Comprehensive load short-term prediction method based on coupling feature matrix time sequence segment analysis

Abstract

The application discloses a kind of based on coupling feature matrix time sequence segment analysis comprehensive load short-term prediction method, comprising:1, original data are collected and normalized, and load data is arranged into the form of load coupling feature matrix, and the time sequence segment sample of load coupling feature matrix and external parameter is constructed by sliding time window reorganization;2, sample time sequence feature and load form coupling feature are extracted simultaneously by 3DCNN network, and clustering analysis is carried out using FCM clustering;3, based on multi-task learning framework combines LSTM network, constructs LSTM-MTL model to realize comprehensive load prediction, and membership degree search mechanism is used to iteratively reconstruct training set and verification set, to improve the final prediction performance.The application is aimed at the problem that energy coupling form is complex and influences comprehensive load prediction performance, by clustering analysis to the time sequence segment sample of coupling feature matrix, the load fluctuation feature extraction capability is enhanced, so as to effectively improve the prediction accuracy of comprehensive load.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of short-term comprehensive load prediction, in particular to a comprehensive load short-term prediction method based on coupling feature matrix time sequence segment analysis. BACKGROUND

[0002] With the continuous development of new energy technology, the comprehensive energy system containing various energy load forms has become one of the most potential key technologies. However, the energy forms in the comprehensive energy system are various and the coupling relationship is close, which increases the uncertainty of load fluctuation and brings great challenges to the formulation of energy system dispatching plan. Therefore, it is necessary to propose a comprehensive load prediction method for the load characteristics of the comprehensive energy system.

[0003] Currently, the comprehensive load prediction is usually based on the historical data of electric, heat and gas loads, and machine learning, neural network and other methods are used to learn the historical change law to train the prediction model to realize the comprehensive load prediction. In order to improve the learning ability of the model, the existing research often uses clustering analysis and other preprocessing methods to obtain clearer data features.

[0004] Most of the existing research results still use the electric, heat and gas load data roughly, or separately establish a prediction model for different load forms, or directly bring the three load forms into the multi-task learning model for direct training. However, the learning of single change law cannot accurately reflect the change trend of the load due to the complex load form of the comprehensive energy system, which leads to low accuracy of the traditional prediction scheme in the field of comprehensive load prediction. In addition, the traditional prediction scheme usually uses one day as the time window length of the sample in clustering preprocessing, which leads to the local change characteristics of the load fluctuation being covered up and difficult to be reflected in the clustering results, and cannot play a substantial role in the prediction process. In addition, the existing scheme usually selects a hard boundary clustering method, which leads to the transition data of high probability type being learned or the data capacity of low probability type being too small, which increases the difficulty of model learning. SUMMARY

[0005] In order to overcome the deficiencies in the prior art, the present application provides a comprehensive load short-term prediction method based on coupling feature matrix time sequence segment analysis, so as to explicitly learn two change laws of historical load data and strengthen the extraction ability of load fluctuation characteristics, thereby improving the accuracy and prediction precision of comprehensive load short-term prediction.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0007] The comprehensive load short-term prediction method based on coupling feature matrix time sequence segment analysis of the present application is characterized in that the following steps are performed:

[0008] Step one, based on sliding time window, construct load coupling feature matrix and time sequence segment data of external parameters;

[0009] Step 1.1, obtain the historical data sequence of comprehensive load and the historical data sequence of related external influencing factors from the integrated energy system, and respectively perform normalization processing, and correspondingly obtain the historical coupling feature matrix and the historical related external influencing factor matrix, wherein the coupling feature matrix at sampling time t is denoted as L t , and L t ={l t (n L1 ,n L2 )|n L1 =1,2,3;n L2 =1,2,3}; the related external influencing factor matrix at sampling time t is denoted as P t , and P t ={p t (n P )|n P =1,2,…,N P}, wherein n L1 and n L2 take values 1, 2, and 3, respectively, corresponding to three load forms, i.e., electric load, heat load, and natural gas load; l t (n L1 ,n L2 ) represents the load amount converted from the nth L1 load form to the nth L2 load form at sampling time t; when n L1 =n L2 , l t (n L1 ,n L2 ) represents the power directly consumed at the user side without conversion of the nth L1 load form at sampling time t; pt(n P ) represents the nth P external influencing factor at sampling time t; N P represents the number of considered external influencing factors;

