Power-economy mapping matching method and system based on improved hierarchical neural network model

Abstract

The invention discloses a power-economic mapping matching method and system based on an improved hierarchical neural network model, and the method comprises the steps: inputting power data and economic data into a feature coding layer for feature extraction, and obtaining context semantic features of the power-economic data; inputting the context semantic features into an event relation modeling layer, constructing a multi-type graph structure by taking time nodes as event nodes in the graph structure, and modeling the multi-type graph structure to obtain a power-economic state sequence; performing feature matching on the electric power-economic state sequence through a preset electric power-economic feature matching mechanism of a feature matching layer to obtain a semantic similarity between the current electric power economic state and a target economic structure; and the output prediction layer constructs a task output structure according to different task requirements on the basis of semantic similarity, and obtains a final conclusion of power-economy mapping in combination with structural features and a matching result. According to the invention, the expression ability and decision-making assistance ability of the model to a complex coupling system are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing technology, and in particular to a power-economy mapping matching method and system based on an improved hierarchical neural network model. BACKGROUND

[0002] As the basis of modern economic operation, the load level, new energy output, and time-of-use electricity price of power, as key indicators, not only reflect the real-time changes in the energy supply and demand pattern, but also affect the industrial layout, enterprise investment, and regional development at the micro level. Meanwhile, the GDP, industrial structure, and electricity intensity in the macro economic system shape the load structure and energy structure of the power system. A highly complex two-way dynamic relationship is formed between the power and economic systems. However, current power-economy mapping analysis mostly relies on static correlation analysis or one-way causal modeling, which cannot effectively express the deep coupling relationship between the two in terms of time, structure, and semantics. For example, predicting GDP growth based only on power load cannot identify the stage disturbance caused by new energy fluctuations, and analyzing the energy structure optimization path based only on annual economic data ignores the real-time dynamics and periodic feedback of electricity consumption behavior.

[0003] In addition, there are significant differences in sampling frequency, numerical distribution, and time requirement between power and economic data. Power data are mostly hourly high-frequency data with strong periodicity and mutability, while economic data are often quarterly or annual low-frequency data. This "heterogeneous frequency" characteristic makes it difficult for modeling methods based on traditional time series models or static models to balance the alignment processing and structural semantic modeling of multi-source data. With the rapid development of artificial intelligence technology, especially the emergence of new generation structure perception models such as Transformer, self-attention mechanism, and graph neural network, a new approach to integrating multi-source time series data and describing complex causal paths is provided. Through deep coding, structure mapping, and feature matching of power and economic multi-dimensional data, it is expected to achieve accurate modeling from data to knowledge and from prediction to matching, providing intelligent support for energy policy making, load scheduling optimization, and economic trend analysis. SUMMARY

[0004] In view of the above existing problems, the present application is proposed. Therefore, the present application provides a power-economy mapping matching method and system based on an improved hierarchical neural network model to solve the problems mentioned in the background art.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] In a first aspect, the present application embodiment provides a power-economy mapping matching method based on an improved hierarchical neural network model, comprising:

[0007] Acquire power data and economic data of the power system, and input the power data and economic data into the feature encoding layer for feature extraction to obtain the contextual semantic features of power-economic data;

[0008] The contextual semantic features are input into the event relationship modeling layer. Multi-type graph structures are constructed using time nodes as event nodes in the graph structure. The multi-type graph structures are then modeled to obtain the power-economic state sequence.

[0009] The power-economic feature matching mechanism preset by the feature matching layer is used to perform feature matching on the power-economic state sequence to obtain the semantic similarity between the current power economic state and the target economic structure.

[0010] Based on the semantic similarity, the output prediction layer constructs a task output structure according to different task requirements, and combines the structural features and matching results to draw the final conclusion of the power-economy mapping.

[0011] As a preferred embodiment of the power-economy mapping matching method based on the improved hierarchical neural network model described in this invention, the feature encoding layer consists of multiple Transformer encoding layers with the same structure. Each Transformer encoding layer also includes a multi-head self-attention mechanism and a feedforward neural network. Each encoding sub-layer is equipped with residual connections and layer normalization.

[0012] As a preferred embodiment of the power-economic mapping matching method based on the improved hierarchical neural network model described in this invention, the power data and its economic data are input into a feature encoding layer for feature extraction to obtain the contextual semantic features of the power-economic data, including:

[0013] Set any time step t, and receive the power-side dataset P. t With economic side dataset E t And concatenate them into the input vector x t ;

[0014] Input vector x t The data is projected onto a unified high-dimensional embedding space for feature embedding mapping and nonlinear enhancement processing. The data at each time step t is represented as a 128-dimensional real vector. The embedding vectors of all time steps are stacked in chronological order to form the input sequence matrix.

[0015] The beneficial effects of this preferred technical solution are that the vectors formed by direct splicing can retain the original structure and numerical features, avoid information loss, and support the synchronous modeling of subsequent neural network structures in the spatial and temporal domains.

[0016] As a preferred embodiment of the power-economy mapping matching method based on the improved hierarchical neural network model described in this invention, it further includes: inputting the embedded sequence into a multi-layer Transformer encoder;

[0017] Let the output of the (l-1)th layer be H (l-1) The initial input is H (0) =X (0) For each head, calculate the trainable parameter matrix at time step t, and obtain the combined weights MHAtt(H) of all heads. (l-1) );

[0018] Input H (l-1) With the output of the network layer MHAtt(H) (l-1) Residual linking is performed, and all samples of each feature are normalized; after stacking L layers, the semantic features of the current time step in the context and multivariate joint space are obtained.

[0019] As a preferred embodiment of the power-economic mapping matching method based on an improved hierarchical neural network model described in this invention, the following steps are taken: The contextual semantic features are input into the event relationship modeling layer; multiple types of graph structures are constructed using time nodes as event nodes in the graph structure; and the power-economic state sequence is obtained by modeling the multiple types of graph structures, including:

[0020] Each time step's context vector is treated as a node in the graph, forming a node set; the influence relationships between events at different time steps are treated as an edge set; and multiple types of graph structures are constructed based on the node set and the edge set.

