CN-122022334-A - Electric power material purchasing demand prediction method and system based on deep learning

CN122022334ACN 122022334 ACN122022334 ACN 122022334ACN-122022334-A

Abstract

The application provides a power material purchasing demand prediction method and a system based on deep learning, and relates to the technical field of material management, wherein the method comprises the steps of acquiring historical purchasing data, project unit operation data and external environment data; and simultaneously, processing external environment and project unit operation data by using a time sequence convolution to obtain time sequence characteristics, splicing the space embedding characteristics and the time sequence characteristic tensor, inputting the space embedding characteristics and the time sequence characteristic tensor into a full-connection deep neural network to perform nonlinear transformation and cross-dimension mapping, and finally outputting a material purchasing demand prediction result in a future period. The prediction accuracy and timeliness of the electric power material purchasing demand are improved.

Inventors

ZHANG CHAOYANG
LI LIN
GUO JING
HU SHENGNAN
ZHAI ZHUOFAN

Assignees

国家能源集团物资有限公司数据科技分公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. The electric power material purchasing demand prediction method based on deep learning is characterized by comprising the following steps of: Acquiring historical material purchasing data, project unit operation data and external environment data; constructing a node connection graph based on project unit organization relations, mapping the historical material purchasing data to corresponding nodes of the node connection graph, and aggregating the characteristics of adjacent nodes in the node connection graph through a graph convolution network to generate node embedding characteristics representing organization dependency relations, wherein the graph convolution network is a branch of deep learning; Performing time sequence convolution processing on the external environment data and the project unit operation data to extract power time sequence characteristics, and performing tensor splicing on the node embedded characteristics and the power time sequence characteristics to form a characteristic fusion tensor; And inputting the feature fusion tensor into a full-connection deep neural network, processing the feature fusion tensor through the full-connection deep neural network, and outputting a purchasing demand prediction result of a specific period in the future, wherein the full-connection deep neural network is the other branch of deep learning.
2. The method of claim 1, wherein the constructing a node connection graph based on project unit organization relationships, mapping the historical material procurement data onto corresponding nodes of the node connection graph, aggregating features of neighboring nodes in the node connection graph through a graph convolution network, and generating node embedded features characterizing organization dependencies comprises: Constructing a node connection diagram taking a project unit as a node and a service or membership as an edge according to the project unit organization relationship; Taking the historical material purchasing data corresponding to each node as an initial feature vector of the node, and loading the initial feature vector onto a corresponding node of the node connection graph to form a feature topological graph; In the characteristic topological graph, the characteristic vector of each node and a first-order adjacent node which is directly connected is aggregated through a first convolution layer of a graph convolution network, so that a first-order neighborhood characteristic is obtained; Aggregating feature vectors of each node and second-order adjacent nodes based on the first-order neighborhood features through a second convolution layer of the graph convolution network to obtain second-order neighborhood features; and fusing the second-order neighborhood characteristics with neighborhood material characteristics of other layers through multi-layer iterative computation of a graph convolution network to generate node embedding characteristics.
3. The method according to claim 2, wherein the generating node embedded features by fusing the second-order neighborhood features with neighborhood material features of other layers through multi-layer iterative computation of a graph convolutional network includes: Inputting the second-order neighborhood characteristics into a third convolution layer of a graph convolution network, and aggregating the material characteristics of each node and third-order adjacent nodes to obtain third-order neighborhood characteristics; Introducing a residual error connection mechanism in the feature fusion process, and performing cross-layer fusion on the first-order neighborhood feature and the third-order neighborhood feature to form a primary fusion feature; Acquiring fourth-order neighborhood characteristics by a fourth convolution layer of the graph convolution network based on the primary fusion characteristics and aggregating the material characteristics of each node and fourth-order adjacent nodes; Establishing a feature pyramid structure, and grouping and fusing neighborhood material features of different levels according to adjacent distances, wherein a first-order neighborhood feature and a second-order neighborhood feature are combined to form a local feature group, and a third-order neighborhood feature and a fourth-order neighborhood feature are combined to form a global feature group; In the feature pyramid structure, calculating importance scores of all nodes based on the purchase amount ratio and the consumption frequency of the nodes; generating gating weight according to the importance score, and adjusting the fusion proportion of the local feature group and the global feature group by using the gating weight through a gating mechanism; And performing feature stitching on the local feature group subjected to the gating weight adjustment and the global feature group subjected to the gating weight adjustment to generate node embedded features.
