that: The Pearson coefficient is used to select the core load characteristic indicators of users, which are as follows: The Pearson coefficient is used to calculate the Pearson coefficients between various indicators and between various indicators and the original user load data, and the indicators with the Pearson coefficients between various indicators and the user load data greater than 0.5 are selected; The Pearson correlation coefficient calculation formula is: In the formula, r represents the Pearson correlation coefficient, x i represents the i-th load data, represents the average value of all user load data, y i represents the load characteristic index value corresponding to the i-th load data, It represents the mean value of all load characteristic index values, i represents the i-th power point, and n represents the number of all power points.

5. The method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model according to claim 1, characterized in that: Step S5 includes the following steps: Step S5.1: Build the FCN network model architecture. The FCN network model architecture includes three layers: convolution layer, global pooling layer and multi-classification layer. The number of filters in the convolution layer is set to k, the convolution kernel size is set to l, and the learning step size is set to 1. Step S5.2: The relationship between the input data and the output data in the convolutional layer is: y ij =f ks ({x si+δi,sj+δj } 0≤δi,δj＜k ) Where: k is the size of the convolution kernel, s is the step size or sampling factor, f ks Determine the type of layer, x ij is the data vector written at position (i, j) in a particular layer, y ij is the data vector output at position (i, j) in a specific layer; x si+δi,sj+δj is the convolution kernel coverage area with (si,sj) as the center and size k×k in the input data matrix x; si is the row coordinate of the center of the current convolution kernel, sj is the column coordinate of the center of the current convolution kernel; the two variables δi and δj represent the offset of the convolution kernel in the row direction and column direction respectively; Step S5.3: The transformation rules of the convolution kernel and step size are as follows: In the above formula: f ks represents a convolution kernel with a size of k and a step size of s; g k′s′ represents a convolution kernel with a size of k′ and a stride of s′; Represents a convolution operation, which means performing a convolution operation on the input data; Indicates the use of convolution kernel f ks Perform a convolution operation on the input data, which will generate a new feature map, and then the convolution kernel g k′s′ Apply it to the feature map and perform another convolution operation; The equivalent convolution kernel size and stride after the composite of the two convolution operations are k′+(k-1)s′, ss′; The upsampling convolution formula is: In the above formula: γ is the upsampling factor, y ij is the user electricity consumption data point with high adjustable potential at position (i, j) after data reconstruction; x represents the input data matrix; α and β are the indexes inside the convolution kernel, which take values ​​0 or 1 and are used to traverse the four pixels around the reference point; i / γ and j / γ are used to map the coordinates i, j of the output data back to the coordinates of the input data; {i / γ} and {i / j} represent the decimal parts of i / γ and i / j, respectively, indicating the offset near the reference point coordinates; [i / γ] and [j / γ] represent the integer parts of i / γ and j / γ, respectively, indicating the reference point coordinates in the input data; |1-α-{i / γ}| and |1-β-{i / j}| calculate the weights of the data points in the horizontal and vertical directions, respectively.

6. The method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model according to claim 1, characterized in that: In step S7, the mean square error objective function is solved based on the improved BKA optimization algorithm to determine the initial number of filters k and the convolution kernel size l of the FCN network model, specifically: Step S7.1: Initialize the position state of the BK individual to vector σ = (k, l), where vectors k and l correspond to the initial number of filters k and convolution kernel size l of the FCN network model respectively. The position of each BK individual combines the information of the initial number of filters k and convolution kernel size l parameters, and each BK individual represents a solution to the mean square error objective function; set the initial parameters of BKA: population size pop, dimension dim, maximum number of iterations T, lower limit of solution space l b 、Find the upper limit u of the solution space b , the parameter p used to control the balance between global and local search, the variable r = rand (0, 1), then the position of the randomly initialized pop BK in the optimization range is y i,j , each BK individual is a potential solution that satisfies the mean square error objective function; Step S7.2: Use the refraction reverse learning strategy to generate the refraction reverse population Step S7.3: Merge the initial population and the refraction reverse population generated by the refraction reverse learning strategy According to the mean square error objective function, the fitness value of the entire population is calculated in the initial stage of iteration, and the fitness values ​​are arranged in descending order. The first pop BK individuals with the best fitness value are selected as the initial population. In this initialization stage, the individual with the best fitness value in the initial population is selected as the leader. The larger the mean square error, the greater the fitness value of the BK individual. Step S7.4: Since BK has two kinds of attack behaviors during flight, which are used for global exploration and local search processes respectively, and these two processes are controlled by constant p∈[0,1], the dynamic switching probability is used to optimize BKA to balance the proportion of local and global searches to achieve a better optimization strategy, and update the position and fitness value of the best BK individual in the current BK population under the current attack strategy; Step S7.5: During the migration of the BK population, if the fitness value of the current population is less than the fitness value of the random population, the leader of the current population will be given leadership and join the migrating population; if the fitness value of the current population is greater than the fitness value of the random population, it will guide the population until it reaches the destination. The position of the leader in each population of the BK individuals changes according to this strategy. Since the range of the global optimal solution has been basically determined in the later search, the Levy flight strategy is used to adjust the leader position, disturb it, and enhance its local escape ability, which can improve the quality of the algorithm's later solutions and enable the algorithm to converge to the global optimum; Step S7.6: Record the BK individual with the best fitness value in each iteration process, repeat the iteration process until the maximum number of iterations T is reached, and select the BK with the highest fitness value from the best BK leader individual in each iteration as the final result of the optimization.

