CN115907118A - Comprehensive load short-term prediction method based on coupling characteristic matrix time sequence segment analysis - Google Patents

Comprehensive load short-term prediction method based on coupling characteristic matrix time sequence segment analysis Download PDF

Info

Publication number
CN115907118A
CN115907118A CN202211409152.8A CN202211409152A CN115907118A CN 115907118 A CN115907118 A CN 115907118A CN 202211409152 A CN202211409152 A CN 202211409152A CN 115907118 A CN115907118 A CN 115907118A
Authority
CN
China
Prior art keywords
load
time
sample
time sequence
coupling
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211409152.8A
Other languages
Chinese (zh)
Inventor
张磊
仇茹嘉
王小明
吴红斌
徐斌
谢毓广
滕越
彭勃
毕锐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd
Hefei University of Technology
Original Assignee
Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd
Hefei University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd, Hefei University of Technology filed Critical Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd
Priority to CN202211409152.8A priority Critical patent/CN115907118A/en
Publication of CN115907118A publication Critical patent/CN115907118A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a comprehensive load short-term prediction method based on coupling feature matrix time sequence segment analysis, which comprises the following steps: 1. acquiring and normalizing original data, arranging load data into a form of a load coupling characteristic matrix, and constructing a time sequence fragment sample of the load coupling characteristic matrix and external parameters through sliding time window recombination; 2. simultaneously extracting sample time sequence characteristics and load form coupling characteristics through a 3DCNN (data communication network), and performing clustering analysis by using FCM (fuzzy c-means) clustering; 3. based on a multi-task learning framework and combined with an LSTM network, an LSTM-MTL model is constructed to realize comprehensive load prediction, a training set and a verification set are iteratively reconstructed by utilizing a membership degree search mechanism, and the final prediction performance is improved. Aiming at the problem that the comprehensive load prediction performance is influenced by the complex energy coupling form, the load fluctuation feature extraction capability is enhanced by carrying out cluster analysis on the time sequence segment samples of the coupling feature matrix, so that the prediction precision of the comprehensive load is effectively improved.

