An integrated framework for load aggregate body prediction method and system

Abstract

The application discloses a load aggregation body prediction method and system of an integrated framework, and the method comprises the following steps: applying a prediction result of a trained load aggregation body prediction model to perform scheduling; wherein the construction of the load aggregation body prediction model comprises the following steps: obtaining load aggregation body historical data; performing external influence factor analysis on the load aggregation body historical data, and screening important external influence factors; applying a K-shape algorithm to cluster the load aggregation body historical data, and dividing the load aggregation body historical data into a plurality of clusters; constructing a load aggregation body prediction model, including a CNN-LSTM-ATTENTION prediction model and a GBDT prediction model; performing load prediction on each cluster through the CNN-LSTM-ATTENTION prediction model by using the screened important external influence factors and the divided plurality of clusters; and applying the GBDT prediction model to perform integrated prediction on the load prediction results of the clusters, thereby obtaining a load aggregation body prediction result; and the application can improve the prediction accuracy of the load aggregation body prediction model.

Description

Technical Field

[0001] This invention relates to the field of load aggregate technology, specifically to an integrated framework method and system for predicting load aggregates. Background Technology

[0002] With the large-scale grid connection of renewable energy and user-side adjustable resources, and the continuous development of the electricity market, demand response, which reflects market competition and demand elasticity, has opened up new possibilities for user-side resources to participate in the electricity market. Against this backdrop, load aggregators, as emerging market players aggregating large amounts of load resources, have developed rapidly. Load aggregators can integrate user-side resources through demand response and thereby provide load smoothing services to the grid to generate revenue. Therefore, research on load aggregator is essential. Furthermore, with the widespread adoption of smart meters, fine-grained load data is readily available.

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

[0004] The paper "Load Aggregator Prediction Method Based on Gated Cyclic Unit Network and Model Fusion" discloses a method for predicting load aggregates by using a distributed spectral clustering algorithm to obtain load groups with similar load characteristics, then performing group predictions, employing GRU as the meta-model to dynamically model the time series, and using a random forest algorithm to fuse multiple GRU networks with different structures to predict the load groups. Finally, the predicted values ​​of each group are summed to obtain the predicted value of the load aggregate. However, in this paper, the final predicted value of the load aggregate is obtained by directly adding the prediction results of different groups. Load aggregation aims to achieve flexible adjustment and response of the power system by centrally managing and scheduling the power demand of multiple small power users. 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 based on the prediction results, the responses of different users may be different, further affecting the scheduling effect of the system. Summary of the Invention

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

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] An integrated framework for load aggregation prediction method, which uses the prediction results of a trained load aggregation prediction model for scheduling;

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

[0009] Obtain historical data of load aggregates;

[0010] External influencing factors were analyzed based on historical data of load aggregates to identify important external influencing factors.

[0011] The historical data of the load aggregates were clustered using the K-shape algorithm and divided into several clusters.

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

[0013] Important external influencing factors and several clusters will be selected, and load prediction will be performed on each cluster separately using the CNN-LSTM-ATTENTION prediction model.

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

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

[0016] Temperature, air pressure, relative humidity, and wind speed were 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, which was used as the correlation index.

[0018] Set a relationship threshold, compare the correlation index with the relationship threshold, and identify the candidate external influencing factors whose correlation index is greater than the relationship threshold as important external influencing factors.

[0019] In one embodiment of the present invention, the correlation index P s It can be obtained through the following formula:

[0020]

[0021] In the formula, R i S i These represent the levels of the observed value i. Let d represent the average rank of the load and the candidate external influencing factors, respectively, and N represent the total number of observations. i =R i -S i It is represented as the rank difference between two paired variables.

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

[0023] The similarity between any two sets of historical load aggregate data is calculated by cross-correlation of the historical load aggregate data.

[0024] The number of clusters was determined using the inflection point method based on historical data of load aggregates;

[0025] Initialize the centroid;

[0026] The centroids are recalculated based on the distance scale of the cross-correlation measurement between the current centroid of each category and all points in the category. Each data point in the historical data of each load aggregate is re-divided into clusters. The historical data of each load aggregate are then reassigned to the clusters closest to the centroid based on the distance between the historical data of each load aggregate and the new centroid. This process is repeated until the labels no longer change, and the regrouped historical data of the load aggregates is obtained.

