Method and system for predicting load polymer of integrated framework

Abstract

The invention discloses a load polymer prediction method and system for an integrated framework, and the method comprises the steps: carrying out the scheduling through employing a prediction result of a trained load polymer prediction model; the construction of the load polymer prediction model comprises the following steps: obtaining historical data of a load polymer; external influence factor analysis is carried out on the historical data of the load polymer, and important external influence factors are screened out; clustering the historical data of the load polymer by applying a K-shape algorithm, and dividing the historical data into a plurality of clusters; constructing a load polymer prediction model, wherein the load polymer prediction model comprises a CNN-LSTM-ATTENTION prediction model and a GBDT prediction model; carrying out the load prediction of each cluster through a CNN-LSTM-ATTENTION prediction model according to the screened important external influence factors and a plurality of divided clusters; and applying the GBDT prediction model to perform integrated prediction on the load prediction results of the clusters to obtain a load polymer prediction result, so that the prediction precision of the load polymer prediction model can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of load aggregates, and in particular to a load aggregate prediction method and system of an integrated framework. Background Art

[0002] With the large-scale grid connection of renewable energy and user-side adjustable resources, and the continuous development of the power market, demand response has opened up a new situation for user-side resources to participate in the power market because it can reflect market competition and demand elasticity. In this context, load aggregators have developed rapidly as emerging market entities that aggregate a large number of load resources. Load aggregators can integrate user-side resources through demand response, and thereby provide load smoothing services to the power grid to obtain profits. Therefore, it is necessary to conduct research on aggregated loads. In addition, with the large-scale popularization of smart meters, fine-grained load data is easy to obtain.

[0003] In the process of research and practice of existing technologies, it is found that most of the current research focuses on establishing load forecasting models, but does not consider the diversity of loads within load aggregates. Therefore, it is urgent to select a load aggregate forecasting method and system that can consider the internal differences of load aggregates.

[0004] The paper "Load Aggregate Prediction Method Based on Gated Recurrent Unit Network and Model Fusion" discloses that a load group with similar load characteristics is obtained through a distributed spectral clustering algorithm, and then group prediction is performed. GRU is used as a meta-model to dynamically model the time series, and a random forest algorithm is used to fuse multiple GRU networks with different structures to achieve the prediction of the load group. Finally, the prediction values ​​of each group are summed up to obtain the load aggregate prediction value. However, in this paper, the final load aggregate prediction value is obtained by directly adding the prediction results of different groups. Load aggregates are achieved by centrally managing and scheduling the power demands of multiple small power users to achieve flexible regulation and response to the power system. However, directly adding the prediction results of different groups may lead to the accumulation of prediction errors, resulting in inaccurate prediction accuracy. At the same time, when scheduling according to the prediction results, the responses of different user terminals may be different, further affecting the scheduling effect of the system. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the existing load aggregation prediction accuracy is low.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A load aggregate prediction method with an integrated framework, which uses the prediction results of the trained load aggregate prediction model for scheduling;

[0008] The construction of the load aggregate prediction model includes:

[0009] Get load aggregate historical data;

[0010] Analyze external influencing factors on historical data of load aggregates and screen out important external influencing factors;

[0011] The load aggregate historical data is clustered using the K-shape algorithm and divided into several clusters;

[0012] Construct load aggregation prediction models, including CNN-LSTM-ATTENTION prediction model and GBDT prediction model;

[0013] The important external influencing factors and several clusters are screened out, and the load forecast is performed for each cluster through the CNN-LSTM-ATTENTION forecasting model;

[0014] The GBDT prediction model is used to integrate the load prediction results of each cluster to obtain the load aggregate prediction results.

[0015] In one embodiment of the present invention, important external influencing factors are screened out and obtained through the following steps:

[0016] Air temperature, air pressure, relative humidity and wind speed are selected as candidate external influencing factors;

[0017] The Spearman correlation coefficient method was used to obtain the relationship between the load and each candidate external influencing factor as a correlation index;

[0018] A relationship threshold is set, the correlation index is compared with the relationship threshold, and the candidate external influencing factors corresponding to the correlation index greater than the relationship threshold are regarded as important external influencing factors.