[0010] Step 1.2, set the time length t TW as the time window length, and slide the time window on the historical load coupling feature matrix; every time the time window is slid by a sampling step t S , a coupling feature time sequence segment data at sampling time t is obtained, wherein t n represents the sampling time under the current time window, and T n =t TW / t S represents the number of sampling points in the time sequence segment data; Indicates the sampling time t under the current time window n The coupling characteristics;

[0011] Step 1.3: Slide the time window across the historical relevant external influencing factor matrix, with each slide taking a sampling step t. S To obtain time series data of external influencing factors at a sampling time t. Indicates the sampling time t under the current time window n External influencing factors;

[0012] Step 1.4: After reconstructing the historical coupling feature matrix and the historical related external influencing factor matrix, the time-series segment sample dataset is obtained. in, T represents a time-series segment sample at sampling time t. D Let T be the total number of sampling times for historical data. S T represents the total number of time-series segment samples, and T S =T D -T n +1;

[0013] Step 2: Perform cluster analysis based on 3D-CNN network and FCM method;

[0014] Step 2.1: Construct a 3D-CNN network, including an input layer, convolutional layers, pooling layers, and an output layer; and use the 3D-CNN network to process the coupled feature temporal segment data at sampling time t. Perform feature extraction:

[0015] Step 2.2: The convolutional layer receives data of dimension 3×3×T through the input layer. n of And using equation (1) to After processing, we obtain a dimension of 2×2×(T) n -2) Adjacent time coupling characteristics out conv ;

[0016]

[0017] In formula (1): The coupling feature out at adjacent time points conv The coupling feature at position i (width), j (height), and k (depth); w I,J,K The weights of the 2×2×3 convolutional kernel in the convolutional layer at positions I (width), J (height), and K (depth); I+i-1,J+j-1,K+k-1 for Data in the position of I+i-1 wide, J+j-1 high, K+k-1 deep; bias is the network bias to be trained; ReLU is an activation function;

[0018] Step 2.3, the pooling layer utilizes formula (2) to the adjacent time coupling feature out conv Processing along the depth K, obtaining the dimension T n One-dimensional coupling feature array out pool , and as a time sequence segment sample of sampling time t Corresponding time sequence segment feature So as to output by the output layer;

[0019]

[0020] In formula (2): The pooling layer utilizes the average pooling method with the dimension of 2x2 to obtain the coupling feature output on the depth k;

[0021] Step 2.4, clustering analysis is performed on the time sequence segment feature Based on the FCM method:

[0022] Step 2.4.1, the Euclidean distance formula is used to calculate the distance between T S Time sequence segment samples in the time sequence segment sample data set Dataset and N c Clustering cores {d mc |m=1,2,…,T S ; c=1,2,…,N c}, wherein d mc Indicates the distance between the mth time sequence segment sample and the cth clustering core, and N c is the total number of clusters;

[0023] Step 2.4.2, formula (3) is used to calculate the membership degree u mc of the mth time sequence segment sample to the cth clustering core;

[0024]

[0025] In formula (3): θ∈[1,∞) is a membership degree weighting coefficient; d mn Indicates the distance between the mth time sequence segment sample and the nth clustering core;

[0026] Step 2.4.3, each time sequence segment sample is divided into corresponding N c classes according to the membership degree, and the feature mean of all samples in each class is calculated, so as to update the clustering core of each class;

[0027] Step 2.5, iteratively calculate according to the process of 2.4 until the target function a shown in formula (4) reaches the minimum value, or the total number of iterations reaches the set threshold, so as to obtain the final mth membership matrix U m = {U mc |c = 1, 2, …, N c}; wherein, U mc represents the final membership of the mth time sequence segment sample to the cth cluster core;

[0028]

[0029] Step 2.6, integrate all membership matrices [U m |m = 1, 2, …, T S ] into the time sequence segment sample data set Dataset, so as to obtain the updated time sequence segment sample data set Dataset′ wherein, represents the coupling feature matrix of the mth time sequence segment sample, represents the external influencing factor matrix of the mth time sequence segment sample;