[0027] In one embodiment of the present invention, the similarity SBD(u,v) between two sets of historical data of load aggregates is calculated using the following formula:

[0028]

[0029] In the formula, SBD(u,v) ranges from [0,2], where 0 indicates that the two sets of load aggregate historical data are most similar, u and v represent the two sets of load aggregate historical data based on time series, and max w (·) represents the computation of w that maximizes u, CC w (u,v) represents the cross-correlation coefficient between 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 plotted between the sum of squared deviations of historical data of each load aggregate and the number of historical data of each load aggregate, and the position of the inflection point is observed. As the historical data of the load aggregate increases, the sum of squared deviations tends to stabilize, and the number of inflection points is taken as the number of clusters.

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

[0032] The CNN network model extracts local features from the input matrix to obtain feature vectors. The input matrix is ​​a matrix formed by K-shape clustering to divide the matrix into several clusters and selecting important external influencing factors. The feature vectors are obtained using the following formula:

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

[0034] The eigenvectors corresponding to all input matrices are used to form the input sequence sample set;

[0035] The input sequence sample set is used as the input to the LSTM network model, and the output is the load prediction result of each cluster.

[0036] The LSTM network model includes input gate i t Forgotten 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] Forgotten Gate f 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 The feature vector represents 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 Serial connection, W i W f W o W s Let b be the weight matrix of the input gate, forget gate, output gate, and memory unit, respectively. Let ⊙ represent element-wise multiplication, and let b be the weight matrix of the memory unit. i b f b o b s Let represent the bias vector parameters of the input gate, forget gate, output gate, and memory unit, respectively; tanh represents the hyperbolic tangent function; and ReLU and sigmoid represent two different activation functions, respectively. k Let W be the feature vector output by the k-th filter. k Let X represent the slope of the k-th filter, and let b represent the input matrix. k The value represents the intercept of the k-th filter, and * represents the convolution operation.

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

[0049] For a given sequence of hidden states in a hidden layer, calculate the attention score for each hidden layer.

[0050] Attention score is determined as follows:

[0051] U m =tan(W·h) m +b);

[0052] In the formula, U m Let h represent the attention score of the m-th hidden layer. m Let W be the m-th hidden layer, W be the attention weight matrix, b be the bias vector parameter of the attention mechanism, and tan be the tangent.

[0053] Based on the attention score, obtain the attention weight probability;

[0054] The attention weight probability is obtained using the following formula:

[0055]

[0056] In the formula, p mLet be the attention weight probability of the m-th hidden layer, exp be the exponential function, and M be the maximum hidden layer.

[0057] Based on the attention weight probabilities, the attention weights of each hidden layer are weighted to obtain the weighted hidden layer; the weighted hidden layer is obtained using the following formula:

[0058]

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

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

[0061] Given a prediction dataset T, T = {(x1,y1),(x2,y2),...(x k ,y k ),...,(x K ,y K )};where, x k Let y represent the prediction result for the k-th cluster, and y represent the prediction result for the load aggregate.

[0062] Initialize the learner F0(x):

[0063]

[0064] Construct M classification and regression trees, and calculate the response corresponding to the m-th tree;

[0065] Wherein, the response r corresponding to the m-th tree m,k It can be obtained through the following formula:

[0066]

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

[0068]

[0069] Update the learner for each tree:

[0070]

[0071] Obtain the final learner:

[0072]

[0073] In the formula, K represents the largest cluster, c represents a constant value, and L represents the loss function. Represented by the differential symbol, R m,jLet c be the sample set of the j-th leaf node of the m-th tree. m,j Let J represent the best-fit value for the j-th leaf node of the m-th tree. m Let I represent the number of leaf nodes of the m-th regression tree, and let I be the indicator function, indicating when x belongs to the leaf node region R. m,j At that time, I ( x∈R m,j =1, otherwise 0.

[0074] An integrated framework-based load aggregation prediction system, applying the aforementioned integrated framework-based load aggregation prediction method, includes:

[0075] The prediction module is used to perform scheduling by applying the prediction results of the trained load aggregation prediction model.

[0076] The module is used to construct the load aggregation prediction model, including: acquiring historical load aggregation data; analyzing the external influencing factors of the historical load aggregation data and screening out important external influencing factors; applying the K-shape algorithm to cluster the historical load aggregation data into several clusters; constructing the load aggregation prediction model, including the CNN-LSTM-ATTENTION prediction model and the GBDT prediction model; using the CNN-LSTM-ATTENTION prediction model to predict the load for each cluster, based on the screened important external influencing factors and the several clusters; and applying the GBDT prediction model to integrate the load prediction results of each cluster to obtain the load aggregation prediction result.

