Method and device for establishing short-term probability prediction model of integrated energy load aggregation

Abstract

The application discloses a kind of comprehensive energy load aggregate short-term probability prediction model establishing method and device.The method includes: the historical data of comprehensive energy load aggregate collected is preprocessed and dimensionality reduction;Grouping is carried out to the data after dimensionality reduction using affinity propagation clustering algorithm based on comprehensive similarity, and the centroid feature of each group is extracted as representative load feature input prediction model;The short-term probability prediction model of comprehensive energy load aggregate that fusion Copula self-adapting correlation analysis, cross feature-time graph neural network and mixed density network is constructed.The precision of short-term probability prediction of comprehensive energy load aggregate is effectively improved by dynamic coupling relationship mining and spatiotemporal feature joint learning mechanism.

Description

Technical Field

[0001] This invention relates to the field of energy big data analysis technology, and in particular to a method and apparatus for establishing a short-term probability prediction model for integrated energy load aggregates. Background Technology

[0002] With the rapid development of integrated energy systems, traditional demand response is gradually shifting towards integrated demand response, making integrated energy load aggregations increasingly important in supply and demand coordination. Accurate forecasting of integrated energy load aggregations is not only crucial for the stable operation of the system but also plays a vital role in market planning and scheduling optimization. Especially in the context of multi-energy complementarity and regional collaborative optimization, load forecasting research has expanded from single energy sources or individual users to the load aggregation level. How to improve forecasting accuracy under complex coupling relationships and high uncertainty has become a critical issue that urgently needs to be addressed.

[0003] However, in the process of researching and practicing existing technologies, the inventors of this invention have found that most current research focuses on the prediction of single energy load aggregates. Although the aggregation effect of load aggregates is considered, the complex coupling relationships and uncertainties between multiple energy loads are not addressed. Therefore, there is an urgent need to propose a short-term probabilistic prediction method and device for integrated energy load aggregates to more effectively address their complex coupling relationships and uncertainties. Summary of the Invention

[0004] The purpose of this invention is to provide a short-term probabilistic prediction model for integrated energy load aggregates, so as to solve the problems of insufficient ability to characterize the complex coupling relationship between integrated energy loads, the dynamic correlation with external influencing factors, and the uncertainty of the correlation between loads in each group in the existing technology, thereby improving the overall prediction performance.

[0005] This invention provides a method for establishing a short-term probabilistic prediction model for integrated energy load aggregates, the method comprising:

[0006] The collected historical data of the integrated energy load aggregate are preprocessed and dimensionality reduced;

[0007] The affinity propagation clustering algorithm based on comprehensive similarity is used to group the dimensionality-reduced data, and the centroid features of each group are extracted as representative load features to be input into the prediction model.

[0008] A short-term probability prediction model for integrated energy load aggregates is constructed, integrating Copula adaptive correlation analysis, cross-feature-time graph neural networks, and hybrid density networks. This model is used for short-term probability prediction of integrated energy load aggregates.

[0009] Furthermore, the preprocessing and dimensionality reduction of the collected historical data of the integrated energy load aggregate includes:

[0010] The collected historical data on comprehensive energy load is preprocessed, including data cleaning and normalization.

[0011] The preprocessed data is dimensionality reduced using the three-segment maximum triangular algorithm.

[0012] Furthermore, the step of grouping the dimensionality-reduced data using an affinity propagation clustering algorithm based on comprehensive similarity, and extracting the centroid features of each group as representative load features input to the prediction model, includes:

[0013] To address the problem that traditional clustering methods, which rely solely on Euclidean distance to measure similarity, fail to comprehensively reflect the overall similarity of integrated energy users across various energy load curves (electricity, cooling, heating, gas, etc.), a comprehensive similarity metric that integrates Euclidean distance and cosine distance is proposed. The formula for calculating the comprehensive similarity distance is as follows:

[0014]

[0015] In the formula, d e d represents the Euclidean distance. cos The distance represents the cosine distance, and n represents the number of energy load types included in the user, taking into account four dimensions: electrical load, cooling load, heating load, and gas load.

[0016] Based on the aforementioned similarity measurement method, a comprehensive similarity matrix among users is constructed. Using this similarity matrix as input, an affinity propagation clustering algorithm is executed, and the centroids of each cluster are extracted as representative load features input to the prediction model.

[0017] Furthermore, the construction of a comprehensive short-term probabilistic prediction model for energy load aggregates, integrating Copula adaptive correlation analysis, cross-feature-time graph neural networks, and hybrid density networks, includes:

[0018] Copula adaptive correlation analysis was used to explore the complex dynamic correlations between integrated energy loads, between loads and external influencing factors, and between different groups;

[0019] We utilize cross-feature-time graph neural networks to jointly model the complex coupling relationships between feature dimension, time dimension and group dimension, and combine them with Bi-SRU for multivariate time series modeling;

[0020] The input features are nonlinearly mapped by a mixture density network to estimate the mean, standard deviation and corresponding weight of each mixture component, and output the combined mixture probability density distribution.

[0021] Furthermore, the method of using Copula adaptive correlation analysis to uncover the complex dynamic correlations between integrated energy loads, between loads and external influencing factors, and between different groups includes:

[0022] A directed graph structure based on load characteristics is constructed, adjacency relationships are defined to characterize the correlation between different load variables, the Copula method is used to model the joint distribution between node pairs, and significant edges are retained based on the correlation threshold.

[0023] A cross-feature map is constructed, and the Copula function is used to model the joint distribution between load features and external factors. An initial feature correlation matrix is ​​constructed, and the weight matrix for graph construction is obtained through regularization. The specific formula is as follows:

[0024]

[0025] In the formula, E F [i,j] represents feature nodes and The relative weights between them.

[0026] A cross-time graph is constructed. Dominant period information is obtained through frequency domain transformation and multi-scale sampling is performed. An initial time correlation matrix is ​​constructed using the Copula function, and a weight matrix for graph construction is obtained through regularization. The specific formula is as follows:

[0027]

[0028] In the formula, E T [i,j] are time nodes. and The relative weights between them.

[0029] A group association graph is constructed, using the centroids of each clustered group as graph nodes. The correlation between different groups is modeled based on the Copula function, and a weight matrix for graph construction is obtained through regularization. The specific formula is as follows:

[0030]

[0031] In the formula, E C [i,j] is the centroid node and The relative weights between them.

[0032] Furthermore, the method of jointly modeling the complex coupling relationships between feature dimensions, time dimensions, and group dimensions using cross-feature-time graph neural networks, and combining this with Bi-SRU for multivariate time series modeling, includes:

[0033] By designing message passing mechanisms in the dimensions of features, time, and groups, the dynamic interaction characteristics of load aggregates in terms of time sequence, features, and groups are modeled to achieve adaptive learning of multidimensional dynamic dependencies.