[0030] Step three, LSTM-MTL short-term prediction model based on membership search;

[0031] Step 3.1, divide the updated time sequence segment sample data set Dataset′ into a training set Train = {Dataset m |m = 1, 2, …, N tra} and a verification set Verify = {Dataset m |m = N tra +1, N tra +2, …, T S}; wherein, N tra represents the number of samples of the training set;

[0032] Step 3.2, divide the training set Train again according to the FCM clustering result:

[0033] Step 3.2.1, initialize the membership threshold β c of the cth cluster center and the error threshold RMSE T，c ;

[0034] Step 3.2.2, compare the membership of each sample in the training set Train with the cth cluster center with β tra,c respectively, and divide the samples greater than β tra,c into the data set Train c corresponding to the cth cluster center;

[0035] Step 3.2.3, according to the process of steps 3.2.1-3.2.2, each sample in the training set Train is divided into the data set where the corresponding cluster center is located, thereby obtaining N c training data sets from the training set Train;

[0036] Step 3.3, the verification set Verify is divided according to the process of step 3.2, thereby obtaining N c verification data sets;

[0037] Step 3.4, the MTL short-term prediction model is established, and the external influence factor matrix in the classified cth training data set is taken as the input, and the three load forms and the 9 load powers formed after mutual conversion in the classified cth training data set are taken as the output, thereby training the MTL short-term prediction model, and obtaining the network model N o,c of the cth class;

[0038] Step 3.5, the cth verification data set is input into the network model N o,c of the cth class for processing, and the obtained prediction result is used to calculate the root mean square error value RMSE c of the cth class;

[0039] Step 3.6, according to the process of steps 3.2-3.5, the RMSE c and the RMSE T，c are compared, the membership threshold is updated after the better root mean square error value is retained, until the maximum iteration number is reached, thereby obtaining the optimal network model of the cth class corresponding to the global optimal root mean square error value and further obtaining the optimal network model of the N c classes and taking it as the short-term prediction model of the comprehensive load.

[0040] The electronic device of the present application comprises a memory and a processor, and the memory is used to store a program 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.

[0041] The computer readable storage medium of the present application stores a computer program, and the computer program is executed by the processor to execute the steps of the comprehensive load short-term prediction method.

[0042] Compared with the prior art, the beneficial effects of the present application are as follows:

[0043] 1The present application is aimed at the characteristics of complex coupling form of integrated energy system, the specific energy interaction conversion process is explicitly characterized in the form of coupling characteristic matrix, and the local variation characteristics of various load forms on the time axis are intercepted by constructing time sequence fragment samples, and the load data demand variation law and coupling variation law are characterized, the feature expression ability of the historical data sample is effectively improved, the problem of complex energy coupling form affecting the performance of integrated load prediction is solved, and the prediction accuracy of integrated load is effectively improved.

[0044] 2The present application uses 3D-CNN to extract the features of time sequence fragments, extracts the load coupling information on the time section in the two-dimensional level, extracts the time sequence variation law in the time dimension, and classifies and learns the samples with similar features through FCM clustering analysis of the extracted features, effectively improves the clarity of the fluctuation characteristics in the sample, and further improves the integrated prediction accuracy under complex coupling form.

[0045] 3The present application iteratively updates the data set used for model training and generalization through the way of membership search, provides sufficient data amount for feature learning while ensuring the clarity of sample features, then establishes LSTM-MTL model for different sample classes respectively, ensures independent analysis of load form characteristics while adding coupling information, and greatly improves the prediction accuracy of integrated load. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a flowchart of integrated load short-term prediction;

[0047] Figure 2 It is a schematic diagram of LSTM-MTL model structure. DETAILED DESCRIPTION