[0077] Compared with existing technologies, the beneficial effects of this invention are: this invention integrates the prediction results of different clusters to obtain the overall prediction result, which can balance the errors in each group's prediction, thereby improving the overall prediction accuracy and robustness. Through ensemble learning, it can eliminate the overall error caused by the failure of a certain group's prediction, ensuring that the final prediction result is more stable and reliable.

[0078] This invention employs the GBDT method to improve prediction accuracy and reduce the impact of prediction errors on overall scheduling by mining the nonlinear relationships between the prediction results of each cluster, thereby ensuring that the prediction of load aggregates is more stable and reliable.

[0079] In terms of clustering, this invention utilizes K-shape clustering, which groups data based on the dynamic shape characteristics of time series, thus more accurately reflecting the changing patterns of load time series. Compared to distributed spectral clustering, distributed spectral clustering has higher computational complexity, especially for high-dimensional time series, and may neglect the dynamic shape characteristics of the time series.

[0080] In terms of prediction models, this invention offers several advantages: CNNs can extract local features and effectively capture local patterns or abrupt changes in load sequences. LSTMs outperform GRUs in capturing long-term dependencies in time series and are suitable for handling complex dynamic changes. Attention mechanisms can identify key time points or patterns in a sequence, assigning higher weights to important hidden states and improving prediction accuracy. In contrast, while Random Forest improves prediction stability and robustness by integrating multiple GRUs with different structures, thus contributing to model parameter tuning, it doesn't offer substantial improvements in feature extraction and complex pattern capture. Furthermore, the method is cumbersome and may not consider all possible scenarios.

[0081] This invention analyzes the different load types within a load aggregate, establishes a load aggregate prediction model, and fully explores the spatiotemporal correlation of the data, which can effectively improve the accuracy of load aggregate prediction. Attached Figure Description

[0082] Figure 1 This is a flowchart illustrating the construction process of the load aggregate prediction model according to an embodiment of the present invention.

[0083] Figure 2 This is a block diagram of a load aggregate prediction system based on an integrated framework according to an embodiment of the present invention. Detailed Implementation

[0084] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0085] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0086] Please see Figure 1 As shown, this invention discloses an integrated framework-based load aggregation prediction method, which applies the prediction results of a trained load aggregation prediction model for scheduling. The construction of the load aggregation prediction model includes:

[0087] S10, Obtain historical data of load aggregate.

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

[0089] S20, analyze the external influencing factors of historical data of load aggregates and screen out important external influencing factors.

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

[0091] Temperature, air pressure, relative humidity, and wind speed were selected as candidate external influencing factors. The Spearman correlation coefficient method was used to analyze the relationship between the load and each external influencing factor, using these as correlation indicators to select important influencing factors as input features for the model. The formula for calculating the Spearman correlation coefficient is as follows:

[0092]

[0093] In the formula, R i S i These represent the levels of the observed value i. Let d represent the average rank of the load and the candidate external influencing factors, respectively, and N represent the total number of observations. i =R i -S i It is represented as the rank difference between two paired variables.

[0094] A relationship threshold is set, and the correlation index is compared with the relationship threshold. Candidate external influencing factors with correlation indices greater than the relationship threshold are considered important external influencing factors. The higher the correlation index value, the stronger the correlation. External influencing factors with correlation indices greater than the relationship threshold are selected as inputs to the prediction model.

[0095] S30 uses the K-shape algorithm to cluster the historical data of the load aggregates into several clusters.

[0096] In one embodiment of the present invention, the processed historical data of load aggregates are clustered using the K-shape algorithm, and the similarity between each pair of load data is calculated by measuring the distance through cross-correlation. To achieve translation invariance, the v sequence is kept constant when calculating the cross-correlation, and u is slid along v, calculating the inner product of each displacement of u. Therefore, the formula for calculating Shape-Based Distance (SBD) is as follows:

[0097]

[0098] In the formula, SBD(u,v) ranges from [0,2], where 0 indicates that the two sets of load aggregate historical data are most similar, u and v represent the two sets of load aggregate historical data based on time series, and max w (·) represents the computation of w that maximizes u, CC w (u,v) represents the cross-correlation coefficient between 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 was clustered using the K-shape algorithm, and the number of clusters was determined by the inflection point method. A line graph was plotted showing the relationship between the sum of squared deviations of each cluster and the number of clusters to observe the location of the inflection point. As the number of clusters increases, the sum of squared deviations tends to stabilize. The sum of squared deviations can be expressed as:

[0100]

[0101] In the formula, E ii Let represent the historical data of the i-th load aggregate, and n represent the maximum value of ii. is represented by the mean of the historical data of the load aggregate, and SS is represented by 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 centroids are recalculated based on the current centroids of each cluster and the distance scale of the cross-correlation measurement of all points within the cluster. The sequences are then reassigned to the clusters closest to the centroids based on their distance from the new centroids. This process is iterated until the labels no longer change.