[0034] In terms of features, the nonlinear interaction between feature nodes is captured through the message passing mechanism in the graph neural network. For feature node v, its message passing process can be represented as:

[0035]

[0036] In the formula, σ is the activation function. It is a learnable parameter matrix. This represents the feature of feature node v at the k-th layer. and Let represent the sets of neighbors that are homogeneous and heterogeneous with the feature node v, respectively, and Norm represents the normalization operation.

[0037] In the time dimension, a stacked K-layer graph neural network structure is used to characterize the dynamic dependencies between time points. The specific process is as follows:

[0038]

[0039] In the formula, σ is the activation function, and W i (k) It is a learnable parameter matrix. N represents the feature of time node i in the kth layer. i Let be the set of time points adjacent to time point i, and Norm is the normalization operation.

[0040] At the group level, considering the complexity of load types within a load aggregate and the heterogeneity between groups, a separate message passing mechanism is defined for each group. Assuming a group node is represented as , its message passing process can be represented as:

[0041]

[0042] In the formula, σ is the activation function. It is a learnable parameter matrix. N represents the feature of group node c at the kth layer. c Let represent the neighbor set of group node c, and Norm is the normalization operation.

[0043] After completing the message passing across the three dimensions of feature, time, and group, the feature-time graph neural network generates multi-step prediction results through the prediction model, as shown in the following formula:

[0044] Y = f P (H final )

[0045] In the formula, f P (·) represents the prediction model; here, the Bi-SRU model is used, H final This represents the combined features output by the feature-time graph neural network.

[0046] Furthermore, the step of performing a nonlinear transformation on the input features through a hybrid density network, estimating the mean, standard deviation, and weight parameters of each basis distribution, and outputting the corresponding hybrid probability density distribution includes:

[0047] Assuming that each component in the mixture follows a Gaussian distribution, the mixture weights are modeled as a class distribution, satisfying the weight normalization condition. The mixture density network outputs the mean, standard deviation, and weight of each component, and these are weighted and summed to form the mixture probability distribution function.

[0048] This invention provides an apparatus for establishing a short-term probabilistic prediction model for integrated energy load aggregates, comprising:

[0049] The preprocessing module is used to preprocess and reduce the dimensionality of the collected historical data of the integrated energy load aggregate.

[0050] The clustering analysis module is used to group the dimensionality-reduced data using an affinity propagation clustering algorithm based on comprehensive similarity, and extract the centroid features of each group as representative load features to input into the prediction model.

[0051] A prediction model module is established to construct a short-term probability prediction model for integrated energy load aggregates, which integrates Copula adaptive correlation analysis, cross-feature-time graph neural networks, and hybrid density networks. This short-term probability prediction model for integrated energy load aggregates is used for short-term probability prediction of integrated energy load aggregates.

[0052] The present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described method for establishing a short-term probabilistic prediction model for integrated energy load aggregates.

[0053] The present invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described method for establishing a short-term probability prediction model for integrated energy load aggregates.

[0054] Compared with existing technologies, this invention fully considers the complex coupling relationships, uncertainties, and dynamic correlations with external factors among multiple energy loads in the short-term probabilistic prediction of integrated energy load aggregates. By introducing the Copula adaptive correlation analysis method, it effectively explores the nonlinear dependency structures between different load types, between loads and external influencing factors, and between different groups. Combined with a cross-feature-time graph neural network, it realizes the modeling of multidimensional coupling relationships between feature dimensions, time dimensions, and group dimensions, and uses a Bi-SRU structure to improve the efficiency and performance of multivariate time series modeling. At the same time, by using a hybrid density network to characterize the probability distribution of the prediction results, it enhances the model's expressive power and the ability to represent prediction uncertainties, thereby significantly improving prediction accuracy and robustness in complex load environments, and possessing stronger engineering adaptability and promotional value.

[0055] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0056] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0057] Figure 1 This is a flowchart of the method for establishing a short-term probability prediction model for integrated energy load aggregates according to an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the device for establishing a short-term probability prediction model for a comprehensive energy load aggregate according to Embodiment 1 of the present invention.

[0059] Figure 3 This is a schematic diagram of the device for establishing a short-term probability prediction model for a comprehensive energy load aggregate according to Embodiment 2 of the present invention. Detailed Implementation

[0060] To overcome the shortcomings of existing technologies, this invention proposes a method and apparatus for short-term probability prediction of integrated energy load aggregates, which is used to perform dimensionality reduction and cluster analysis on integrated energy load aggregate data and establish a short-term probability prediction model for integrated energy load aggregates.

[0061] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. It should be understood that the described embodiments are only a part of the present invention, and not all embodiments. For those skilled in the art, various other embodiments made without departing from the concept of the present invention should be included within the protection scope of the present invention.

[0062] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0063] Furthermore, 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 indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0064] The technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0065] This invention provides a method for short-term probabilistic prediction of integrated energy load aggregates based on an adaptive Copula cross-feature-time graph hybrid density network model. The flowchart of the method is shown below. Figure 1 As shown, the specific processing includes the following:

[0066] Step 1 (S101) involves preprocessing and dimensionality reduction of the collected historical data of the integrated energy load aggregate. Specifically, the following processing methods can be employed:

[0067] Step 1.1 involves preprocessing the collected load data. The specific process is as follows:

[0068] (1) Clean the user load data to remove outliers and missing data;

[0069] (2) Normalize the user load data to ensure that the dimensions of each feature are consistent.

[0070] Step 1.2: The preprocessed data is then subjected to dimensionality reduction using the three-segment maximum triangulation algorithm. The specific process is as follows:

[0071] The original time series data is divided into several equal segments, and in each segment, the most important point (denoted as point F) is selected based on the largest effective area to represent all points in the current segment, so as to achieve dimensionality compression.

[0072] Step 2 (S102) involves using an affinity propagation clustering algorithm based on comprehensive similarity to group the dimensionality-reduced data, extracting the centroid features of each group as representative load features to input into the prediction model. Specifically, the following processing can be performed:

[0073] Step 2.1: Calculate the similarity between any two user energy consumption data sequences using a comprehensive similarity metric. The comprehensive similarity includes the following two aspects:

[0074] a. Distance metric: Euclidean distance is used to calculate the distance between energy consumption data sequences of different users;

[0075] b. Shape fluctuation metric: Cosine distance is used to measure the shape similarity of energy consumption data sequences among different users.

