Electric power material purchasing demand prediction method and system based on deep learning

Abstract

The invention provides an electric power material purchasing demand prediction method and system based on deep learning, and relates to the technical field of material management, and the method comprises the steps: obtaining historical purchasing data, project unit operation data and external environment data; constructing a node connection graph based on a project unit organization relationship, and aggregating adjacent node features by using a graph convolutional network to generate spatial embedding features; and processing external environment and project unit operation data by using time sequence convolution to obtain time sequence features, splicing space embedded features and time sequence feature tensors, inputting the spliced space embedded features and time sequence feature tensors into a full-connection deep neural network to carry out nonlinear transformation and cross-dimension mapping, and finally outputting a material purchase demand prediction result in a future period. According to the invention, the prediction accuracy and timeliness of the electric power material purchasing demand are improved.

Description

Technical Field

[0001] This application relates to the field of materials management technology, and in particular to a method and system for predicting the procurement demand of power materials based on deep learning. Background Technology

[0002] In the supply chain management of large energy enterprises such as China Energy Investment Corporation, accurately predicting the material procurement needs of sectors such as thermal power, coal mining, and transportation plays an important role in ensuring stable production, optimizing inventory, and reducing costs and increasing efficiency.

[0003] With the diversification of business segments, the wide geographical distribution of project units, and the volatile market environment, traditional forecasting methods have limitations in dealing with the transmission of demand between different organizational units and the combined impact of production and operation data and external environmental factors.

[0004] However, these time-series analysis-based material forecasting schemes have shortcomings in handling inter-organizational relationships. In particular, they fail to effectively characterize the spatial transmission effect of material consumption caused by service and affiliation relationships between project units and regional synergy. At the same time, the complex dynamic coupling relationship between the actual operational data of project units and external environmental data is not fully reflected in the forecasting process. As a result, in complex business scenarios involving multiple organizations and factors, there is a significant deviation between the forecast results and actual needs. Summary of the Invention

[0005] This application provides a method and system for predicting the procurement demand of power materials based on deep learning, in order to solve the problems of low accuracy and poor timeliness in the prediction of procurement demand of power materials in the prior art.

[0006] To address the aforementioned technical problems, firstly, this application provides a deep learning-based method for predicting the procurement demand of power materials, comprising:

[0007] Acquire historical material procurement data, project unit operation data, and external environment data;

[0008] Based on the organizational relationships of project units, a node connection graph is constructed, and the historical material procurement data is mapped to the corresponding nodes in the node connection graph. The features of adjacent nodes in the node connection graph are aggregated through a graph convolutional network to generate node embedding features that represent organizational dependencies. The graph convolutional network is a branch of deep learning.

[0009] Temporal convolution processing is performed on the external environment data and the project unit operation data to extract power time-series features. The node embedding features and the power time-series features are then concatenated using tensors to form a feature fusion tensor.

[0010] The feature fusion tensor is input into a fully connected deep neural network, which processes the feature fusion tensor to output the procurement demand prediction result for a specific future period. The fully connected deep neural network is another branch of deep learning.

[0011] Secondly, this application provides a deep learning-based power material procurement demand forecasting system, comprising:

[0012] The acquisition module is used to acquire historical material procurement data, project unit operation data, and external environment data;

[0013] The module is used to construct a node connection graph based on the organizational relationship of project units, map the historical material procurement data to the corresponding nodes of the node connection graph, and aggregate the features of adjacent nodes in the node connection graph through a graph convolutional network to generate node embedding features that represent organizational dependencies. The graph convolutional network is a branch of deep learning.

[0014] The extraction module is used to perform temporal convolution processing on the external environment data and the project unit operation data to extract power time-series features, and to concatenate the node embedding features with the power time-series features to form a feature fusion tensor.

[0015] The input module is used to input the feature fusion tensor into a fully connected deep neural network, process the feature fusion tensor through the fully connected deep neural network, and output the procurement demand prediction result for a specific future period. The fully connected deep neural network is another branch of deep learning.

[0016] Thirdly, this application provides an electronic device, comprising:

[0017] Memory, used to store computer programs;

[0018] A processor is configured to execute the computer program to implement the steps of the deep learning-based power material procurement demand forecasting method described in the first aspect above.

[0019] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the deep learning-based power material procurement demand forecasting method described in the first aspect above.

[0020] The technical solution provided in this application has the following beneficial effects:

[0021] This application first provides a complete business data foundation for subsequent in-depth analysis by integrating multi-source business data. On this basis, it effectively captures the spatial dependence (or organizational association) relationships between various nodes in the organizational relationship of project units, constructs an inter-organizational transmission model of material demand, and extracts the time dynamic features from operational data and external environment data to capture the temporal change patterns of production activities and the market environment. Furthermore, it realizes the deep interaction between organizational association features and temporal features, and mines potential association patterns through nonlinear transformation, ultimately generating prediction results that meet actual business needs and providing a direct basis for material procurement decisions.

[0022] Furthermore, the method proposed in this application can systematically characterize the spatial transmission effect of material demand in the organizational network of project units, and accurately reflect the impact of changes in organizational hierarchy and collaborative relationships on material consumption through multi-level neighborhood feature aggregation, thereby effectively improving the spatial perception capability of demand forecasting under complex group organizational structures.