Description

Comprehensive load short-term prediction method based on coupling characteristic matrix time sequence fragment analysis
Technical Field
The invention relates to the technical field of short-term comprehensive load prediction, in particular to a comprehensive load short-term prediction method based on coupling characteristic matrix time sequence fragment analysis.
Background
With the continuous development of new energy technology, an integrated energy system comprising various energy load forms becomes one of the most potential key technologies. However, the energy forms in the comprehensive energy system are various, the coupling relationship is tight, the uncertainty of load fluctuation is increased, and great challenges are brought to the formulation of the energy system scheduling plan. Therefore, it is necessary to provide an integrated load prediction method for the load characteristics of the integrated energy system.
Currently, the comprehensive load prediction is usually based on historical data of electric, thermal and gas loads, and the historical change rule is learned by methods such as machine learning and neural network, so that a prediction model is trained to realize the comprehensive load prediction. In order to improve the learning capability of the model, the prior research usually adopts preprocessing methods such as cluster analysis to obtain clearer data characteristics.
Most of the existing research results still have rough utilization of three load data of electricity, heat and gas, or different load forms are separated to independently establish a prediction model, or the three load forms are directly brought into a multi-task learning model for direct training. However, the complex load form of the comprehensive energy system and the learning of a single change rule cannot accurately reflect the change trend of the load, so that the traditional prediction scheme has low precision in the field of comprehensive load prediction. In addition, the conventional prediction scheme usually adopts the time window length of one day as a sample during the clustering preprocessing, so that the local change characteristic of load fluctuation is covered, the local change characteristic is difficult to reflect in a clustering result, and the prediction process cannot be substantially helped. In the prediction method, the conventional scheme selects a hard boundary clustering method more, so that the transitional data of a high probability type is learned, or the data capacity of a low probability type is too small, and the difficulty is increased for model learning.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, and provides a comprehensive load short-term prediction method based on coupled feature matrix time sequence segment analysis, so that two change rules of historical load data can be explicitly learned, the extraction capability of load fluctuation features is enhanced, and the accuracy and the prediction precision of the comprehensive load short-term prediction are improved.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
the invention relates to a comprehensive load short-term prediction method based on coupling characteristic matrix time sequence fragment analysis, which is characterized by comprising the following steps of:
constructing time sequence fragment data of a load coupling characteristic matrix and external parameters based on a sliding time window;
step 1.1, acquiring a historical data sequence of a comprehensive load and a historical data sequence of relevant external influence factors from a comprehensive energy system, respectively carrying out normalization processing, and correspondingly acquiring a historical coupling characteristic matrix and a historical relevant external influence factor matrix, wherein the coupling characteristic matrix of a sampling time t is marked as L t And L is t ＝{l t (n L1 ,n L2 )|n L1 ＝1,2,3；n L2 =1,2,3}; the matrix of the relevant external influence factors at the sampling time t is denoted as p t And P is t ＝{p t (n P )|n P ＝1,2,…,N P In which n is L1 And n L2 The values are 1,2 and 3 respectively, and correspondingly represent three load forms, namely an electric load, a heat load and a natural gas load; l t (n L1 ,n L2 ) Indicating that the sample time t is counted by the nth L1 Variety of load forms to the nth L2 The amount of the seed load form converted; when n is L1 ＝n L2 When l is turned on t (n L1 ,n L2 ) Indicating the sampling instant tth L1 The power consumed by the load form on the user side without conversion; pt (n) P ) N-th of the sampling time t P (ii) seed external influencing factors; n is a radical of hydrogen P Representing the number of types of external influencing factors considered;
step 1.2, setting time length t TW As the length of the time window, sliding the time window on the historical load coupling characteristic matrix; each sliding by one sampling step t S Obtaining the coupled characteristic time sequence segment data of a sampling time t
Figure BDA0003937770570000021
Wherein, t n Representing the sampling instant, T, under the current time window n ＝t TW /t S Representing the number of sample points in the time series fragment data; />
Figure BDA0003937770570000022
Representing the sampling instant t in the current time window n The coupling feature of (a);
step 1.3, the time window is set atSliding on the historical related external influence factor matrix, wherein each sliding is performed by one sampling step length t S Obtaining external influence factor time sequence segment data of a sampling time t
Figure BDA0003937770570000023
Figure BDA0003937770570000024
Representing the sampling instant t in the current time window n External influence factors of (1);
step 1.4, reconstructing the historical coupling characteristic matrix and the historical related external influence factor matrix to obtain a time sequence fragment sample data set
Figure BDA0003937770570000025
Wherein it is present>
Figure BDA0003937770570000026