[0076] Combining the Euclidean distance and cosine distance mentioned above, the comprehensive similarity distance formula is as follows:

[0077]

[0078] In the formula, n represents the number of energy load types included in a single user, considering four dimensions: electrical load, cooling load, heating load, and gas load. d e For Euclidean distance, d cos The cosine distance is used. An N*N comprehensive similarity matrix S is obtained by calculation (where N is the total number of particles), and s(j,j) is set as the median of the similarity matrix.

[0079] Step 2.2 involves clustering via message passing. During this process, two types of messages are exchanged: attraction information and attribution information. Attraction information r t (i,j) reflects the suitability of data point j as the cluster center of data point i at time t, and represents the message from i to j; the attribution information a t (i,j) reflects the suitability of data point i for selecting data point j as the cluster center at time t, and represents the message from j to i. Based on the above two types of information, the message passing steps are as follows:

[0080] Attracting information r t+1 (i,j) is iterated according to the following formula to calculate the attractiveness of data point i as the cluster center of data point j, taking into account the influence of all other cluster centers.

[0081] r t+1 (i,j)=s(i,j)-max{a t (i,j′)+s(i,j′)},stj′∈{1,2,…,N},j′≠j

[0082] In the formula, s(i,j) represents the similarity between data point i and data point j. t (i,j′) represents the cluster center affiliation information of data point i at time t, where data point j′ is chosen as the cluster center. s(i,j′) represents the similarity between data point i and data point j′.

[0083] Attribution information a t+1 (i,j) is iterated according to the following formula to determine whether data point i selects data point j as the cluster center.

[0084]

[0085] a t+1 (j,j)=∑ i′≠j max{0,r t (i′,j)}

[0086] In the formula, r t (j,j) represents the attraction information between data point j at time t and itself, r t (i′,j) represents the attraction information between data point i′ and data point j at time t. During the iteration process of the algorithm, if the changes in attraction information and attribution information tend to stabilize, i.e., no longer change significantly, or if the number of iterations exceeds the preset maximum number, the algorithm will terminate.

[0087] To avoid oscillations, the affinity propagation clustering algorithm introduces a decay coefficient λ (ranging from 0 to 1) during information updates. The updated value of each piece of information is a weighted average of the previous iteration value and the current iteration value, where the weight of the previous iteration value is the decay coefficient λ, and the weight of the current iteration value is (1-λ). Therefore, in the (t+1)th iteration, the attracted information r... t+1 (i,j) and attribution information a t+1 The updated values ​​for (i,j) are as follows:

[0088] r′ t+1 (i,j)=(1-λ)r t+1 (i,j)+λr t (i,j)

[0089] a′ t+1 (i,j)=(1-λ)a t+1 (i,j)+λa t (i,j)

[0090] Step 3 (S103) involves constructing a comprehensive short-term probabilistic prediction model for energy load aggregates, integrating Copula adaptive correlation analysis, cross-feature-time graph neural networks, and hybrid density networks. Specifically, this can be achieved through:

[0091] Step 3.1: Through Copula adaptive correlation analysis, the complex dynamic correlations between integrated energy loads, between loads and external influencing factors, and between different groups are explored. The specific process is as follows:

[0092] Assume the input matrix composed of multivariate time series is χ=[X 1 ,X 2 ,...,X L ],in, D represents the number of variables. Construct a directed graph structure whose topological relationships are represented by an adjacency matrix A, where each element a... ij The definition is as follows:

[0093]

[0094] In the formula, w i and w j χ represents the features of the i-th and j-th rows, and g is the threshold for judging the significance of correlation.

[0095] The construction of a directed graph G can be represented as:

[0096] G = (V, E)

[0097] In the formula, V = {v1, v2, ..., v} n} is a node set containing n nodes. Denotes an ordered set of edges, consisting of ordered pairs (v i ,v j Composed of, (v i ,v j ) indicates from node v i Pointing to node v j A directed edge.

[0098] The cross feature map is represented as G F =(V F E F ),in D is the set of feature nodes. i For the number of features, Let represent the j-th feature node. The Copula function is used to model the correlation between any two nodes, constructing an initial feature correlation matrix R. F :

[0099] R F (i,j)=Copula(Z :,i Z :,j ), i,j∈{1,2,…,D i}

[0100] The original correlation matrix is ​​Softmax normalized to obtain the initialized feature correlation matrix.

[0101]

[0102] set up Represents nodes Related positive neighbor set, The set of negative neighbors is represented, and the corresponding homogeneous and heterogeneous correlation weights are renormalized as follows:

[0103]

[0104] In the formula, E F [i,j] is and The correlation weights between edges are determined. This process retains homogeneous and heterogeneous edges while eliminating other types of edges. The weights of homogeneous edges are positively correlated with the correlation score, while the weights of heterogeneous edges are negatively correlated with the correlation score. Finally, after normalization, the decomposed homogeneous and heterogeneous correlation cross-feature maps are constructed.

[0105] Cross-time plot is represented as G T =(V T E T ),in It is a set of time nodes, L h It refers to a time range. This represents the i-th time node. This represents the correlation weight matrix between time nodes. To effectively extract the latent features of multi-feature time series at different time scales and reduce the interference of noise on the correlation modeling between time nodes, a combination of frequency domain analysis, Copula correlation modeling, and graph structure optimization strategies is used to dynamically adjust the adjacency relationships of time nodes, thereby constructing the adjacency matrix of the time graph.

[0106] The Fast Fourier Transform is used to extract the frequency domain features of the input time series, the frequency domain amplitude M of each feature is calculated, and the mean value is taken along the feature dimension.

[0107]

[0108] In the formula, Amp(·) represents the amplitude calculation. Let represent the average amplitude at each frequency, and Q be the number of frequency components. Select the 's' frequency components with the largest amplitudes; their corresponding period lengths are... Downsampling of the original sequence:

[0109] X (s) =AvgPool(X,kernel=q) s stride = q s )

[0110] The downsampled sequence is in This is the length of the downsampled sequence. The downsampled sequences from all time scales are concatenated to generate a multi-scale time feature sequence.

[0111] X′=Concat(X (1) ,X (2) ,...,X (s) )

[0112] In the formula, This is the total length after concatenation. The correlation between time nodes of X′ is calculated using the Copula method to construct the initial time correlation matrix R. T :

[0113] R T (i,j)=Copula(Z′ i ,Z′ j ), i,j∈{1,2,...,L′}

[0114] E is obtained from the original Copula correlation matrix. T initial value

[0115]

[0116] In the formula, R T This is the original Copula time correlation matrix. ReLU(·) is the regularization function, and Softmax(·) ensures that the sum of the weights of all nodes related to a specific time node is 1. To capture time trend features, the edge connections between time nodes and their adjacent preceding and following nodes are preserved. The trend neighbor set, represented as a time node, is defined as follows:

[0117]

[0118] In the formula, The trend neighbor set consists of adjacent time nodes sharing the same scale (i.e., |ij|≤1). The final relevance weights are renormalized as follows:

[0119]

[0120] In the formula, E T [i,j] is and The correlation weights between nodes are determined. This step removes connections with insignificant correlations and retains a limited set of neighboring nodes for each node. Subsequently, the retained relevant weights are renormalized to complete the construction of the cross-time correlation graph.