[0023] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart illustrating a deep learning-based method for predicting the procurement demand of power materials, provided as an embodiment of this application;

[0026] Figure 2 A schematic diagram illustrating a specific implementation of a deep learning-based method for predicting the procurement demand of power materials, provided in this application embodiment;

[0027] Figure 3 This is a schematic diagram of the structure of a deep learning-based power material procurement demand forecasting system provided in an embodiment of this application. Detailed Implementation

[0028] To address the shortcomings of existing technologies in responding to the impact of organizational structures and diverse business operations, this application proposes a deep learning-based method for predicting energy material procurement demand. The core idea of ​​this method is to construct a predictive model that integrates spatial and temporal features through the fusion of multi-source information, including organizational relationships among project units, production and operation data, and external environmental factors. Specifically, this method first utilizes graph convolutional networks to extract organizational relationship features between nodes in each project unit to capture the impact of organizational structure and business collaboration on material demand. Simultaneously, it uses temporal convolutional networks to analyze the dynamic correlation between production and operation data and external environmental factors. Finally, it achieves comprehensive judgment of multi-dimensional information through feature fusion and deep neural networks. Therefore, this method effectively overcomes the limitations of traditional solutions in responding to changes in organizational structure and complex business factors, thereby improving the accuracy and adaptability of demand forecasting in complex business environments involving multiple sectors and organizations.

[0029] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] The core of this application is to provide a method for predicting the procurement demand of power materials based on deep learning. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0031] Step 101: Obtain historical material procurement data, project unit operation data, and external environment data.

[0032] In step 101, historical material procurement data refers to the quantity, amount, and time information of various materials procured by project units within the group, such as thermal power plants, coal mines, and transportation sections, in the past period, including limestone, valves, anchor bolts, and fans; project unit operation data refers to the core indicators reflecting the production activities of each project unit, such as the power generation and utilization hours of thermal power plants, the raw coal output and tunneling footage of coal mines, and the railway transportation volume and travel kilometers of transportation sections; external environment data refers to data such as temperature and humidity provided by meteorological departments, as well as external impact data such as market prices and policy indices that change over time.

[0033] For example, taking the limestone forecasting task of a thermal power plant as an example, the historical procurement data from 2020 to 2022, including the limestone procurement quantity and procurement amount, were first obtained from the material procurement system. At the same time, the actual power generation sequence of the power plant's units was obtained from the production management system. In addition, daily average temperature, humidity and local limestone market price data were obtained from external data sources. All data were aligned to the time on a monthly basis.

[0034] Step 102: Based on the organizational relationship of the project units, construct a node connection graph, map the historical material procurement data to the corresponding nodes of the node connection graph, and aggregate the features of adjacent nodes in the node connection graph through a graph convolutional network to generate node embedding features that represent organizational dependencies. The graph convolutional network is a branch of deep learning.

[0035] In step 102, the node connection graph is a network graph composed of project units as nodes and service relationships, affiliation relationships or geographical proximity relationships as edges. Nodes represent the main body of material demand, and edges represent the relationship between organizations. The node embedding feature is a feature representation of organizational dependency relationship generated after integrating multi-level neighborhood information.

[0036] It should be noted that this embodiment does not impose specific limitations on the specific structure or training process of the graph convolutional network, and can be set accordingly based on the actual situation.

[0037] In this embodiment, a graph structure is first constructed based on the service affiliation between project units and procurement agencies within the group, and the historical procurement data of each node is loaded into the graph as node features. Then, the features of each node and its direct neighbors are aggregated through the first layer of the graph convolutional network to obtain first-order neighborhood features. Then, second-order neighbor features are aggregated based on the first-order features through the second layer to obtain second-order neighborhood features. Finally, node embedding features containing complete multi-order neighborhood information are generated through iterative fusion of multi-layer networks.

[0038] Step 103: Perform temporal convolution processing on the external environment data and the project unit operation data to extract power time-series features, and concatenate the node embedding features with the power time-series features to form a feature fusion tensor.

[0039] In step 103, the feature fusion tensor is a multidimensional data tensor formed by splicing organizational correlation features and temporal features.

[0040] In this embodiment, the monthly power generation sequence of the power plant, the local monthly average temperature sequence, and the market price sequence are first time-aligned to form a unified time-aligned input sequence. Then, the short-term time-series pattern of the device operation is extracted through the first layer of the temporal convolutional network, the medium-term time-series pattern under the influence of environmental factors is extracted through the second layer, and the long-term time-series pattern at the system level is extracted through the third layer by fusing multi-source features. Finally, the time-series features of these three levels are concatenated and tensor-concatenated with the node embedding features generated in the previous steps to form the final feature fusion tensor.

[0041] Step 104: Input the feature fusion tensor into a fully connected deep neural network, process the feature fusion tensor through the fully connected deep neural network, and output the procurement demand prediction result for a specific future period. The fully connected deep neural network is another branch of deep learning.

[0042] In step 104, the future specific period refers to a continuous time period pre-set according to the material procurement management needs. The procurement demand forecast result is the predicted procurement quantity or amount of various materials within the future specific period output by the neural network.

[0043] It should be noted that this embodiment does not impose specific limitations on the specific structure or training process of the fully connected deep neural network, and can be set accordingly based on the actual situation.

[0044] In this embodiment, the feature fusion tensor is input into a fully connected deep neural network. Basic nonlinear transformation is performed through the first hidden layer, and then a more complex nonlinear transformation is performed through the second hidden layer. Subsequently, the interaction between multi-dimensional features is realized in the third hidden layer. Finally, the high-dimensional features processed above are mapped to the material prediction space through the output layer to output the specific predicted quantity or amount of different material categories or total procurement amount.

[0045] The embodiments of this application construct a demand forecasting model that comprehensively considers spatial dependence and temporal variation patterns, thereby improving the accuracy and practicality of power material procurement demand forecasting.

