Distributed resource power prediction method and device

Abstract

The invention provides a distributed resource power prediction method and apparatus. The method comprises the steps of obtaining a bus power signal, an electricity price and environmental factor data; calculating the response capability, depth, duration, fluctuation entropy and time shift coefficient of virtual aggregation; determining four types of virtual aggregation feature centers through fuzzy C-means clustering; decomposing the bus power signal by using a multi-variable variational mode decomposition algorithm to obtain an intrinsic mode component; determining four types of virtual aggregation power proportions according to the matching degree; independent power components are generated through rapid independent component analysis; constructing a long short-term memory neural network prediction model; and generating advanced power prediction results of four types of virtual aggregation. According to the method, the independent source signals related to the virtual aggregation resources are extracted, power signals of different virtual aggregation types are effectively separated, interference is reduced, and the power prediction accuracy is improved.

Description

Technical Field

[0001] This application belongs to the field of power calculation, and particularly relates to a distributed resource power prediction method and device. Background Art

[0002] In order to match the resource aggregation work with the scheduling requirements and support the business observability and controllability of the scheduling control center in a large number of distributed resources, for the key business scenarios where distributed resources participate in demand response adjustment, it is necessary to accurately predict the power of distributed resources. However, traditional power prediction methods often ignore the diversity and complexity of distributed resources and are difficult to accurately reflect their actual operating characteristics. Summary of the Invention

[0003] The purpose of this application is to overcome the defects in the above-mentioned prior art and provide a distributed resource power prediction method and device.

[0004] This application provides a distributed resource power prediction method, including:

[0005] Obtain bus power signals, electricity price data, and environmental factor data;

[0006] Calculate the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient of virtual aggregation according to the bus power signal;

[0007] Based on the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient, determine the characteristic clustering centers of four types of virtual aggregations through fuzzy C-means clustering;

[0008] Decompose the bus power signal through a multivariate variational mode decomposition algorithm to obtain multiple intrinsic mode components;

[0009] Determine the power proportion of four types of virtual aggregations according to the matching degree between the intrinsic mode components and the characteristic clustering centers;

[0010] Generate independent power components of four types of virtual aggregations from the intrinsic mode components through a fast independent component analysis algorithm;

[0011] Based on the independent power components, the electricity price data, and the environmental factor data, construct a long short-term memory neural network prediction model;

[0012] Generate an advanced power prediction result of four types of virtual aggregations through the long short-term memory neural network prediction model.

[0013] Optionally, determining the characteristic clustering centers of four types of virtual aggregations through fuzzy C-means clustering includes:

[0014] Initialize the random clustering center vectors for the four types of virtual aggregations;

[0015] Update the membership matrix according to the distance between the characteristic coefficients of the sample power curve and the clustering center;

[0016] Make the change amount of the clustering center less than the preset threshold through iterative calculation, and output the final characteristic clustering center.

[0017] Optionally, when decomposing the bus power signal by the multivariate variational mode decomposition algorithm, it further includes:

[0018] Perform correlation detection on the intrinsic mode components and the original bus power signal;

[0019] Select the intrinsic mode components with a correlation higher than the preset threshold as effective components.

[0020] Optionally, when generating independent power components by the fast independent component analysis algorithm, it specifically includes:

[0021] Construct an observation signal matrix, the column vectors of which are the selected intrinsic mode components;

[0022] Maximize the non-Gaussianity of the independent components by iteratively updating the separation matrix;

[0023] When the change amount of the separation matrix converges, output the independent power components of the four types of virtual aggregations.

[0024] Optionally, when constructing a long short-term memory neural network prediction model, it further includes:

[0025] Determine the number of hidden layer neurons, learning rate, and number of iterations based on the sparrow search algorithm;

[0026] Use the historical independent power components, electricity price data, and environmental factor data as input features;

[0027] Use the independent power components in the future time period as output labels to train the model parameters.