[0048] In this embodiment, a comprehensive short-term load forecasting method based on time-series segment analysis of coupling feature matrices is constructed. Addressing the problem of implicitly utilizing load demand and coupling patterns, this method proposes an explicit representation of load patterns using a load coupling feature matrix. Based on normalization, a sliding window is used to reconstruct time-series segment samples of the load coupling feature matrix and external parameters. Then, a 3D convolutional neural network (3D-CNN) is used to extract both the time-series features of the samples and the load form coupling features. Fuzzy C-means (FCM) is used to perform cluster analysis on the time-series segment samples to avoid fuzzy fluctuations in transitional types within the samples. Finally, a membership search mechanism is used to iteratively reconstruct the training and validation sets, constructing an LSTM-MTL model that combines a multitask learning (MTL) framework with long short-term memory networks (LSTM). By independently analyzing the load form variation patterns in the coupling feature matrix and then combining the coupling information, comprehensive load forecasting is achieved, effectively improving the accuracy of comprehensive load forecasting. The specific process is as follows: Figure 1 As shown, the procedure is as follows:

[0049] Step 1: Construct time-series segment data of load coupling feature matrix and external parameters based on sliding time window;

[0050] Step 1.1: Obtain historical data sequences of the integrated load and related external influencing factors from the integrated energy system, and perform normalization processing respectively to obtain the historical coupling feature matrix and the historical related external influencing factor matrix. The coupling feature matrix at sampling time t is denoted as L. t And L t ={l t (n L1 ,n L2 )|n L1 =1,2,3; n L2 =1,2,3}; the matrix of relevant external influencing factors at sampling time t is denoted as P. t And P t ={p t (n P )|n P =1,2,…,N P}, where n L1 and n L2 The values ​​1, 2, and 3 represent three load types: electrical load, heat load, and natural gas load, respectively. t (n L1 ,n L2 ) indicates that at sampling time t, the nth...L1 a load form to the nth L2 a load form; when n L1 = n L2 , l t (n L1 , n L2 ) represents the power consumed directly on the user side without conversion of the nth L1 load form at sampling time t; taking electrical load as an example, for a general electric-thermal-gas coupled comprehensive energy system, the electrical load data should include three kinds of time series data, which are non-conversion load data of power-consuming equipment, electric-thermal equipment load data (heat pump, electric heater), and electric-gas equipment load data (electrolytic cell, P2G), and correspond to l t (1,1), l t (1,2) and l t (1,3) respectively; similarly, l t (2,1), l t (2,2) and l t (2,3) represent heat-to-electricity equipment load data, non-conversion load data of heat-consuming equipment, and heat-to-gas equipment load data; l t (3,1), l t (3,2) and l t (3,3) represent gas-to-electricity equipment load data, gas-to-thermal equipment load data, and non-conversion load data of gas-consuming equipment; it should be noted that if the above load matrix form does not exist in the actual system, 0 can be used at all times, but the integrity of the matrix structure should be ensured; p t (n P ) represents the nth P external influencing factor at sampling time t; N P represents the number of external influencing factors considered; external influencing factors include but are not limited to temperature, rainfall, light intensity, wind speed, electricity price, calendar information, and sampling time;

[0051] For comprehensive load forecasting, the main difficulty lies in the independent demand change law and the change law of the coupling relationship of different load forms; the two are related but not consistent, and the intuitive fluctuation data will simultaneously imply the above two change laws, which will interfere with the prediction of the comprehensive load, so the construction of the coupling feature matrix can explicitly reflect multiple laws and improve the prediction accuracy.

[0052] Step 1.2, set the time length t TW as the time window length, and slide the time window on the historical load coupling feature matrix; every time the time window is slid by a sampling step t S , a coupling feature time series segment data at sampling time t nT represents the sampling time within the current time window. n =t TW / t S This indicates the number of sampling points in the time-series data segment; Indicates the sampling time t under the current time window n The coupling characteristics;

[0053] Step 1.3: Slide the time window across the historical relevant external influencing factor matrix, with each slide taking a sampling step t. S To obtain time series data of external influencing factors at a sampling time t. Indicates the sampling time t under the current time window n External influencing factors;

[0054] Step 1.4: After reconstructing the historical coupling feature matrix and the historical related external influencing factor matrix, the time-series segment sample dataset is obtained. in, T represents a time-series segment sample at sampling time t. D Let T be the total number of sampling times for historical data. S T represents the total number of time-series segment samples, and T S =T D -T n +1; This processing method ensures that each sample contains both coupled feature information and temporal feature information simultaneously;

[0055] Step 2: Perform cluster analysis based on 3D-CNN network and FCM method;