[0046] Optionally, such as Figure 2 As shown, step 102: Based on the organizational relationships of project units, construct a node connection graph, map the historical material procurement data to the corresponding nodes in the node connection graph, and aggregate the features of adjacent nodes in the node connection graph through a graph convolutional network to generate node embedding features representing organizational dependencies, including:

[0047] Step 201: Based on the organizational relationships of the project units, construct a node connection graph with project units as nodes and service or affiliation relationships as edges.

[0048] In this embodiment, the organizational attributes and hierarchical information of all project units and procurement agencies are first identified based on the group's organizational structure data and business service relationships, and a corresponding set of nodes is established. Then, based on the actual service coverage, management affiliation, or geographical region collaboration relationships, edges are established between the interconnected organizational nodes, ultimately forming a complete node connection graph.

[0049] Step 202: Use the historical material procurement data corresponding to each node as the initial feature vector of the node, and load the initial feature vector onto the corresponding node of the node connection graph to form a feature topology graph.

[0050] In step 202, the initial feature vector is a numerical vector composed of historical material procurement data corresponding to each organization node; the feature topology graph is a node connection graph loaded with these material feature vectors, which allows each node in the topology graph to carry the corresponding material consumption information.

[0051] In this embodiment, the quantity or amount of key materials purchased by each node in the past period is first used as the initial feature vector of the node. For example, the feature vector of a thermal power plant node can be represented as [tons of limestone, amount of valves, liters of lubricating oil]. Then, these feature vectors are loaded onto each node of the node connection graph according to the node correspondence, thereby forming a feature topology graph that contains both organizational relationships and material feature information.

[0052] Step 203: In the feature topology graph, the feature vectors of each node and its directly connected first-order neighboring nodes are aggregated through the first convolutional layer of the graph convolutional network to obtain the first-order neighborhood features.

[0053] In step 203, first-order adjacent nodes are the neighboring nodes that are directly connected to the current node in the node connection graph; the first-order neighborhood feature is a comprehensive feature representation obtained by aggregating the feature vectors of these directly adjacent nodes, reflecting the material consumption of the nearest neighboring region of the current node.

[0054] In this embodiment of the application, in the feature topology graph, each node is processed by the first convolutional layer of the graph convolutional network. First, the feature vectors of all directly connected first-order neighboring nodes of the node are collected. Then, a weighted average algorithm is used to aggregate the feature vectors of these neighboring nodes to obtain first-order neighborhood features containing direct neighborhood information.

[0055] Step 204: Using the second convolutional layer of the graph convolutional network, based on the first-order neighborhood features, aggregate the feature vectors of each node and its second-order neighboring nodes to obtain the second-order neighborhood features.

[0056] In step 204, a second-order adjacency node refers to a node that is indirectly connected through an intermediate node in the organizational graph. That is, a node needs to go through two edges to reach an adjacent node. These nodes constitute the second-level neighbor relationship of the node.

[0057] Second-order neighborhood features are extended feature representations obtained by further aggregating the features of these indirectly adjacent nodes based on first-order neighborhood features. They reflect the impact of material consumption on organizational units that are far from the current node.

[0058] In this embodiment, the first-order neighborhood features are input into the second convolutional layer of the graph convolutional network. Based on each node, the feature vectors of the second-order adjacent nodes that are two edges away from the node are further collected. Based on the closeness of business association between nodes, such as the frequency of material allocation and the degree of supply chain collaboration, different aggregation weights are set. Then, the feature vectors of the second-order adjacent nodes are fused with the first-order neighborhood features to obtain the second-order neighborhood features.

[0059] Step 205: Through multi-layer iterative computation of the graph convolutional network, the second-order neighborhood features are fused with the neighborhood material features of other levels to generate node embedding features.

[0060] In step 205, other levels refer to higher-order adjacency relationships besides first-order and second-order, including third-order adjacency nodes, fourth-order adjacency nodes, etc., where third-order adjacency nodes need to be connected by three edges, and fourth-order adjacency nodes need to be connected by four edges. These higher-order adjacency relationships together constitute the multi-hop neighborhood structure of the node, which is used to capture a wider range of organizational network influences.

[0061] In this embodiment, second-order neighborhood features are input into subsequent convolutional layers through multi-layer iterative computation of graph convolutional networks to continue aggregating higher-order neighbor node features. During feature fusion, a residual connection mechanism is used to retain shallow feature information. Through multi-layer feature transfer and fusion, node embedding features containing multi-scale organizational association information are finally generated.

[0062] It should be noted that this embodiment does not impose specific limitations on the specific structural design of each layer of the graph convolutional network, and can be set accordingly based on the actual situation.

[0063] In this embodiment, a complete model of material demand transmission in the project unit organizational network is established through multi-level neighborhood feature aggregation, which improves the accuracy and adaptability of organizational collaborative perception of demand forecasting under complex group organizational structures and diversified business segments.

[0064] To further improve the ability of node embedding features to represent multi-level dependencies in organizational networks, in some embodiments, step 205: through multi-layer iterative computation of a graph convolutional network, the second-order neighborhood features are fused with neighborhood material features at other levels to generate node embedding features, including:

[0065] Step 301: Input the second-order neighborhood features into the third convolutional layer of the graph convolutional network, and aggregate the material features of each node and its third-order neighboring nodes to obtain the third-order neighborhood features.

[0066] In step 301, a third-order adjacency node is a node that is indirectly connected to the current node through two intermediate nodes in the node connection graph; the third-order neighborhood feature is an extended feature representation obtained by aggregating the features of these indirectly connected nodes that are far apart, reflecting the impact of the current node on material consumption in a broader organizational network.