4. The method of claim 1, wherein the inputting the feature fusion tensor into the fully-connected deep neural network, processing the feature fusion tensor through the fully-connected deep neural network, and outputting the purchasing demand prediction result of the specific future period comprises: inputting the feature fusion tensor into a full-connection depth neural network, and performing first nonlinear transformation on the feature fusion tensor through a first hidden layer of the full-connection depth neural network to obtain a first abstract feature representation; Performing a second nonlinear transformation on the first abstract feature representation through a second hidden layer of the fully connected deep neural network to obtain a second abstract feature representation; performing multidimensional feature interaction on the second abstract feature representation through a third hidden layer of the fully-connected deep neural network to obtain a third abstract feature representation; And carrying out target dimension mapping on the third abstract feature representation through an output layer of the fully-connected deep neural network, and outputting a purchasing demand prediction result of a specific future period.
5. The method of claim 4, wherein the mapping, by the output layer of the fully-connected deep neural network, the target dimension for the third abstract feature representation, and outputting the procurement requirements prediction result for the specific future period comprise: inputting the third abstract feature representation into a plurality of predicted branches of an output layer, respectively; In each prediction branch, mapping the third abstract feature representation to the purchase quantity or the amount space of different material categories respectively through a preset weight matrix to obtain a prediction value of each material; and carrying out weighted fusion on the predicted values of the materials to form a purchasing demand predicted result.
6. The method of claim 1, further comprising, after outputting the procurement requirements forecast result for the future specific period: acquiring a business rule base, wherein the business rule base comprises at least one of a purchasing budget upper limit, a safety stock threshold value, a supplier supply capacity constraint and a purchasing period constraint; And carrying out constraint optimization processing on the purchasing demand prediction result according to the rules in the business rule base to generate an actual executable purchasing plan conforming to the business rules.
7. The method of claim 1, wherein the performing a time-series convolution process on the external environment data and the project unit operation data to extract a power time-series feature comprises: Performing time dimension alignment on the generating capacity and the output sequence in the project unit operation data and the temperature and market price sequences in the external environment data to obtain a time sequence input sequence; The time sequence input sequence is input into a time sequence processing network, wherein the time sequence processing network comprises a plurality of expansion convolution layers which are sequentially connected; performing feature extraction on the first time sequence feature by combining a second expansion convolution layer of the time sequence processing network with a preset matching strategy to obtain a second time sequence feature; Processing the second time sequence feature by a third expansion convolution layer of the time sequence processing network by adopting a cross-modal attention fusion mechanism to obtain a third time sequence feature; and based on a preset evaluation rule, performing feature dimension stitching on the first time sequence feature, the second time sequence feature and the third time sequence feature to form a power time sequence feature.
8. Electric power material purchasing demand prediction system based on deep learning, characterized by comprising: the acquisition module is used for acquiring historical material purchasing data, project unit operation data and external environment data; The construction module is used for constructing a node connection graph based on project unit organization relations, mapping the historical material purchasing data to corresponding nodes of the node connection graph, aggregating the characteristics of adjacent nodes in the node connection graph through a graph convolution network, and generating node embedding characteristics representing organization dependency relations, wherein the graph convolution network is a branch of deep learning; The extraction module is used for carrying out time sequence convolution processing on the external environment data and the project unit operation data so as to extract power time sequence characteristics, and carrying out tensor splicing on the node embedded characteristics and the power time sequence characteristics to form a characteristic fusion tensor; The input module is used for inputting the feature fusion tensor into the full-connection depth neural network, processing the feature fusion tensor through the full-connection depth neural network, and outputting a purchasing demand prediction result in a specific period in the future, wherein the full-connection depth neural network is the other branch of the deep learning.
9. An electronic device, comprising: A memory for storing a computer program; a processor for implementing the steps of the deep learning-based electric power material procurement requirements prediction method according to any one of claims 1 to 7 when executing the computer program.
10. A computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and when executed by a processor, the computer program is capable of implementing the electric power material purchasing demand prediction method based on deep learning as claimed in any one of claims 1 to 7.