Time-sequential fractional samples representing sampling instants T D Let T be the total number of sampling time of the historical data S Represents the total number of time-series fragment samples, and T S ＝T D -T n +1；
Step two, performing cluster analysis based on a 3D-CNN network and an FCM method;
step 2.1, constructing a 3D-CNN network, and comprising a data input layer, a convolution layer, a pooling layer and an output layer; and adopting the 3D-CNN network to carry out coupling characteristic time sequence fragment data of the sampling time t
Figure BDA0003937770570000027
And (3) carrying out feature extraction:
step 2.2, the receiving dimension of the convolution layer through the input layer is 3 multiplied by T n Is/are as follows
Figure BDA0003937770570000028
And use the pair of formula (1)
Figure BDA0003937770570000029
Treated to obtain dimension of 2 × 2 × (T) n -2) ofAdjacent moment coupling characteristic out conv ；
Figure BDA00039377705700000210
In formula (1):
Figure BDA00039377705700000211
coupling features out for adjacent time instants conv The middle width is the coupling characteristic of the position where i is high and j is deep and k is deep; w is a I,J,K Weights for positions in the convolutional layer where the convolutional kernels are of the form 2 × 2 × 3, and the width is I, the height is J, and the depth is K; in I+i-1,J+j-1,K+k-1 Is->
Figure BDA0003937770570000031
Data with a middle width of I + I-1, a height of J + J-1 and a depth of K + K-1; bias is the network bias to be trained; reLU is an activation function;
step 2.3, the pooling layer utilizes the adjacent time coupling characteristic out of the formula (2) conv Processing along the depth K to obtain a dimension T n One-dimensional array of coupling characteristics out of-2 pool And as time-sequential fractional samples of the sampling instant t
Figure BDA0003937770570000032
Corresponding time-sequential segment characteristic->
Figure BDA0003937770570000033
To be output by the output layer;
Figure BDA0003937770570000034
in formula (2):
Figure BDA0003937770570000035
obtaining the coupling characteristic output on the depth k for the pooling layer by utilizing an average pooling mode with the dimensionality of 2 multiplied by 2;
step 2.4, time sequence fragment characteristics are determined based on FCM method
Figure BDA0003937770570000036
Performing cluster analysis:
step 2.4.1, calculating T in the sample data set Dataset of the time sequence segment by adopting the Euclidean distance formula S A time sequence of fragment samples and N c Distance between clustering kernels d mc |m＝1,2,…,T S ；c＝1,2,…,N c In which d is mc Represents the distance, N, between the mth time series fragment sample and the c-th cluster core c Is the total cluster number;
step 2.4.2, calculating the membership degree u of the mth time sequence fragment sample to the c clustering core by using the formula (3) mc ；
Figure BDA0003937770570000037
In formula (3): theta belongs to [1, ∞) as a membership degree weighting coefficient; d mn Representing a distance between the mth time series segment sample and the nth clustering core;
step 2.4.3, dividing each time sequence fragment sample into corresponding N according to membership degree c In each class, calculating the characteristic mean value of all samples in each class, thereby updating the clustering core of each class;
step 2.5, iterative computation is carried out according to the process of 2.4 until the objective function alpha shown in the formula (4) reaches the minimum value or the total iteration times reaches a set threshold value, so that the final mth membership matrix U is obtained m ＝{U mc |c＝1,2,…,N c }; wherein, U mc Representing the final degree of membership of the mth time series fragment sample to the mth clustering core;
Figure BDA0003937770570000038
step 2.6, all membership degree matrixes U m |m＝1,2,…,T S ]Integrating the time sequence fragment sample data set Dataset to obtain the updated time sequence fragment sampleData set
Figure BDA0003937770570000039
Wherein the content of the first and second substances,
Figure BDA00039377705700000310
a coupled characteristic matrix representing the mth time series segment sample, based on the sample value>
Figure BDA0003937770570000041
An external influence factor matrix representing an mth time series segment sample;
step three, searching an LSTM-MTL short-term prediction model based on the membership degree;
step 3.1, dividing the updated time sequence fragment sample data set Dataset' into a training set Train = { Dataset = (Dataset) m |m＝1,2,...,N tra And verification set Verify = { Dataset = } m |m＝N tra +1,N tra +2,...,T S }; wherein, N tra Representing the number of samples of the training set;
step 3.2, the training set Train is divided again according to the FCM clustering result:
step 3.2.1, initializing membership threshold beta of the c-th clustering center c Error threshold RMSE T，c ；
Step 3.2.2, the membership degree of each sample in the training set Train and the c clustering center is respectively compared with beta tra,c Making a comparison and comparing the value of beta tra,c The sample is divided into a data set Train corresponding to the c clustering center c The preparation method comprises the following steps of (1) performing;
step 3.2.3, according to the process of the step 3.2.1-3.2.2, dividing each sample in the training set Train into the data set where the corresponding clustering center is located, thereby obtaining N from the training set Train c A training data set;
step 3.3, dividing the verification set Verify according to the process of the step 3.2, thereby obtaining N c A verification data set;
step 3.4, establishing an MTL short-term prediction model, taking an external influence factor matrix in the c-th classified training data set as input,taking the three load forms in the classified c training data set and 9 load powers formed after mutual conversion as outputs, and training the MTL short-term prediction model to obtain a c network model N o,c ；