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

[0104] The similarity between each pair of power load data is calculated by measuring the distance of cross-correlation based on shape. The number of clusters is determined by the inflection point method, and the centroids are initialized. The centroids are recalculated based on the current centroid of each group and the distance scale of the cross-correlation measurement of all points in the group. Each data point in the historical data of each load aggregate is re-divided into clusters, and they are reassigned to the clusters closest to the centroids based on the distance of each group to the new centroids. This process is iterated until the labels no longer change.

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

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

[0107] S50 will select important external influencing factors and divide them into several clusters, and use the CNN-LSTM-ATTENTION prediction model to predict the load of each cluster separately.

[0108] In one embodiment of the present invention, local features are extracted using a CNN network model. CNN networks can extract local dependencies between variables and short-term patterns in the time dimension. The convolutional layers of the network consist of multiple kernels, and the k-th filter sweeps through the input matrix, which can be represented as:

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

[0110] In the formula, h k Let W be the feature vector output by the k-th filter. k Let X represent the slope of the k-th filter, and let b represent the input matrix. k denoted as the intercept of the k-th filter, and * denotes the convolution operation. The input matrix is ​​a matrix formed by K-shape clustering to divide the data into clusters and selecting important external influencing factors. The output of the convolutional layer is a matrix of size d. c A matrix of ×TT, where d c The number of filters is represented by TT, where TT represents the time dimension. The activation function is ReLU(xx) = max(xx, 0), where xx ∈ X, X ∈ R. D×TT D represents the dimension of the multivariate variables.

[0111] LSTM extracts time information. Recurrent Neural Networks (RNNs) are frequently used to solve time series prediction problems. LSTM, as an improvement on RNNs, can address the long-range dependency problem during RNN training and avoid gradient explosion or gradient dispersion. Each LSTM unit consists of a memory unit with state st at time t, controlled by three sigmoid gates: input gate i... t Forgotten Gate t and output gate o t If we take the eigenvectors corresponding to all input matrices and form an input sequence sample set, then the computational 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] Forgotten Gate f 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 The feature vector represents 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 Serial connection, W i W f W o W s Let b be the weight matrix of the input gate, forget gate, output gate, and memory unit, respectively. Let ⊙ represent element-wise multiplication, and let b be the weight matrix of the memory unit. i b f b o b sThese represent the bias vector parameters of the input gate, forget gate, output gate, and memory unit, respectively; tanh represents the hyperbolic tangent function; and ReLU and sigmoid represent two different activation functions.

[0123] In one embodiment of the present invention, weights are assigned using 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 the hidden states, the importance of the hidden states is distinguished, thereby improving the accuracy of prediction, including the following steps:

[0124] For a given sequence of hidden states in a hidden layer, calculate the attention score for each hidden layer.

[0125] Attention score is determined as follows:

[0126] U m =tan(W·h) m +b);

[0127] In the formula, U m Let h represent the attention score of the m-th hidden layer. m Let W be the m-th hidden layer, W be the attention weight matrix, b be the bias vector parameter of the attention mechanism, and tan be the tangent.

[0128] Based on the attention score, obtain the attention weight probability.

[0129] The attention weight probability is obtained using the following formula:

[0130]

[0131] In the formula, p m Let be the attention weight probability of the m-th hidden layer, exp be the exponential function, and M be the maximum hidden layer.

[0132] Based on the attention weight probability, the attention weights of each hidden layer are weighted to obtain the weighted hidden layer.

[0133] The weighted hidden layer is obtained using the following formula:

[0134]

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

[0136] In one embodiment of the present invention, the Adam algorithm is further employed 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 based on the training data. The Adam algorithm combines the best performance of the AdaGrad and RMSProp algorithms, providing an optimization method to address sparse gradient and noise problems.

[0137] S60 uses the GBDT prediction model to integrate the load prediction results of each cluster and obtain the load aggregate prediction result.

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

[0139] Given a prediction dataset T, T = {(x1,y1),(x2,y2),...(x k ,y k ),...,(x K ,y K )};where, x k Let represent the prediction result for the k-th cluster, and y represent the prediction result for the load aggregate. Wherein, the predicted values ​​for all clusters are x. k We need to approximate this target value through weighted fusion; therefore, y1 = y2 = ... = y K =y.