Description

Electric power material purchasing demand prediction method and system based on deep learning Technical Field The application relates to the technical field of material management, in particular to a power material purchasing demand prediction method and system based on deep learning. Background In the material supply chain management of large-scale energy enterprises such as national energy resource groups, accurate prediction of material purchasing demands of plates such as thermal power, coal mines, transportation and the like plays an important role in guaranteeing production stability, optimizing inventory, reducing cost and enhancing efficiency. Along with diversification of business plates, wide geographical distribution of project units and changeable market environment, the traditional prediction method has limitations in coping with the requirement conduction among different organization units and the comprehensive influence of production operation data and external environment factors. However, the material prediction scheme based on time sequence analysis has the defects in the aspect of processing the correlation among organizations, particularly the space conduction effect generated by the service, membership and regional synergistic effect among project units on the material consumption cannot be effectively described, and meanwhile, the complex dynamic coupling relationship between the actual operation data of the project units and the external environment data is also insufficiently reflected in the prediction process, so that a relatively obvious deviation exists between the prediction result and the actual demand under the complex business scene of cross-organization and multiple factors. Disclosure of Invention The application provides a power material purchasing demand prediction method and system based on deep learning, which are used for solving the problems of low prediction accuracy and poor timeliness of power material purchasing demands in the prior art. In order to solve the above technical problems, in a first aspect, the present application provides a method for predicting demand for purchasing electric power supplies based on deep learning, including: Acquiring historical material purchasing data, project unit operation data and external environment data; constructing a node connection graph based on project unit organization relations, mapping the historical material purchasing data to corresponding nodes of the node connection graph, and aggregating the characteristics of adjacent nodes in the node connection graph through a graph convolution network to generate node embedding characteristics representing organization dependency relations, wherein the graph convolution network is a branch of deep learning; Performing time sequence convolution processing on the external environment data and the project unit operation data to extract power time sequence characteristics, and performing tensor splicing on the node embedded characteristics and the power time sequence characteristics to form a characteristic fusion tensor; And inputting the feature fusion tensor into a full-connection deep neural network, processing the feature fusion tensor through the full-connection deep neural network, and outputting a purchasing demand prediction result of a specific period in the future, wherein the full-connection deep neural network is the other branch of deep learning. In a second aspect, the present application provides a power material procurement demand prediction system based on deep learning, including: the acquisition module is used for acquiring historical material purchasing data, project unit operation data and external environment data; The construction module is used for constructing a node connection graph based on project unit organization relations, mapping the historical material purchasing data to corresponding nodes of the node connection graph, aggregating the characteristics of adjacent nodes in the node connection graph through a graph convolution network, and generating node embedding characteristics representing organization dependency relations, wherein the graph convolution network is a branch of deep learning; The extraction module is used for carrying out time sequence convolution processing on the external environment data and the project unit operation data so as to extract power time sequence characteristics, and carrying out tensor splicing on the node embedded characteristics and the power time sequence characteristics to form a characteristic fusion tensor; The input module is used for inputting the feature fusion tensor into the full-connection depth neural network, processing the feature fusion tensor through the full-connection depth neural network, and outputting a purchasing demand prediction result in a specific period in the future, wherein the full-connection depth neural network is the other branch of the deep learning. In a third aspect, the prese