[0021] Features within the multi-type graph structure domain are aggregated to enable the current node to obtain context information of its neighborhood; the node states of the multi-type graph structure are iteratively updated through a gated graph neural network to obtain the final state fusion power-economic state sequence.

[0022] The beneficial effects of this preferred technical solution are that it achieves iterative updates of node states through a gated graph neural network, which belongs to the spatial domain message passing model; it controls the flow of information through a gating mechanism, solving the gradient vanishing problem in traditional graph neural networks when they are deeply stacked.

[0023] As a preferred embodiment of the power-economic mapping matching method based on the improved hierarchical neural network model described in this invention, the power-economic state sequence is subjected to feature matching through a preset power-economic feature matching mechanism in the feature matching layer to obtain the semantic similarity between the current power economic state and the target economic structure, including:

[0024] Define the current time step's power-economic state sequence as the sequence to be matched, and the historical economic state sequence as the target sequence;

[0025] When the power system state responds to the economic state in a local time slice, the similarity between the semantic representation vectors at the same position in the two time slices is calculated. If the evolution paths of the two states are similar, the corresponding vectors tend to be the same, and the cosine similarity tends to be close to 1; if the change directions are opposite, the cosine similarity tends to be close to -1; the average similarity is calculated as the local matching score.

[0026] The two sequences are subjected to average pooling to obtain global state vectors, and the cosine similarity between the two average vectors is calculated. The local and global matching indices are then fused to form a unified matching score.

[0027] As a preferred embodiment of the power-economic mapping matching method based on the improved hierarchical neural network model described in this invention, the following steps are taken: Constructing a task output structure according to different task requirements, and combining the structural features with the matching results to derive the final conclusion of the power-economic mapping include:

[0028] Based on the two information sources output by the feature matching layer and the original data as input data, economic indicators are predicted using the current power system state. The input data treasure trove includes the structure-enhanced representation vector of the current time point, the matching score between the current state and the target economic structure, and the original economic and power datasets.

[0029] The cross-entropy loss function is used to determine the degree of matching between the current power system state and the target economic structure. When the label is 1, it means that the state matches the target structure, and when the label is 0, it means that the state does not match the target structure.

[0030] The economic status is divided into several categories, and the status of any indicator is judged based on the probability distribution of each category.

[0031] Secondly, the present invention provides a power-economy mapping matching system based on an improved hierarchical neural network model, comprising:

[0032] The feature encoding module is used to acquire power data and economic data of the power system, and input the power data and economic data into the feature encoding layer for feature extraction to obtain the contextual semantic features of power-economic data.

[0033] The event relationship modeling module is used to input the context semantic features into the event relationship modeling layer, construct multi-type graph structures with time nodes as event nodes in the graph structure, and model the multi-type graph structures to obtain the power-economic state sequence.

[0034] The feature matching module is used to perform feature matching on the power-economic state sequence through a preset power-economic feature matching mechanism to obtain the semantic similarity between the current power economic state and the target economic structure.

[0035] The output prediction module is used to construct a task output structure based on the semantic similarity and according to different task requirements, and to draw the final conclusion of the power-economy mapping by combining the structural features and matching results.

[0036] Thirdly, the present invention provides an electronic device, comprising:

[0037] Memory and processor;

[0038] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the power-economy mapping matching method based on the improved hierarchical neural network model.

[0039] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the power-economy mapping matching method based on the improved hierarchical neural network model.

[0040] Compared with existing technologies, the advantages of this invention are as follows: This invention constructs a "semantic-structural" dual-driven mapping framework oriented towards structural cognition and task diversity by integrating the Transformer encoding mechanism and the gated graph neural network propagation mechanism through an improved hierarchical neural network architecture. Based on a multi-layer Transformer, it achieves unified temporal context encoding for high-frequency power data and low-frequency economic data, effectively aligning heterogeneous input dimensions by combining position embedding and upsampling mechanisms. Simultaneously, it models cross-period causal paths between non-adjacent time steps on the event graph structure through a graph neural network, dynamically capturing potential feedback relationships between system variables. The semantic matching module utilizes the fusion of local cosine alignment and global embedding to achieve consistency measurement between the current state and the target economic structure, providing structure-aware support for downstream prediction or discrimination tasks. The multi-task output module further enhances the model's application scalability, adapting to various task types such as continuous value regression, state classification, and structural matching. The overall technical solution covers steps such as multimodal data preprocessing, embedding encoding, graph structure construction, feature alignment and output mapping. While maintaining interpretability, it significantly improves the model's ability to express complex coupled systems and its decision-making support capabilities, providing intelligent support for new energy scheduling optimization, industrial structure prediction and regional economic regulation. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0042] Fig. 1 This is a schematic diagram of the process flow of a power-economy mapping matching method based on an improved hierarchical neural network model according to an embodiment of the present invention;

[0043] Fig. 2 This is a schematic diagram of the multi-head attention mechanism of an electricity-economy mapping matching method based on an improved hierarchical neural network model according to an embodiment of the present invention;

[0044] Fig. 3 This is a schematic diagram illustrating the implementation process of a power-economy mapping matching method based on an improved hierarchical neural network model, according to an embodiment of the present invention. Detailed Implementation

[0045] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0046] Example 1, referring to Figs. 1-3 As one embodiment of the present invention, this embodiment provides a power-economic mapping matching method based on an improved hierarchical neural network model, such as... Fig. 1 As shown, it includes:

[0047] S100: Acquire power data and economic data of the power system, and input the power data and economic data into the feature encoding layer for feature extraction to obtain the contextual semantic features of power-economic data;

[0048] S200: Input the contextual semantic features into the event relationship modeling layer, construct multi-type graph structures with time nodes as event nodes in the graph structure, and model the multi-type graph structures to obtain the power-economic state sequence;

[0049] S300: Through the power-economic feature matching mechanism preset by the feature matching layer, feature matching is performed on the power-economic state sequence to obtain the semantic similarity between the current power economic state and the target economic structure;

[0050] S400: The output prediction layer is based on semantic similarity. It constructs the task output structure according to different task requirements and combines the structural features with the matching results to draw the final conclusion of the power-economy mapping.