[0019] In one embodiment of the present invention, the correlation index P s , obtained by the following formula:

[0020]

[0021] In the formula, R i , S i They represent the levels of the values ​​of observation i, They are the average levels of load and candidate external influencing factors, N is the total number of observations, and d i =R i -S i Expressed as the number of rank differences between two columns of paired variables.

[0022] In one embodiment of the present invention, it is divided into several clusters, including:

[0023] The distance between the load aggregate historical data is measured by cross-correlation, and the similarity between each two groups of load aggregate historical data is calculated;

[0024] The number of clusters is determined by using the inflection point method for the historical data of load aggregates;

[0025] Initialize the centroid;

[0026] The centroid is recalculated based on the distance scale of the cross-correlation measurement between the current centroid of each category and all points in the class, and each data in each group of load aggregate historical data is re-divided to form a cluster; and each group of load aggregate historical data is redistributed to the cluster closest to the centroid based on the distance between it and the new centroid. The cycle is iterated until the label no longer changes, and the regrouped load aggregate historical data is obtained.

[0027] In one embodiment of the present invention, the similarity SBD(u,v) of each two groups of load aggregate historical data is calculated by the following formula:

[0028]

[0029] In the formula, the value range of SBD(u,v) is [0,2], 0 means that the historical data of two sets of load aggregation are most similar, u,v means the historical data of two sets of load aggregation based on time series, max w (·) represents the calculation of w that maximizes u, CC w (u,v) represents the mutual correlation coefficient of u and v, R0(u,u) represents the autocorrelation coefficient of u, and R0(v,v) represents the autocorrelation coefficient of v.

[0030] In one embodiment of the present invention, a line graph is drawn between the sum of the squared deviations of each group of load aggregate historical data and the number of each group of load aggregate historical data to observe the position of the inflection point change. As the load aggregate historical data increases, the sum of the squared deviations tends to be stable, and the number of inflection points is used as the number of clusters.

[0031] In one embodiment of the present invention, the load prediction is performed on each cluster using the CNN-LSTM-ATTENTION prediction model, including:

[0032] The local features of the input matrix are extracted through the CNN network model to obtain the feature vector; wherein the input matrix is ​​a matrix composed of several clusters divided by the K-shape algorithm and important external influencing factors screened out; and the feature vector is obtained by the following formula:

[0033] h k =ReLU(W k *X+bk );

[0034] The eigenvectors corresponding to all input matrices form an input sequence sample set;

[0035] The input sequence sample set is used as the input of the LSTM network model, and the load forecast results of each cluster are obtained as output;

[0036] Among them, the LSTM network model includes input gate i t 、Forget gate t and output gate o t :

[0037] Input Gate i t :

[0038] i t =sigmoid(W i [h t-1 ;x t ]+b i );

[0039] Forget Gate t :

[0040] f t =sigmoid(W f [h t-1 ;x t ]+b f );

[0041] Output gate o t :

[0042] o t =sigmoid(W o [h t-1 ;x t ]+b o );

[0043] Memory unit update:

[0044] s t =f t ⊙s t-1 +i t ⊙tanh(W s [h t-1 ;x t ]+b s );

[0045] Output:

[0046] h t =o t ⊙tanh(s t );

[0047] In the formula, x t is a feature vector, representing the current input; [h t-1 ;x t ] represents the hidden state h at time t-1 t-1 and the current input x t The series connection, W i , W f , W o , W s They are respectively represented as the weight matrices of the input gate, forget gate, output gate and memory unit, ⊙ represents element-by-element multiplication, and b i , b f , b o , b s They represent the bias vector parameters of the input gate, forget gate, output gate and memory unit respectively, tanh represents the hyperbolic tangent function, ReLU and sigmoid represent two different activation functions respectively; h k Represented as the feature vector output by the kth filter, W k is represented as the slope of the kth filter, X is represented as the input matrix, b k It is represented as the intercept of the kth filter and * represents the convolution operation.