[0028] This application also provides a distributed resource power prediction device, which is characterized by including:

[0029] An acquisition module that acquires the bus power signal, electricity price data, and environmental factor data;

[0030] A calculation module that calculates the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient of the virtual aggregation according to the bus power signal;

[0031] The clustering module determines the characteristic clustering centers of four types of virtual aggregations through fuzzy C-means clustering based on the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient;

[0032] The decomposition module decomposes the bus power signal through the multivariate variational mode decomposition algorithm to obtain multiple intrinsic mode components;

[0033] The matching module determines the power proportion of four types of virtual aggregations according to the matching degree between the intrinsic mode components and the characteristic clustering centers;

[0034] The component module generates independent power components of four types of virtual aggregations from the intrinsic mode components through the fast independent component analysis algorithm;

[0035] The modeling module constructs a long short-term memory neural network prediction model based on the independent power components, the electricity price data, and the environmental factor data;

[0036] The prediction module generates the leading power prediction results of four types of virtual aggregations through the long short-term memory neural network prediction model.

[0037] Optionally, the clustering module determines the characteristic clustering centers of four types of virtual aggregations through fuzzy C-means clustering, including:

[0038] Initializing the random clustering center vectors of four types of virtual aggregations;

[0039] Updating the membership matrix according to the distance between the characteristic coefficients of the sample power curve and the clustering centers;

[0040] Making the change amount of the clustering centers less than a preset threshold through iterative calculation, and outputting the final characteristic clustering centers.

[0041] Optionally, when the decomposition module decomposes the bus power signal through the multivariate variational mode decomposition algorithm, it further includes:

[0042] Performing correlation detection on the intrinsic mode components and the original bus power signal;

[0043] Selecting the intrinsic mode components with a correlation higher than a preset threshold as effective components.

[0044] Optionally, when the component module generates independent power components through the fast independent component analysis algorithm, it specifically includes:

[0045] Constructing an observation signal matrix, the column vectors of which are the selected intrinsic mode components;

[0046] Maximizing the non-Gaussianity of the independent components by iteratively updating the separation matrix;

[0047] When the change amount of the separation matrix converges, output the independent power components of the four types of virtual aggregations.

[0048] Optionally, when the modeling module constructs the long short-term memory neural network prediction model, it further includes:

[0049] Determine the number of hidden layer neurons, learning rate, and number of iterations based on the sparrow search algorithm;

[0050] Use the historical independent power components, electricity price data, and environmental factor data as input features;

[0051] Use the independent power components in future time periods as output labels to train the model parameters.

[0052] The beneficial effects of this application are:

[0053] This application provides a distributed resource power prediction method, including: obtaining bus power signals, electricity price data, and environmental factor data; calculating the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient of virtual aggregation according to the bus power signals; based on the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient, determine the characteristic clustering centers of the four types of virtual aggregations through fuzzy C-means clustering; decompose the bus power signals through the multivariate variational mode decomposition algorithm to obtain multiple intrinsic mode components; determine the power ratios of the four types of virtual aggregations according to the matching degrees between the intrinsic mode components and the characteristic clustering centers; generate the independent power components of the four types of virtual aggregations from the intrinsic mode components through the fast independent component analysis algorithm; based on the independent power components, the electricity price data, and the environmental factor data, construct a long short-term memory neural network prediction model; generate the advanced power prediction results of the four types of virtual aggregations through the long short-term memory neural network prediction model. This application obtains bus power signals and virtual aggregation resource classification parameters, and uses multivariate variational mode decomposition and independent component analysis technologies to extract independent source signals related to virtual aggregation resources from complex bus power signals, effectively separating the power signals of different virtual aggregation types and reducing interference between signals, thereby improving the accuracy of power prediction. Description of the Drawings

[0054] Figure 1 is the schematic diagram of the distributed resource power prediction process in this application;

[0055] Figures 2 - 5 is the power curves of the four types of virtual aggregation resources before and after response in this application;

[0056] Figure 6 is the calculation process of SSA-LSTM in this application;

[0057] Figure 7 It is a schematic diagram of signal decomposition in this application;

[0058] Figure 8 It is a flowchart of the algorithm in this application;

[0059] Figure 9 It is a schematic diagram of data simulation generation in this application;

[0060] Figure 10 It is the first schematic diagram of the performance of the prediction model in this application;

[0061] Figure 11 It is the second schematic diagram of the performance of the prediction model in this application. Detailed implementation manners

[0062] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it can be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, the embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be completely conveyed to those skilled in the art.