[0051] It should be noted that current power-economic mapping analyses mostly rely on static correlation analysis or one-way causal modeling, which struggles to effectively express the deep coupling relationship between the two in terms of time, structure, and semantics. For example, predicting GDP growth solely based on electricity load fails to identify the phased disturbances caused by fluctuations in new energy sources; similarly, analyzing energy structure optimization paths based solely on annual economic data ignores the real-time dynamics and periodic feedback of electricity consumption behavior. Power and economic data differ significantly in sampling frequency, numerical distribution, and timeliness requirements. Power data is mostly hourly high-frequency data with strong periodicity and abrupt changes; while economic data is often quarterly or annual low-frequency data, significantly influenced by policies and macroeconomic structures. This "heterogeneous frequency" characteristic makes it difficult for traditional time-series or static modeling methods to simultaneously handle the alignment of multi-source data and structural semantic modeling. This invention proposes a power-economic mapping matching method based on an improved hierarchical neural network model, providing a systematic solution to several key technical bottlenecks in current power-economic system modeling. First, to address the challenge of traditional time series models failing to capture deep causal relationships between non-adjacent time points, a graph neural network structure is introduced, combined with a multi-type edge construction mechanism, to explicitly model cross-period and cross-variable structural dependencies in time series data. Second, considering the significant differences in sampling frequency and data granularity between power data and economic data, an upsampling fusion mechanism and embedding alignment strategy are designed to achieve joint representation and unified encoding of multi-source heterogeneous data. Furthermore, to overcome the insufficient response of existing methods to local anomalies or global trends, a multi-scale matching mechanism combining local semantic alignment and global trend matching is constructed, thereby achieving accurate semantic measurement of the relationship between the current state of the power system and the target economic structure. Finally, addressing the limitations of traditional models' single structure and poor scalability, this invention supports multiple task types such as regression prediction, structural discrimination, and state classification through modular design, exhibiting good task generalization ability and application adaptability.

[0052] It should be noted that this invention achieves refined data modeling and intelligent matching through four core modules;

[0053] The feature encoding layer integrates high-frequency electricity data such as time-of-use pricing, output curves, and electricity load with low-frequency economic data such as GDP and industrial structure. Through upsampling and Transformer multi-layer encoding, it achieves contextual modeling with a unified spatiotemporal resolution. The event relationship modeling layer uses time nodes as event nodes in a graph structure, constructing multiple types of graph structures including time evolution edges, periodic edges, and logical edges. It introduces a gated graph neural network multi-round propagation mechanism to explicitly express causal links across time domains and modalities. The feature matching layer uses a dual mechanism of local and global matching to measure the semantic similarity between the current power economic state and the target economic structure, balancing trend consistency and instantaneous responsiveness. The output prediction layer addresses multi-task requirements, constructing three task output structures: regression, discrimination, and classification. Combining structural features and matching results, it outputs indicator values, risk probabilities, or category labels, achieving a closed loop from modeling to decision-making. The method of this invention has strong scalability, interpretability and robustness, and is applicable to various scenarios such as smart grids, regional load forecasting, energy policy evaluation and economic regulation decision support. It fills the technical gap in the field of power-economic multi-source coupled data modeling and semantic matching prediction.

[0054] In this embodiment of the invention, the feature encoding layer in step S100 consists of multiple Transformer encoding layers with the same structure. Each Transformer encoding layer also includes a multi-head self-attention mechanism and a feedforward neural network. Each encoding sub-layer is equipped with residual connections and layer normalization.

[0055] It should be noted that the core of multi-layer Transformer encoding processing lies in its unique architectural design. By stacking multiple encoding layers, and utilizing multi-head self-attention mechanisms and feedforward neural networks, deep feature extraction and relation modeling are performed on the input data.

[0056] In an optional embodiment, to map the relationship between electricity data and economic data, electricity data (such as power generation, electricity load, etc.) and economic data (such as GDP growth rate, changes in industrial structure, etc.) are used as inputs. After processing through multiple Transformer encoding layers, the global dependencies and local features between the data can be captured. The multi-head self-attention mechanism allows the model to simultaneously focus on the interactions between different data dimensions, while the feedforward neural network further performs nonlinear transformations on the extracted features, enhancing the model's expressive power.

[0057] In this embodiment of the invention, before step S100, which involves inputting power data and economic data into the feature encoding layer for feature extraction to obtain the contextual semantic features of the power-economic data, the method further includes: data initialization and preprocessing, processing the input power and economic data, forming location codes, and finally obtaining a time series X carrying location codes. (0)This data is then input into the multi-layer Transformer encoder.

[0058] In this embodiment of the invention, step S100, which inputs power data and its economic data into a feature encoding layer for feature extraction to obtain the contextual semantic features of the power-economic data, includes:

[0059] Set any time step t, and receive the power-side dataset P. t With economic side dataset E t And concatenate them into the input vector x t ;

[0060] Input vector x t The data is projected onto a unified high-dimensional embedding space for feature embedding mapping and nonlinear enhancement processing. The data at each time step t is represented as a 128-dimensional real vector. The embedding vectors of all time steps are stacked in chronological order to form the input sequence matrix.

[0061] In this embodiment of the invention, step S100 further includes: inputting the embedded sequence into a multilayer Transformer encoder;

[0062] Let the output of the (l-1)th layer be H (l-1) The initial input is H (0) =X (0) For each head, calculate the trainable parameter matrix at time step t, and obtain the combined weights MHAtt(H) of all heads. (l-1) );

[0063] Input H (l-1) With the output of the network layer MHAtt(H) (l-1) Residual linking is performed, and all samples of each feature are normalized; after stacking L layers, the semantic features of the current time step in the context and multivariate joint space are obtained.

[0064] Specifically, refer to Fig. 2 Let the output of the (l-1)th layer be The initial input is H (0) =X (0) .