[0048] In one embodiment of the present invention, an attention mechanism is introduced into the hidden state of the hidden layer in the LSTM network model to assign different attention weights to the hidden state of each hidden layer, including the following steps:

[0049] For a given hidden state sequence of a hidden layer, calculate the attention score of each hidden layer;

[0050] Among them, the attention score is obtained by:

[0051] U m =tan(W·h m + b);

[0052] Where U m It is represented as the attention score of the mth hidden layer, h m represents the mth hidden layer, W represents the attention weight matrix, b represents the bias vector parameter of the attention mechanism, and tan represents the tangent;

[0053] According to the attention score, obtain the attention weight probability;

[0054] Among them, the attention weight probability is obtained by the following formula:

[0055]

[0056] In the formula, p mIt is represented as the attention weight probability of the mth hidden layer, exp is represented as the exponential function, and M is represented as the maximum hidden layer;

[0057] According to the attention weight probability, the attention weight of each hidden layer is weighted to obtain the weighted hidden layer; the weighted hidden layer is obtained by the following formula:

[0058]

[0059] In the formula, represents the weighted hidden layer.

[0060] In one embodiment of the present invention, the load aggregate prediction result F is obtained. M (x) including:

[0061] Given a prediction data set T, T = {(x1, y1), (x2, y2), ... (x k ,y k ),...,(x K ,y K )}; where x k It is represented as the prediction result of the kth cluster, and y is represented as the prediction result of the load aggregate;

[0062] Initialize the learner F0(x):

[0063]

[0064] Establish M classification and regression trees and calculate the response corresponding to the mth tree;

[0065] Among them, the response r corresponding to the mth tree m,k , obtained by the following formula:

[0066]

[0067] For each leaf node, calculate the best fitting value;

[0068]

[0069] Update the learner of each tree:

[0070]

[0071] Get the final learner:

[0072]

[0073] In the formula, K represents the largest cluster, c represents a constant value, and L represents the loss function. Expressed as a differential symbol, R m,jRepresented as the sample set of the jth leaf node of the mth tree, c m,j Represented as the best fitting value of the jth leaf node of the mth tree, J m It is represented by the number of leaf nodes of the mth regression tree, and I is represented by the indicator function. When x belongs to the leaf node region R m,j When ( x∈R m,j )=1, otherwise it is 0.

[0074] A load aggregate prediction system of an integrated framework, applying the load aggregate prediction method of the integrated framework described above, comprises:

[0075] The prediction module is used to apply the prediction results of the trained load aggregate prediction model for scheduling;

[0076] A construction module is used to construct a load aggregate prediction model, including: obtaining historical data of load aggregates; analyzing external influencing factors on historical data of load aggregates and screening out important external influencing factors; clustering historical data of load aggregates using the K-shape algorithm and dividing them into several clusters; constructing a load aggregate prediction model, including a CNN-LSTM-ATTENTION prediction model and a GBDT prediction model; screening out important external influencing factors and dividing several clusters, and performing load prediction on each cluster separately through the CNN-LSTM-ATTENTION prediction model; and using the GBDT prediction model to integrate the load prediction results of each cluster and obtain the load aggregate prediction results.

[0077] Compared with the prior art, the present invention has the following beneficial effects: the present invention integrates the prediction results of different clusters to obtain the overall prediction result, which can balance the errors in the predictions of each group, thereby improving the accuracy and robustness of the overall prediction. Through integrated learning, the overall error caused by the prediction error of a certain group can be eliminated, ensuring that the final prediction result is more stable and reliable.

[0078] The present invention adopts the GBDT method to improve the prediction accuracy by mining the nonlinear relationship between the prediction results of each cluster, reduce the impact of prediction errors on the overall scheduling, and ensure that the prediction of load aggregates is more stable and reliable.

[0079] In terms of clustering, the present invention: K-shape clustering is based on the dynamic characteristics of the time series shape, which can more accurately reflect the changing rules of the load time series. Compared with distributed spectral clustering, the computational complexity of distributed spectral clustering is higher, especially in high-dimensional time series, and the dynamic shape characteristics of the time series may be ignored.