[0063] Please refer to Figure 1 、 8 As shown, this application provides a distributed resource power prediction method, including:

[0064] S101. Obtain bus power signals, electricity price data, and environmental factor data;

[0065] Collect bus power signals through smart meters, obtain time-of-use electricity price data from the electricity market database, and obtain environmental factor data such as temperature, humidity, and light intensity from weather stations to form a time series data set.

[0066] S102. Calculate the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient of virtual aggregation according to the bus power signal;

[0067] According to the operating characteristics and demand response characteristics of distributed resources, they are divided into 4 categories: discrete adjustable resources, flexible adjustable resources, time-varying resources, and non-adjustable resources.

[0068] One type of aggregation includes discrete adjustable resources: having two-way adjustment capabilities, which can not only quickly adjust the working power but also help to approximately instantaneously adjust the working power. The adjustment is flexible but the adjustment range is small, and it cannot be adjusted frequently.

[0069] Discrete adjustable resources have the characteristics of discrete adjustment, which are manifested as:

[0070] (1) Having a certain adjustable space;

[0071] (2) Fast adjustment speed within the limited adjustment capacity range;

[0072] (3) Once adjusted, it must ensure to maintain power within a limited time and cannot respond to frequent adjustment needs.

[0073] The second - type aggregation includes flexible adjustable resources: with characteristics such as two - way continuous adjustment ability, flexible resource adjustment, and smooth ramping.

[0074] Flexible adjustable resources have the characteristic of continuous adjustment, which is specifically manifested as:

[0075] (1) Can be continuously adjusted within a certain capacity adjustment range;

[0076] (2) There is no constraint on the power stable duration requirement;

[0077] (3) Can respond to adjustment demands at any time.

[0078] The third - type aggregation includes time - varying resources: which can achieve daily adjustment. According to the requirements of grid dispatching, the response speed of its power adjustment can reach the millisecond level. The response speed of time - shift resources is very fast, and there is no adjustment time constraint for this. Time - shift resources can achieve power transfer by changing the operation cycle. One of its characteristics is that the total electric energy within a day is constant, only the operation power cycle changes.

[0079] The third - type aggregation includes time - shift resources with adjustable characteristics, which are specifically manifested as:

[0080] (1) Continuously adjustable within a certain capacity adjustment range;

[0081] (2) Can respond to adjustment demands. By adjusting the capacity of the resources, the adjustment speed can be quickly and flexibly adjusted;

[0082] (3) Within the capacity adjustment range of the resources, the capacity output transfer of the resources can be adjusted by adjusting the capacity of the resources, and the load electricity consumption time can be shifted in response to peak - valley electricity prices;

[0083] (4) Can ensure that the total electricity consumption remains unchanged before and after the load transfer.

[0084] The first - type virtual aggregation and the second - type virtual aggregation increase or decrease the electricity consumption in a specific time period through different forms of adjustment (discrete adjustment, continuous adjustment, etc.), while time - shift resources transfer the power consumption in a specific time period.

[0085] The fourth - type aggregation includes non - adjustable resources: generally not participating in demand response, and its operating characteristics do not change under electricity prices or incentives. The working curves before and after adjustment are an overlapping curve.

[0086] Please refer to Figures 2 - 5As shown, according to the above analysis, the biggest difference between the first three types of virtual polymers and the fourth type of virtual polymer is that: the operating power curves of the fourth type of virtual polymer have no difference before and after demand response adjustment; at the same time, the operating power curves of the other three types of virtual polymers have changed.