[0121] The intergroup association diagram is represented by G. C =(V C E C ),in It is the set of centroid nodes of the clusters, where N represents the number of clusters. Let j be the j-th centroid node. This represents the correlation weight matrix between centroid nodes. The Copula method is used to calculate the correlation between centroid nodes, constructing an initial inter-group correlation matrix R. C :

[0122] R C (i,j)=Copula(C i C j ), i,j∈{1,2,…,N}

[0123] E is obtained from the original Copula correlation matrix. C initial value

[0124]

[0125] use Represents nodes Related positive neighbor set, The set of negative neighbors is represented, and the corresponding homogeneous and heterogeneous correlation weights are renormalized as follows:

[0126]

[0127] In the formula, E C [i,j] is and The weights of the edges were determined. During the decoupling process, only the weights of homogeneous and heterogeneous edges were retained and normalized respectively. Finally, a correlation graph between homogeneous and heterogeneous groups was constructed, effectively characterizing the complex correlations between groups.

[0128] Step 3.2 involves jointly modeling the complex coupling relationships between the feature dimension, time dimension, and group dimension using a cross-feature-time graph neural network, and combining this with Bi-SRU for multivariate time series modeling. Specifically, this includes:

[0129] By designing a message passing mechanism that integrates features, time, and group dimensions, a comprehensive expression of the dynamic interaction characteristics of the load aggregate in terms of time series, features, and groups is achieved. This enables adaptive learning of the multi-level complex dependency structure between various dimensions, thereby enhancing prediction performance.

[0130] In terms of features, the nonlinear interactions between feature nodes are captured through the message passing mechanism in graph neural networks. For feature node v, its message passing process can be represented as:

[0131]

[0132] In the formula, σ is the activation function. It is a learnable parameter matrix. This represents the feature of feature node v at the k-th layer. and Let represent the sets of neighbors that are homogeneous and heterogeneous with the feature node v, respectively, and Norm denotes the normalization operation. This mechanism achieves effective modeling of complex relationships between features by aggregating the features of positive and negative neighbor nodes.

[0133] In the time dimension, a stacked K-layer graph neural network structure is used to characterize the dynamic dependencies between time points. The specific process is as follows:

[0134]

[0135] In the formula, σ is the activation function, and W i (k) It is a learnable parameter matrix. N represents the feature of time node i in the kth layer. i Let be the set of time points adjacent to time point i, and Norm be the normalization operation. By aggregating the feature information of time neighbor nodes, the model updates the time node features layer by layer, and finally outputs the result of K layers.

[0136] At the group level, considering the complex load types within load aggregates and the significant heterogeneity between groups, the feature-time graph neural network defines a separate message passing mechanism for each group. Assuming a group node is represented as , its message passing process can be represented as:

[0137]

[0138] In the formula, σ is the activation function. It is a learnable parameter matrix. N represents the feature of group node c at the kth layer. c Let represent the neighbor set of node c in group , and Norm be the normalization operation. Through the interactions between groups, the feature-time graph neural network can effectively capture the correlation between loads in different groups, thereby improving the overall prediction performance of the load aggregate.

[0139] After completing the message passing across the three dimensions of feature, time, and group, the feature-time graph neural network generates multi-step prediction results through the prediction model, as shown in the following formula:

[0140] Y = f P (H final )

[0141] In the formula, f P (·) represents the prediction model; here, the Bi-SRU model is used, H final This represents the comprehensive features output by the feature-time graph neural network. By performing temporal dependency modeling on the hidden layer features, intermediate prediction results of the graph neural network are generated, and then probabilistic prediction modeling of the load is carried out.

[0142] The forward propagation calculation formula for the Bi-SRU model is as follows:

[0143] f t =σ(W f x t +V f ⊙c t-1 +b f )

[0144] r t =σ(W r x t +V r ⊙c t-1 +b r )

[0145] c t =f t ⊙c t-1 +(1-f t )⊙(W c x t )

[0146] h t =r t ⊙c t +(1-r t )⊙x t

[0147] In the formula, f t The output of the forget gate at time t represents the information c from the previous time step. t-1Whether to continue propagating, partially propagate, or remove; σ represents the Sigmoid activation function in the model; x t Represents the input at time t; r t This indicates that the gate output is reset at time t, determining the information c from the previous time step. t-1 How much is written to the current candidate set; c t h represents the new information value generated at time t; t W represents the state of the hidden layer at time t; f W r V f V r W c b represents the corresponding weighting coefficient. f b r The corresponding bias term; the symbol ⊙ represents element-wise multiplication, also known as the Hadamard product.

[0148] The Bi-SRU network model processes information in two directions: the forward SRU unit generates the preceding information, i.e., the hidden state is produced. in, The SRU unit generates subsequent information, i.e., it generates a hidden state. in, Connecting these two pieces of information (forward information and backward information) produces the following joint information:

[0149]

[0150] In the formula, and These represent the hidden states of the forward and backward SRU cells at time t, respectively.

[0151] Step 3.3 involves performing a nonlinear mapping on the input features using a mixture density network to estimate the mean, standard deviation, and corresponding weight of each mixture component, and outputting the combined mixture probability density distribution. Specifically, this includes:

[0152] Assuming the conditional distribution of the mixture components comes from a family of Gaussian distributions, and the mixture weights δ k ∈[0,1] M The model is represented as a class distribution with M possible states. δ k satisfy Posterior distribution P(δ) k |k) can be embedded through deterministic routing. k To calculate:

[0153]

[0154] In the formula, It is a linear model f δThe softmax function τ forms a nonlinear transformation. The parameters output by the hybrid density network include the mean, standard deviation, and weights of each basic distribution. By weighted summing of these distributions, a hybrid probability distribution function is finally formed, as shown below:

[0155]

[0156] In the formula, μ m (k) and σ m (k) represent the mean and variance parameters of the m-th mixture component, respectively. The parameters of the output conditional density function can be further expressed as:

[0157]

[0158] In the formula, It is a linear model. Hybrid density networks can effectively handle uncertainties in load forecasting by providing decision-makers with more reliable load change expectations by giving a probability density function for each forecast value.