[0080] In terms of the prediction model of the present invention: CNN can extract local features and can effectively capture local patterns or mutation trends in load sequences. LSTM is superior to GRU in capturing long-term dependencies in time series and is suitable for processing complex dynamic changes. The Attention mechanism can identify key time points or patterns in the sequence, assign higher weights to important hidden states, and improve prediction accuracy. Compared with the random forest method, the random forest method improves the stability and robustness of the prediction by integrating multiple GRUs with different structures, which plays a role in tuning the parameters of the model, but there is no substantial improvement in feature extraction and complex pattern capture, and the method is relatively cumbersome, and the circumstances considered are not necessarily comprehensive.

[0081] The present invention analyzes the division of different load types within a load aggregate, establishes a load aggregate prediction model, and fully mines the temporal and spatial correlation of data, which can effectively improve the accuracy of load aggregate prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a flow chart for constructing a load aggregation prediction model according to an embodiment of the present invention.

[0083] Figure 2 A block diagram of a load aggregation prediction system with an integrated framework according to an embodiment of the present invention. DETAILED DESCRIPTION

[0084] In order to facilitate those skilled in the art to understand the technical solution of the present invention, the technical solution of the present invention is further described in conjunction with the accompanying drawings of the specification.

[0085] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0086] See also Figure 1 As shown, the present invention discloses a load aggregate prediction method of an integrated framework, which uses the prediction results of a trained load aggregate prediction model to perform scheduling. The construction of the load aggregate prediction model includes:

[0087] S10, obtaining load aggregate historical data.

[0088] In one embodiment of the present invention, the acquired load aggregate historical data is preprocessed, including cleaning the load aggregate historical data for incomplete data, cleaning the load aggregate historical data for data format, and cleaning the load aggregate historical data for erroneous data. Undoubtedly, in steps S20 and S30, the load aggregate historical data used is the preprocessed data.

[0089] S20, analyzing external influencing factors on historical data of load aggregates, and screening out important external influencing factors.

[0090] In one embodiment of the present invention, external influencing factors are analyzed on the pre-processed load aggregate historical data to obtain analysis results, including:

[0091] Air temperature, air pressure, relative humidity and wind speed are selected as candidate external influencing factors. The Spearman correlation coefficient method is used to analyze the relationship between the load and each external influencing factor as a correlation index, and important influencing factors are selected as input features of the model. The calculation formula of the Spearman correlation coefficient is as follows:

[0092]

[0093] In the formula, R i , S i They represent the levels of the values ​​of observation i, They are the average levels of load and candidate external influencing factors, N is the total number of observations, and d i =R i -S i Expressed as the number of rank differences between two columns of paired variables.

[0094] Set a relationship threshold, compare the correlation index with the relationship threshold, and take the candidate external influencing factors corresponding to the correlation index greater than the relationship threshold as important external influencing factors. The larger the value of the correlation index, the stronger the correlation, and the external influencing factors with correlation index greater than the relationship threshold are selected as the input of the prediction model.

[0095] S30, clustering the load aggregate historical data using a K-shape algorithm to divide the data into a plurality of clusters.

[0096] In one embodiment of the present invention, the processed load aggregate historical data is clustered using the K-shape algorithm, and the similarity of each two sets of load data is calculated by measuring the distance through cross-correlation. In order to achieve translation invariance, the v sequence is kept unchanged when calculating the cross-correlation, and u is slid on v to calculate the inner product of each displacement of u. Therefore, the calculation formula of shape-based distance (SBD) is as follows:

[0097]

[0098] In the formula, the value range of SBD(u,v) is [0,2], 0 means that the historical data of two sets of load aggregation are most similar, u,v means the historical data of two sets of load aggregation based on time series, max w (·) represents the calculation of w that maximizes u, CC w (u,v) represents the mutual correlation coefficient of u and v, R0(u,u) represents the autocorrelation coefficient of u, and R0(v,v) represents the autocorrelation coefficient of v.

[0099] The processed power load data is clustered using the K-shape algorithm, and the number of clusters is determined by the inflection point method. A line graph is drawn between the sum of the squared deviations of each cluster and the number of clusters to observe the position of the inflection point change. As the number of clusters increases, the sum of the squared deviations tends to be stable. The sum of squared deviations can be expressed as:

[0100]

[0101] In the formula, E ii It represents the historical data of the load aggregate of group ii, and n represents the maximum value of ii. It is represented as the mean of the historical data of the load aggregation, and SS is represented as the sum of squared deviations.

[0102] In one embodiment of the present invention, the processed load data is clustered using the K-shape algorithm, the centroid is recalculated based on the distance scale of the current centroid of each cluster and the cross-correlation measurement of all points in the cluster, and each sequence is redistributed to the cluster closest to the centroid based on its distance from the new centroid, and the cycle is iterated until the label no longer changes.

[0103] In other words, the specific analysis process of K-shape clustering is as follows:

[0104] The processed power load data is used to calculate the similarity between each two groups of power load data through shape-based cross-correlation measurement distance. The number of clusters is determined by the inflection point method, and the centroid is initialized. The centroid is recalculated based on the distance scale of the current centroid of each group and the cross-correlation measurement of all points in the group. Each data in each group of load aggregate historical data is re-divided to form a cluster, and each group is redistributed to the cluster closest to the centroid according to the distance from the new centroid. The cycle is iterated until the label no longer changes.

[0105] In one embodiment of the present invention, the order of steps S20 and S30 is adjustable.

[0106] S40, construct a load aggregation prediction model, including a CNN-LSTM-ATTENTION prediction model and a GBDT prediction model.

[0107] S50, screen out important external influencing factors and divide them into several clusters, and perform load forecasting on each cluster through the CNN-LSTM-ATTENTION prediction model.

[0108] In one embodiment of the present invention, local features are extracted through a CNN network model. The CNN network can extract local dependencies between variables and short-term patterns in the time dimension. The convolution layer of the network consists of multiple kernels, and the kth filter sweeps over the input matrix, which can be expressed as:

[0109] h k =ReLU(W k *X+b k );

[0110] In the formula, h k Represented as the feature vector output by the kth filter, W k is represented as the slope of the kth filter, X is represented as the input matrix, b k It is represented as the intercept of the kth filter, and * represents the convolution operation. The input matrix is ​​a matrix composed of several clusters divided by the K-shape algorithm and important external influencing factors. The output of the convolution layer is a matrix of size d c ×TT matrix, where d c Represents the number of filters, TT represents the time dimension. Activation function ReLU(xx)=max(xx,0), xx∈X, X∈R D×TT , D represents the multivariate dimension.

[0111] Extract time information through LSTM. Recurrent Neural Networks (RNN) are often used to solve time series prediction problems. As an improvement of RNN, LSTM can solve the long-distance dependency problem of RNN during training and avoid gradient explosion or gradient dispersion. Each LSTM unit consists of a memory unit with a state of st at time t, which is controlled by three sigmoid gates: input gate i t 、Forget gate t and output gate o t The eigenvectors corresponding to all input matrices form an input sequence sample set, and the calculation expression in the LSTM model is as follows:

[0112] Input Gate i t :

[0113] i t =sigmoid(W i[h t-1 ;x t ]+b i );

[0114] Forget Gate t :

[0115] f t =sigmoid(W f [h t-1 ;x t ]+b f );

[0116] Output gate o t :

[0117] o t =sigmoid(W o [h t-1 ;x t ]+b o );

[0118] Memory unit update:

[0119] s t =f t ⊙s t-1 +i t ⊙tanh(W s [h t-1 ;x t ]+b s );

[0120] Output:

[0121] h t =o t ⊙tanh(s t );

[0122] In the formula, x t is a feature vector, representing the current input; [h t-1 ;x t ] represents the hidden state h at time t-1 t-1 and the current input x t The series connection, W i , W f , W o , W s They are respectively represented as the weight matrices of the input gate, forget gate, output gate and memory unit, ⊙ represents element-by-element multiplication, and b i , b f , b o , b sThey represent the bias vector parameters of the input gate, forget gate, output gate and memory unit respectively, tanh represents the hyperbolic tangent function, ReLU and sigmoid represent two different activation functions respectively.

[0123] In one embodiment of the present invention, weights are assigned through an attention mechanism. The attention layer is used to represent the correlation between the input sequence and the output result. By assigning different attention weights to hidden states, the importance of hidden states is distinguished, and the accuracy of prediction is improved, including the following steps:

[0124] For a given hidden state sequence of a hidden layer, calculate the attention score of each hidden layer;

[0125] Among them, the attention score is obtained by:

[0126] U m =tan(W·h m + b);

[0127] Where U m It is represented as the attention score of the mth hidden layer, h m represents the mth hidden layer, W represents the attention weight matrix, b represents the bias vector parameter of the attention mechanism, and tan represents the tangent;

[0128] According to the attention score, get the attention weight probability.

[0129] Among them, the attention weight probability is obtained by the following formula:

[0130]

[0131] In the formula, p m It is represented as the attention weight probability of the mth hidden layer, exp is represented as the exponential function, and M is represented as the maximum hidden layer;

[0132] According to the attention weight probability, the attention weight of each hidden layer is weighted to obtain the weighted hidden layer.

[0133] Among them, the weighted hidden layer is obtained by the following formula:

[0134]

[0135] In the formula, represents the weighted hidden layer.

[0136] In one embodiment of the present invention, further, the Adam algorithm is used to optimize the CNN-LSTM-ATTENTION prediction model. The Adam algorithm is a first-order optimization algorithm that can replace the traditional stochastic gradient descent process. It can iteratively update the weights of the neural network according to the training data. The Adam algorithm combines the optimal performance of the AdaGrad and RMSProp algorithms and can provide an optimization method to solve the sparse gradient and noise problems.

[0137] S60, using the GBDT prediction model to integrate the load prediction results of each cluster to obtain a load aggregate prediction result.

[0138] In one embodiment of the present invention, the GBDT (GradientBoostingDecisionTree) prediction model integrates the prediction results of each cluster to obtain the load aggregate prediction result F M (x) including:

[0139] Given a prediction data set T, T = {(x1, y1), (x2, y2), ... (x k ,y k ),...,(x K ,y K )}; where x k The predicted value x of all clusters is represented by k It is necessary to approach this target value through weighted fusion, so y1 = y2 = ... = y K =y.

[0140] Initialize the learner F0(x):

[0141]

[0142] Establish M classification and regression trees and calculate the response corresponding to the mth tree;

[0143] Among them, the response r corresponding to the mth tree m,k , obtained by the following formula:

[0144]

[0145] For each leaf node, calculate the best fitting value;

[0146]

[0147] Update the learner of each tree:

[0148]

[0149] Get the final learner:

[0150]

[0151] In the formula, K represents the largest cluster, c represents a constant value, and L represents the loss function. Expressed as a differential symbol, R m,j Represented as the sample set of the jth leaf node of the mth tree, c m,j Represented as the best fitting value of the jth leaf node of the mth tree, J m It is represented by the number of leaf nodes of the mth regression tree, and I is represented by the indicator function. When x belongs to the leaf node region R m,j When I(x∈R m,j )=1, otherwise it is 0.

[0152] See also Figure 2 As shown, the present invention also provides a load aggregate prediction system of an integrated framework, which applies the load aggregate prediction method of the integrated framework described above, including:

[0153] The prediction module is used to apply the prediction results of the trained load aggregate prediction model for scheduling.

[0154] A construction module is used to construct a load aggregate prediction model, including: obtaining historical data of load aggregates; analyzing external influencing factors on historical data of load aggregates and screening out important external influencing factors; clustering historical data of load aggregates using the K-shape algorithm and dividing them into several clusters; constructing a load aggregate prediction model, including a CNN-LSTM-ATTENTION prediction model and a GBDT prediction model; screening out important external influencing factors and dividing several clusters, and performing load prediction on each cluster separately through the CNN-LSTM-ATTENTION prediction model; and using the GBDT prediction model to integrate the load prediction results of each cluster and obtain the load aggregate prediction results.