7. The method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model according to claim 6 is characterized in that: In step S7.1, the position state of the BK individual is initialized as follows: Initialize the population as follows: In the above formula: BK ij is the i-th BK in the j-th dimension; Evenly distribute the position of each BK as follows: X i =BK lb +rand(UK ub -BK lb ) Where: X i is the position of the i-th black-winged kite individual; BK lb and BK ub are the lower and upper limits of the i-th black-winged kite individual in the j-th dimension; rand is a random value between [0,1]; During the initialization process, BKA selects the individual with the best fitness value in the initial population as the leader X L , which is considered the best position for BK, and is selected as follows: f best =min(f(X i )) X L =X(find(f best ==f(X i ))) Where: f(X i ) represents the fitness value of the i-th black-winged kite individual; f best Indicates the best fitness value; X represents the spatial position; X L Indicates the position of the black-winged kite individual with the best fitness value (the leader in the population); The mathematical model of BK attack behavior is as follows: In the above formula: and denotes the position of the i-th BK in the j-th dimension in the t-th and (t+1)-th iteration steps respectively; is a random number ranging from 0 to 1; T is the total number of iterations, t is the number of iterations completed so far; r is a random number between 0 and 1; The mathematical model of BK migration behavior is as follows: m=2×sin(r+π / 2) In the above formula: denotes the leading scorer of BK in the jth dimension at the tth iteration so far; and represents the position of the i-th BK in the j-th dimension in the t-th and (t+1)-th iteration steps respectively; F i represents the fitness value of the current position in the jth dimension obtained by any BK in t iterations; F ri represents the fitness value of the random position of the jth dimension obtained from any BK in the tth iteration; C(0,1) represents the Cauchy mutation; The one-dimensional Cauchy distribution is a continuous probability distribution with two parameters. The probability density function of the one-dimensional Cauchy distribution is: x: is a random variable, representing the value of calculating its probability density; δ: is the scale parameter of the distribution, which controls the width of the distribution, that is, the degree of diffusion of the distribution; μ: is the location parameter of the distribution, which determines the center position of the distribution, that is, the location of the peak in the Cauchy distribution; When δ=1,μ=0, its probability density function becomes the standard form, and the following is the exact formula:

8. The method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model according to claim 6 is characterized in that : In step S7.2, the refraction reverse learning strategy is adopted to generate the refraction reverse population Specifically: According to the principle of refraction reverse learning, the reverse learning formula can be derived as follows: * =v+uy, Extending to the high-dimensional space of BKA, setting n1 = 1, we can get the following: In the above formula, y i,j is the position of the i-th BK in the population in the j-dimensional position (i=1,2,...,D; j=1,2,...,N), D is the population number, N is the dimension; for y i,j The refraction reverse position; y represents the position of the initialization BK individual, y * represents the refraction reverse position of y, u j represents the upper limit of the optimal range of the black-winged kite individual in the j-dimensional position, v j represents the lower limit of the optimal search range of the black-winged kite individual in the j-dimensional position, ε represents the scaling factor, and ω represents the refractive index.

9. The method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model according to claim 6, characterized in that: Dynamic switching probability is used to optimize BKA, specifically: The formula of dynamic switching probability p is: p=0.6-0.1×(MaxIter-t) / MaxIter, where: MaxIter is the maximum number of iterations, and t is the current number of iterations.

10. The method for extracting flexible load high adjustable potential data based on an improved fully convolutional network model according to claim 6, characterized in that: The leader position is adjusted using the Levy flight strategy, specifically: In the above formula: represents the position of the jth dimension of the ith individual at time t; C(0,1) represents a uniformly distributed random number from 0 to 1; represents the Levy distribution factor of the jth dimension at time t; L (λ) represents the scaling factor of the Levy distribution; F i represents the fitness value of the i-th individual; F ri represents the reference fitness value of the i-th individual; m is a constant; L (λ) Random search path for Levy flight: L(s,λ)=η~s -λ ,1<λ≤3 Among them, s is the random step length of Levy flight; η is a probability density function in power form with λ as the power; since the Levy distribution is very complex, the Mantegna algorithm is used to simulate it, and the step length calculation formula is as follows: In the formula, η and ν are two normal random variables with standard deviations, σ η and σ ν is the standard deviation of its corresponding normal distribution, The calculation formula is as follows:

11. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the flexible load highly adjustable potential data extraction method based on the improved fully convolutional network model as described in any one of claims 1-10.

12. An electronic device comprising: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, wherein the instructions are executed by the at least one processor so that the at least one processor can execute the flexible load high-adjustable potential data extraction method based on the improved fully convolutional network model as described in any one of claims 1-10.

Images