[0159] In other words, the specific process of short-term probability prediction for integrated energy load aggregates is as follows:

[0160] (1) Prediction process

[0161] The dataset selected in this embodiment comes from the US Building Data Set, which contains electricity, cooling, heating, and gas load data for 12,004 residents in Arizona, covering the period from 00:00 on January 1, 2018 to 00:00 on January 1, 2019. The dataset also includes hourly weather data for the corresponding time, covering external factors such as temperature, relative humidity, wind speed, wind direction, global horizontal radiation, direct normal radiation, and diffuse horizontal radiation, which may affect the load. Given the limited data volume, to fully utilize the data and optimize model performance, this study divides the data for each season into training, validation, and test sets in a 7:2:1 ratio. Regarding the prediction model parameters, dropout is 0.01, the initial learning rate is 0.0001, the number of iterations is 100, and the batch size is 80. The model contains two encoder layers and one decoder layer. Each layer has 8 heads for multi-head attention, and the hidden layer dimension is 64.

[0162] (2) Predictive performance evaluation

[0163] To evaluate the prediction performance, the pinball loss function and the Winkler score were selected to evaluate the effectiveness of the probabilistic prediction model.

[0164] The formula for calculating pinball loss is as follows:

[0165]

[0166] In the formula, Let y represent the q-th quantile of the load forecast for the i-th sample. i This is the actual observed load value. The smaller the bouncing loss fraction, the better the prediction effect.

[0167] The Winkler score is calculated using the following formula:

[0168]

[0169] In the formula, δ i =U i -L i U represents the prediction interval width for the i-th sample. i and L i These are the upper and lower bounds, respectively. When the actual value falls within the prediction range, the Winkler score depends only on the response width; otherwise, a penalty term is added to reflect the deviation from covering the actual value. The smaller the Winkler score, the better the prediction performance.

[0170] In this embodiment, the proposed method (denoted as ACCFTG-MDN) is compared with the prediction results of six other prediction models for the next hour: the adaptive Copula cross-feature-time graph quantile model (denoted as ACCFTG-MQ), the adaptive Copula cross-feature-time graph Gaussian model (denoted as ACCFTG-Gaussian), the adaptive Copula cross-feature-time graph Bayesian model (denoted as ACCFTG-Bayes), the adaptive Copula cross-feature-time graph convolutional hybrid density network model (denoted as ACCFTGCN-MDN), the adaptive Copula cross-feature-time graph cyclic hybrid density network model (denoted as ACCFTGRN-MDN), and the adaptive Copula cross-feature-time convolutional attention mechanism bidirectional gated cyclic hybrid density network model (denoted as ACC-CNN-BiGRU-ATT-MDN). The analysis results are shown in Table 1.

[0171] Table 1

[0172]

[0173]

[0174] In summary, this invention proposes a method for short-term probability prediction of integrated energy load aggregates. It uses a three-segment maximum triangulation algorithm for dimensionality reduction and an affinity propagation clustering algorithm based on comprehensive similarity for load aggregate grouping and centroid feature extraction. By constructing a model that integrates Copula adaptive correlation analysis, feature-time graph neural network, and hybrid density network, the accuracy of short-term probability prediction of integrated energy load aggregates is effectively improved.

[0175] Device Example 1

[0176] According to an embodiment of the present invention, an apparatus for establishing a short-term probabilistic prediction model for integrated energy load aggregates is provided. Figure 2 This is a schematic diagram of the device for establishing a short-term probability prediction model for integrated energy load aggregates according to an embodiment of the present invention, as shown below. Figure 2 As shown, the apparatus for establishing a short-term probability prediction model for a comprehensive energy load aggregate according to an embodiment of the present invention specifically includes: a preprocessing module, a cluster analysis module, and a prediction model establishment module, thereby obtaining the result of the short-term probability prediction of the comprehensive energy load aggregate. Specifically:

[0177] The preprocessing module 60 is used to preprocess and reduce the dimensionality of the collected historical data of the integrated energy load aggregate. Specifically, the preprocessing module 60 is used for:

[0178] The collected historical data on comprehensive energy load is preprocessed, including data cleaning and normalization.

[0179] The preprocessed data is dimensionality reduced using the three-segment maximum triangular algorithm.

[0180] The clustering analysis module 62 is used to group the dimensionality-reduced data using an affinity propagation clustering algorithm based on comprehensive similarity, and extract the centroid features of each group as representative load features to input into the prediction model. Specifically, the clustering analysis module 62 is used for:

[0181] Calculate the comprehensive similarity measure between any two user energy consumption data sequences.

[0182] Clustering is achieved using the affinity propagation clustering algorithm.

[0183] Furthermore, the calculation of the comprehensive similarity measure between any two user energy consumption data sequences specifically includes:

[0184] The overall similarity includes the following two aspects:

[0185] a. Distance metric: Euclidean distance is used to calculate the distance between energy consumption data sequences of different users;

[0186] b. Shape fluctuation metric: Cosine distance is used to measure the shape similarity of energy consumption data sequences among different users.

[0187] Combining the Euclidean distance and cosine distance mentioned above, the comprehensive similarity distance formula is as follows:

[0188]

[0189] In the formula, n represents the number of energy load types included in a single user, considering four dimensions: electrical load, cooling load, heating load, and gas load. d e For Euclidean distance, d cos The cosine distance is used. An N*N comprehensive similarity matrix S is obtained by calculation (where N is the total number of particles), and s(j,j) is set as the median of the similarity matrix.

[0190] Furthermore, the clustering using the affinity propagation clustering algorithm specifically includes:

[0191] Clustering is achieved through message passing. In this process, two types of messages are exchanged: attraction information and attribution information. Attraction information r t (i,j) reflects the suitability of data point j as the cluster center of data point i at time t, and represents the message from i to j; the attribution information a t (i,j) reflects the suitability of data point i for selecting data point j as the cluster center at time t, and represents the message from j to i. Based on the above two types of information, the message passing steps are as follows:

[0192] Attracting information r t+1 (i,j) is iterated according to the following formula to calculate the attractiveness of data point i as the cluster center of data point j, taking into account the influence of all other cluster centers.

[0193] r t+1 (i,j)=s(i,j)-max{a t (i,j′)+s(i,j′)},stj′∈{1,2,…,N},j′≠j

[0194] In the formula, s(i,j) represents the similarity between data point i and data point j. t (i,j′) represents the cluster center affiliation information of data point i at time t, where data point j′ is chosen as the cluster center. s(i,j′) represents the similarity between data point i and data point j′.

[0195] Attribution information a t+1 (i,j) is iterated according to the following formula to determine whether data point i selects data point j as the cluster center.

[0196]

[0197] a t+1 (j,j)=Σ i′≠j max{0,r t (i′,j)}

[0198] In the formula, r t (j,j) represents the attraction information between data point j at time t and itself, r t (i′,j) represents the attraction information between data point i′ and data point j at time t. During the iteration process of the algorithm, if the changes in attraction information and attribution information tend to stabilize, i.e., no longer change significantly, or if the number of iterations exceeds the preset maximum number, the algorithm will terminate.