[0087] Therefore, responsiveness I1 is designed as the first characteristic index, which is defined as the Pearson correlation coefficient of the power curves of the virtual polymer before and after response. In statistics, the Pearson correlation coefficient can represent the strength of the linear correlation between two variables:

[0088]

[0089] where cov() is the covariance function; E is the mathematical expectation function; E is the mean square deviation; U and V are two variables. If the absolute value of the correlation coefficient is greater than 0.5, it is considered that these two variables have a strong correlation. According to the definition of the Pearson correlation coefficient, the correlation coefficient between the pre-adjustment standardized virtual polymer operating power Pa and the post-adjustment standardized virtual polymer operating power Pb can be obtained:

[0090] I1 = r(P a ,P b )

[0091] The response depth I2 is defined as the maximum value of the power change during the operation of the virtual polymer before and after adjustment, and this is mathematically expressed as follows:

[0092]

[0093] where is the standardized virtual polymer operating power adjusted for the t time period; is the standardized virtual polymer operating power during the adjusted time period t; and T is the total number of time slots in a day.

[0094] From the response characteristics of the first and second types of virtual polymers, the obvious difference between the two is that the first type of virtual polymer can maintain stable operation for a period of time after response, while the second type of virtual polymer can be flexibly adjusted under large power changes.

[0095] Therefore, the response duration I3 is selected as the third index, and its mathematical expression is as follows:

[0096]

[0097] In the formula, I3 is the average value of the virtual polymer operating power before and after adjustment during the adjustment period s; m is the total number of time periods when the difference between the two curves is greater than 0.01; is the time period when the difference between the two curves is greater than 0.01; ε = 10-6 is a positive number with a particularly small value to avoid the case of a zero denominator.

[0098] Considering that the operating power curve fluctuations of various virtual polymers are different before and after adjustment, the response fluctuation entropy is introduced as the fourth index I4.

[0099] First, introduce the statistic (t) as follows:

[0100]

[0101] The statistic τ(t) reflects the fluctuation degree of the virtual polymer's operating power curve before and after adjustment, and its value range is [0, 1]. If the value range [0, 1] is divided into k equal parts, then the k intervals can be expressed as:

[0102]

[0103] The statistical data can be viewed at each moment of the day. Assuming that the moment point distributed in the k-th interval is Φ, then the probability that Φ(t) is within this interval is as follows:

[0104] v(t) = m k / T

[0105] Define the response fluctuation entropy I4 as:

[0106]

[0107] According to the definition of the response fluctuation entropy, when the response fluctuation entropy (t) all falls within the same interval, the response fluctuation entropy is the smallest, and its value is equal to 0. When the response fluctuation entropy of any period is not within the same interval, the response fluctuation entropy is the largest.

[0108] The response time shift coefficient I5 is defined as the ratio of the sum of the power changes in each period before and after the adjustment of the virtual polymer to the sum of the absolute values of the power changes in each period of the virtual polymer. This reflects the changes of the two curves in each period. Its mathematical expression is as follows:

[0109]

[0110] Considering that the absolute sum of the power changes of the four types of virtual polymers before and after regulation is zero, to avoid the case of a zero denominator, a minimum positive number ε is added to both the numerator and denominator of the expression. Therefore, take ε = 10 -6 . In this case, the response time shift coefficient is 1. For three types of virtual polymers, the time shift coefficient is close to 0, while for the other three types of virtual polymers, the time shift coefficient is close to 1.

[0111] S103. Based on the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient, determine the characteristic clustering centers of four types of virtual aggregations through fuzzy C-means clustering;

[0112] Select the power curve samples before and after adjusting the typical virtual aggregates, and calculate the demand response characteristic coefficients of each sample.