[0155] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting from any point of view, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0156] The above-described embodiments merely represent implementation methods of the invention. The protection scope of the present invention is not limited to the above-described embodiments. For those skilled in the art, several modifications and improvements may be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A load aggregate prediction method with an integrated framework, characterized in that: Apply the prediction results of the trained load aggregate prediction model to perform scheduling; The construction of the load aggregate prediction model includes: Get load aggregate historical data; Analyze external influencing factors on historical data of load aggregates and screen out important external influencing factors; The load aggregate historical data is clustered using the K-shape algorithm and divided into several clusters; Construct load aggregation prediction models, including CNN-LSTM-ATTENTION prediction model and GBDT prediction model; The important external influencing factors and several clusters are screened out, and the load forecast is performed for each cluster through the CNN-LSTM-ATTENTION forecasting model; The GBDT prediction model is used to integrate the load prediction results of each cluster to obtain the load aggregate prediction results.

2. The load aggregate prediction method of the integrated framework according to claim 1, characterized in that: Screen out important external influencing factors and obtain them through the following steps: Air temperature, air pressure, relative humidity and wind speed are selected as candidate external influencing factors; The Spearman correlation coefficient method was used to obtain the relationship between the load and each candidate external influencing factor as a correlation index; A relationship threshold is set, the correlation index is compared with the relationship threshold, and the candidate external influencing factors corresponding to the correlation index greater than the relationship threshold are regarded as important external influencing factors.

3. The load aggregate prediction method of the integrated framework according to claim 2, characterized in that: Correlation index P s , obtained by the following formula: In the formula, R i , S i They represent the levels of the observation value i, They are the average levels of load and candidate external influencing factors, N is the total number of observations, and d i =R i -S i Expressed as the number of rank differences between two columns of paired variables.

4. The load aggregate prediction method of the integrated framework according to claim 1, characterized in that: Divided into several clusters, including: The distance between the load aggregate historical data is measured by cross-correlation, and the similarity between each two groups of load aggregate historical data is calculated; The number of clusters is determined by using the inflection point method for the historical data of load aggregates; Initialize the centroid; The centroid is recalculated based on the distance scale of the cross-correlation measurement between the current centroid of each category and all points in the class, and each data in each group of load aggregate historical data is re-divided to form a cluster; and each group of load aggregate historical data is redistributed to the cluster closest to the centroid based on the distance between it and the new centroid. The cycle is iterated until the label no longer changes, and the regrouped load aggregate historical data is obtained.

5. The load aggregate prediction method of the integrated framework according to claim 4, characterized in that: Calculate the similarity SBD(u,v) of each two groups of load aggregate historical data using the following formula: In the formula, the value range of SBD(u,v) is [0,2], 0 means that the historical data of two sets of load aggregates are most similar, u,v means the historical data of two sets of load aggregates based on time series, max w (·) represents the calculation of w that maximizes u, CC w (u,v) represents the mutual correlation coefficient of u and v, R0(u,u) represents the autocorrelation coefficient of u, and R0(v,v) represents the autocorrelation coefficient of v.

6. The load aggregate prediction method of the integrated framework according to claim 4, characterized in that: Draw a line graph between the sum of squared deviations of each group of load aggregate historical data and the number of each group of load aggregate historical data, and observe the position of the inflection point change. As the load aggregate historical data increases, the sum of squared deviations tends to be stable, and the number of inflection points is used as the number of clusters.