Step 3.5, inputting the c-th verification data set into the c-th class network model N o,c The obtained prediction result is used for calculating the root mean square error value RMSE of the c-th class c ；
Step 3.6, iterating according to the process of steps 3.2-3.5, and comparing RMSE c And RMSE T，c After the optimal root mean square error value is reserved, the membership threshold is updated until the maximum iteration times are reached, and therefore the optimal network model of the c-th class corresponding to the global optimal root mean square error value is obtained
Figure BDA0003937770570000042
Further obtain N c And (4) performing an optimal network model of each class and serving as a short-term prediction model of the full-type comprehensive load.
The electronic device comprises a memory and a processor, and is characterized in that the memory is used for storing a program for supporting the processor to execute the comprehensive load short-term prediction method, and the processor is configured to execute the program stored in the memory.
The invention relates to a computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, performs the steps of the method for short-term prediction of an integrated load.
Compared with the prior art, the invention has the beneficial effects that:
the method aims at the characteristic that the coupling form of the comprehensive energy system is complex, the specific energy interaction conversion process is explicitly represented in the form of the coupling feature matrix, the local change features of various load forms on a time axis are intercepted in a time sequence fragment sample constructing mode, and meanwhile, the load data demand change rule and the coupling change rule are represented, so that the feature expression capability of a historical data sample is effectively improved, the problem that the complex energy coupling form influences the comprehensive load prediction performance is solved, and the prediction precision of the comprehensive load is effectively improved.
The method utilizes the 3D-CNN to extract the characteristics of the time sequence segment, extracts the load coupling information on the time section on a two-dimensional layer, extracts the time sequence change rule on a time dimension, performs cluster analysis on the extracted characteristics through FCM, and performs classification learning on samples with similar characteristics, thereby effectively improving the definition of the fluctuation characteristics in the samples and further improving the comprehensive prediction precision under the complex coupling form.
According to the invention, through a membership degree searching mode, a data set used for model training and generalization is iteratively updated, the clear sample characteristics are ensured, meanwhile, enough data volume is provided for characteristic learning, then, LSTM-MTL models are respectively established for different sample classes, the independent analysis of load form characteristics is ensured, meanwhile, coupling information is added, and the prediction precision of the comprehensive load is greatly improved.
Drawings
FIG. 1 is a schematic flow chart of a short-term prediction of an integrated load;
FIG. 2 is a schematic structural diagram of the LSTM-MTL model.
Detailed Description
In the embodiment, a comprehensive load short-term prediction method based on coupling characteristic matrix time sequence fragment analysis is constructed, and aims at the problem of implicit utilization of load requirements and coupling rules, the method is characterized in that the load coupling characteristic matrix explicit representation load rules are provided, and time sequence fragment samples of the load coupling characteristic matrix and external parameters are constructed through sliding time window recombination on the basis of normalization processing. Then, a 3D convolutional neural network (3D-CNN) is adopted to extract load form coupling characteristics while extracting sample time sequence characteristics, and a fuzzy mean (FCM) method is used for carrying out cluster analysis on time sequence fragment samples to avoid transition type fuzzy fluctuation characteristics in the samples. And finally, iteratively reconstructing the training set and the verification set by using a membership degree search mechanism, and constructing an LSTM-MTL (least squares-maximum likelihood models) model of a multi-task learning (MTL) framework combined with a long-short-term memory network (LSTM). The load form change rule in the coupling characteristic matrix is independently analyzed and then the comprehensive load prediction is realized by combining the coupling information, so that the prediction precision of the comprehensive load is effectively improved. The specific process, as shown in fig. 1, is performed as follows:
constructing time sequence fragment data of a load coupling characteristic matrix and external parameters based on a sliding time window;
step 1.1, acquiring a historical data sequence of a comprehensive load and a historical data sequence of relevant external influence factors from a comprehensive energy system, respectively carrying out normalization processing, and correspondingly acquiring a historical coupling characteristic matrix and a historical relevant external influence factor matrix, wherein the coupling characteristic matrix of a sampling time t is marked as L t And L is t ＝{l t (n L1 ,n L2 )|n L1 ＝1,2,3；n L2 1,2,3}; the matrix of the relevant external influence factors at the sampling time t is denoted as P t And P is t ＝{p t (n P )|n P ＝1,2,…,N P In which n is L1 And n L2 The values are 1,2 and 3 respectively, and correspondingly represent three load forms, namely an electric load, a heat load and a natural gas load; l t (n L1 ,n L2 ) Indicating that the sample time t is counted by the nth L1 From the load form to the nth L2 The amount of load converted from the seed load form; when n is L1 ＝n L2 When l is turned on t (n L1 ,n L2 ) Indicating the sampling instant tth L1 The power consumed by the load form directly at the user side without conversion; taking