[0065] For each head, a query Q, key K, and value V are generated at time step t, using the following formula:

[0066]

[0067] Where h is the number of heads, W Q W K WV These are the trainable parameter matrices;

[0068] The formula for outputting the weights after merging all heads is:

[0069] MHAtt(H (l-1) = Concat(head1,…,head) h W O

[0070] Among them, W O These are the trainable parameter matrices;

[0071] The formula for calculating the attention weight of each head is:

[0072]

[0073] In a preferred embodiment, the multi-head self-attention mechanism is one of the core ideas of Transformer. Its purpose is to enable the model to pay attention to different parts of the input sequence at the same time and capture information in different subspaces. In this embodiment, the number of heads of the multi-head self-attention mechanism can be set to 8.

[0074] Furthermore, in this embodiment, residual linking refers to linking the input H... (l-1) The output MHTt(H) is directly added to the network layer. (l-1) The purpose of this approach is to alleviate the vanishing gradient problem during deep network training, making the network easier to train. LayerNorm is a layer normalization function that normalizes all samples for each feature, helping to stabilize and accelerate the network training process. The calculation formula is shown below:

[0075]

[0076] Furthermore, in the Transformer model, the feedforward neural network, as part of the encoder and decoder layers, performs independent nonlinear transformations at each position in the sequence through a position-wise fully connected network. It consists of two linear transformation layers with a nonlinear activation function layer in between. In the feedforward neural network, the input data first undergoes the first linear transformation layer, then is processed by a nonlinear activation function, and finally passes through the second linear transformation layer for output. The calculation formula is shown below:

[0077] FFN(x)=W2·φ(W1x+b1)+b2

[0078] in, and Here, b1 and b2 are learnable weight matrices, b1 and b2 are bias terms, and φ is the activation function.

[0079] The updated residual connectivity and layer normalization formulas yield the following:

[0080]

[0081] After stacking L layers, the final result is obtained. It is the semantic representation of the current time step in the context and multivariate joint space, which will be used in subsequent graph modeling, matching, prediction and other modules.

[0082] It should be noted that in this embodiment of the invention, the feedforward neural network uses the Gaussian Error Linear Unit (GELU) as the activation function. Compared with other activation functions, the GELU activation function does not completely set the gradient to zero with negative input values. In the subsequent backpropagation process, more neurons can participate in the gradient update, thereby accelerating the convergence of the model; at the same time, it provides a smoother gradient, which helps to stabilize the training process, especially suitable for the Transformer structure.

[0083] In this embodiment of the invention, step S200 involves inputting contextual semantic features into the event relationship modeling layer, constructing multi-type graph structures using time nodes as event nodes in the graph structure, and modeling the multi-type graph structures to obtain the power-economic state sequence, including:

[0084] Each time step's context vector is treated as a node in the graph, forming a node set; the influence relationships between events at different time steps are treated as an edge set; and multiple types of graph structures are constructed based on the node set and the edge set.

[0085] Features within the multi-type graph structure domain are aggregated to enable the current node to obtain contextual information about its neighborhood; the node states of the multi-type graph structure are iteratively updated through a gated graph neural network to obtain the final state fusion sequence of power-economic states.

[0086] Specifically, the context vector U at each time step t t Consider it as a node v in the graph t The set of nodes is represented as:

[0087] V = {v1, v2, ..., v} T},v t ≡U t

[0088] Where T represents the length of the time series, and the node features are initialized as follows:

[0089]

[0090] Construct a directed graph G = (V, E) and its edge set. It represents the influence relationship between events at different time steps. The constructed event graph structure G has multiple edge forms such as time, logic, and physical structure, which can fully express the causal chain of events in the power economic system.

[0091] In an optional embodiment, to fully represent the power economic evolution structure, the edge construction types are shown in Table 1:

[0092] Table 1 Edge Type Construction

[0093]

[0094] It should be noted that significant nonlinear dependencies and lagged effects exist between variables in multivariate time-series power systems and macroeconomic systems. For example, fluctuations in renewable energy output may delay the impact on industrial GDP, and changes in electricity load may only reflect industrial activity several days later. Therefore, relying solely on time-series adjacency information is insufficient to capture such long-range, cross-dimensional causal relationships. Gated graph neural networks, as a structural modeling mechanism, propagate structural information through multiple rounds based on a pre-constructed "event graph," enabling the display of nonlinear causal paths and semantically enhanced associations between modeling time points. Combining the advantages of graph topological representation and gating memory mechanisms, it is currently the most effective, convergent, and interpretable solution for modeling time-series graph structures.

[0095] Furthermore, gated graph neural networks (GNNs) achieve iterative updates of node states through recurrent neural networks, belonging to the spatial domain message passing model. Their design goal is to control information flow through gating mechanisms, solving the gradient vanishing problem in deep stacked graph neural networks. Each round of propagation in a gated graph neural network consists of three steps: adjacency aggregation, gating update, and state fusion.

[0096] In an optional embodiment, the state update process of a certain layer includes:

[0097] Information is collected from adjacent nodes and summed using weighted averages (W):

[0098]

[0099] in, Represents all adjacent nodes of node t; The state of node j in the previous round; W∈R 128 To share the adjacency transformation matrix, this step achieves feature aggregation within the structural domain, enabling the current node to obtain contextual information from its neighborhood.

[0100] A GRU-style gating structure is introduced to control the degree of node state updates, where the update gate is used to determine how much historical information needs to be discarded, and the reset gate is used to determine how much new information should be adopted.

[0101] The updated gate formula is:

[0102]

[0103] The formula for resetting the door is:

[0104]

[0105] Where σ is the Sigmoid function, W z U z W r U r ∈R 128 These are learnable parameters.

[0106] Candidate state:

[0107]

[0108] Last Update:

[0109]

[0110] The above process makes node v t It can adaptively control the trade-off between memory retention and absorption of new structures.