[0056] Step 2.1: Construct a 3D-CNN network, including an input layer, convolutional layers, pooling layers, and an output layer; and use the 3D-CNN network to process the coupled feature temporal segment data at sampling time t. Perform feature extraction:

[0057] Step 2.2: The convolutional layer receives data of dimension 3×3×T through the input layer. n of And using equation (1) to The data is processed, and the input depth is equal to T. n Each depth section is a 3×3 coupled feature matrix; resulting in a dimension of 2×2×(T) n -2) Adjacent time coupling characteristics out conv ;

[0058]

[0059] In formula (1): The coupling feature out at adjacent time pointsconv out is the coupling feature at position i wide, j high, and k deep; w I,J,K is the weight at position i wide, j high, and k deep in the convolution kernel of the form 2 x 2 x 3 in the convolution layer; in I+i-1,J+j-1,K+k-1 is the input data at position i wide, j high, and k deep; bias is the network bias to be trained; and ReLU is the activation function. is the data at position i wide, j high, and k deep; bias is the network bias to be trained; and ReLU is the activation function.

[0060] Step 2.3, the pooling layer uses formula (2) to process the coupling feature out conv along the depth k to obtain a one-dimensional coupling feature array out n of dimension T pool , which is taken as the time sequence segment sample at sampling time t corresponding to the time sequence segment feature , so as to be output by the output layer.

[0061]

[0062] In formula (2), out is the coupling feature output at depth k obtained by the pooling layer using the average pooling method of dimension 2 x 2.

[0063] Step 2.4, clustering analysis is performed on the time sequence segment feature based on the FCM method.

[0064] Step 2.4.1, the Euclidean distance formula is used to calculate the distance d S between T c time sequence segment samples in the time sequence segment sample data set Dataset and N mc cluster cores, where m = 1, 2, …, T S ; c = 1, 2, …, N c , wherein d mc represents the distance between the mth time sequence segment sample and the cth cluster core, and N c is the total number of clusters.

[0065] Step 2.4.2, formula (3) is used to calculate the membership degree u mc of the mth time sequence segment sample to the cth cluster core.

[0066]

[0067] In formula (3), θ ∈ [1, ∞) is a membership degree weighting coefficient, and is usually taken as 2; d mn represents the distance between the mth time sequence segment sample and the nth cluster core.

[0068] Step 2.4.3, divide each time segment sample into corresponding N c classes according to the membership, calculate the feature mean of all samples in each class, and update the clustering core of each class;

[0069] Step 2.5, iterate the calculation according to the process of 2.4 until the objective function a shown in formula (4) reaches the minimum value, or the total iteration number reaches the set threshold, thereby obtaining the final mth membership matrix U m = {U mc |c = 1, 2, …, N c}; wherein U mc represents the final membership of the mth time segment sample to the cth clustering core;

[0070]

[0071] Step 2.6, integrate all membership matrices [U m |m = 1, 2, …, T S ] into the time segment sample dataset Dataset, thereby obtaining the updated time segment sample dataset Dataset′ wherein, represents the coupling feature matrix of the mth time segment sample, represents the external influencing factor matrix of the mth time segment sample;

[0072] Step three, LSTM-MTL short-term prediction model based on membership search;

[0073] Step 3.1, divide the updated time segment sample dataset Dataset′ into a training set Train = {Dataset m |m = 1, 2, …, N tra} and a verification set Verify = {Dataset m |m = N tra +1, N tra +2, …, T S}; wherein N tra represents the number of samples in the training set; in order to ensure sufficient training data, the data proportion in the training set and the verification set should be greater than 4:1;

[0074] Step 3.2, divide the training set Train again according to the FCM clustering result:

[0075] Step 3.2.1, initialize the membership threshold β c of the cth clustering center and the error threshold RMSE T，cThe initial value of the membership threshold should be lower than 0.1, and the error threshold should be greater than 100, to ensure the smooth execution of the iteration process;

[0076] Step 3.2.2, compare the membership of each sample in the training set Train and the c th cluster center with β tra,c respectively, and divide the sample greater than β tra,c into the data set Train c corresponding to the c th cluster center; any sample may appear in different types repeatedly, or be excluded from any sub-data set;