[0140] Initialize the learner F0(x):

[0141]

[0142] Construct M classification and regression trees, and calculate the response corresponding to the m-th tree;

[0143] Wherein, the response r corresponding to the m-th tree m,k It can be obtained through the following formula:

[0144]

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

[0146]

[0147] Update the learner for each tree:

[0148]

[0149] Obtain the final learner:

[0150]

[0151] In the formula, K represents the largest cluster, c represents a constant value, and L represents the loss function. Represented by the differential symbol, R m,j Let c be the sample set of the j-th leaf node of the m-th tree. m,j Let J represent the best-fit value for the j-th leaf node of the m-th tree. m Let I represent the number of leaf nodes of the m-th regression tree, and let I be the indicator function, indicating when x belongs to the leaf node region R. m,j When, I(x∈R) m,j =1, otherwise 0.

[0152] Please see Figure 2 As shown, the present invention also provides an integrated framework load aggregation prediction system, which applies the above-described integrated framework load aggregation prediction method, including:

[0153] The prediction module is used to perform scheduling by applying the prediction results of the trained load aggregation prediction model.

[0154] The module is used to construct the load aggregation prediction model, including: acquiring historical load aggregation data; analyzing the external influencing factors of the historical load aggregation data and screening out important external influencing factors; applying the K-shape algorithm to cluster the historical load aggregation data into several clusters; constructing the load aggregation prediction model, including the CNN-LSTM-ATTENTION prediction model and the GBDT prediction model; using the CNN-LSTM-ATTENTION prediction model to predict the load for each cluster, based on the screened important external influencing factors and the several clusters; and applying the GBDT prediction model to integrate the load prediction results of each cluster to obtain the load aggregation prediction result.

[0155] It will be apparent 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 invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0156] The above-described embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above-described embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for predicting load aggregates using an integrated framework, characterized in that, Considering the internal differences within the load aggregate, the prediction results of the trained load aggregate prediction model are used for scheduling; The construction of the load aggregation prediction model includes: Obtain historical data of load aggregates; External influencing factors were analyzed based on historical data of load aggregates to identify important external influencing factors. The historical data of load aggregates were clustered into several clusters based on the dynamic characteristics of time series shape using the K-shape algorithm. Construct load aggregation prediction models, including CNN-LSTM-ATTENTION prediction models and GBDT prediction models; Important external influencing factors and several clusters will be selected, and load prediction will be performed on each cluster separately using the CNN-LSTM-ATTENTION prediction model. The GBDT forecasting model is applied to integrate the load forecasting results of each cluster, explore the nonlinear relationships between the load forecasting results of each cluster, and obtain the load aggregate forecasting result. ,include: Given a prediction dataset , ;in, Let y represent the prediction result for the k-th cluster, and y represent the prediction result for the load aggregate. Initialize the learner : ； Establish M Tree of classification and regression, calculate the first The response corresponding to each tree; Among them, the The response for each tree It can be obtained through the following formula: ； For each leaf node, calculate the best fit value; ； Update the learner for each tree: ； Obtain the final learner: ； In the formula, K represents the largest cluster. Represented as a constant value, Represented as a loss function, Represented as the differential symbol, Represented as the first Tree No. A sample set of leaf nodes, Represented as the first Tree No. The best-fit value for each leaf node Represented as the first The number of tree nodes in a regression tree. Represented as an indicator function, when Belonging to the leaf node region hour, Otherwise, it is 0.

2. The load aggregation prediction method of the integrated framework according to claim 1, characterized in that, Key external influencing factors were identified through the following steps: Temperature, air pressure, relative humidity, and wind speed were 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, which was used as the correlation index. Set a relationship threshold, compare the correlation index with the relationship threshold, and identify the candidate external influencing factors whose correlation index is greater than the relationship threshold as important external influencing factors.

3. The load aggregation prediction method of the integrated framework according to claim 2, characterized in that, Correlation indicators It can be obtained through the following formula: ； In the formula, , These represent the levels of the observed value i. , These represent the average rank of the load and the candidate external influencing factors, respectively. Represented as the total number of observations, It is represented as the rank difference between two paired variables.