[0199] To avoid oscillations, the affinity propagation clustering algorithm introduces a decay coefficient λ (ranging from 0 to 1) during information updates. The updated value of each piece of information is a weighted average of the previous iteration value and the current iteration value, where the weight of the previous iteration value is the decay coefficient λ, and the weight of the current iteration value is (1-λ). Therefore, in the (t+1)th iteration, the attracted information r... t+1 (i,j) and attribution information a t+1 The updated values ​​for (i,j) are as follows:

[0200] r′ t+1 (i,j)=(1-λ)r t+1 (i,j)+λr t (i,j)

[0201] a′ t+1 (i,j)=(1-λ)a t+1 (i,j)+λa t (i,j)

[0202] The prediction model building module 64 is used to construct a short-term probability prediction model for the integrated energy load aggregate, which integrates Copula adaptive correlation analysis, cross-feature-time graph neural network, and hybrid density network. This short-term probability prediction model is used to perform short-term probability prediction of the integrated energy load aggregate. Specifically, the prediction model building module 64 is used for:

[0203] Copula adaptive correlation analysis was used to explore the complex dynamic correlations between integrated energy loads, between loads and external influencing factors, and between different groups;

[0204] We utilize cross-feature-time graph neural networks to jointly model the complex coupling relationships between feature dimension, time dimension and group dimension, and combine them with Bi-SRU for multivariate time series modeling;

[0205] The input features are nonlinearly mapped by a mixture density network to estimate the mean, standard deviation and corresponding weight of each mixture component, and output the combined mixture probability density distribution.

[0206] Furthermore, the complex dynamic correlations between integrated energy loads, between loads and external influencing factors, and between different groups are explored through Copula adaptive correlation analysis. The specific process is as follows:

[0207] Assume the input matrix composed of multivariate time series is χ=[X 1 ,X 2 ,...,X L ],in, D represents the number of variables. Construct a directed graph structure whose topological relationships are represented by an adjacency matrix A, where each element a... ij The definition is as follows:

[0208]

[0209] In the formula, w i and w j χ represents the features of the i-th and j-th rows, and g is the threshold for judging the significance of correlation.

[0210] The construction of a directed graph G can be represented as:

[0211] G = (V, E)

[0212] In the formula, V = {v1, v2, ..., v} n} is a node set containing n nodes. Denotes an ordered set of edges, consisting of ordered pairs (v i ,v j Composed of, (v i ,v j ) indicates from node v i Pointing to node v j A directed edge.

[0213] The cross feature map is represented as G F =(V F E F ),in D is the set of feature nodes. i For the number of features, Let represent the j-th feature node. The Copula function is used to model the correlation between any two nodes, constructing an initial feature correlation matrix R. F :

[0214] R F (i,j)=Copula(Z :,i Z :,j), i,j∈{1,2,...,D i}

[0215] The original correlation matrix is ​​Softmax normalized to obtain the initialized feature correlation matrix.

[0216]

[0217] set up Represents nodes Related positive neighbor set, The set of negative neighbors is represented, and the corresponding homogeneous and heterogeneous correlation weights are renormalized as follows:

[0218]

[0219] In the formula, E F [i,j] is and The correlation weights between edges are determined. This process retains homogeneous and heterogeneous edges while eliminating other types of edges. The weights of homogeneous edges are positively correlated with the correlation score, while the weights of heterogeneous edges are negatively correlated with the correlation score. Finally, after normalization, the decomposed homogeneous and heterogeneous correlation cross-feature maps are constructed.

[0220] Cross-time plot is represented as G T =(V T E T ),in It is a set of time nodes, L h It refers to a time range. This represents the i-th time node. This represents the correlation weight matrix between time nodes. To effectively extract the latent features of multi-feature time series at different time scales and reduce the interference of noise on the correlation modeling between time nodes, a combination of frequency domain analysis, Copula correlation modeling, and graph structure optimization strategies is used to dynamically adjust the adjacency relationships of time nodes, thereby constructing the adjacency matrix of the time graph.

[0221] The Fast Fourier Transform is used to extract the frequency domain features of the input time series, the frequency domain amplitude M of each feature is calculated, and the mean value is taken along the feature dimension.

[0222]

[0223] In the formula, Amp(·) represents the amplitude calculation. Let represent the average amplitude at each frequency, and Q be the number of frequency components. Select the 's' frequency components with the largest amplitudes; their corresponding period lengths are... Downsampling of the original sequence:

[0224] X(s) =AvgPool(X,kernel=q) s stride = q s )

[0225] The downsampled sequence is in This is the length of the downsampled sequence. The downsampled sequences from all time scales are concatenated to generate a multi-scale time feature sequence.

[0226] X′=Concat(X (1) ,X (2) ,...,X (s) )

[0227] In the formula, This is the total length after concatenation. The correlation between time nodes of X′ is calculated using the Copula method to construct the initial time correlation matrix R. T :

[0228] R T (i,j)=Copula(Z′ i ,Z′ j ), i,j∈{1,2,...,L′}

[0229] E is obtained from the original Copula correlation matrix. T initial value

[0230]

[0231] In the formula, R T This is the original Copula time correlation matrix. ReLU(·) is the regularization function, and Softmax(·) ensures that the sum of the weights of all nodes related to a specific time node is 1. To capture time trend features, the edge connections between time nodes and their adjacent preceding and following nodes are preserved. The trend neighbor set, represented as a time node, is defined as follows:

[0232]

[0233] In the formula, The trend neighbor set consists of adjacent time nodes sharing the same scale (i.e., |ij|≤1). The final relevance weights are renormalized as follows:

[0234]

[0235] In the formula, E T [i,j] is and The correlation weights between nodes are determined. This step removes connections with insignificant correlations and retains a limited set of neighboring nodes for each node. Subsequently, the retained relevant weights are renormalized to complete the construction of the cross-time correlation graph.

[0236] The intergroup association diagram is represented by G. C =(V C E C ),in It is the set of centroid nodes of the clusters, where N represents the number of clusters. Let j be the j-th centroid node. This represents the correlation weight matrix between centroid nodes. The Copula method is used to calculate the correlation between centroid nodes, constructing an initial inter-group correlation matrix R. C :

[0237] R C (i,j)=Copula(C i C j ), i,j∈{1,2,...,N}

[0238] E is obtained from the original Copula correlation matrix. C initial value

[0239]

[0240] use Represents nodes Related positive neighbor set, The set of negative neighbors is represented, and the corresponding homogeneous and heterogeneous correlation weights are renormalized as follows:

[0241]

[0242] In the formula, E C [i,j] is and The weights of the edges were determined. During the decoupling process, only the weights of homogeneous and heterogeneous edges were retained and normalized respectively. Finally, a correlation graph between homogeneous and heterogeneous groups was constructed, effectively characterizing the complex correlations between groups.