the electrical load as an example, for a general electric-thermal-electric coupling comprehensive energy system, the electrical load data should include 3 kinds of time series data, which are respectively non-conversion load data of power consumption equipment, load data of electric-to-heat equipment (heat pump, electric heater) and load data of electric-to-gas equipment (electrolytic bath, P2G), and respectively correspond to l t (1,1)、l t (1, 2) and l t (1, 3); same principle l t (2,1)、l t (2, 2) and l t (2, 3) representing load data of heat-to-power equipment, non-conversion load data of heat consumption equipment and load data of heat-to-gas equipment; l. the t (3,1)、l t (3, 2) and l t (3, 3) gas-to-electric apparatusLoad data, gas-to-heat equipment load data and gas consumption equipment non-conversion load data; it should be noted that if the load matrix form does not exist in the actual system, the load matrix form can be represented by 0 at all times, but the integrity of the matrix structure is still to be ensured; p is a radical of formula t (n P ) N-th of the sampling time t P (ii) seed external influencing factors; n is a radical of hydrogen P Representing the number of types of external influencing factors considered; external influencing factors including but not limited to temperature, rainfall, illumination intensity, wind speed, electricity price, calendar information, sampling time;
for comprehensive load prediction, the main difficulty of prediction lies in that different load forms have independent change rules of demand and change rules of coupling relation; the two are related but not consistent, the two change rules can be simultaneously hidden by the visual fluctuation data, and the prediction of the comprehensive load is interfered, so that various rules can be explicitly reflected by constructing a coupling characteristic matrix, and the prediction precision is improved.
Step 1.2, setting time length t TW As the length of the time window, sliding the time window on the historical load coupling characteristic matrix; each sliding by one sampling step t S Obtaining the coupled characteristic time sequence segment data of a sampling time t
Figure BDA0003937770570000061
Wherein, t n Representing the sampling instant, T, under the current time window n ＝t TW /t S Representing the number of sample points in the time series fragment data; />
Figure BDA0003937770570000062
Representing the sampling instant t in the current time window n The coupling feature of (a);
step 1.3, sliding the time window on a historical related external influence factor matrix, wherein each sliding is performed by one sampling step length t S Obtaining external influence factor time sequence segment data of a sampling time t
Figure BDA0003937770570000063
Figure BDA0003937770570000064
Representing the sampling instant t in the current time window n External influence factors of (c);
step 1.4, reconstructing the historical coupling characteristic matrix and the historical related external influence factor matrix to obtain a time sequence fragment sample data set
Figure BDA0003937770570000065
Wherein it is present>
Figure BDA0003937770570000066
Time-sequential fractional samples representing sampling instants T D Let T be the total number of sampling time of the historical data S Represents the total number of time-series fragment samples, and T S ＝T D -T n +1; by the processing mode, each sample can be ensured to simultaneously contain the coupling characteristic information and the time sequence characteristic information;
step two, performing cluster analysis based on a 3D-CNN network and an FCM method;
step 2.1, constructing a 3D-CNN network, and comprising a data input layer, a convolution layer, a pooling layer and an output layer; and adopting the 3D-CNN network to carry out coupling characteristic time sequence fragment data of the sampling time t
Figure BDA0003937770570000071
And (3) carrying out feature extraction:
step 2.2, the receiving dimension of the convolution layer through the input layer is 3 multiplied by T n Is/are as follows
Figure BDA0003937770570000072
And utilizes the formula (1) to make a pair>
Figure BDA0003937770570000073
Processing is carried out with a data input depth equal to T n Each depth profile is a 3 × 3 coupling feature matrix; the resulting dimension was 2X (T) n Adjacent time coupling characteristic out of-2) conv ；/>
Figure BDA0003937770570000074
In formula (1):
Figure BDA0003937770570000075
coupling features out for adjacent time instants conv The middle width is the coupling characteristic of the position with i height being j depth being k; w is a I,J,K Weights for convolution kernels of the form 2 × 2 × 3 in the convolutional layer at positions of width I height J depth K; in I+i-1,J+j-1,K+k-1 Is->
Figure BDA0003937770570000076
The middle width is data of the position with I + I-1 height of J + J-1 depth of K + K-1; bias is the network bias to be trained; reLU is an activation function;
step 2.3, the pooling layer utilizes the adjacent time coupling characteristic out of the formula (2) conv Processing along the depth K to obtain a dimension T n One-dimensional array of coupling characteristics out of-2 pool And as time-series fragment samples of the sampling time t
Figure BDA0003937770570000077
Corresponding chronological segment characteristic>
Figure BDA0003937770570000078
To be output by the output layer;
Figure BDA0003937770570000079
in the formula (2):
Figure BDA00039377705700000710
obtaining the coupling characteristic output on the depth k for the pooling layer by using an average pooling mode with the dimensionality of 2 multiplied by 2;
step 2.4, time sequence fragment characteristics are determined based on FCM method
Figure BDA00039377705700000711
Performing cluster analysis:
step 2.4.1, calculating T in the time sequence fragment sample data set Dataset by adopting an Euclidean distance formula S A time sequence of fragment samples and N c Distance between individual cluster cores { d mc |m＝1,2,…,T S ；c＝1,2,…,N c In which d is mc Represents the distance, N, between the mth time series fragment sample and the c-th cluster core c Is the total cluster number;