[0067] In this embodiment, second-order neighborhood features are input into the third convolutional layer of the graph convolutional network. Based on each node, feature vectors of third-order adjacent nodes that are three edges away from the node are collected. The corresponding aggregation weights are set according to the business collaboration strength or management span between nodes. Then, the feature vectors of the third-order adjacent nodes are weighted and fused to obtain third-order neighborhood features.

[0068] Step 302: In the feature fusion process, a residual connection mechanism is introduced to perform cross-layer fusion of the first-order neighborhood features and the third-order neighborhood features to form primary fused features.

[0069] In step 302, the primary fusion feature is an intermediate feature representation obtained by cross-layer fusion of first-order neighborhood features and third-order neighborhood features, which retains both local detail information and global perspective.

[0070] In this embodiment, a residual connection mechanism is introduced during the feature fusion process. The first-order neighborhood features output by the first convolutional layer are connected across layers with the third-order neighborhood features output by the third convolutional layer. The local features and global features are initially fused by feature weighted summation to form a primary fused feature.

[0071] Step 303: Using the fourth convolutional layer of the graph convolutional network, based on the primary fusion features, and aggregating the material features of each node and its fourth-order neighboring nodes, the fourth-order neighborhood features are obtained.

[0072] In step 303, the fourth-order adjacency node is a node that is indirectly connected to the current node through three intermediate nodes in the node connection graph; the fourth-order neighborhood feature is a global feature representation obtained by further aggregating the features of these farthest nodes based on the primary fusion feature, reflecting the impact of the current node's material consumption within the entire organizational network.

[0073] In this embodiment, the primary fusion features are extracted through the fourth convolutional layer of the graph convolutional network. At the same time, the material features of each node and its fourth-order neighboring nodes are aggregated. Different feature aggregation weights are set according to the hierarchy and management relationship of the organizational network, and finally, the fourth-order neighborhood features containing the most global information are obtained.

[0074] Step 304: Establish a feature pyramid structure and group and merge the neighborhood material features of different levels according to their adjacency distance. The first-order neighborhood features and the second-order neighborhood features are combined into a local feature group, and the third-order neighborhood features and the fourth-order neighborhood features are combined into a global feature group.

[0075] In step 304, the feature pyramid structure is a multi-level feature architecture that groups and organizes features at different levels according to their receptive field size. The feature pyramid structure and the graph convolutional network are independent and work together.

[0076] Local feature groups are sets of detailed features composed of first- and second-order neighborhood features, while global feature groups are sets of macroscopic features composed of third- and fourth-order neighborhood features. Adjacency distance refers to the number of edges traversed by the shortest path between two nodes in an organizational graph. This distance is calculated on the node connection graph using the shortest path algorithm in graph theory.

[0077] In this embodiment of the application, a feature pyramid structure is established, and the neighborhood material features of different levels are grouped according to their adjacency distance. The first-order neighborhood features and the second-order neighborhood features are divided into local feature groups, while the third-order neighborhood features and the fourth-order neighborhood features are divided into global feature groups, thereby forming a multi-level feature organization system.

[0078] Step 305: In the feature pyramid structure, calculate the importance score of each node based on the proportion of the purchase amount and the consumption frequency of the node.

[0079] In step 305, the procurement amount percentage refers to the proportion of the historical procurement amount of this node in the procurement of its respective procurement agency or the overall procurement of the group, reflecting the importance of the procurement scale of this node; the consumption frequency refers to the frequency of material outbound or requisition by this node per unit time, reflecting the operational activity and material dependence of this node.

[0080] The importance score is a quantitative score of the node's importance calculated by combining the proportion of the total procurement amount and the frequency of consumption.

[0081] In this embodiment of the application, in the feature pyramid structure, the importance score of each node in the organizational network is calculated based on the proportion of the purchase amount and the consumption frequency of the node. The proportion of the purchase amount is assessed based on the node's historical purchase share within its respective purchasing agency, while the consumption frequency is determined based on the density of the node's material outbound records. Finally, the two indicators are weighted and averaged to obtain the final importance score.

[0082] Step 306: Generate gating weights based on the importance score, and adjust the fusion ratio of local feature groups and global feature groups through a gating mechanism using the gating weights.

[0083] In step 306, the gating weights are control parameters generated based on importance scores to adjust the feature fusion ratio.

[0084] In this embodiment, a gating weight is first generated based on the importance score. The gating mechanism is used to adjust the fusion ratio of local feature groups and global feature groups. That is, for nodes with high importance, a larger weight of global feature groups is given during feature fusion, while for nodes with low importance, they are made to rely more on the information of local feature groups.

[0085] The specific implementation process involves using the sigmoid function to map the importance score to the interval between 0 and 1 as the weight of the global feature group, while the weights of the local feature groups are their complements. The calculation formula is: Global weight = ,in This is the slope parameter, such as a value of 10. For example, importance score. At that time, global weight = The local weight is 0.06.

[0086] Step 307: Concatenate the local feature group and the global feature group after gating weight adjustment to generate node embedding features.

[0087] In this embodiment, the local feature group and the global feature group, which have been adjusted by gating weights, are spliced ​​together. The local feature group retains the detailed material consumption pattern of the neighboring area of ​​the node, while the global feature group provides the overall impact information within the organizational network. Through this feature splicing, node embedding features containing multi-scale spatial information are finally generated.

[0088] In this embodiment, a feature representation method that can dynamically adjust the weights of local and global information is constructed through multi-level neighborhood feature fusion and adaptive gating mechanism, which improves the representation accuracy of node embedding features on multi-level dependencies of organizational networks, thereby providing more accurate and comprehensive organizational association feature support for subsequent demand prediction.