[0243] Furthermore, the method of jointly modeling the complex coupling relationships between feature dimensions, time dimensions, and group dimensions using cross-feature-time graph neural networks, and combining this with Bi-SRU for multivariate time series modeling, specifically includes:

[0244] By designing a message passing mechanism that integrates features, time, and group dimensions, a comprehensive expression of the dynamic interaction characteristics of the load aggregate in terms of time series, features, and groups is achieved. This enables adaptive learning of the multi-level complex dependency structure between various dimensions, thereby enhancing prediction performance.

[0245] In terms of features, the nonlinear interactions between feature nodes are captured through the message passing mechanism in graph neural networks. For feature node v, its message passing process can be represented as:

[0246]

[0247] In the formula, σ is the activation function. It is a learnable parameter matrix. This represents the feature of feature node v at the k-th layer. and Let represent the sets of neighbors that are homogeneous and heterogeneous with the feature node v, respectively, and Norm denotes the normalization operation. This mechanism achieves effective modeling of complex relationships between features by aggregating the features of positive and negative neighbor nodes.

[0248] In the time dimension, a stacked K-layer graph neural network structure is used to characterize the dynamic dependencies between time points. The specific process is as follows:

[0249]

[0250] In the formula, σ is the activation function, and W i (k) It is a learnable parameter matrix. N represents the feature of time node i in the kth layer. i Let be the set of time points adjacent to time point i, and Norm be the normalization operation. By aggregating the feature information of time neighbor nodes, the model updates the time node features layer by layer, and finally outputs the result of K layers.

[0251] At the group level, considering the complex load types within load aggregates and the significant heterogeneity between groups, the feature-time graph neural network defines a separate message passing mechanism for each group. Assuming a group node is represented as , its message passing process can be represented as:

[0252]

[0253] In the formula, σ is the activation function. It is a learnable parameter matrix. N represents the feature of group node c at the kth layer. cLet represent the neighbor set of node c in group , and Norm be the normalization operation. Through the interactions between groups, the feature-time graph neural network can effectively capture the correlation between loads in different groups, thereby improving the overall prediction performance of the load aggregate.

[0254] After completing the message passing across the three dimensions of feature, time, and group, the feature-time graph neural network generates multi-step prediction results through the prediction model, as shown in the following formula:

[0255] Y = f P (H final )

[0256] In the formula, f P (·) represents the prediction model; here, the Bi-SRU model is used, H final This represents the comprehensive features output by the feature-time graph neural network. By performing temporal dependency modeling on the hidden layer features, intermediate prediction results of the graph neural network are generated, and then probabilistic prediction modeling of the load is carried out.

[0257] The forward propagation calculation formula for the Bi-SRU model is as follows:

[0258] f t =σ(W f x t +V f ⊙c t-1 +b f )

[0259] r t =σ(W r x t +V r ⊙c t-1 +b r )

[0260] c t =f t ⊙c t-1 +(1-f t )⊙(W c x t )

[0261] h t =r t ⊙c t +(1-r t )⊙x t

[0262] In the formula, f t The output of the forget gate at time t represents the information c from the previous time step. t-1 Whether to continue propagating, partially propagate, or remove; σ represents the Sigmoid activation function in the model; x tRepresents the input at time t; r t This indicates that the gate output is reset at time t, determining the information c from the previous time step. t-1 How much is written to the current candidate set; c t h represents the new information value generated at time t; t W represents the state of the hidden layer at time t; f W r V f V r W c b represents the corresponding weighting coefficient. f b r The corresponding bias term; the symbol ⊙ represents element-wise multiplication, also known as the Hadamard product.

[0263] The Bi-SRU network model processes information in two directions: the forward SRU unit generates the preceding information, i.e., the hidden state is produced. in, The SRU unit generates subsequent information, i.e., it generates a hidden state. in, Connecting these two pieces of information (forward information and backward information) produces the following joint information:

[0264]

[0265] In the formula, and These represent the hidden states of the forward and backward SRU cells at time t, respectively.

[0266] Furthermore, the step of performing nonlinear mapping on the input features through a hybrid density network, estimating the mean, standard deviation, and corresponding weight of each hybrid component, and outputting the combined hybrid probability density distribution specifically includes:

[0267] Assuming the conditional distribution of the mixture components comes from a family of Gaussian distributions, and the mixture weights δ k ∈[0,1] M The model is represented as a class distribution with M possible states. δ k satisfy Posterior distribution P(δ) k |k) can be embedded through deterministic routing. k To calculate:

[0268]

[0269] In the formula, It is a linear model f δThe softmax function τ forms a nonlinear transformation. The parameters output by the hybrid density network include the mean, standard deviation, and weights of each basic distribution. By weighted summing of these distributions, a hybrid probability distribution function is finally formed, as shown below:

[0270]

[0271] In the formula, μ m (k) and σ m (k) represent the mean and variance parameters of the m-th mixture component, respectively. The parameters of the output conditional density function can be further expressed as:

[0272]

[0273] In the formula, It is a linear model. Hybrid density networks can effectively handle uncertainties in load forecasting by providing decision-makers with more reliable load change expectations by giving a probability density function for each forecast value.

[0274] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operation of each module can be understood with reference to the description of the method embodiments, and will not be repeated here.

[0275] Device Example 2

[0276] This invention provides an electronic device, such as... Figure 3 As shown, it includes: a memory 70, a processor 72, and a computer program stored in the memory 70 and executable on the processor 72, wherein the computer program, when executed by the processor 72, performs the steps as described in the method embodiment.

[0277] Device Example 3

[0278] This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 72, performs the steps described in the method embodiment.

[0279] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0280] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0281] In the 1930s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many improvements to the methodology today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that an improvement to the methodology cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0282] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0283] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0284] For ease of description, the above apparatus is described by dividing it into various functional units. Of course, when implementing the embodiments of this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0285] Those skilled in the art will understand that one or more embodiments of this specification can be provided as a method, system, or computer program product. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0286] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0287] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0288] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0289] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0290] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0291] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0292] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0293] One or more embodiments of this specification can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a particular task or implement a particular abstract data type. One or more embodiments of this specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0294] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0295] The above description is merely an embodiment of this document and is not intended to limit the scope of this document. Various modifications and variations can be made to this document by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this document should be included within the scope of the claims of this document.