step 2.4.2, calculating the membership degree u of the mth time sequence fragment sample to the c clustering core by using the formula (3) mc ；
Figure BDA00039377705700000712
In formula (3): theta belongs to [1, ∞) as a membership weighting coefficient, and the value is usually 2; d is a radical of mn Representing a distance between the mth time series segment sample and the nth clustering core;
step 2.4.3, dividing each time sequence fragment sample into corresponding N according to membership degree c In each class, calculating the characteristic mean value of all samples in each class, thereby updating the clustering core of each class;
step 2.5, iterative computation is carried out according to the process of 2.4 until the objective function alpha shown in the formula (4) reaches the minimum value or the total iterative times reach a set threshold value, so as to obtain the final mth membership matrix U m ＝{U mc |c＝1,2,…,N c }; wherein, U mc Representing the final membership degree of the mth time sequence fragment sample to the c clustering core;
Figure BDA0003937770570000081
step 2.6, all membership degree matrixes U m |m＝1,2,…,T S ]Integrating the time sequence fragment sample data set Dataset to obtain an updated time sequence fragment sample data set
Figure BDA0003937770570000082
Wherein the content of the first and second substances,
Figure BDA0003937770570000083
a coupled characteristic matrix representing the mth time series segment sample, based on the sample value>
Figure BDA0003937770570000084
An external influence factor matrix representing an mth time series segment sample;
step three, searching an LSTM-MTL short-term prediction model based on the membership degree;
step 3.1, dividing the updated time sequence fragment sample data set Dataset' into a training set Train = { Dataset = (Dataset) m |m＝1,2,...,N tra And verification set Verify = { Dataset = } m |m＝N tra +1,N tra +2,...,T S }; wherein, N tra Representing the number of samples of the training set; in order to ensure that the training data amount is enough, the data ratio of the training set to the verification set is ensured to be more than 4;
step 3.2, the training set Train is divided again according to the FCM clustering result:
step 3.2.1, initializing a membership threshold beta of the c clustering center c Error threshold RMSE T，c (ii) a Setting the initial value of the membership threshold to be lower than 0.1, and setting the error threshold to be larger than 100 to ensure the smooth execution of the iterative process;
step 3.2.2, the membership degree of each sample in the training set Train and the c clustering center is respectively compared with beta tra,c Comparing and comparing the two with each other to obtain a value greater than beta tra,c The sample is divided into a data set Train corresponding to the c-th clustering center c The preparation method comprises the following steps of (1) performing; any sample may be repeated in different types, or may be culled and not exist in any sub data set;
step 3.2.3, according to the process of the step 3.2.1-3.2.2, dividing each sample in the training set Train into the data set where the corresponding clustering center is located, thereby obtaining N from the training set Train c A training data set;
step 3.3, dividing the verification set Verify according to the process of the step 3.2, thereby obtaining N c A verification data set;
step 3.4, as shown in fig. 2, establishing an MTL short-term prediction model, wherein the MTL model can realize independent training of different training tasks, firstly, learning input time sequence characteristics through an LSTM module, and transmitting an output result to a sharing layer; then, the sharing layer is responsible for splicing the output results of the 9 prediction tasks, and finally, the output results are transmitted to two subsequent full-connection layer networks of the 9 prediction tasks for establishing the connection between the sharing layer and the target output; taking the external influence factor matrix in the classified c-th training data set as input, taking three load forms in the classified c-th training data set and 9 load powers formed after mutual conversion of the three load forms as output, training the MTL short-term prediction model, and obtaining the c-th network model N o,c ；
Step 3.5, inputting the c-th verification data set into the c-th class of network model N o,c The obtained prediction result is used for calculating the root mean square error value RMSE of the c-th class shown in the formula (5) c ；
Figure BDA0003937770570000091
In formula (5):
Figure BDA0003937770570000092
to verify the nth of the set ver N th of sample L An actual measurement of seed load patterns; />
Figure BDA0003937770570000093
For verifying the nth ver N th of sample L A predicted value of a seed load form; n is a radical of ver Total number of samples in the validation set;
step 3.6, if RMSE c Less than RMSE T，c Then RMSE will be applied c Assign value to RMSE T,c And is beta c +0.05 assignment to β c Recording the network model at the time as the optimal network model
Figure BDA0003937770570000094
Iterating according to the process of steps 3.2-3.5 until the maximum iteration times are reached, thereby obtaining the optimal network model(s) in the class c corresponding to the global optimal root mean square error value>
Figure BDA0003937770570000095
To obtain N c And (4) performing optimization network model of each class and serving as a short-term prediction model of the full-type comprehensive load.
In this embodiment, an electronic device includes a memory for storing a program that supports a processor to execute the above-described integrated load short-term prediction method, and a processor configured to execute the program stored in the memory.
In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer program executes the steps of the method for short-term prediction of an integrated load.