[0113] The vector Ri of the characteristic coefficients of the demand response of the i-th power curve sample is as follows:

[0114]

[0115] Calculate the clustering centers of the demand response characteristic coefficients of all N samples of a typical virtual aggregate. In this paper, the fuzzy C-clustering algorithm is used, and its basic steps are as follows:

[0116] Select the number of clustering centers.

[0117] Randomly initialize the clustering centers from the sample points, as follows:

[0118]

[0119] where v j is the j-th clustering center vector.

[0120] Initialize the membership matrix:

[0121] Update the cluster centers and the membership relationship matrix, as shown in the following formula:

[0122]

[0123]

[0124] where m is the fuzzy parameter (usually greater than 1).

[0125] When the change of the clustering center is less than the set threshold, or reaches the preset maximum number of iterations, the algorithm terminates; otherwise, continue the iteration.

[0126] To sum up, by setting the number of clustering centers to C = 1, the demand response characteristics of various typical virtual aggregates are extracted using the fuzzy C-clustering algorithm.

[0127] S104. Decompose the bus power signal through the multivariate variational mode decomposition algorithm to obtain multiple intrinsic mode components;

[0128] As Figure 7 shown, perform multivariate variational mode decomposition on the bus power signal, and obtain 6 IMF components through Gaussian window transformation and iterative balancing.

[0129] Multivariate Variational Mode Decomposition (MVMD) separates modal components from multivariate time series data. Traditional Variational Mode Decomposition.

[0130] VMD is a signal decomposition technique based on an optimization method that decomposes a time series into a series of Intrinsic Mode Functions (IMFs). MVMD is a specialized extension of VMD, particularly suitable for analyzing multivariate time series data. The main concept behind MVMD is to treat the multivariate time series as a matrix and use optimization techniques to decompose this matrix into a set of modal components. The main goal of MVMD optimization is to reduce the correlation between modal components and achieve decomposition by gradually balancing the interactions between different components. The decomposition process of MVMD includes the following steps:

[0131] (1) Perform a Gaussian window transform on the multivariate signal, deconstruct it, and obtain a set of multivariate window functions.

[0132] (2) The initial decomposition process generates initial Intrinsic Mode Functions (IMFs). The iterative steps enhance the matching degree of each IMF with the original signal and then minimize the error through iterative steps.

[0133] (3) Subtract the original signal from the IMFs obtained from the previous decomposition to obtain the residual as the input for the subsequent decomposition.

[0134] (4) Repeat steps 2 and 3 until convergence.

[0135] S105. Determine the power proportion of the four types of virtual aggregations according to the matching degree between the intrinsic mode components and the feature clustering centers;

[0136] The power proportion of each category can be statistically obtained by calculating the Euclidean distance between each IMF component and the four types of clustering centers and classifying the IMFs according to the principle of the minimum distance.

[0137] S106. Generate independent power components of the four types of virtual aggregations from the intrinsic mode components through the Fast Independent Component Analysis algorithm;

[0138] Construct the screened IMFs into an observation matrix X, and through the formula

[0139]

[0140] Iteratively update the separation matrix. When W * new -W * old <1e - 6, terminate and output the four types of independent power components.

[0141] S107. Based on the independent power components, the electricity price data, and the environmental factor data, construct a long short-term memory neural network prediction model;

[0142] As Figure 6 shown, the input dimension is [power component, electricity price, temperature, light], the output is the power for the next 24 periods, and the number of neurons in the hidden layer is determined by SSA optimization.

[0143] The Sparrow Search Algorithm (SSA) is a heuristic algorithm that simulates the foraging activities of sparrows. It can perform global and local searches. When using a neural network, it is crucial to carefully select the optimal number of neurons, learning rate, and iteration interval to achieve the best performance of the model.

[0144] This application constructs a suitable prediction model with multi-feature input and single-variable output by optimizing the number of neurons, learning rate, and iteration period in the hidden layer of the LSTM model.

[0145] S108. Generate the leading power prediction results of four types of virtual aggregations through the long short-term memory neural network prediction model.

[0146] Input the real-time data into the trained SSA-LSTM model, and output the 96-point power prediction curves of four types of virtual aggregations.

[0147] Furthermore, determine the characteristic clustering centers of the four types of virtual aggregations through fuzzy C-means clustering, including:

[0148] Initialize the random clustering center vectors of the four types of virtual aggregations: for example, randomly generate 4 five-dimensional vectors vj = [vj1,..., vj5].

[0149] After determining the value range of each dimension, update the membership matrix according to the distance between the characteristic coefficients of the sample power curve and the clustering center.

[0150] Make the change amount of the clustering center less than the preset threshold through iterative calculation, and output the final characteristic clustering center.

[0151] Furthermore, when decomposing the bus power signal through the multivariate variational mode decomposition algorithm, it further includes:

[0152] Perform correlation detection on the intrinsic mode components and the original bus power signal: for example, calculate the Pearson correlation coefficient between each IMF component IMFk and the original bus power Pbus.

[0153] Select the intrinsic mode components with a correlation higher than the preset threshold as the effective components, eliminate the high-frequency noise components with a correlation coefficient lower than the threshold, and construct a set of effective components.

[0154] Furthermore, when generating independent power components through the Fast Independent Component Analysis algorithm, it specifically includes:

[0155] Construct an observation signal matrix, whose column vectors are the screened intrinsic mode components: For example, arrange the 3 screened IMF components by column and supplement a column of pseudo-random noise to form a 4-dimensional observation signal.

[0156] Maximize the non-Gaussianity of the independent components by iteratively updating the separation matrix: For example, use negative entropy as the objective function for iterative update.

[0157] When the change amount of the separation matrix converges, output the independent power components of four types of virtual aggregation: For example, stop the iteration when the change rate of the Frobenius norm of the separation matrix W < 1e-6, and output the independent components S = WX corresponding to the four types of virtual aggregation power.

[0158] Furthermore, when constructing a long short-term memory neural network prediction model, it further includes:

[0159] Determine the number of hidden layer neurons, learning rate, and number of iterations based on the sparrow search algorithm:

[0160] For example: Initialize the sparrow population position as [number of neurons (20 - 200), learning rate (0.001 - 0.1), number of iterations (50 - 500)].

[0161] Use historical independent power components, electricity price data, and environmental factor data as input features:

[0162] For example: Use historical 72-hour data as input (including four types of power components, electricity price, temperature), predict the power for the next 24 hours, adopt Dropout = 0.2 to prevent overfitting, and set the batch size to 64.

[0163] Finally, use the independent power components in the future time period as output labels to train the model parameters.

[0164] Please refer to Figure 9 , and determine the learning rate and number of iterations of the LSTM model.

[0165] The learning rate determines the magnitude of the adjustment of the model weights during the entire update process. A too high learning rate will cause the model to diverge, while a too low learning rate will cause the model to converge slowly. The number of iterations represents the frequency of training the model on the entire dataset. Too many iterations will lead to overfitting, while insufficient iterations will lead to underfitting. By using the SSA technology, the optimal learning rate and number of iterations can be determined, thereby enhancing the performance of the model.

[0166] Finally, an appropriate prediction model must be developed that accepts multiple input features and produces a single output variable. Utilizing multi-feature input involves using numerous features to predict the dependent variable, thereby improving the prediction accuracy of the model. A single dependent variable output means that there is only one dependent variable for prediction, which can reduce the complexity of the model. By fine-tuning the number of neurons in the hidden layer, adjusting the learning rate and number of iterations of the LSTM model using SSA, and adopting a fitting prediction model with multi-input features and a single output variable, a model with superior performance can be obtained.

[0167] Please refer to Figure 10 、 11 and tested the prediction effect of the improved LSTM neural network prediction model. Due to the improvements proposed in this paper, the LSTM model has mined the coupling correlation of time series multivariate data, so it has superior advantages in prediction compared with traditional neural network models.

[0168] This application also provides a distributed resource power prediction device, which is characterized by including:

[0169] An acquisition module that acquires bus power signals, electricity price data, and environmental factor data;

[0170] A calculation module that calculates the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient of virtual aggregation according to the bus power signal;

[0171] A clustering module that determines the characteristic clustering centers of four types of virtual aggregation through fuzzy C-means clustering based on the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient;

[0172] A decomposition module that decomposes the bus power signal through a multivariate variational mode decomposition algorithm to obtain multiple intrinsic mode components;

[0173] A matching module that determines the power proportion of four types of virtual aggregation according to the matching degree between the intrinsic mode components and the characteristic clustering centers;

[0174] A component module that generates independent power components of four types of virtual aggregation from the intrinsic mode components through a fast independent component analysis algorithm;

[0175] A modeling module that constructs a long short-term memory neural network prediction model based on the independent power components, the electricity price data, and the environmental factor data;

[0176] A prediction module that generates an advanced power prediction result of four types of virtual aggregation through the long short-term memory neural network prediction model.

[0177] Further, the clustering module determines the characteristic clustering centers of four types of virtual aggregations through fuzzy C-means clustering, including:

[0178] Initializing the random clustering center vectors of four types of virtual aggregations;

[0179] Updating the membership matrix according to the distance between the characteristic coefficients of the sample power curve and the clustering centers;

[0180] Making the change amount of the clustering centers less than a preset threshold through iterative calculation, and outputting the final characteristic clustering centers.

[0181] Further, when the decomposition module decomposes the bus power signal through the multivariate variational mode decomposition algorithm, it further includes:

[0182] Performing correlation detection on the intrinsic mode components and the original bus power signal;

[0183] Selecting the intrinsic mode components with a correlation higher than a preset threshold as effective components.

[0184] Further, when the component module generates independent power components through the fast independent component analysis algorithm, it specifically includes:

[0185] Constructing an observation signal matrix, whose column vectors are the selected intrinsic mode components;

[0186] Maximizing the non-Gaussianity of the independent components by iteratively updating the separation matrix;

[0187] When the change amount of the separation matrix converges, outputting the independent power components of four types of virtual aggregations.

[0188] Further, when the modeling module constructs a long short-term memory neural network prediction model, it further includes:

[0189] Determining the number of hidden layer neurons, learning rate and number of iterations based on the sparrow search algorithm;

[0190] Taking historical independent power components, electricity price data and environmental factor data as input features;

[0191] Taking the independent power components in future time periods as output labels to train the model parameters.

[0192] The above description of the embodiments is for those of ordinary skill in the art in this technical field to understand and apply the present invention. Those who are familiar with the technology in this field can obviously make various modifications to the above embodiments easily, and apply the general principles described herein to other embodiments without creative labor. Therefore, the present invention is not limited to the above embodiments, and all improvements and modifications made by those skilled in the art according to the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A distributed resource power prediction method, characterized in that, Including: Obtain the bus power signal, electricity price data, and environmental factor data; Calculate the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient of virtual aggregation according to the bus power signal; Based on the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient, determine the characteristic clustering centers of four types of virtual aggregation through fuzzy C-means clustering; Decompose the bus power signal through the multivariate variational mode decomposition algorithm to obtain multiple intrinsic mode components; Determine the power proportion of four types of virtual aggregation according to the matching degree between the intrinsic mode components and the characteristic clustering centers; Generate the independent power components of four types of virtual aggregation from the intrinsic mode components through the fast independent component analysis algorithm; Construct a long short-term memory neural network prediction model based on the independent power components, the electricity price data, and the environmental factor data; Generate the advanced power prediction results of four types of virtual aggregation through the long short-term memory neural network prediction model.

2. The distributed resource power prediction method according to claim 1, wherein Determine the characteristic clustering centers of four types of virtual aggregation through fuzzy C-means clustering, including: Initialize the random clustering center vectors of four types of virtual aggregation; Update the membership matrix according to the distance between the characteristic coefficients of the sample power curve and the clustering centers; Make the change amount of the clustering centers less than the preset threshold through iterative calculation, and output the final characteristic clustering centers.

3. The distributed resource power prediction method according to claim 1, wherein When decomposing the bus power signal through the multivariate variational mode decomposition algorithm, it further includes: Perform correlation detection on the intrinsic mode components with the original bus power signal; Select the intrinsic mode components with a correlation higher than the preset threshold as effective components.

4. A distributed resource power prediction method according to claim 1, characterized in that When generating independent power components through the fast independent component analysis algorithm, it specifically includes: Construct an observation signal matrix, the column vectors of which are the selected intrinsic mode components; Maximize the non-Gaussianity of the independent components by iteratively updating the separation matrix; When the change amount of the separation matrix converges, output the independent power components of four types of virtual aggregation.

5. A distributed resource power prediction method according to claim 1, characterized in that, When constructing a long short-term memory neural network prediction model, it further includes: Determine the number of hidden layer neurons, learning rate, and number of iterations based on the sparrow search algorithm; Use the historical independent power components, electricity price data, and environmental factor data as input features; Use the independent power components in the future time period as output labels to train the model parameters.

6. A distributed resource power prediction device, characterized in that, Including: An acquisition module that acquires the bus power signal, electricity price data, and environmental factor data; A calculation module that calculates the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient of virtual aggregation according to the bus power signal; A clustering module that determines the characteristic clustering centers of four types of virtual aggregation through fuzzy C-means clustering based on the response ability coefficient, response depth coefficient, response duration coefficient, response fluctuation entropy coefficient, and response time shift coefficient; A decomposition module that decomposes the bus power signal through the multivariate variational mode decomposition algorithm to obtain multiple intrinsic mode components; A matching module that determines the power proportion of four types of virtual aggregation according to the matching degree between the intrinsic mode components and the characteristic clustering centers; The component module generates four types of virtual aggregated independent power components from the intrinsic mode components through the fast independent component analysis algorithm; The modeling module constructs a long short-term memory neural network prediction model based on the independent power components, the electricity price data, and the environmental factor data; The prediction module generates four types of virtual aggregated advanced power prediction results through the long short-term memory neural network prediction model.

7. A distributed resource power prediction device according to claim 6, characterized in that The clustering module determines the characteristic clustering centers of the four types of virtual aggregations through fuzzy C-means clustering, including: Initializing the random clustering center vectors of the four types of virtual aggregations; Updating the membership matrix according to the distance between the characteristic coefficients of the sample power curve and the clustering center; Making the change amount of the clustering center less than the preset threshold through iterative calculation, and outputting the final characteristic clustering center.

8. The distributed resource power prediction device as claimed in claim 6, wherein, When the decomposition module decomposes the bus power signal through the multivariate variational mode decomposition algorithm, it further includes: Performing correlation detection on the intrinsic mode components and the original bus power signal; Selecting the intrinsic mode components with a correlation higher than the preset threshold as effective components.

9. The distributed resource power prediction device according to claim 6, characterized in that, When the component module generates independent power components through the fast independent component analysis algorithm, it specifically includes: Constructing an observation signal matrix, the column vectors of which are the selected intrinsic mode components; Maximizing the non-Gaussianity of the independent components by iteratively updating the separation matrix; When the change amount of the separation matrix converges, outputting the independent power components of the four types of virtual aggregations.

10. The distributed resource power prediction device according to claim 6, wherein When the modeling module constructs a long short-term memory neural network prediction model, it further includes: Determining the number of hidden layer neurons, the learning rate, and the number of iterations based on the sparrow search algorithm; Taking the historical independent power components, electricity price data, and environmental factor data as input features; Taking the independent power components in the future time period as output labels to train the model parameters.

Images