4. The load aggregation prediction method of the integrated framework according to claim 1, characterized in that, Divided into several clusters, including: The similarity between any two sets of historical load aggregate data is calculated by cross-correlation of the historical load aggregate data. The number of clusters was determined using the inflection point method based on historical data of load aggregates; Initialize the centroid; The centroids are recalculated based on the distance scale of the cross-correlation measurement between the current centroid of each category and all points in the category. Each data point in the historical data of each load aggregate is re-divided into clusters. The historical data of each load aggregate are then reassigned to the clusters closest to the centroid based on the distance between the historical data of each load aggregate and the new centroid. This process is repeated until the labels no longer change, and the regrouped historical data of the load aggregates is obtained.

5. The load aggregation prediction method of the integrated framework according to claim 4, characterized in that, Calculate the similarity of historical data for each pair of load aggregates. The formula is as follows: ； In the formula, The value range is [0,2], where 0 indicates that the two sets of load aggregate historical data are most similar. This represents historical data of two load aggregates based on time series. Represented as a calculation that makes u The largest w , Represented as u and v cross-correlation coefficients, Represented as u The autocorrelation coefficient, Represented as v The autocorrelation coefficient.

6. The load aggregation prediction method of the integrated framework according to claim 4, characterized in that, Plot a line graph showing the sum of squared deviations of historical data for each load aggregate group and the number of historical data for each load aggregate group. Observe the location of the inflection point. As the historical data of the load aggregate group increases, the sum of squared deviations tends to stabilize. The number of inflection points is taken as the number of clusters.

7. The load aggregation prediction method of the integrated framework according to claim 1, characterized in that, Load prediction is performed for each cluster using the CNN-LSTM-ATTENTION prediction model, including: The CNN network model extracts local features from the input matrix to obtain feature vectors. The input matrix is ​​a matrix formed by K-shape clustering to divide the matrix into several clusters and selecting important external influencing factors. The feature vectors are obtained using the following formula: ； The eigenvectors corresponding to all input matrices are used to form the input sequence sample set; The input sequence sample set is used as the input to the LSTM network model, and the output is the load prediction result of each cluster. The LSTM network model includes an input gate. Forgotten Gate and output gate : Input gate : ； Forgotten Gate : ； Output gate : ； Memory unit update: ； Output: ； In the formula, The feature vector represents the current input; Represented as Hidden state of time and current input Serial connection, W i , W f , W o , W s These are represented by the weight matrices for the input gate, forget gate, output gate, and memory unit, respectively. This is represented as element-wise multiplication, or as... b i , b f , b o , b s These represent the bias vector parameters for the input gate, forget gate, output gate, and memory unit, respectively. Represented as the hyperbolic tangent function, , These are represented as two different activation functions; Represented as the first The feature vectors output by each filter Represented as the first The slope of each filter, Represented as the input matrix, Represented as the first The intercept of each filter, This is represented as a convolution operation.

8. The load aggregation prediction method of the integrated framework according to claim 7, characterized in that, In LSTM network models, an attention mechanism is introduced into the hidden states of the hidden layers to assign different attention weights to the hidden states of each hidden layer. This includes the following steps: For a given sequence of hidden states in a hidden layer, calculate the attention score for each hidden layer. Attention score is determined as follows: ； In the formula, Let the attention score of the m-th hidden layer be denoted as . This is represented as the m-th hidden layer. Represented as an attention weight matrix, This is represented as the bias vector parameter of the attention mechanism. This is represented as tangent; Based on the attention score, obtain the attention weight probability; The attention weight probability is obtained using the following formula: ； In the formula, Let be the attention weight probability of the m-th hidden layer, and exp be the exponential function. Represented as the maximum hidden layer; Based on the attention weight probability, the attention weights of each hidden layer are weighted to obtain the weighted hidden layer; The weighted hidden layer is obtained using the following formula: ； In the formula, represents the weighted hidden layer.

9. A load aggregation prediction system with an integrated framework, characterized in that, The load aggregation prediction method using the integrated framework of any one of claims 1-8 includes: The prediction module is used to perform scheduling by applying the prediction results of the trained load aggregation prediction model. The module is used to construct the load aggregation prediction model, including: acquiring historical load aggregation data; analyzing the external influencing factors of the historical load aggregation data and screening out important external influencing factors; applying the K-shape algorithm to cluster the historical load aggregation data into several clusters; constructing the load aggregation prediction model, including the CNN-LSTM-ATTENTION prediction model and the GBDT prediction model; using the CNN-LSTM-ATTENTION prediction model to predict the load for each cluster, based on the screened important external influencing factors and the several clusters; and applying the GBDT prediction model to integrate the load prediction results of each cluster to obtain the load aggregation prediction result.

Images