Claims (3)

1. A comprehensive load short-term prediction method based on coupling feature matrix time sequence segment analysis is characterized by comprising the following steps:
firstly, constructing time sequence fragment data of a load coupling characteristic matrix and external parameters based on a sliding time window;
step 1.1, acquiring a historical data sequence of a comprehensive load and a historical data sequence of relevant external influence factors from a comprehensive energy system, respectively carrying out normalization processing, and correspondingly acquiring a historical coupling characteristic matrix and a historical relevant external influence factor matrix, wherein the coupling characteristic matrix of a sampling time t is marked as L t And L is t ＝{l t (n L1 ,n L2 )|n L1 ＝1,2,3；n L2 1,2,3}; the matrix of the relevant external influence factors at the sampling time t is denoted as p t And P is t ＝{p t (n P )|n P ＝1,2,…,N P In which n is L1 And n L2 The values of 1,2 and 3 represent three load forms correspondinglyElectrical, thermal and natural gas loads; l. the t (n L1 ,n L2 ) Indicating that the sample time t is counted by the nth L1 From the load form to the nth L2 The amount of load converted from the seed load form; when n is L1 ＝n L2 When l is turned on t (n L1 ,n L2 ) Indicating the sampling instant tth L1 The power consumed by the load form directly at the user side without conversion; pt (n) P ) N-th indicating the sampling time t P (ii) seed external influencing factors; n is a radical of P Representing the number of types of external influencing factors considered;
step 1.2, setting time length t TW As the length of the time window, sliding the time window on the historical load coupling characteristic matrix; every sliding of one sampling step t S Obtaining the coupled characteristic time sequence segment data of a sampling time t
Figure FDA0003937770560000011
Wherein, t n Representing the sampling instant, T, under the current time window n ＝t TW /t S Representing the number of sample points in the time series fragment data; />
Figure FDA0003937770560000012
Representing the sampling instant t in the current time window n The coupling feature of (a);
step 1.3, sliding the time window on the historical relevant external influence factor matrix, wherein each sliding is performed by one sampling step length t S Obtaining external influence factor time sequence segment data of a sampling time t
Figure FDA0003937770560000013
Figure FDA0003937770560000014
Representing the sampling instant t in the current time window n External influence factors of (1);
step 1.4, after the historical coupling characteristic matrix and the historical related external influence factor matrix are reconstructed, a time sequence fragment sample data set is obtained
Figure FDA0003937770560000015
Wherein it is present>
Figure FDA0003937770560000016
Time-sequential fractional samples representing the sampling instants T D Let T be the total number of sampling time of the historical data S Represents the total number of time-series fragment samples, and T S ＝T D -T n +1；
Step two, performing cluster analysis based on a 3D-CNN network and an FCM method;
step 2.1, constructing a 3D-CNN network, and comprising a data input layer, a convolution layer, a pooling layer and an output layer; and adopting the 3D-CNN network to carry out coupling characteristic time sequence fragment data of the sampling time t
Figure FDA0003937770560000017
And (3) carrying out feature extraction:
step 2.2, the receiving dimension of the convolution layer through the input layer is 3 multiplied by T n Is/are as follows
Figure FDA0003937770560000018
And utilizes formula (1) to->
Figure FDA0003937770560000019
Treated to obtain dimension of 2 × 2 × (T) n Adjacent time coupling characteristic out of-2) conv ；
Figure FDA0003937770560000021
In formula (1):
Figure FDA0003937770560000022
coupling features out for adjacent time instants conv The middle width is the coupling characteristic of the position where i is high and j is deep and k is deep; w is a I,J,K Is a convolution of the form 2 × 2 × 3 in the convolution layerThe weight of the kernel at the position with width I, height J, depth K; in I+i-1,J+j-1,K+k-1 Is->
Figure FDA0003937770560000023
Data with a middle width of I + I-1, a height of J + J-1 and a depth of K + K-1; bias is the network bias to be trained; reLU is an activation function;
step 2.3, the pooling layer utilizes the adjacent time coupling characteristic out of the formula (2) conv Processing along the depth K to obtain dimension T n One-dimensional array of coupling characteristics out of-2 pool And as time-sequential fractional samples of the sampling instant t
Figure FDA0003937770560000024
Corresponding time-sequential fragment characteristics F t TSS To be output by the output layer;
Figure FDA0003937770560000025
in the formula (2):
Figure FDA0003937770560000026
obtaining the coupling characteristic output on the depth k for the pooling layer by utilizing an average pooling mode with the dimensionality of 2 multiplied by 2;
step 2.4, feature F of time sequence segment based on FCM method t TSS Performing cluster analysis:
step 2.4.1, calculating T in the sample data set Dataset of the time sequence segment by adopting the Euclidean distance formula S A time sequence of fragment samples and N c Distance between clustering kernels d mc |m＝1,2,…,T S ；c＝1,2,…,N c In which d is mc Represents the distance, N, between the mth time series fragment sample and the c-th cluster core c Is the total cluster number;
step 2.4.2, calculating the membership degree u of the mth time sequence fragment sample to the c clustering core by using the formula (3) mc ；
Figure FDA0003937770560000027
In formula (3): theta belongs to [1, ∞) as a membership degree weighting coefficient; d mn Representing a distance between the mth time-series fragment sample and the nth clustering core;
step 2.4.3, dividing each time sequence fragment sample into corresponding N according to membership degree c In each class, calculating the characteristic mean value of all samples in each class, thereby updating the clustering core of each class;
step 2.5, iterative computation is carried out according to the process of 2.4 until the objective function alpha shown in the formula (4) reaches the minimum value or the total iteration times reaches a set threshold value, so that the final mth membership matrix U is obtained m ＝{U mc |c＝1,2,…,N c }; wherein, U mc Representing the final degree of membership of the mth time series fragment sample to the mth clustering core;
Figure FDA0003937770560000031
step 2.6, all membership degree matrixes U m |m＝1,2,…,T S ]Integrating the time sequence fragment sample data set Dataset to obtain an updated time sequence fragment sample data set
Figure FDA0003937770560000032
Wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003937770560000033
a coupled characteristic matrix representing an mth sequential segment sample, based on the characteristic value of the sample in the sequential segment>
Figure FDA0003937770560000034
An external influence factor matrix representing an mth time series segment sample;
step three, searching an LSTM-MTL short-term prediction model based on the membership degree;
step 3.1,Dividing the updated time sequence segment sample data set Dataset' into a training set Train = { Dataset = { (Dataset) m |m＝1,2,...,N tra And verification set Verify = { Dataset = } m |m＝N tra +1,N tra +2,...,T S }; wherein N is tra Representing the number of samples of the training set;
step 3.2, the training set Train is divided again according to the FCM clustering result:
step 3.2.1, initializing membership threshold beta of the c-th clustering center c Error threshold value RMSE T，c ；
Step 3.2.2, respectively relating the membership degree of each sample in the training set Train and the c clustering center to beta tra,c Making a comparison and comparing the value of beta tra,c The sample is divided into a data set Train corresponding to the c clustering center c Performing the following steps;
step 3.2.3, according to the process of the step 3.2.1-3.2.2, dividing each sample in the training set Train into the data set where the corresponding clustering center is located, thereby obtaining N from the training set Train c A training data set;
step 3.3, dividing the verification set Verify according to the process of the step 3.2, thereby obtaining N c A verification data set;
step 3.4, establishing an MTL short-term prediction model, taking the external influence factor matrix in the classified c-th training data set as input, taking three load forms in the classified c-th training data set and 9 load powers formed after mutual conversion of the three load forms as output, and training the MTL short-term prediction model to obtain a c-th network model N o,c ；
Step 3.5, inputting the c-th verification data set into the c-th class network model N o,c The obtained prediction result is used for calculating the root mean square error value RMSE of the c-th class c ；
Step 3.6, iterating according to the process of steps 3.2-3.5, and comparing RMSE c And RMSE T，c After the better root mean square error value is reserved, the membership threshold is updated until the maximum iteration times are reached, and therefore the global optimum is obtainedOptimal network model of c class corresponding to root error value
Figure FDA0003937770560000035
Further obtain N c And (4) performing an optimal network model of each class and serving as a short-term prediction model of the full-type comprehensive load.
2. An electronic device comprising a memory and a processor, wherein the memory is configured to store a program that enables the processor to perform the integrated load short term prediction method of claim 1, and the processor is configured to execute the program stored in the memory.
3. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method for short-term prediction of integrated loads according to claim 1.
CN202211409152.8A 2022-11-11 2022-11-11 Comprehensive load short-term prediction method based on coupling characteristic matrix time sequence segment analysis Pending CN115907118A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211409152.8A CN115907118A (en) 2022-11-11 2022-11-11 Comprehensive load short-term prediction method based on coupling characteristic matrix time sequence segment analysis