[0077] Step 3.2.3, according to the process of steps 3.2.1-3.2.2, each sample in the training set Train is divided into the data set corresponding to the cluster center, so that N c training data sets are obtained from the training set Train;

[0078] Step 3.3, divide the verification set Verify according to the process of step 3.2, so as to obtain N c verification data sets;

[0079] Step 3.4, as shown in Figure 2 , an MTL short-term prediction model is established, which can realize separate training for different training tasks. First, learn the time sequence characteristics of the input through the LSTM module, and pass the output result to the shared layer; then, the shared layer is responsible for splicing the output results of the 9 prediction tasks, and finally, the subsequent two fully connected layer networks of the 9 prediction tasks are transmitted, which is used to establish the connection between the shared layer and the target output; the external influence factor matrix in the classified c th training data set is taken as the input, and the three load forms in the classified c th training data set and the 9 load powers formed after mutual conversion are taken as the output, so as to train the MTL short-term prediction model, and obtain the network model N o,c of the c th class;

[0080] Step 3.5, input the c th verification data set into the network model N o,c of the c th class for processing, and the obtained prediction result is used to calculate the root mean square error value RMSE c of the c th class as shown in formula (5);

[0081]

[0082] In formula (5): is the measured value of the n ver th load form of the n L th sample in the verification set; is the measured value of the n verthe n L th prediction value of the load form; N ver is the total number of samples in the validation set;

[0083] Step 3.6, if RMSE c is less than RMSE T，c , then RMSE c is assigned to RMSE T,c , and β c +0.05 is assigned to β c , and the network model at this time is recorded as the optimal network model According to the processes of steps 3.2-3.5, iteration is performed until the maximum number of iterations is reached, thereby obtaining the optimal network model of the cth class corresponding to the global optimal root mean square error value and further obtaining the optimal network models of N c classes and taking them as the short-term comprehensive load prediction model.

[0084] In this embodiment, an electronic device includes a memory for storing a program supporting a processor to execute the above-mentioned short-term comprehensive load prediction method, and the processor is configured to execute the program stored in the memory.

[0085] In this embodiment, a computer-readable storage medium has a computer program stored thereon, and the computer program, when executed by a processor, performs the steps of the above-mentioned short-term comprehensive load prediction method.

Claims

1. A method for short-term load forecasting based on coupled feature matrix time series segment analysis, characterized in that, is carried out as follows: Step one, constructing historical coupling feature matrix and time sequence segment data of external parameters based on sliding time window; Step 1.1, obtain the historical data sequence of the integrated load and the historical data sequence of the related external influencing factors from the integrated energy system, and normalize them respectively, and correspondingly obtain the historical coupling feature matrix and the historical related external influencing factor matrix, wherein the historical coupling feature matrix at the sampling time t is denoted as , and ; the related external influencing factor matrix at the sampling time t is denoted as , and , wherein, and take values 1, 2, 3 respectively, which correspond to three load forms, namely, electric load, heat load and natural gas load; represents the load amount converted from the th load form to the th load form at the sampling time t; when , represents the power consumed directly at the user side without conversion of the th load form at the sampling time t; represents the th external influencing factor at the sampling time t; represents the number of external influencing factors considered; Step 1.2, set the length of time As the length of time window, the time window is slid on the historical coupling feature matrix; each sliding a sampling step , a coupling feature time series segment data of a sampling time t is obtained ; wherein, indicates the sampling time under the current time window, indicates the number of sampling points in the time series segment data; indicates the coupling feature of the sampling time under the current time window. Step 1.3, sliding the time window on the historical related external influence factor matrix, every sliding a sample step , obtaining an external influence factor time sequence segment data of a sampling time t ; representing the external influence factor of the sampling time under the current time window ; Step 1.