[0089] To further improve the accuracy and reliability of power material procurement demand forecasting, in some embodiments, step 104 involves inputting the feature fusion tensor into a fully connected deep neural network, processing the feature fusion tensor through the fully connected deep neural network, and outputting the procurement demand forecast result for a specific future period, including:

[0090] Step 401: Input the feature fusion tensor into a fully connected deep neural network, and perform a first nonlinear transformation on the feature fusion tensor through the first hidden layer of the fully connected deep neural network to obtain a first abstract feature representation.

[0091] In step 401, the first nonlinear transformation is the process of nonlinearly transforming the input features through the activation function in the neural network; the first abstract feature representation is the intermediate feature obtained after the first nonlinear transformation, which contains the preliminary abstraction and combination information of the input features.

[0092] In this embodiment, the feature fusion tensor is input into the first hidden layer of the fully connected deep neural network. Through the fully connected relationships between neurons in this layer and the action of the activation function, the input features are linearly combined and nonlinearly transformed to extract the preliminary correlation patterns between features and obtain the first abstract feature representation with reduced dimensionality.

[0093] Step 402: The first abstract feature representation is subjected to a second nonlinear transformation through the second hidden layer of the fully connected deep neural network to obtain the second abstract feature representation.

[0094] In step 402, the second nonlinear transformation is a process of further deepening the feature abstraction based on the first transformation; the second abstract feature representation is a deeper level feature obtained after the second nonlinear transformation, which contains more complex feature interaction information.

[0095] In this embodiment, the first abstract feature representation is further processed by the second hidden layer of a fully connected deep neural network. By utilizing the denser neuron connections and activation functions of this layer, the first-level features are combined and abstracted at a deeper level, thereby uncovering more complex nonlinear relationships between features and obtaining a further dimensionality-reduced second abstract feature representation.

[0096] Step 403: Through the third hidden layer of the fully connected deep neural network, perform multi-dimensional feature interaction on the second abstract feature representation to obtain the third abstract feature representation.

[0097] In step 403, the third abstract feature representation is the final abstract feature obtained after multi-dimensional feature interaction, which contains comprehensive information after deep fusion of organizational correlation features and temporal features;

[0098] In this embodiment of the application, the second abstract feature representation is subjected to multi-dimensional feature interaction through the third hidden layer of the fully connected deep neural network. In this layer, the deep cross-fusion of organizational dependence features and temporal change features is realized, and the deep correlation between different dimensional features is captured by the complex connection relationship between neurons, thereby obtaining a highly abstract third abstract feature representation.

[0099] Step 404: Through the output layer of the fully connected deep neural network, the target dimension is mapped to the third abstract feature representation, and the procurement demand prediction result for a specific future period is output.

[0100] In this embodiment, the output layer of a fully connected deep neural network maps the third abstract feature representation to the target dimension. The linear transformation function of the output layer is used to transform the high-dimensional feature space into the material procurement quantity or amount space. By setting output neurons corresponding to the predicted target dimension, the predicted values ​​of procurement quantity or procurement amount for different material categories such as limestone, lubricating oil, and fans are directly output.

[0101] In the embodiments of this application, a precise conversion from a complex feature space to a specific prediction target is achieved through multi-level nonlinear transformation and cross-dimensional feature mapping.

[0102] To further improve the adaptability of the prediction results to operational risks, in some embodiments, step 404: mapping the third abstract feature representation to the target dimension through the output layer of the fully connected deep neural network, and outputting the procurement demand prediction results for a specific future period, including:

[0103] Step 501: Input the third abstract feature representation into multiple prediction branches of the output layer respectively.

[0104] In step 501, the prediction branch is an independent computation path set up in the output layer of the fully connected deep neural network for different material categories or prediction dimensions. Each branch performs specialized calculations for a specific prediction target.

[0105] In the embodiments of this application, the third abstract feature is represented as multiple prediction branches of the input and output layers respectively, so as to realize the special prediction of different material categories such as bulk materials and spare parts, or different prediction dimensions such as purchase quantity and purchase amount.

[0106] Step 502: In each prediction branch, the third abstract feature representation is mapped to the purchase quantity or amount space of different material categories through a preset weight matrix to obtain the predicted value of each material.

[0107] In this embodiment of the application, in each prediction branch, the third abstract feature representation is linearly transformed by the preset weight matrix of the branch and mapped to its corresponding prediction space, such as the space of limestone purchase tonnage and the space of valve purchase amount, to obtain the prediction value of the branch. Each weight matrix is ​​learned by backpropagation algorithm during the training of the fully connected deep neural network.

[0108] Step 503: Weight and merge the predicted values ​​of each material to form the procurement demand prediction result.

[0109] It should be understood that step 503 includes steps b1 to b2, wherein step b1: calculates the dynamic adjustment coefficient based on the project unit's real-time operational data and external market fluctuation data, dynamically adjusts the output of each prediction branch based on the dynamic adjustment coefficient, and calculates the attention weight of each material category or prediction dimension.

[0110] In step a1, the real-time operational data of the project unit includes indicators reflecting production intensity such as power generation, output, and equipment operating load rate; external market fluctuation data includes key raw material price indices and market demand change rates; the dynamic adjustment coefficient is a quantitative value of risk or opportunity calculated by combining operational data and market fluctuation data; and the attention weight is the importance weight of each prediction target in the final fusion, determined based on the dynamic adjustment coefficient.

[0111] In this embodiment, real-time operational data and external market fluctuation data are first acquired, and dynamic adjustment coefficients corresponding to each material are calculated through a predefined fusion model. Then, these coefficients are converted into attention weights through the softmax function, thereby realizing the function of dynamically adjusting the prediction focus according to the real-time business status and external environment.

[0112] For example, when the power plant's load factor increases significantly and the limestone market price fluctuates more, the dynamic adjustment coefficient of the limestone forecast branch increases, and its corresponding attention weight increases accordingly, making the final forecast result more focused on the accurate prediction of limestone demand.

[0113] Step a2: Based on the attention weights, the predicted values ​​of each branch are weighted and fused to form the procurement demand prediction result.

[0114] In this embodiment of the application, the predicted values ​​of each branch are weighted and summed based on the calculated attention weights to form the final comprehensive procurement demand prediction result.

[0115] In this embodiment, the prediction model is adaptively adjusted to dynamic business environment and external risks through a multi-branch prediction structure and an attention mechanism based on real-time business data.

[0116] Optionally, after outputting the procurement demand forecast for a specific future period, the following may also be included:

[0117] Step b1: Obtain the business rule base, which includes at least one of the following: procurement budget limit, safety stock threshold, supplier supply capacity constraint, and procurement cycle constraint.

[0118] In step b1, the business rules base is a set of predefined rigid or flexible conditions used to ensure the feasibility and compliance of the procurement plan; the procurement budget limit refers to the maximum amount of procurement set for a specific project unit, material category or time period; the safety stock threshold is the minimum inventory level set to prevent stockouts; the supplier supply capacity constraint reflects the maximum quantity of materials that the supplier can stably provide within the agreed period; and the procurement cycle constraint specifies the shortest or standard time required from plan submission to material warehousing.

[0119] In this embodiment, management strategies scattered across multiple business systems are centralized and structured. Specifically, the process begins by extracting annual / monthly budget data for each project unit or material category from the financial management system, obtaining current inventory levels and preset safety stock strategies from the warehouse management system, analyzing and determining the reliable supply capabilities of each supplier from the supplier management system or historical cooperation data, and clarifying the standard procurement lead time for various materials from the procurement process. Subsequently, these multi-source, heterogeneous business parameters are formatted according to preset rule templates and stored in a unified rule base, forming a dynamically queryable and flexibly configurable business constraint knowledge base, providing clear input conditions for subsequent optimization calculations.

[0120] For example, taking the valve procurement scenario of a thermal power plant as an example, the system retrieves information from the finance module that "the power plant's maximum budget for valve procurement in the third quarter of 2024 is 1.2 million yuan"; from the inventory module, it obtains information that "currently, the power plant has 15 units of key valve models in stock, while the safety stock level determined according to the equipment maintenance plan should be 30 units"; from the supplier evaluation module, it extracts information that "core supplier A's maximum committed supply of commonly used valves this quarter is 200 units"; and, based on the procurement process, it sets "the standard procurement cycle for this type of valve (from order placement to warehousing) is 20 working days." The system integrates these parameters to form a complete business rule for "a power plant - valves," containing multi-dimensional constraints, and stores it in the rule library.

[0121] Step b2: Based on the rules in the business rule base, perform constraint optimization on the procurement demand forecast results to generate an actual executable procurement plan that conforms to the business rules.

[0122] In this embodiment, the initial procurement demand forecast is first used as the basis for the optimization objective, such as the predicted quantity or amount of various materials. Then, all relevant constraints in the business rule base are transformed into mathematical conditions for the optimization model, such as transforming the procurement budget limit into a total amount inequality constraint, the safety stock threshold into an inventory balance equation, and the supplier capacity into a supply limit constraint. Then, a suitable optimization algorithm, such as linear programming, integer programming, or a heuristic rule engine, is used to solve the model. The algorithm searches within the "feasible region" composed of all constraints to find the procurement quantity combination that maximizes the achievement of the forecast objective. Finally, the series of adjusted procurement quantities obtained constitute the actual executable procurement plan.

[0123] To further improve the ability of time-series features to represent complex operating modes of power systems, in some embodiments, step 103 involves performing time-series convolution processing on the external environment data and the project unit's operating data to extract power time-series features, including:

[0124] Step 701: Align the power generation and output sequences in the project unit's operational data with the temperature and market price sequences in the external environment data in terms of time dimension to obtain the time-series input sequence.

[0125] In step 701, the time series input sequence is a unified data sequence formed by aligning multi-source time series data in the time dimension, providing regular input data for subsequent time series analysis.

[0126] In this embodiment of the application, the key indicator sequence in the external environment data is first aligned with the project unit's operational data sequence in terms of time dimension, and the integrity of the timestamps of each data sequence is checked. Missing time points are interpolated to ensure that all data have corresponding values ​​at the same time point, and finally a regular time-series input sequence is formed.

[0127] Step 702: The temporal input sequence is input into a temporal processing network, which includes multiple dilated convolutional layers connected in sequence. Features are extracted from the temporal input sequence through the first dilated convolutional layer of the temporal processing network to obtain first temporal features.

[0128] In step 702, the first time-series feature is a feature representation extracted from the time-series input sequence that reflects the change law of the operating state of power equipment over time.

[0129] In this embodiment of the application, the time-series input sequence is input into the time-series processing network, and the input sequence is convolved by the first dilated convolutional layer. The large receptive field characteristic of the dilated convolution is used to capture the long-term dependency relationship of the device's operating state, thereby extracting the feature pattern that characterizes the device's operating rules, and finally obtaining the first time-series feature.

[0130] Step 703: Through the second dilated convolutional layer of the temporal processing network, combined with a preset matching strategy, feature extraction is performed on the first temporal feature to obtain the second temporal feature.

[0131] In step 703, the preset matching strategy refers to matching the current external environment data with the historical patterns of the same period to identify the impact of recurring external environment patterns on production operations and material consumption; the second time series feature is a feature representation that is closely related to environmental factors extracted from the first time series feature.

[0132] In this embodiment of the application, the second dilated convolutional layer of the temporal processing network, combined with a preset matching strategy, firstly calculates the similarity between the current meteorological data and the historical meteorological patterns of the same period, and then adjusts the weight allocation of the convolutional kernel based on the similarity calculation results, thereby separating the feature components that are strongly related to environmental factors from the first temporal features, and finally obtaining the second temporal features.

[0133] Step 704: The second temporal feature is processed through the third dilated convolutional layer of the temporal processing network and a cross-modal attention fusion mechanism is adopted to obtain the third temporal feature.

[0134] In step 704, the third time-series feature is a feature representation that reflects the overall operating status of the business system and external influences after fusing multi-source information.

[0135] In this embodiment, the cross-modal attention fusion mechanism is adopted through the third dilated convolutional layer of the temporal processing network to calculate the cross-attention weight between device operation features and environmental features. Then, the two types of features are dynamically weighted and fused according to the weight, thereby extracting system-level features that contain both device operation information and environmental influence, and finally obtaining the third temporal feature.

[0136] Step 705: Based on the preset evaluation rules, the first time-series feature, the second time-series feature, and the third time-series feature are concatenated by feature dimensions to form power time-series features.

[0137] In step 705, the preset evaluation rule refers to a comprehensive evaluation criterion based on feature variance and mutual information. Specifically, it involves calculating the variance of each feature dimension and the mutual information with the target variable, selecting features with a variance greater than 0.1 and a mutual information greater than 0.05 to participate in the final assembly, ensuring that features with discriminative power and relevance can be retained.

[0138] It should be noted that this embodiment does not specifically limit the content of the evaluation rules, and can be set accordingly based on the actual situation.

[0139] In this embodiment, based on preset evaluation rules, the importance score of each feature dimension in the first, second, and third time-series features is first calculated. Then, important feature dimensions are selected based on the score results. Finally, the selected features are concatenated to form power time-series features that retain key information while avoiding dimensional redundancy. The concatenation step is performed after the time-series processing network has completed the processing of all dilated convolutional layers and is not part of the internal computation of the time-series processing network.

[0140] In this embodiment, a time-series feature representation that can simultaneously capture equipment operating patterns, environmental influencing factors, and the overall system state is constructed through multi-level time-series feature extraction and intelligent feature fusion. This enhances the power time-series features' ability to represent complex operating scenarios, thereby providing richer time-series information support for subsequent material demand forecasting.

[0141] Figure 3 A schematic diagram of a deep learning-based power material procurement demand forecasting system provided in this application embodiment is shown below. Figure 3 As shown, the system includes:

[0142] The acquisition module 31 is used to acquire historical material procurement data, project unit operation data, and external environment data.

[0143] The construction module 32 is used to construct a node connection graph based on the organizational relationship of project units, map the historical material procurement data to the corresponding nodes of the node connection graph, and aggregate the features of adjacent nodes in the node connection graph through a graph convolutional network to generate node embedding features that represent organizational dependencies. The graph convolutional network is a branch of deep learning.

[0144] The extraction module 33 is used to perform temporal convolution processing on the external environment data and the project unit operation data to extract power time-series features, and to concatenate the node embedding features with the power time-series features to form a feature fusion tensor.

[0145] The input module 34 is used to input the feature fusion tensor into a fully connected deep neural network, process the feature fusion tensor through the fully connected deep neural network, and output the procurement demand prediction result for a specific future period. The fully connected deep neural network is another branch of deep learning.

[0146] The deep learning-based power material procurement demand forecasting system of this application embodiment is used to implement the aforementioned deep learning-based power material procurement demand forecasting method. Therefore, the specific implementation of the deep learning-based power material procurement demand forecasting system can be found in the embodiment section of the deep learning-based power material procurement demand forecasting method above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0147] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the deep learning-based power material procurement demand forecasting method described above.

[0148] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described deep learning-based methods for predicting the procurement demand of power materials.

[0149] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0150] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the deep learning-based power material procurement demand prediction method.

[0151] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0152] The above provides a detailed description of a deep learning-based method and system for predicting power material procurement demand. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of these embodiments are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for predicting power material procurement demand based on deep learning, characterized in that, include: Acquire historical material procurement data, project unit operation data, and external environment data; Based on the organizational relationships of project units, a node connection graph is constructed, and the historical material procurement data is mapped to the corresponding nodes in the node connection graph. The features of adjacent nodes in the node connection graph are aggregated through a graph convolutional network to generate node embedding features that represent organizational dependencies. The graph convolutional network is a branch of deep learning. Temporal convolution processing is performed on the external environment data and the project unit operation data to extract power time-series features. The node embedding features and the power time-series features are then concatenated using tensors to form a feature fusion tensor. The feature fusion tensor is input into a fully connected deep neural network, which processes the feature fusion tensor to output the procurement demand prediction result for a specific future period. The fully connected deep neural network is another branch of deep learning.

2. The method according to claim 1, characterized in that, The process involves constructing a node connection graph based on the organizational relationships of project units, mapping the historical material procurement data to corresponding nodes in the node connection graph, and aggregating the features of adjacent nodes in the node connection graph using a graph convolutional network to generate node embedding features representing organizational dependencies, including: Based on the organizational relationships of the project units, construct a node connection graph with project units as nodes and service or affiliation relationships as edges; The historical material procurement data corresponding to each node is used as the initial feature vector of the node. The initial feature vector is loaded onto the corresponding node of the node connection graph to form a feature topology graph. In the feature topology graph, the feature vectors of each node and its directly connected first-order neighboring nodes are aggregated through the first convolutional layer of the graph convolutional network to obtain the first-order neighborhood features. By using the second convolutional layer of the graph convolutional network, based on the first-order neighborhood features, the feature vectors of each node and its second-order neighboring nodes are aggregated to obtain the second-order neighborhood features; Through multi-layer iterative computation of graph convolutional networks, the second-order neighborhood features are fused with neighborhood material features at other levels to generate node embedding features.

3. The method according to claim 2, characterized in that, The process involves multi-layer iterative computation using a graph convolutional network to fuse the second-order neighborhood features with neighborhood material features from other levels, generating node embedding features, including: The second-order neighborhood features are input into the third convolutional layer of the graph convolutional network, and the material features of each node and its third-order neighboring nodes are aggregated to obtain the third-order neighborhood features. A residual connection mechanism is introduced during the feature fusion process to fuse the first-order neighborhood features and the third-order neighborhood features across layers to form primary fused features; By using the fourth convolutional layer of the graph convolutional network, based on the primary fusion features, and aggregating the material features of each node and its fourth-order neighboring nodes, fourth-order neighborhood features are obtained. A feature pyramid structure is established, and the neighborhood material features of different levels are grouped and fused according to their adjacency distance. The first-order neighborhood features and the second-order neighborhood features are combined into a local feature group, and the third-order neighborhood features and the fourth-order neighborhood features are combined into a global feature group. In the feature pyramid structure, the importance score of each node is calculated based on the proportion of the purchase amount and the consumption frequency of the node. Based on the importance score, a gating weight is generated, and the fusion ratio of local feature groups and global feature groups is adjusted using the gating mechanism. The local feature group and the global feature group, which have been adjusted by gating weights, are concatenated to generate node embedding features.

4. The method according to claim 1, characterized in that, The step of inputting the feature fusion tensor into a fully connected deep neural network, processing the feature fusion tensor through the fully connected deep neural network, and outputting the procurement demand forecast result for a specific future period includes: The feature fusion tensor is input into a fully connected deep neural network. The first nonlinear transformation of the feature fusion tensor is performed through the first hidden layer of the fully connected deep neural network to obtain the first abstract feature representation. The first abstract feature representation is subjected to a second nonlinear transformation through the second hidden layer of the fully connected deep neural network to obtain the second abstract feature representation. The third abstract feature representation is obtained by performing multi-dimensional feature interaction on the second abstract feature representation through the third hidden layer of the fully connected deep neural network. The output layer of the fully connected deep neural network is used to map the target dimension of the third abstract feature representation, and output the procurement demand prediction result for a specific future period.

5. The method according to claim 4, characterized in that, The step of mapping the third abstract feature representation to the target dimension through the output layer of the fully connected deep neural network to output the procurement demand forecast result for a specific future period includes: The third abstract feature is represented as input to multiple prediction branches of the output layer; In each prediction branch, the third abstract feature representation is mapped to the purchase quantity or amount space of different material categories through a preset weight matrix to obtain the predicted value of each material. The predicted values ​​of each material are weighted and merged to form the procurement demand forecast result.

6. The method according to claim 1, characterized in that, After outputting the procurement demand forecast results for a specific future period, it also includes: Obtain a business rule base, which includes at least one of the following: procurement budget limit, safety stock threshold, supplier supply capacity constraint, and procurement cycle constraint; Based on the rules in the business rule base, the procurement demand forecast results are constrained and optimized to generate a procurement plan that conforms to the business rules and is actually executable.

7. The method according to claim 1, characterized in that, The step of performing time-series convolution processing on the external environment data and the project unit's operational data to extract power time-series features includes: The power generation and output sequences in the project unit's operational data are aligned with the temperature and market price sequences in the external environment data to obtain a time-series input sequence. The temporal input sequence is input into a temporal processing network, which includes multiple dilated convolutional layers connected in sequence; features are extracted from the temporal input sequence through the first dilated convolutional layer of the temporal processing network to obtain first temporal features; The second temporal feature is obtained by extracting features from the first temporal feature through the second dilated convolutional layer of the temporal processing network and in combination with a preset matching strategy. The second temporal feature is processed through the third dilated convolutional layer of the temporal processing network, and a cross-modal attention fusion mechanism is used to obtain the third temporal feature. Based on preset evaluation rules, the first time-series feature, the second time-series feature, and the third time-series feature are concatenated by feature dimensions to form power time-series features.

8. A power material procurement demand forecasting system based on deep learning, characterized in that, include: The acquisition module is used to acquire historical material procurement data, project unit operation data, and external environment data; The module is used to construct a node connection graph based on the organizational relationship of project units, map the historical material procurement data to the corresponding nodes of the node connection graph, and aggregate the features of adjacent nodes in the node connection graph through a graph convolutional network to generate node embedding features that represent organizational dependencies. The graph convolutional network is a branch of deep learning. The extraction module is used to perform temporal convolution processing on the external environment data and the project unit operation data to extract power time-series features, and to concatenate the node embedding features with the power time-series features to form a feature fusion tensor. The input module is used to input the feature fusion tensor into a fully connected deep neural network, process the feature fusion tensor through the fully connected deep neural network, and output the procurement demand prediction result for a specific future period. The fully connected deep neural network is another branch of deep learning.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the deep learning-based power material procurement demand forecasting method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the deep learning-based method for predicting power material procurement demand as described in any one of claims 1 to 7.

Images