Claims

1. A method for establishing a short-term probabilistic prediction model for integrated energy load aggregates, characterized in that, include: S101 preprocesses and reduces the dimensionality of the collected historical data of the integrated energy load aggregate; S102 uses an affinity propagation clustering algorithm based on comprehensive similarity to group the dimensionality-reduced data and extracts the centroid features of each group as representative load features to input into the prediction model; S103 constructs a short-term probability prediction model for integrated energy load aggregates that integrates Copula adaptive correlation analysis, cross-feature-time graph neural network, and hybrid density network; wherein, the short-term probability prediction model for integrated energy load aggregates is used to perform short-term probability prediction of integrated energy load aggregates. Specifically, it includes: Copula adaptive correlation analysis was used to explore the complex dynamic correlations between integrated energy loads, between loads and external influencing factors, and between different groups; This paper utilizes a cross-feature-time graph neural network to jointly model the complex coupling relationships between feature dimensions, time dimensions, and group dimensions, and combines this with Bi-SRU for multivariate time series modeling. Specifically, this includes: By designing message passing mechanisms in the dimensions of features, time, and groups, the dynamic interaction characteristics of load aggregates in terms of time sequence, features, and groups are modeled to achieve adaptive learning of multidimensional dynamic dependencies. In terms of feature dimension, the non-linear interaction relationships between feature nodes are captured through the message passing mechanism in graph neural networks. v The message passing process is represented as follows: ， ， In the formula, It is an activation function. It is a learnable parameter matrix. Represents feature nodes v In the k Features of the layer and Representing the feature nodes respectively v The set of homogeneous and heterogeneous neighbors, where Norm represents the normalization operation; In the time dimension, through stacking K A layered graphical neural network structure is used to characterize the dynamic dependencies between time points. The specific process is as follows: In the formula, It is an activation function. It is a learnable parameter matrix. Indicates time node i In the k Features of the layer To align with time nodes i The set of adjacent time points, where Norm is a normalization operation; At the group level, considering the complexity of load types within a load aggregate and the heterogeneity between groups, a separate message passing mechanism is defined for each group; assuming a group node is represented as , its message passing process is represented as: ， ， In the formula, It is an activation function. It is a learnable parameter matrix. Group node c In the k Features of the layer Representing group nodes respectively c The set of neighbors, where Norm is the normalization operation; After completing the message passing across the three dimensions of feature, time, and group, the feature-time graph neural network generates multi-step prediction results through the prediction model, as shown in the following formula: In the formula, For the prediction model, the Bi-SRU model is used here. The combined features represented by the feature-time graph neural network output; The input features are nonlinearly mapped by a mixture density network to estimate the mean, standard deviation and corresponding weight of each mixture component, and output the combined mixture probability density distribution.

2. The method for establishing a short-term probabilistic prediction model for integrated energy load aggregates according to claim 1, characterized in that, An affinity propagation clustering algorithm based on comprehensive similarity is used to group the dimensionality-reduced data, and the centroid features of each group are extracted as representative load features to be input into the prediction model. Specifically, this includes: A comprehensive similarity metric that integrates Euclidean distance and cosine distance is proposed. The formula for calculating the comprehensive similarity distance is as follows: ， In the formula, Represents Euclidean distance. Represents the cosine distance. n This indicates the number of energy load types included in the user's data, taking into account four dimensions: electrical load, cooling load, heating load, and gas load. Based on the aforementioned similarity measurement method, a comprehensive similarity matrix among users is constructed; using the similarity matrix as input, an affinity propagation clustering algorithm is executed, and the centroids of each cluster are extracted as representative load features input to the prediction model.

3. The method for establishing a short-term probabilistic prediction model for integrated energy load aggregates according to claim 1, characterized in that, Through Copula adaptive correlation analysis, the complex dynamic correlations between integrated energy loads, between loads and external influencing factors, and between different groups are explored, specifically including: A directed graph structure based on load characteristics is constructed, adjacency relationships are defined to characterize the correlation between different load variables, the Copula method is used to model the joint distribution between node pairs, and significant edges are retained based on the correlation threshold. A cross-feature map is constructed, and the Copula function is used to model the joint distribution between load features and external factors. An initial feature correlation matrix is ​​constructed, and the weight matrix for graph construction is obtained through regularization. The specific formula is as follows: ， In the formula, Represents feature nodes and The correlation weights between them; A cross-time graph is constructed. Dominant period information is obtained through frequency domain transformation and multi-scale sampling is performed. An initial time correlation matrix is ​​constructed using the Copula function, and a weight matrix for graph construction is obtained through regularization. The specific formula is as follows: ， In the formula, It is a time node and The correlation weights between them; A group association graph is constructed, using the centroids of each clustered group as graph nodes. The correlation between different groups is modeled based on the Copula function, and a weight matrix for graph construction is obtained through regularization. The specific formula is as follows: ， In the formula, It is the centroid node and The relative weights between them.

4. The method for establishing a short-term probabilistic prediction model for integrated energy load aggregates according to claim 1, characterized in that, The input features are nonlinearly mapped using a mixture density network to estimate the mean, standard deviation, and corresponding weight of each mixture component, and output the combined mixture probability density distribution, specifically including: Assuming that each component of the mixture follows a Gaussian distribution, the mixture weights are modeled as a class distribution that satisfies the weight normalization condition. The mixture density network outputs the mean, standard deviation and weight of each component, and the weighted sum is used to form the mixture probability distribution function.

5. A device for establishing a short-term probabilistic prediction model for integrated energy load aggregates, characterized in that, The method for establishing a short-term probabilistic prediction model for a comprehensive energy load aggregate as described in any one of claims 1 to 4, when executed, includes: The preprocessing module is used to preprocess and reduce the dimensionality of the collected historical data of the integrated energy load aggregate. The clustering analysis module is used to group the dimensionality-reduced data using an affinity propagation clustering algorithm based on comprehensive similarity, and extract the centroid features of each group as representative load features to input into the prediction model. A prediction model module is established to construct a short-term probabilistic prediction model for integrated energy load aggregates, which combines Copula adaptive correlation analysis, cross-feature-time graph neural network and hybrid density network. A short-term probabilistic prediction model for integrated energy load aggregates is constructed by combining Copula adaptive correlation analysis, cross-feature-time graph neural network, and hybrid density network. The input features are nonlinearly mapped by a mixture density network to estimate the mean, standard deviation and corresponding weight of each mixture component, and output the combined mixture probability density distribution. Based on the aforementioned similarity measurement method, a comprehensive similarity matrix among users is constructed; using the similarity matrix as input, an affinity propagation clustering algorithm is executed, and the centroids of each cluster are extracted as representative load features input to the prediction model.

6. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for establishing a short-term probabilistic prediction model for an integrated energy load aggregate as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the method for establishing a short-term probabilistic prediction model for an integrated energy load aggregate as described in any one of claims 1 to 4.

Images