[0111] The gated graph neural network propagation process executes L=3 iterations, and the final representation of each node is:

[0112]

[0113] This vector compared to the initial U t The results obtained by GGNN incorporate structural causal paths, giving it a stronger ability to model long-range dependencies, and can be used as input data for the feature matching module and the output prediction module.

[0114] In this embodiment of the invention, step S300 uses a preset power-economic feature matching mechanism in the feature matching layer to perform feature matching on the power-economic state sequence, obtaining the semantic similarity between the current power economic state and the target economic structure, including:

[0115] Define the current time step's power-economic state sequence as the sequence to be matched, and the historical economic state sequence as the target sequence;

[0116] When the power system state responds to the economic state in a local time slice, the similarity between the semantic representation vectors at the same position in the two time slices is calculated. If the evolution paths of the two states are similar, the corresponding vectors tend to be the same, and the cosine similarity tends to be close to 1; if the change directions are opposite, the cosine similarity tends to be close to -1; the average similarity is calculated as the local matching score.

[0117] The two sequences are subjected to average pooling to obtain global state vectors, and the cosine similarity between the two average vectors is calculated. The local and global matching indices are then fused to form a unified matching score.

[0118] In this embodiment of the invention, step S400, which involves constructing a task output structure based on different task requirements and combining the structural features with the matching results to arrive at the final conclusion of the power-economic mapping, includes:

[0119] Based on the two information sources output by the feature matching layer and the original data as input data, economic indicators are predicted using the current power system state. The input data treasure trove includes the structure-enhanced representation vector of the current time point, the matching score between the current state and the target economic structure, and the original economic and power datasets.

[0120] The cross-entropy loss function is used to determine the degree of matching between the current power system state and the target economic structure. When the label is 1, it means that the state matches the target structure, and when the label is 0, it means that the state does not match the target structure.

[0121] The economic status is divided into several categories, and the status of any indicator is judged based on the probability distribution of each category.

[0122] Reference Fig. 3 The implementation process of the power-economy mapping matching method includes steps A1 to A5:

[0123] Step A1: Data Preprocessing; The data sources of the power system and the economic system have significant structural differences and temporal attributes. Power-side data has high-frequency characteristics, such as electricity prices, output, and electricity consumption collected hourly or minutely; economic-side data are mostly low-frequency indicators, such as quarterly or annual GDP and industrial added value. To achieve coupled modeling of power-side and economic-side data, they are first uniformly coded at the structural level.

[0124] Specifically, at a certain time step t, the received power-side dataset is... Data sets from the economic side Here, m and n represent the number of data points on the electricity side and the economic side, respectively. The economic side data includes industrial electricity consumption, energy prices, etc. To achieve joint modeling, the model directly concatenates the two variable vectors into a complete input vector x. t :

[0125] x t =Concat(P t E t )∈R m+n

[0126] It should be noted that the vectors formed by direct concatenation can preserve the original structure and numerical features, avoid information loss, and support the synchronous modeling of subsequent neural network structures in the spatial and temporal domains.

[0127] Furthermore, since electricity data typically has a high sampling frequency on an hourly basis, while economic data such as GDP and PPI are sampled on a monthly or quarterly basis; the concatenated input vector x t High dimensionality, coupled with heterogeneous variable sources and inconsistent distributions, can negatively impact subsequent modeling efficiency and convergence speed. To further enhance the representation capability of original features in high-dimensional space, a linear embedding layer and activation function are introduced during the data processing stage to transform x... t Projecting onto a unified high-dimensional embedding space, this mapping method helps the model learn the intrinsic relationships between different data, thereby reducing heterogeneity caused by differences in time frequency, etc. The formula is as follows:

[0128]

[0129] Among them, W x ∈R d×(m+n) b is a learnable weight matrix; x ∈R d d is the bias vector; d = 128 is the embedding vector dimension; the activation function is GELU (Gaussian Error Linear Unit), whose mathematical expression is as follows:

[0130]

[0131] After feature embedding mapping and nonlinear enhancement, the data at each time step t is represented as a 128-dimensional real vector. Used for subsequent model training and prediction.

[0132] Stack the embedding vectors of all time steps in chronological order to form the input sequence matrix:

[0133]

[0134] Where T represents the length of the input sequence (e.g., for hourly data of one month, T is 24 × 30 = 720). However, since the standard feedforward neural network structure does not have time-aware capabilities, the time step number of each position must be explicitly labeled.

[0135] Step A2: Multi-layer Transformer encoding processing; To enable the model to have order awareness in the sequence, this invention adopts a learnable positional encoding mechanism. Compared with the traditional sine and cosine fixed encoding method, learnable positional encoding is more suitable for non-periodic data, especially for macroeconomic variables with uneven sampling. The specific formula is:

[0136] PE = [p1, p2, ..., p T ],p t ∈R 128

[0137] Adding it to the original embedding sequence is represented as:

[0138] X (0) =X (0) +PE

[0139] The presence of location coding allows the meaning of the same economic or electricity data to be distinguished at different time steps, thereby enhancing the model's ability to express the temporal evolution structure.

[0140] To achieve context modeling across time steps, the embedded sequence input is fed into a multi-layer Transformer encoder. This invention sets the number of Transformer layers to L=4, with each layer containing the following structure:

[0141] The multi-head self-attention mechanism consists of 8 heads, each with a dimension of 16. By computing multiple attention mechanisms in parallel, the model can simultaneously capture the dependencies between different positions in the sequence, enhancing the model's ability to understand sequence data.

[0142] Residual connection layers can directly add input to output, which helps to alleviate the gradient vanishing or gradient explosion problems in deep network training, makes it easier to form identity mappings, accelerates the training process and improves model performance.

[0143] Layer normalization normalizes each feature of each sample to reduce internal covariate bias, which helps stabilize and accelerate the training process.

[0144] A feedforward fully connected network, with a dimension of 512, can further process features and enhance the nonlinear expressive power of the model. It typically contains two linear transformations with a GELU activation function in between.

[0145] The recursive calculation for each level is as follows:

[0146] X (l) =Transformer l (X (l-1) ), l=1,2,…,L

[0147] Finally, the feature representation at each time step t is updated as follows:

[0148]

[0149] The input data at each time step t is processed by four Transformer layers to generate a 128-dimensional feature vector U. tVector U t It integrates contextual information such as electricity price changes, power output fluctuations, electricity consumption structure and economic variable evolution, which forms the basis for subsequent graph modeling and matching determination.

[0150] Step A3: Relationship Modeling. While Transformer can capture long-distance dependencies, it cannot characterize causal paths or hierarchical relationships between variables. Therefore, a structural modeling module, namely the event relationship modeling layer, is introduced after the feature encoding layer. This layer constructs an event graph and uses a gated graph neural network to propagate relationships between event nodes, further enhancing the model's ability to model causal structures, time backtracking, and potential path dependencies. Compared to ordinary graph neural networks, gated graph neural networks introduce a gating mechanism, which can automatically control the degree to which the current node retains its old state and receives information from new neighbors, making it particularly suitable for time series-structure hybrid problems with long-term dependencies or lag effects.

[0151] Based on the constructed node set V and the directed graph G, the completed event graph structure G has multiple edge forms such as time, logic, and physical structure, which can fully express the causal chain of events in the power economic system.

[0152] Secondly, assuming the number of propagation rounds is L = 3, perform 3 rounds of information diffusion on the graph structure as shown in 4.1.2. Each round of information diffusion includes three steps: adjacency aggregation, gating update, and state fusion, ultimately outputting u. t Compared to the Transformer output vector, u t It further incorporates structural causal relationship information and can model long-term chain effects across time series, serving as input data for matching and output layers.

[0153] Step A4: Input definition and construction. The input to the feature matching layer consists of two time series segments:

[0154] The main sequence (the sequence to be matched), derived from the current time period's electricity-economic status sequence, is represented as:

[0155]

[0156] The target sequence (reference segment) is derived from a known historical sequence of economic states or policy objective constraints.

[0157]

[0158] Among them, l a , l b These represent the lengths of the two sequences (which can be the same or different).

[0159] Step A4 can define a function in the feature space to measure the degree of matching between the two. The larger the value, the closer the structures are and the more consistent the economic evolution trend.

[0160] The local matching submodule assumes that the evolution of the power system state can induce a rapid response in the economic state within a local time slice. Therefore, the short-time alignment matching method is defined as follows:

[0161] For each pair of local alignment positions i∈[1,l], calculate the cosine similarity:

[0162]

[0163] This formula is used to calculate the similarity between semantic representation vectors at the same location in two time periods. If the evolution paths of two states are similar, the corresponding vectors tend to be the same. i If the change direction is opposite, then Sim approaches 1; i Approaching -1.

[0164] Let the matching window length be l = min(l a ,l b If the overall local matching degree is the average value, then the overall local matching degree is:

[0165]

[0166] This matching value reflects the structural similarity of the two time periods in their hourly state evolution.

[0167] The global matching submodule can verify the consistency between two time periods in terms of overall distribution, trend changes, or long-term evolution mechanisms. The specific steps are as follows:

[0168] Perform average pooling on each sequence to obtain the global state vector:

[0169]

[0170] Calculate the cosine similarity between the two:

[0171]

[0172] Unlike local matching, global matching ignores specific temporal locations and emphasizes the consistency of structural average states. For example, it can be used to measure whether the overall electricity price and power output structure meet the future GDP growth structure.

[0173] By integrating local and global matching metrics, a unified judgment score is formed:

[0174]

[0175] Here, α controls the relative weights of local and global values, and is set according to the specific application scenario, as shown in Table 2. The final matching score... It can be used as input to subsequent prediction modules or directly for classification / regression output.

[0176] Table 2. Standards for setting α in different application scenarios

[0177]

[0178] Step A5: Based on different application needs, the mapping task between the power and economic systems can be roughly divided into the following three categories: Structure matching judgment (discrimination task): determining whether the current power operation status matches a certain economic structure; Economic variable prediction (regression task): predicting indicators such as GDP and energy intensity based on power-side input data; Indicator change classification (categorization task): determining whether a certain indicator will be in a state of growth / decline / stagnation in the future.

[0179] Furthermore, during the training phase, this invention employs a specific optimization strategy to improve model performance. The AdamW algorithm, based on adaptive estimation, is selected as the optimizer, combining the advantages of the Adam optimizer with improvements to weight decay. The initial learning rate is set to 1×10⁻⁶. -4 This helps the model converge stably in the early stages of training. To prevent gradient explosion, gradients are pruned, and the maximum gradient norm is set to 1.0. Additionally, an early stopping mechanism is introduced: if the performance metrics on the validation set do not improve after 10 consecutive training iterations, the training process will automatically terminate, aiming to avoid overfitting and save computational resources.

[0180] Furthermore, in regression-type tasks, the goal is to predict the specific values ​​of one or more continuous-value economic indicators using the current power system state. This task uses two information sources from the feature matching layer output and the original data as input data, including the structure-enhanced representation vector at the current time point. Matching score between the current state and the target economic structure and the original economic and power dataset

[0181] The concatenated input is then fed into a fully connected regression network (MLP):

[0182]

[0183] Among them, f reg A regression network composed of a multilayer sensing mechanism; k is the output dimension.

[0184] Loss function and optimization objective: The loss is expressed using the standard mean square error loss as follows:

[0185]

[0186] The loss function measures the average squared error between the predicted and actual values; a smaller error indicates better model performance. The training process continuously minimizes this loss function, optimizing model parameters and improving prediction accuracy.

[0187] It should be noted that in practical applications, this technology can be used to predict key continuous indicators of a region's economic system, such as Gross Domestic Product (GDP), energy consumption per unit of GDP, and industrial investment intensity. With the combined input of high-frequency monitoring data from the power system and macroeconomic reference data, the model can output continuous numerical prediction results for the corresponding time window. Ultimately, it can achieve a quantitative assessment of macroeconomic trends; support dynamic evaluation of the economic benefits of renewable energy output and load reconfiguration; embed itself in the energy dispatch strategy evaluation process to form economically oriented control objectives; and replace or assist traditional economic forecasting methods, improving real-time performance and structural adaptability.

[0188] Furthermore, discriminative tasks do not aim for numerical prediction, but rather to determine whether the current power system state matches a given target economic structure. Similarly, the concatenated vector is used as input to the classification and discriminative network p:

[0189]

[0190] A threshold of 0.5 can be set, meaning that when p ≥ 0.5, the two sets of data are considered to match; and when p < 0.5, the two sets of data are considered to be mismatched.

[0191] Using the standard binary classification cross-entropy loss function

[0192]

[0193] The smaller the loss value calculated by this function, the more accurate the model's judgment of the current sample; the larger the loss, the greater the deviation between the model output and the actual label. This loss serves as a signal for backpropagation, guiding the model parameters to continuously adjust and improve the discrimination accuracy. Here, y∈[0,1] is the true label. When y=1, it means that the state matches the target structure; when y=0, it means that the state does not match the target structure. logp and log(1-p) are logarithmic transformations used to penalize the confidence level of errors. If the true label y is 1 but p approaches 0, the loss approaches positive infinity, resulting in a strong penalty.

[0194] The model utilizes semantically enhanced representations obtained from graph neural networks, performing bidirectional matching with the target state. The output layer outputs the degree of matching or whether the current state meets the judgment label through logistic regression or classification networks. In practical applications, this technology identifies whether the current power system operating state has a strong match or consistency with a preset target economic structure (such as green and low-carbon, resource-efficient, and structurally optimized). The output is a matching probability value or a binary state judgment result. Ultimately, this technology can support the adaptability analysis of real-time operating states to economic development strategies; provide early warning mechanisms for structural anomalies and supply-demand mismatches; assist in the compliance analysis of regional operating structures and aid administrative supervision; and achieve closed-loop adaptation between economic goals and power dispatch.

[0195] Furthermore, categorization tasks are suitable for forecasting economic indicators that are not continuous but rather categorized into stages. For example, GDP may be in a state of "high-speed growth," "stable development," or "decline"; energy consumption per unit may be in a state of "high energy consumption," "acceptable," or "low carbon"; and PPI or CPI inflationary pressure may be in a state of "risk / neutral / safe." This type of task can simplify modeling and improve identification accuracy, making it suitable for high-level strategy formulation or fuzzy control.

[0196] Assume there are C state categories, and the input data is graph neural network-enhanced features. Structure matching score Output vector for:

[0197]

[0198] in, This represents the output probability of the Cth class of results, reflecting the likelihood that the system state belongs to each class.

[0199] For the classification of indicator intervals, appropriate reference criteria are selected based on the changing patterns (trend / volatility / interval) of different economic / electricity variables, as shown in Table 3.

[0200] Table 3 Reference Table for Setting Classification Thresholds for Different Types of Variables

[0201]

[0202] Employing a multi-class cross-entropy loss function

[0203]

[0204] This loss function measures the deviation between the model's predicted output and the true label, guiding model parameter updates and thus gradually improving classification accuracy. During training, the model outputs the probability corresponding to each class. For the correct class, the loss is greater when the model predicts a low probability and smaller when the model predicts a high probability. Ultimately, the goal is to make the predicted probability of the correct class as close to 1 as possible and the probability of the incorrect class as close to 0, thus completing the learning of the classification boundary.

[0205] This invention utilizes an improved hierarchical neural network architecture, integrating Transformer encoding and gated graph neural network propagation mechanisms to construct a "semantic-structure" dual-driven mapping framework oriented towards structural cognition and task diversity. Based on a multi-layered Transformer, it achieves unified temporal context encoding for high-frequency power data and low-frequency economic data, effectively aligning heterogeneous input dimensions through positional embedding and upsampling mechanisms. Simultaneously, it models cross-period causal paths between non-adjacent time steps on an event graph structure using a graph neural network, dynamically capturing potential feedback relationships between system variables. The semantic matching module leverages local cosine alignment and global embedding fusion to achieve consistency measurement between the current state and the target economic structure, providing structure-aware support for downstream prediction or discrimination tasks. The multi-task output module further enhances the model's application scalability, adapting to various task types such as continuous value regression, state classification, and structural matching. The overall technical solution encompasses multimodal data preprocessing, embedding encoding, graph structure construction, feature alignment, and output mapping, significantly improving the model's expressive power and decision-making support capabilities for complex coupled systems while maintaining interpretability, providing intelligent support for new energy dispatch optimization, industrial structure prediction, and regional economic regulation.

[0206] Example 2: The above example is an illustrative scheme of a power-economic mapping matching method based on an improved hierarchical neural network model. It should be noted that the technical solution of this power-economic mapping matching system based on an improved hierarchical neural network model belongs to the same concept as the technical solution of the power-economic mapping matching method based on an improved hierarchical neural network model described above. Details not described in detail in this example of the power-economic mapping matching system based on an improved hierarchical neural network model can be found in the description of the technical solution of the power-economic mapping matching method based on an improved hierarchical neural network model described above.

[0207] This embodiment discloses a power-economic mapping matching system based on an improved hierarchical neural network model, comprising: a feature encoding module, used to acquire power data and economic data of the power system, and input the power data and economic data into the feature encoding layer for feature extraction to obtain the contextual semantic features of the power-economic data;

[0208] The event relationship modeling module is used to input contextual semantic features into the event relationship modeling layer, construct multi-type graph structures with time nodes as event nodes in the graph structure, and model the multi-type graph structures to obtain the power-economic state sequence.

[0209] The feature matching module is used to perform feature matching on the power-economic state sequence through a preset power-economic feature matching mechanism to obtain the semantic similarity between the current power economic state and the target economic structure.

[0210] The output prediction module is used to construct the task output structure based on semantic similarity and different task requirements, and to draw the final conclusion of the power-economy mapping by combining structural features and matching results.

[0211] This embodiment also provides an electronic device applicable to the power-economy mapping matching method based on an improved hierarchical neural network model, including:

[0212] The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement the power-economy mapping matching method based on the improved hierarchical neural network model as proposed in the above embodiments.

[0213] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the power-economy mapping matching method based on an improved hierarchical neural network model as proposed in the above embodiments.

[0214] The storage medium proposed in this embodiment and the power-economy mapping matching method based on the improved hierarchical neural network model proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0215] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0216] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A power-economic mapping matching method based on an improved hierarchical neural network model, characterized in that, include: Acquire power data and economic data of the power system, and input the power data and economic data into the feature encoding layer for feature extraction to obtain the contextual semantic features of power-economic data; The contextual semantic features are input into the event relationship modeling layer. Multi-type graph structures are constructed using time nodes as event nodes in the graph structure. The multi-type graph structures are then modeled to obtain the power-economic state sequence. The power-economic feature matching mechanism preset by the feature matching layer is used to perform feature matching on the power-economic state sequence to obtain the semantic similarity between the current power economic state and the target economic structure. Based on the semantic similarity, the output prediction layer constructs a task output structure according to different task requirements, and combines the structural features and matching results to draw the final conclusion of the power-economy mapping.

2. The power-economic mapping matching method based on an improved hierarchical neural network model as described in claim 1, characterized in that, The feature encoding layer consists of multiple Transformer encoding layers with the same structure. Each Transformer encoding layer also includes a multi-head self-attention mechanism and a feedforward neural network. Each encoding sub-layer is equipped with residual connections and layer normalization.

3. The power-economic mapping matching method based on an improved hierarchical neural network model as described in claim 2, characterized in that, Electricity data and its economic data are input into a feature encoding layer for feature extraction, resulting in the following contextual semantic features of the electricity-economic data: Set any time step t, and receive the power-side dataset P. t With economic side dataset E t And concatenate them into the input vector x t ; Input vector x t The data is projected onto a unified high-dimensional embedding space for feature embedding mapping and nonlinear enhancement processing. The data at each time step t is represented as a 128-dimensional real vector. The embedding vectors of all time steps are stacked in chronological order to form the input sequence matrix.

4. The power-economic mapping matching method based on an improved hierarchical neural network model as described in claim 3, characterized in that, Also includes: The embedded sequence is input into a multi-layer Transformer encoder; Let the output of the (l-1)th layer be H (l-1) The initial input is H (0) =X (0) ; For each head, the trainable parameter matrix at time step t is calculated, and the attention weights of each head are obtained by merging the weights of all heads, MHAtt(H). (l-1) ); Input H (l-1) With the output of the network layer MHAtt(H) (l-1) Residual linking is performed, and all samples of each feature are normalized; after stacking L layers, the semantic features of the current time step in the context and multivariate joint space are obtained.

5. The power-economic mapping matching method based on an improved hierarchical neural network model as described in claim 4, characterized in that, The contextual semantic features are input into the event relationship modeling layer. Multiple types of graph structures are constructed using time nodes as event nodes in the graph structure. Modeling these multiple types of graph structures yields the power-economic state sequence, including: Each time step's context vector is treated as a node in the graph, forming a node set; the influence relationships between events at different time steps are treated as an edge set; and multiple types of graph structures are constructed based on the node set and the edge set. Features within the multi-type graph structure domain are aggregated to enable the current node to obtain context information of its neighborhood; the node states of the multi-type graph structure are iteratively updated through a gated graph neural network to obtain the final state fusion power-economic state sequence.

6. The power-economic mapping matching method based on an improved hierarchical neural network model as described in claim 5, characterized in that, The power-economic feature matching mechanism preset in the feature matching layer is used to perform feature matching on the power-economic state sequence to obtain the semantic similarity between the current power economic state and the target economic structure, including: Define the current time step's power-economic state sequence as the sequence to be matched, and the historical economic state sequence as the target sequence; When the power system state responds to the economic state in a local time slice, the similarity between the semantic representation vectors at the same position in the two time slices is calculated. If the evolution paths of the two states are similar, the corresponding vectors tend to be the same, and the cosine similarity tends to be close to 1; if the change directions are opposite, the cosine similarity tends to be close to -1; the average similarity is calculated as the local matching score. The two sequences are subjected to average pooling to obtain global state vectors, and the cosine similarity between the two average vectors is calculated. The local and global matching indices are then fused to form a unified matching score.

7. The power-economic mapping matching method based on an improved hierarchical neural network model as described in claim 6, characterized in that, Based on different task requirements, task output structures are constructed. Combining structural features and matching results, the final conclusions of the electricity-economic mapping are derived, including: Based on the two information sources output by the feature matching layer and the original data as input data, economic indicators are predicted using the current power system state. The input data treasure trove includes the structure-enhanced representation vector of the current time point, the matching score between the current state and the target economic structure, and the original economic and power datasets. The cross-entropy loss function is used to determine the degree of matching between the current power system state and the target economic structure. When the label is 1, it means that the state matches the target structure, and when the label is 0, it means that the state does not match the target structure. The economic status is divided into several categories, and the status of any indicator is judged based on the probability distribution of each category.

8. A power-economic mapping matching system based on an improved hierarchical neural network model, applied to the method described in any one of claims 1-7, characterized in that, include: The feature encoding module is used to acquire power data and economic data of the power system, and input the power data and economic data into the feature encoding layer for feature extraction to obtain the contextual semantic features of power-economic data. The event relationship modeling module is used to input the context semantic features into the event relationship modeling layer, construct multi-type graph structures with time nodes as event nodes in the graph structure, and model the multi-type graph structures to obtain the power-economic state sequence. The feature matching module is used to perform feature matching on the power-economic state sequence through a preset power-economic feature matching mechanism to obtain the semantic similarity between the current power economic state and the target economic structure. The output prediction module is used to construct a task output structure based on the semantic similarity and according to different task requirements, and to draw the final conclusion of the power-economy mapping by combining the structural features and matching results.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the power-economy mapping matching method based on the improved hierarchical neural network model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the power-economy mapping matching method based on an improved hierarchical neural network model as described in any one of claims 1 to 7.

Images