7. The load aggregate prediction method of the integrated framework according to claim 1, characterized in that: The CNN-LSTM-ATTENTION prediction model is used to perform load prediction for each cluster, including: The local features of the input matrix are extracted through the CNN network model to obtain the feature vector; wherein the input matrix is ​​a matrix composed of several clusters divided by the K-shape algorithm and important external influencing factors screened out; and the feature vector is obtained by the following formula: h k =ReLU(W k *X+b k ); The eigenvectors corresponding to all input matrices form an input sequence sample set; The input sequence sample set is used as the input of the LSTM network model, and the load forecast results of each cluster are obtained as output; Among them, the LSTM network model includes input gate i t 、Forget gate t and output gate o t : Input Gate i t : i t =sigmoid(W i [h t-1 ;x t ]+b i ); Forget Gate t : f t =sigmoid(W f [h t-1 ;x t ]+b f ); Output gate o t : o t =sigmoid(W o [h t-1 ;x t ]+b o ); Memory unit update: s t =f t ⊙s t-1 +i t ⊙tanh(W s [h t-1 ;x t ]+b s ); Output: h t =o t ⊙tanh(s t ); In the formula, x t is a feature vector, representing the current input; [h t-1 ;x t ] represents the hidden state h at time t-1 t-1 and the current input x t The series connection, W i , W f , W o , W s They are respectively represented as the weight matrices of the input gate, forget gate, output gate and memory unit, ⊙ represents element-by-element multiplication, and b i , b f , b o , b s They represent the bias vector parameters of the input gate, forget gate, output gate and memory unit respectively, tanh represents the hyperbolic tangent function, ReLU and sigmoid represent two different activation functions respectively; h k Represented as the feature vector output by the kth filter, W k is represented as the slope of the kth filter, X is represented as the input matrix, b k It is represented as the intercept of the kth filter and * represents the convolution operation.

8. The load aggregate prediction method of the integrated framework according to claim 7, characterized in that: The attention mechanism is introduced into the hidden state of the hidden layer in the LSTM network model to assign different attention weights to the hidden state of each hidden layer, including the following steps: For a given hidden state sequence of a hidden layer, calculate the attention score of each hidden layer; Among them, the attention score is obtained by: U m =tan(W·h m +b); Where U m It is represented as the attention score of the mth hidden layer, h m represents the mth hidden layer, W represents the attention weight matrix, b represents the bias vector parameter of the attention mechanism, and tan represents the tangent; According to the attention score, obtain the attention weight probability; Among them, the attention weight probability is obtained by the following formula: In the formula, p m It is represented as the attention weight probability of the mth hidden layer, exp is represented as the exponential function, and M is represented as the maximum hidden layer; According to the attention weight probability, the attention weight of each hidden layer is weighted to obtain the weighted hidden layer; Among them, the weighted hidden layer is obtained by the following formula: In the formula, represents the weighted hidden layer.

9. The load aggregate prediction method of the integrated framework according to claim 1, characterized in that: Get the load aggregate prediction result F M (x) including: Given a prediction data set T, T = {(x1, y1), (x2, y2), ... (x k ,y k ),...,(x K ,y K )}; where x k It is represented as the prediction result of the kth cluster, and y is represented as the prediction result of the load aggregate; Initialize the learner F0(x): Establish M classification and regression trees and calculate the response corresponding to the mth tree; Among them, the response r corresponding to the mth tree m,k , obtained by the following formula: For each leaf node, calculate the best fitting value; Update the learner of each tree: Get the final learner: In the formula, K represents the largest cluster, c represents a constant value, and L represents the loss function. Expressed as a differential symbol, R m,j Represented as the sample set of the jth leaf node of the mth tree, c m,j Represented as the best fitting value of the jth leaf node of the mth tree, J m It is represented by the number of leaf nodes of the mth regression tree, and I is represented by the indicator function. When x belongs to the leaf node region R m,j When I(x∈R m,j )=1, otherwise it is 0.

10. A load aggregation prediction system with an integrated framework, characterized in that: A load aggregate prediction method using the integrated framework described in any one of claims 1 to 9, comprising: The prediction module is used to apply the prediction results of the trained load aggregate prediction model for scheduling; A construction module is used to construct a load aggregate prediction model, including: obtaining historical data of load aggregates; analyzing external influencing factors on historical data of load aggregates and screening out important external influencing factors; clustering historical data of load aggregates using the K-shape algorithm and dividing them into several clusters; constructing a load aggregate prediction model, including a CNN-LSTM-ATTENTION prediction model and a GBDT prediction model; screening out important external influencing factors and dividing several clusters, and performing load prediction on each cluster separately through the CNN-LSTM-ATTENTION prediction model; and using the GBDT prediction model to integrate the load prediction results of each cluster and obtain the load aggregate prediction results.

Images