4. After reconstructing the historical coupling feature matrix and the historical related external impact factor matrix, a time series segment sample dataset is obtained wherein, denotes a time series segment sample at sampling time t, is the total number of sampling times of the historical data, and denotes the total number of time series segment samples, and ; Step two, clustering analysis based on 3D-CNN network and FCM method; Step 2.1, constructing a 3D-CNN network, and including a data input layer, a convolutional layer, a pooling layer, and an output layer; and using the 3D-CNN network to process the coupled feature time sequence segment data of the sampling time t Feature extraction is performed: Step 2.2: The convolutional layer receives data of dimension 3×3× through the input layer. of And using equation (1) to After processing, we obtain a dimension of 2×2× Adjacent time coupling characteristics ; (1) In formula (1): is the coupling feature at the adjacent time is the coupling feature at the position of width i, height j, and depth k; is the weight of the convolution kernel in the form of 2×2×3 in the convolution layer at the position of width I+i-1, height J+j-1, and depth K+k-1; is is the data at the position of width I+i-1, height J+j-1, and depth K+k-1; is the network bias to be trained; ReLU is the activation function; Step 2.3, the pooling layer utilizes formula (2) to couple features of adjacent time along depth K to obtain a one-dimensional coupled feature array with dimensions and as a time sequence segment sample of sampling time t corresponding time sequence segment feature , so as to be output by the output layer;​ (2) In formula (2): The pooling layer uses an average pooling method with a dimension of 2x2 to obtain the output of the coupling feature in the depth k. Step 2.

4. Feature of time series segment based on FCM method Carrying out cluster analysis: Step 2.4.

1. Calculate the distance between the time-segment sample dataset using the Euclidean distance formula In the distance between the mth time-segment sample and the cth cluster core In wherein, denotes the distance between the mth time-segment sample and the cth cluster core, is the total number of clusters; Step 2.4.

2. Calculate the membership of the mth time segment sample to the cth cluster core using formula (3) ; (3) In formula (3): is a membership weight coefficient; denotes the distance between the mth time segment sample and the nth cluster core. Step 2.4.3, divide each time-sequence fragment sample into corresponding classes according to the membership, and calculate the feature mean of all samples in each class to update the clustering core of each class. Step 2.5, iteratively calculate according to the process of 2.4 until the objective function shown in formula (4) reaches a minimum value, or the total number of iterations reaches a set threshold, thereby obtaining the final mth membership matrix ; wherein, indicates the final membership of the mth time sequence segment sample to the cth cluster core. (4) Step 2.6, combine all the membership matrices into the time series segment sample dataset , thereby obtaining an updated time series segment sample dataset ; wherein, represents the historical coupling feature matrix of the mth time series segment sample, represents the external influence factor matrix of the mth time series segment sample; Step three, LSTM-MTL short-term prediction model based on membership search. Step 3.1: Update the time-series segment sample dataset Divided into training set and verification set ;in, Indicates the number of samples in the training set; Step 3.2, Partitioning on the training set According to the FCM clustering results, the data is again divided respectively: Step 3.2.1, initialize membership threshold for cth cluster center , error threshold RMSE T，c ; Step 3.2.2, compare the membership of each sample in the training set to the cth cluster center, respectively, and divide the samples with membership greater than into the data set corresponding to the cth cluster center ; Step 3.2.

3. Divide each sample in the training set into the data set where the corresponding cluster center is located, so that T training data sets are obtained from the training set ​​ Step 3.3.3, on the validation set The division is performed according to the procedure of step 3.2, resulting in validation data sets; Step 3.4, an MTL short-term prediction model is established, and the external influence factor matrix in the classified cth training data set is taken as the input, and the three load forms in the classified cth training data set and the 9 kinds of load power formed after mutual conversion are taken as the output, so as to train the MTL short-term prediction model, and obtain the network model of the cth class ; Step 3.5, input the cth verification data set into the network model of the cth class The prediction result obtained by processing in the computer is used to calculate the root mean square error value RMSE of the cth class c ; Step 3.

6. Iterating according to the process of steps 3.2-3.5, comparing the RMSE c With the RMSE T，c And retaining the better root mean square error value, updating the membership threshold until the maximum number of iterations is reached, so as to obtain the optimal network model of the cth class corresponding to the global optimal root mean square error value ; and further obtaining Optimal network models of the classes and serving as a short-term comprehensive load prediction model of all types.

2. An electronic device comprising a memory and a processor, characterized in that The memory is configured to store a program supporting the processor to execute the comprehensive 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 having stored thereon a computer program, characterized in that The computer program is configured to execute the steps of the comprehensive load short-term prediction method of claim 1 when executed by the processor.

Images