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211409152.8A CN115907118A (en) 2022-11-11 2022-11-11 Comprehensive load short-term prediction method based on coupling characteristic matrix time sequence segment analysis

Publications (1)

Publication Number Publication Date
CN115907118A true CN115907118A (en) 2023-04-04

Family

ID=86479886

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211409152.8A Pending CN115907118A (en) 2022-11-11 2022-11-11 Comprehensive load short-term prediction method based on coupling characteristic matrix time sequence segment analysis

Country Status (1)

Country Link
CN (1) CN115907118A (en)

Similar Documents

Publication Publication Date Title
CN111860982A (en) Wind power plant short-term wind power prediction method based on VMD-FCM-GRU
CN110969290B (en) Runoff probability prediction method and system based on deep learning
CN108022001A (en) Short term probability density Forecasting Methodology based on PCA and quantile estimate forest
CN106228185A (en) A kind of general image classifying and identifying system based on neutral net and method
CN109492748B (en) Method for establishing medium-and-long-term load prediction model of power system based on convolutional neural network
CN111091247A (en) Power load prediction method and device based on deep neural network model fusion
CN112434848A (en) Nonlinear weighted combination wind power prediction method based on deep belief network
CN112613536A (en) Near infrared spectrum diesel grade identification method based on SMOTE and deep learning
CN111275168A (en) Air quality prediction method of bidirectional gating circulation unit based on convolution full connection
CN113361785A (en) Power distribution network short-term load prediction method and device, terminal and storage medium
CN111539482B (en) RBF kernel function-based space multidimensional wind power data dimension reduction and reconstruction method
CN116187835A (en) Data-driven-based method and system for estimating theoretical line loss interval of transformer area
CN109993208A (en) A kind of clustering processing method having noise image
CN115759389A (en) Day-ahead photovoltaic power prediction method based on weather type similar day combination strategy
CN115660182A (en) Photovoltaic output prediction method based on maximum expected sample weighted neural network model
CN114004152B (en) Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network
CN117786441A (en) Multi-scene photovoltaic user electricity consumption behavior analysis method based on improved K-means clustering algorithm
CN113033898A (en) Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network
CN115688982B (en) Building photovoltaic data complement method based on WGAN and whale optimization algorithm
CN116561569A (en) Industrial power load identification method based on EO feature selection and AdaBoost algorithm
CN116777039A (en) Double-layer neural network wind speed prediction method based on training set segmentation and error correction
CN116167465A (en) Solar irradiance prediction method based on multivariate time series ensemble learning
CN116343032A (en) Classification method combining Gaussian regression mixed model and MRF hyperspectral function data
CN115423091A (en) Conditional antagonistic neural network training method, scene generation method and system
CN115759343A (en) E-LSTM-based user electric quantity prediction method and device

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination