A power material demand prediction method and device based on deep learning and a medium

Abstract

This invention discloses a method, device, and medium for power material demand forecasting based on deep learning, relating to the field of deep learning technology. The method includes: acquiring a material demand sequence; constructing a pyramid hierarchy graph based on predefined material hierarchical relationships; constructing a warehouse-network supply graph based on predefined supply network relationships; and encapsulating the material demand sequence, pyramid hierarchy graph, and warehouse-network supply graph to form a graph sequence joint input package; aligning the graph sequence joint input package with the pyramid hierarchy graph and the warehouse-network supply graph to form a graph fusion representation sequence package; performing graph convolutional demand propagation on the graph fusion representation sequence package through the warehouse-network supply graph to output a demand forecast result package; and publishing the demand forecast result package to the power material business chain in real time to drive business execution. This invention improves the accuracy of demand forecasting and optimizes inventory management strategies by constructing a pyramid hierarchy graph and a warehouse-network supply graph.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method, device and medium for predicting power material demand based on deep learning. Background Technology

[0002] Power supply demand forecasting, as a core component supporting power grid planning, construction, and operation, directly impacts supply chain response efficiency and resource allocation rationality. In practical business scenarios, the power supply system exhibits a significant multi-level structure, forming a complex pyramid of material classifications, from main equipment to sub-components and from broad categories to specific models. Simultaneously, the supply process relies on a multi-level warehousing network comprised of central warehouses, regional warehouses, and field warehouses, creating a topologically constrained warehouse-network supply relationship. Material demand not only manifests as a dynamic sequence over time but is also embedded within the heterogeneous graph system formed by the aforementioned hierarchical structure and supply network. Therefore, constructing a forecasting mechanism capable of collaboratively integrating material semantic hierarchy, warehouse-network topology, and historical demand timelines has significant engineering value for achieving high-precision, interpretable, and executable material demand projection.

[0003] While existing deep learning-based prediction methods have made some progress in time series modeling, they still fall short in handling the unique hierarchical structure and warehouse network topology of power materials. In particular, they lack the ability to jointly model the semantics of material hierarchy and supply path constraints, leading to significant deviations in prediction results for fine-grained materials or edge warehouse scenarios. Current technologies typically mitigate these problems by introducing manual rules or post-processing corrections, such as setting proportional coefficients for hierarchical allocation based on experience or using independent graph neural networks to process structural information separately. However, these methods struggle to achieve end-to-end collaborative optimization. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for predicting power material demand to address the problem of lacking the ability to jointly model material hierarchical semantics and supply path constraints.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting electricity demand based on deep learning, comprising, Obtain the material demand sequence, construct a pyramid hierarchy diagram based on predefined material hierarchical relationships, construct a warehouse network supply diagram based on predefined supply network relationships, and encapsulate the material demand sequence, pyramid hierarchy diagram, and warehouse network supply diagram to form a graph sequence joint input package; Perform pyramid-level graph alignment and warehouse supply graph alignment on the graph sequence joint input package to form a graph fusion representation sequence package; Using the warehouse supply map, graph convolution demand propagation is performed on the graph fusion representation sequence package to output the demand forecast result package; The demand forecast results package is published to the power materials business chain in real time to drive business execution, collect business feedback data generated by business execution, and update the graph sequence joint input package for the next forecast period based on the business feedback data.

[0007] As a preferred embodiment of the deep learning-based power resource demand forecasting method of the present invention, wherein: the construction of the pyramid hierarchy graph specifically involves, Extract power material codes and warehouse identifiers from the material demand sequence; Define the power material code as the bottom layer node and the warehouse identifier as the middle layer node; Establish hierarchical connection edges along the subordinate paths from the bottom-level nodes to the middle-level nodes to form a multi-level node set; Perform intra-level node embedding aggregation and inter-level path attention weight allocation on a multi-level node set to generate a pyramid hierarchy graph.

[0008] As a preferred embodiment of the deep learning-based power resource demand forecasting method of the present invention, wherein: the formation of the graph sequence joint input packet specifically comprises, Map the demand records for each time slice in the material demand sequence to the bottom nodes of the pyramid hierarchy diagram and the network nodes of the warehouse supply diagram. The hierarchical connection edges of the pyramid hierarchy diagram and the directional edges of the warehouse supply diagram are uniformly identified and coded to generate unified connection edges; The mapped demand records and unified connection edges are jointly encapsulated to generate a graph sequence joint input package.

[0009] As a preferred embodiment of the deep learning-based power resource demand forecasting method of the present invention, wherein: the formation of the graph fusion representation sequence package specifically comprises, The decomposition diagram sequence combined with the input package yields the material demand sequence, the pyramid hierarchy diagram, and the warehouse supply diagram. Based on the hierarchical structure of the pyramid diagram, the material demand sequence is aggregated from bottom to top to generate a pyramid-level feature sequence. Based on the connection structure of the warehouse network supply graph, the pyramid-level feature sequence is propagated along the network edges to generate the warehouse network supply feature sequence. By fusing the pyramid-level feature sequence and the warehouse network supply feature sequence, a node fusion feature sequence is generated; The encapsulated node fusion feature sequence, pyramid hierarchy graph, and warehouse supply graph are combined to form a graph fusion representation sequence package.

[0010] As a preferred embodiment of the deep learning-based power resource demand forecasting method of the present invention, wherein: the output demand forecasting result package specifically comprises, Extract node fusion feature sequences and warehouse supply graphs from graph fusion representation sequence packages; Based on the connection structure of the warehouse supply map, neighborhood feature aggregation based on edge type is performed on the node fusion feature sequence to generate a spatial propagation feature sequence; Perform multiple rounds of neighborhood feature aggregation on the spatially propagated feature sequence to generate a depth-propagated feature sequence; Map the deep propagation feature sequence to predicted demand records for future time slices; Based on the node identification of the warehouse supply map, the predicted value of demand is recorded and packaged into a demand forecast result package.

[0011] As a preferred embodiment of the deep learning-based power resource demand forecasting method of the present invention, wherein: updating the graph sequence joint input packet for the next forecast period based on business feedback data specifically involves, Parse the demand forecast results package and, based on predefined replenishment strategies and inventory rules, convert the parsed forecast demand records into a formatted business instruction set; The formatted business instruction set is published to the power materials business chain to drive the execution of procurement and warehousing operations; Collect business feedback data generated after the execution of procurement and warehousing operations; The graph sequence joint input package for the next forecast period is updated using business feedback data.

[0012] As a preferred embodiment of the deep learning-based power material demand forecasting method of the present invention, the step of updating the graph sequence joint input package for the next forecast period is to replace the material demand sequence in the graph sequence joint input package for the next forecast period based on business feedback data, and to correct the supply network relationship of the warehouse supply map, and output the updated graph sequence joint input package.

[0013] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the deep learning-based power material demand forecasting method described in the first aspect of the present invention.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the deep learning-based power material demand forecasting method described in the first aspect of the present invention.

[0015] The beneficial effects of this invention are as follows: By constructing a pyramid hierarchy diagram and a warehouse network supply diagram, multi-dimensional feature extraction and structured representation of power material demand are achieved. Utilizing the hierarchical relationship between power material codes and warehouse identifiers, as well as the supply network relationships defined by replenishment paths and transportation protocols between warehouses, the time-series information and structured information of material demand are deeply integrated. This integration not only enhances the demand forecasting model's understanding of the complex hierarchical structure of power materials but also improves the accuracy of capturing supply chain network topology constraints. It can more accurately simulate the propagation path of material demand in the hierarchical structure and supply network, thereby improving the accuracy of demand forecasting, optimizing inventory management strategies, reducing inventory surpluses or shortages caused by inaccurate forecasts, supporting reasonable business decision-making, and enhancing the responsiveness and operational efficiency of the entire supply chain. In particular, by jointly modeling the semantics of material hierarchy and supply path constraints, the problem of fragmented hierarchical structures in traditional methods is effectively solved. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0017] Figure 1 This is a flowchart of a deep learning-based method for predicting electricity demand.

[0018] Figure 2 A flowchart for constructing a warehouse supply network.

[0019] Figure 3 A flowchart for constructing a pyramid hierarchy diagram.

[0020] Figure 4 This is a flowchart illustrating the propagation of convolutional requirements. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for predicting power resource demand based on deep learning, including the following steps: S1. Obtain the material demand sequence, construct a pyramid hierarchy graph based on the predefined material hierarchy relationship, construct a warehouse supply graph based on the predefined supply network relationship, and encapsulate the material demand sequence, pyramid hierarchy graph and warehouse supply graph to form a graph sequence joint input package.

[0025] The material demand sequence is obtained by collecting historical procurement records, inventory consumption records, and project demand records in the power material business chain. The historical procurement records, inventory consumption records, and project demand records contain the demand record for each time slot, as well as the corresponding power material code and warehouse identifier.

[0026] Extract the power material code and warehouse identifier from the material demand sequence; define the power material code as the bottom layer node and the warehouse identifier as the middle layer node; establish hierarchical connection edges along the subordinate paths from the bottom layer node to the middle layer node to form a multi-level node set.

[0027] Specifically, based on the predefined material hierarchy, the subordinate warehouse identifier corresponding to each power material code is identified as the parent node. For each subordinate path from the power material code to the subordinate warehouse identifier, an undirected connection edge is added. All power material codes are used as bottom-level nodes, all warehouse identifiers are used as middle-level nodes, and all added undirected connection edges are combined to form a multi-level node set containing bottom-level nodes and middle-level nodes.

[0028] The predefined material hierarchy is based on the organizational structure and classification standards in the power material business chain. It is a tree-like hierarchical relationship, with power material codes subordinate to warehouse identifiers. The depth of the hierarchy depends on the actual business scale. For example, in a typical power grid company, it is usually 2 to 4 levels. In the power grid company with the smallest business scale, the hierarchy depth is usually 2 levels, and in the power grid company with the largest business scale, the hierarchy depth is usually 4 levels. Business scale refers to the power grid coverage area, the number of warehouses managed, and the complexity of material classification and organization of the power grid company. This is to achieve gradual subdivision from broad categories to specific materials and code uniqueness. The classification element codes in the standard indicate the actual application of the hierarchical relationship.

[0029] Perform intra-level node embedding aggregation and inter-level path attention weight allocation on a multi-level node set to generate a pyramid hierarchy graph.

[0030] Specifically, an initial node embedding is assigned to each bottom-level node and each middle-level node in the multi-level node set. The initial node embedding is the initial feature vector representation of each bottom-level node and each middle-level node. For the power material code corresponding to each bottom-level node, a bottom-level feature vector is generated using one-hot encoding, where the position corresponding to the power material code is marked as 1 and the other positions are marked as 0. For the warehouse identifier corresponding to each middle-level node, a middle-level feature vector is generated using one-hot encoding, where the position corresponding to the warehouse identifier is marked as 1 and the other positions are marked as 0, to ensure that each node has a unique and distinguishable vector representation as the input basis for subsequent aggregation and attention weight allocation.

[0031] It should be noted that the marker 1 represents the activation of the corresponding power material code category or warehouse identification category, and the marker 0 represents the non-activation of other power material code categories or warehouse identification categories. According to the unique hot coding standard, the binary value 1 is used to activate the corresponding power material code category or warehouse identification category, and the binary value 0 is used to non-activate other power material code categories or warehouse identification categories, thereby ensuring that the embedding vector of each initial node is unique, distinguishable, and suitable for neural network feature input.

[0032] For each bottom node in the bottom node hierarchy, the initial node embeddings of other directly connected bottom nodes are aggregated to form a bottom aggregate embedding. For each middle node in the middle node hierarchy, the initial node embeddings of other directly connected middle nodes are aggregated to form a middle aggregate embedding.

[0033] The attention mechanism is used to determine the path attention weight for each hierarchical connection edge on the subordinate path from the bottom node to the corresponding middle node. The attention score corresponding to the path attention weight is determined by the similarity between the initial node embedding of the bottom node and the initial node embedding of the corresponding middle node. The attention scores on the same subordinate path are normalized to obtain the path attention weight of each level connection edge.

[0034] The influence strength of the bottom-level aggregate embedding is adjusted based on path attention weights. The path attention weights reflect the similarity between the initial node embedding of the bottom-level node and the initial node embedding of the corresponding middle-level node. The higher the similarity, the greater the path attention weights, and the greater the influence strength of the bottom-level aggregate embedding. The adjustment is up to the similarity ratio determined by the path attention weights. The adjusted bottom-level aggregate embeddings are then integrated into the corresponding middle-level aggregate embeddings, that is, the adjusted bottom-level aggregate embeddings and the middle-level aggregate embeddings are concatenated by vectors to form the enhanced node embeddings. The pyramid hierarchy graph consists of a multi-level node set containing bottom-level and middle-level nodes, the enhanced node embeddings, and hierarchical connection edges with path attention weights.

[0035] The training process for intra-level node embedding aggregation and inter-level path attention weight allocation is as follows: Historical material demand sequences are collected as training inputs for the intra-level node embedding aggregation and inter-level path attention weight allocation model. Based on the power material codes and warehouse identifiers recorded in the historical material demand sequences, a multi-level node set containing bottom-level nodes and middle-level nodes is constructed. Bottom-level nodes represent power material code nodes, and middle-level nodes represent warehouse identifier nodes. The multi-level node set is arranged sequentially according to the tree-like hierarchical relationship of power material codes belonging to warehouse identifiers. The number of levels is determined based on the material classification depth, generally including two hierarchical structures: bottom-level node level and middle-level node level. Among them, the intra-level node embedding aggregation and inter-level path attention weight allocation model performs intra-level node embedding aggregation and inter-level path attention weight allocation on the multi-level node set to generate a pyramid hierarchy graph. It is the training object used in the pyramid hierarchy graph to generate the enhanced node embeddings and output the pyramid hierarchy graph.

[0036] Initial node embeddings are assigned to bottom-level and mid-level nodes respectively to form the initial feature representation of the nodes. Through forward propagation, the node embedding aggregation within the layer and the path attention weight allocation between layers are executed sequentially from bottom to top. The embedding features of directly connected nodes are summarized in each layer, and path attention weights are applied along the layer connection edges to pass feature information layer by layer. Through multiple rounds of forward propagation, an enhanced node embedding containing the influence of path attention weights is formed. The enhanced nodes are embedded into the input feature mapping layer, and the embedded nodes are mapped to the predicted demand values ​​for future time slices. The predicted demand values ​​are compared with the actual historical demand records in the historical material demand sequence to obtain the difference information between the prediction results and the actual records. The prediction difference represents the degree of inconsistency between the prediction results and the historical records. When the prediction difference tends to stabilize and no longer changes significantly, the training process is terminated when the trend of the prediction difference remains consistent and the magnitude of the change is in a convergent state during continuous training iterations. During training, backpropagation uses the prediction difference transmission signal to adjust the path attention weight parameters and node embedding aggregation parameters until the prediction difference stabilizes and the outputs of the intra-level node embedding aggregation and inter-level path attention weight allocation model converge. The resulting enhanced node embedding can accurately characterize the hierarchical association and path influence between bottom-level and mid-level nodes, providing a stable feature foundation for subsequent demand prediction.

[0037] Forward propagation refers to the process of starting from the initial node embedding, performing intra-level node embedding aggregation from bottom to top according to the order of intra-level node embedding aggregation and inter-level path attention weight allocation, applying path attention weights along the hierarchical connection edges to form enhanced node embeddings, and then inputting the enhanced node embeddings into the feature mapping layer to obtain the predicted demand values ​​for future time slices.

[0038] Backpropagation refers to the process of transmitting prediction difference information from the direction of the predicted demand value to the direction of the initial node embedding, based on the prediction difference between the predicted demand value and the actual historical demand record, updating the path attention weight parameters and the node embedding aggregation parameters, until the prediction difference enters a stable state.

[0039] Extract warehouse identifiers from the material demand sequence; map the warehouse identifiers to the middle-level nodes in the pyramid hierarchy diagram to form network nodes.

[0040] Specifically, for each warehouse identifier corresponding to a mid-level node in the pyramid hierarchy diagram, a one-to-one match is performed with the warehouse identifier extracted from the material demand sequence. The mid-level node corresponding to the successfully matched warehouse identifier is directly defined as a network node, forming a set of network nodes containing all matched warehouse identifiers.

[0041] Based on predefined supply network relationships, directional edges are established between network nodes to form a network structure with directional edges.

[0042] Specifically, the predefined supply network relationship is defined based on the inter-warehouse replenishment path and transportation agreement in the power materials business chain, which is a supply direction relationship between a pair of warehouse identifiers. For each pair of network nodes corresponding to warehouse identifiers, the direction from the supply source warehouse identifier to the supply target warehouse identifier is determined according to the predefined supply network relationship. A directional edge from the supply source network node to the supply target network node is added, and all network nodes are combined with all added directional edges to form a network structure with directional edges.

[0043] Perform heterogeneous edge type differentiation aggregation and dynamic weight calculation of adjacent nodes on the network structure with directional edges to generate a warehouse network supply graph.

[0044] Specifically, based on the predefined supply network relationship, each directional edge in the network structure with directional edges is marked as a heterogeneous edge type, including replenishment type and transportation type; for each directional edge of the replenishment type, the initial features of the source network nodes connected to the same network node are summarized to form a replenishment aggregation feature; for each directional edge of the transportation type, the initial features of the source network nodes connected to the same network node are summarized to form a transportation aggregation feature.

[0045] The adjusted replenishment aggregation features and the adjusted transportation aggregation features are integrated into the initial features of the current network nodes to form enhanced network node features; the warehouse supply graph consists of a network structure with directional edges, heterogeneous edge type labels, and enhanced network node features.

[0046] The expression for calculating the dynamic weight of adjacent nodes is as follows: ; In the formula, For the current network node and adjacent nodes Dynamic weights, For the current network node The initial feature vector (i.e., the initial features of the current network node). Adjacent nodes The initial feature vector (i.e., the initial features of each neighboring node). The initial features of all nodes are uniformly projected using a learnable linear transformation matrix to improve feature representation. The attention vectors are learnable, and the importance score of the concatenated attention vectors is calculated by dot product. For transpose, specifically for the attention vector The transpose of the vector is used to convert a row vector into a column vector, so that a dot product operation can be performed with the concatenated vector to obtain the attention score. The importance score is converted to a positive value using an exponential function for softmax normalization. For the current network node The set of all adjacent nodes, For the current network node The index variable in the set of all adjacent nodes (traversing all adjacent nodes, including...) (and other adjacent nodes), used for the summation operation in the denominator, to normalize the attention score over all adjacent nodes. The LeakyReLU activation function is used. The negative input slope of the LeakyReLU activation function is usually set to 0.2. A negative input slope of 0.2 is a commonly used preset value for the LeakyReLU activation function in deep learning implementations to improve training stability. The negative input slope is a configurable value during implementation. The LeakyReLU activation function is used to perform a non-linear transformation on the attention score and maintain gradient propagation in the negative interval, preventing the attention score from becoming inactive in the negative interval.

[0047] The demand records for each time slice in the material demand sequence are mapped to the bottom nodes of the pyramid hierarchy diagram and the network nodes of the warehouse supply diagram.

[0048] Specifically, the process involves: traversing each time slice in the material demand sequence; for each demand record within each time slice, identifying the corresponding bottom-level node in the pyramid hierarchy diagram based on the power material code in the demand record, and directly assigning the value of the demand record to the time slice feature position of the bottom-level node; identifying the corresponding network node in the warehouse network supply diagram based on the warehouse identifier in the demand record, and directly assigning the value of the demand record to the time slice feature position of the network node; after all time slice demand records have been assigned values, the bottom-level node features of the pyramid hierarchy diagram and the network node features of the warehouse network supply diagram contain the demand record information for the corresponding time slice, serving as the input basis for subsequent sequence joint encapsulation.

[0049] The mapped demand records and unified connection edges are jointly encapsulated to generate a graph sequence joint input package.

[0050] Specifically, the following steps are taken: the demand records mapped from all time slices, the node embeddings enhanced from the pyramid hierarchy graph, the hierarchical connection edges with path attention weights, the network node features enhanced from the warehouse supply graph, the directional edges with heterogeneous edge type labels, and the unified connection edge set are combined and encapsulated in chronological order to form a graph sequence joint input package containing a multi-time slice graph structure.

[0051] S2. Perform pyramid-level graph alignment and warehouse supply graph alignment on the graph sequence joint input package to form a graph fusion representation sequence package.

[0052] The decomposition graph sequence is combined with the input package to obtain the material demand sequence, the pyramid hierarchy graph, and the warehouse supply graph. Based on the subordinate structure of the pyramid hierarchy graph, the material demand sequence is aggregated from bottom to top to generate the pyramid hierarchy feature sequence.

[0053] Specifically, iterate through each time slice in the joint input bag of the graph sequence. For each time slice, starting from the bottom node of the pyramid hierarchy graph, use the hierarchical connection edges with path attention weights to identify the middle node corresponding to each bottom node. For each middle node, identify all directly subordinate bottom nodes, and summarize the mapping demand features of the bottom nodes in the current time slice and the enhanced node embeddings to form the bottom summary features.

[0054] The influence intensity of each bottom-level aggregate feature is adjusted based on path attention weights. Specifically, path attention weights are applied along the hierarchical connection edges with path attention weights, the influence intensity of the bottom-level aggregate features is adjusted, and the path attention-adjusted bottom-level aggregate features are output. The termination condition for the influence intensity adjustment is that the path attention weights enter a stable state and maintain consistency, and the mid-level aggregate features enter a stable state and maintain consistency. A stable state means that during continuous training iterations, the changing trends of the path attention weights and the mid-level aggregate features are consistent and the magnitude of change is in a convergent state. The adjusted bottom-level aggregate features are integrated into the enhanced node embeddings of the corresponding mid-level nodes to form mid-level aggregate features. The mid-level aggregate features of all time slices are combined in chronological order to form a pyramid-level feature sequence, which contains the mid-level node aggregate representation of each time slice.

[0055] Based on the connection structure of the warehouse network supply graph, the pyramid-level feature sequence is propagated along the network edges to generate the warehouse network supply feature sequence.

[0056] Specifically, iterate through each time slice. For each time slice, directly assign the mid-level aggregated features in the pyramid-level feature sequence to the feature positions of the corresponding network nodes in the warehouse supply map to form the initial propagation features of each network node.

[0057] For each network node, identify all directional edges pointing to the network node that belong to the replenishment type, and summarize the initial propagation features of the source network node corresponding to the replenishment type directional edges to form the replenishment propagation features.

[0058] For each network node, identify all directional edges pointing to the network node that belong to the transportation type, and summarize the initial propagation features of the source network node corresponding to the transportation type directional edges to form transportation propagation features.

[0059] The adjusted replenishment propagation features and the adjusted transportation propagation features are integrated into the initial propagation features of the current network nodes to form the propagated network node features. The propagated network node features of all time slices are combined in chronological order to form the warehouse network supply feature sequence, which contains the network node representations propagated through the neighborhood in each time slice.

[0060] The pyramid-level feature sequence and the warehouse supply feature sequence are merged to generate a node fusion feature sequence. Specifically, for each time slice, the mid-level aggregation feature in the pyramid-level feature sequence and the propagated network node feature of the corresponding network node in the warehouse supply feature sequence are concatenated as vectors to form a node fusion feature vector. The node fusion feature vectors of all time slices are combined in chronological order to form a node fusion feature sequence.

[0061] The encapsulated node fusion feature sequence, pyramid hierarchy graph, and warehouse supply graph are combined to form a graph fusion representation sequence package.

[0062] S3. Using the warehouse supply map, perform graph convolution demand propagation on the graph fusion representation sequence package and output the demand forecast result package.

[0063] Extract node fusion feature sequences and warehouse supply graphs from the graph fusion representation sequence package; based on the connection structure of the warehouse supply graph, perform neighborhood feature aggregation based on edge type on the node fusion feature sequences to generate spatial propagation feature sequences.

[0064] Specifically, this involves extracting the node fusion feature sequences of all time slices and the connection structure of the warehouse supply graph from the graph fusion representation sequence package.

[0065] For each network node in the warehouse supply diagram, within each time slice, the node fusion feature vector corresponding to the time slice in the node fusion feature sequence is directly assigned to the feature position of the network node to form an initial fusion feature set.

[0066] For the directional edges of replenishment types, identify all replenishment source network nodes pointing to the current network node, and summarize the initial fusion features of replenishment source network nodes in each time slice to form a neighborhood feature set of replenishment types.

[0067] For each transport type of directional edge, identify all transport source network nodes pointing to the current network node, and summarize the initial fusion features of the transport source network nodes in each time slice to form a transport type neighborhood feature set.

[0068] For the neighborhood feature set of replenishment type, the initial fusion feature of the current network node in the initial fusion feature set is compared with the initial fusion feature of each replenishment source network node in the neighborhood feature set of replenishment type. The similarity comparison is obtained by dot product to obtain the original attention score.

[0069] The LeakyReLU activation function is applied to all the original attention scores to obtain the adjusted attention scores; the exp function is applied to the adjusted attention scores to convert them into positive scores; softmax normalization is applied to the positive scores to ensure that the sum of the dynamic weights of all replenishment types is 1, thereby generating the dynamic weights of replenishment types. The dynamic weights of replenishment types represent the proportion of influence of the initial fusion features of each replenishment source network node on the initial fusion features of the current network node.

[0070] For the transportation type neighborhood feature set, the initial fusion features of the current network node in the initial fusion feature set are compared with the initial fusion features of each transportation source network node in the transportation type neighborhood feature set. The similarity comparison is used to obtain the original attention score through dot product.

[0071] The LeakyReLU activation function is applied to all original attention scores to obtain adjusted attention scores; the exp function is applied to the adjusted attention scores to convert them into positive scores; softmax normalization is applied to the positive scores to ensure that the sum of the dynamic weights of all transportation types is 1, thereby generating dynamic weights for transportation types. The dynamic weights for transportation types represent the proportion of influence of the initial fusion features of each transportation source network node on the initial fusion features of the current network node.

[0072] The dynamic weight of replenishment type is used to adjust the influence strength of the neighborhood feature set of replenishment type, and the dynamic weight of transportation type is used to adjust the influence strength of the neighborhood feature set of transportation type.

[0073] The adjusted replenishment type neighborhood feature set and the adjusted transportation type neighborhood feature set are integrated into the initial fusion feature of the current network node to generate a spatially propagated network node feature set. The spatially propagated network node feature sets of all time slices are reorganized according to the original time sequence to form a spatially propagated feature sequence. The spatially propagated feature sequence contains the network node representation of each time slice after a single round of edge type neighborhood aggregation.

[0074] It should be noted that the sum of the dynamic weights of all transportation types and the sum of the dynamic weights of all replenishment types are 1 in order to convert the original attention scores into a probability distribution through softmax normalization, thereby ensuring that the influence ratio of the initial fusion features of each source network node on the initial fusion features of the current network node is reasonably distributed and the total influence intensity remains consistent.

[0075] Perform multiple rounds of neighborhood feature aggregation on the spatially propagated feature sequence to generate a depth-propagated feature sequence.

[0076] Specifically, the spatial propagation feature sequence is used as the input to the initial deep feature set in the first round. The connection structure of the warehouse supply graph is used to perform neighborhood feature aggregation based on edge type to generate the network node feature set after the first round of deep propagation. The neighborhood feature aggregation based on edge type refers to the process of summarizing the initial fusion features of the source network nodes in the warehouse supply graph according to the replenishment type directional edge and the transportation type directional edge to form different types of neighborhood feature sets. An attention mechanism is used to generate dynamic weights to adjust the influence intensity of the types, and the adjusted neighborhood feature sets of different types are integrated into the initial fusion features of the current network nodes.

[0077] The network node feature set after the first round of depth propagation is used as the input of the initial deep feature set for the second round. The neighborhood feature aggregation based on edge type is repeatedly performed to generate the network node feature set after the second round of depth propagation.

[0078] The network node feature set after the second round of depth propagation is used as the input of the initial deep feature set for the third round. The neighborhood feature aggregation based on edge type is repeatedly performed to generate the network node feature set after the third round of depth propagation.

[0079] The feature set of network nodes after the third round of deep propagation of all time slices is reorganized according to the original time sequence to form a deep propagation feature sequence. The deep propagation feature sequence contains the network node representation of each time slice after three rounds of edge type neighborhood aggregation.

[0080] Map the deep propagation feature sequence to predicted demand records for future time slices.

[0081] Specifically, the process involves: extracting the network node feature set after the third round of deep propagation for all time slices from the deep propagation feature sequence; for each time slice in the deep propagation feature sequence, traversing each network node in the warehouse supply map, and inputting the feature vector of the corresponding network node in the network node feature set after the third round of deep propagation into a fully connected layer, which converts the feature vector into a scalar predicted value of demand for the next time slice; using the converted scalar predicted value of demand as the predicted value of demand record for the network node in the next time slice; and organizing the predicted values ​​of demand record for all network nodes in the next time slice according to the node identifier of the warehouse supply map to form the predicted value of demand record for the next time slice.

[0082] Based on the node identification of the warehouse supply map, the predicted value of demand is recorded and packaged into a demand forecast result package.

[0083] Specifically, the process involves: traversing all network nodes in the warehouse supply diagram, associating and combining the predicted demand value for the future time slice corresponding to each network node with the warehouse identifier of the network node to form a set of predicted demand values ​​with warehouse identifiers, and encapsulating all sets of predicted demand values ​​with warehouse identifiers in chronological order to form a demand forecast result package.

[0084] It should be noted that this step utilizes the graph convolutional demand propagation mechanism of deep learning, employing attention mechanisms and neighborhood feature aggregation based on edge type to perform multiple rounds of deep propagation on the warehouse supply graph. This significantly enhances the expressive power of node fusion feature sequences, enabling dynamic diffusion simulation of demand in the supply network. Ultimately, it generates high-precision future time-slice demand record predictions, greatly improving prediction accuracy, capturing heterogeneous dependencies between replenishment type and transportation type directional edges, optimizing power material inventory allocation, reducing the risk of surplus or shortage, supporting real-time business decisions, and improving the overall efficiency and economy of the supply chain.

[0085] S4. Publish the demand forecast results package to the power materials business chain in real time to drive business execution, collect business feedback data generated by business execution, and update the graph sequence joint input package for the next forecast period based on the business feedback data.

[0086] Parse the demand forecast results package and, based on predefined replenishment strategies and inventory rules, convert the parsed forecast demand records into a formatted set of business instructions.

[0087] Specifically, this involves extracting the forecast values ​​of all demand records with warehouse identifiers from the demand forecast results package.

[0088] For each warehouse identifier, the predicted demand record is compared with the current inventory level for future time slices. This involves numerically comparing the predicted demand record for each warehouse identifier with the current inventory level. Based on the comparison results, an inventory coverage determination result is output, including an inventory satisfaction flag and a replenishment trigger flag. The predefined replenishment strategy is based on inventory management and supply chain optimization requirements within the power materials business chain. Supply chain optimization aims to improve overall supply chain efficiency and economy by minimizing total inventory holding costs, ordering costs, and stockout costs while ensuring continuous supply and timely response. The predefined replenishment strategies include a minimum inventory threshold replenishment strategy and an economic order quantity replenishment strategy. The minimum inventory threshold replenishment strategy is set based on historical demand fluctuations, replenishment lead time, and safety stock rules in the power materials business chain. The minimum inventory threshold is usually set to a range of 1 to 3 standard deviations of the average demand. The variation of the range of 1 to 3 standard deviations of the average demand is determined based on the fluctuation range of historical demand and replenishment lead time. The higher the fluctuation range of historical demand and replenishment lead time, the closer the minimum inventory threshold is to 3 standard deviations. The lower the fluctuation range of historical demand and replenishment lead time, the closer the minimum inventory threshold is to 1 standard deviation, so as to cover more than 95% of demand uncertainty and avoid stockout risk.

[0089] The predefined inventory rules are based on supply chain optimization requirements and risk management definitions in the power materials business chain. They are a set of binding rules that include safety stock rules to address demand uncertainty and maximum inventory capacity rules to limit storage space and capital occupation. The predefined inventory rules include safety stock rules and maximum inventory capacity rules. Safety stock rules are minimum additional inventory buffer rules set to cope with demand fluctuations, replenishment lead time uncertainty, and supply disruption risks. The buffer amount is determined based on historical demand standard deviation and service level (such as 95% coverage) to ensure that stockouts do not occur in most scenarios. Maximum inventory capacity rules are inventory upper limit constraints set to control storage space occupation, capital occupation, and inventory holding costs. The maximum allowable inventory amount is determined based on warehouse physical space, capital budget constraints, or economic order quantity models to avoid excess inventory.

[0090] If the predicted demand exceeds the sum of the current inventory level and the safety stock defined by the safety stock rule, a replenishment order will be generated.

[0091] Based on the minimum inventory threshold replenishment strategy, the replenishment quantity is set to cover the difference between the predicted demand record and the current inventory level, and at the same time, it is replenished to the safety stock level defined by the safety stock rule, thereby generating a replenishment business instruction with the corresponding warehouse identifier and power material code.

[0092] Based on the economic order quantity replenishment strategy, the replenishment quantity is adjusted to an integer multiple of the economic order quantity, while ensuring that it does not exceed the maximum inventory capacity defined by the maximum inventory capacity rule.

[0093] The replenishment quantity is associated with the corresponding warehouse identifier and power material code to form a replenishment business instruction; all replenishment business instructions are combined into a purchase instruction set and a warehouse adjustment instruction set to form a formatted business instruction set, which includes the replenishment quantity with warehouse identifier and execution time window.

[0094] The standardized business instruction set is issued to the power materials business chain to drive the execution of procurement and warehousing operations. Specifically, after the standardized business instruction set is issued to the power materials business chain, the procurement department initiates supplier replenishment orders according to the procurement instruction set, and the warehousing department executes inventory entry and location adjustment according to the warehousing adjustment instruction set, thereby driving the execution of procurement and warehousing operations.

[0095] Collect business feedback data generated after the execution of procurement and warehousing operations.

[0096] The graph sequence joint input package for the next forecast period is updated using business feedback data. This update involves replacing the material demand sequence in the graph sequence joint input package based on the business feedback data and correcting the supply network relationships in the warehouse supply map. Specifically, based on the actual procurement path records and warehousing and transportation execution records in the business feedback data, the direction of directional edges or the heterogeneous edge type markings in the warehouse supply map are adjusted. If the business feedback data shows a new replenishment path, a corresponding directional edge is added; if the business feedback data shows an existing transportation path is interrupted, a corresponding directional edge is removed. This generates a corrected warehouse supply map and outputs the updated graph sequence joint input package.

[0097] This embodiment also provides a computer device applicable to the deep learning-based power material demand forecasting method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning-based power material demand forecasting method proposed in the above embodiment.

[0098] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0099] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the deep learning-based power resource demand forecasting method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0100] In summary, this invention achieves multi-dimensional feature extraction and structured representation of power material demand by constructing a pyramid hierarchy diagram and a warehouse network supply diagram. It deeply integrates time-series information and structured information of material demand by utilizing the hierarchical relationship between power material codes and warehouse identifiers, as well as the supply network relationships defined by inter-warehouse replenishment paths and transportation protocols. This integration not only enhances the demand forecasting model's understanding of the complex hierarchical structure of power materials but also improves the accuracy of capturing supply chain network topology constraints. It can more accurately simulate the propagation path of material demand in the hierarchical structure and supply network, thereby improving the accuracy of demand forecasting, optimizing inventory management strategies, reducing inventory surpluses or shortages caused by inaccurate forecasts, supporting reasonable business decision-making, and enhancing the responsiveness and operational efficiency of the entire supply chain. In particular, by jointly modeling the semantics of material hierarchy and supply path constraints, it effectively solves the problem of fragmented hierarchical structures in traditional methods.

[0101] 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 method for predicting electricity demand based on deep learning, characterized in that: include, Obtain the material demand sequence, construct a pyramid hierarchy diagram based on predefined material hierarchical relationships, construct a warehouse network supply diagram based on predefined supply network relationships, and encapsulate the material demand sequence, pyramid hierarchy diagram, and warehouse network supply diagram to form a graph sequence joint input package; Perform pyramid-level graph alignment and warehouse supply graph alignment on the graph sequence joint input package to form a graph fusion representation sequence package; Using the warehouse supply map, graph convolution demand propagation is performed on the graph fusion representation sequence package to output the demand forecast result package; The demand forecast results package is published to the power materials business chain in real time to drive business execution, collect business feedback data generated by business execution, and update the graph sequence joint input package for the next forecast period based on the business feedback data.

2. The deep learning-based power resource demand forecasting method as described in claim 1, characterized in that: The construction of the pyramid hierarchy diagram specifically involves, Extract power material codes and warehouse identifiers from the material demand sequence; Define the power material code as the bottom layer node and the warehouse identifier as the middle layer node; Establish hierarchical connection edges along the subordinate paths from the bottom-level nodes to the middle-level nodes to form a multi-level node set; Perform intra-level node embedding aggregation and inter-level path attention weight allocation on a multi-level node set to generate a pyramid hierarchy graph.

3. The deep learning-based power resource demand forecasting method as described in claim 1, characterized in that: The construction of the warehouse supply network map specifically involves, Extract warehouse identifiers from the material demand sequence; Map warehouse identifiers to mid-level nodes in the pyramid hierarchy diagram to form network nodes; Based on predefined supply network relationships, directional edges are established between network nodes to form a network structure with directional edges; Perform heterogeneous edge type differentiation aggregation and dynamic weight calculation of adjacent nodes on the network structure with directional edges to generate a warehouse network supply graph.

4. The power resource demand forecasting method based on deep learning as described in claim 1, characterized in that: The formation of the graph sequence joint input packet specifically includes, Map the demand records for each time slice in the material demand sequence to the bottom nodes of the pyramid hierarchy diagram and the network nodes of the warehouse supply diagram. The hierarchical connection edges of the pyramid hierarchy diagram and the directional edges of the warehouse supply diagram are uniformly identified and coded to generate unified connection edges; The mapped demand records and unified connection edges are jointly encapsulated to generate a graph sequence joint input package.

5. The deep learning-based power resource demand forecasting method as described in claim 1, characterized in that: The formation of the graph fusion representation sequence package specifically includes: The decomposition diagram sequence combined with the input package yields the material demand sequence, the pyramid hierarchy diagram, and the warehouse supply diagram. Based on the hierarchical structure of the pyramid diagram, the material demand sequence is aggregated from bottom to top to generate a pyramid-level feature sequence. Based on the connection structure of the warehouse network supply graph, the pyramid-level feature sequence is propagated along the network edges to generate the warehouse network supply feature sequence. By fusing the pyramid-level feature sequence and the warehouse network supply feature sequence, a node fusion feature sequence is generated; The encapsulated node fusion feature sequence, pyramid hierarchy graph, and warehouse supply graph are combined to form a graph fusion representation sequence package.

6. The power resource demand forecasting method based on deep learning as described in claim 1, characterized in that: The output demand prediction result package is specifically as follows: Extract node fusion feature sequences and warehouse supply graphs from graph fusion representation sequence packages; Based on the connection structure of the warehouse supply map, neighborhood feature aggregation based on edge type is performed on the node fusion feature sequence to generate a spatial propagation feature sequence; Perform multiple rounds of neighborhood feature aggregation on the spatially propagated feature sequence to generate a depth-propagated feature sequence; Map the deep propagation feature sequence to predicted demand records for future time slices; Based on the node identification of the warehouse supply map, the predicted value of demand is recorded and packaged into a demand forecast result package.

7. The deep learning-based power resource demand forecasting method as described in claim 1, characterized in that: The process of updating the graph sequence joint input package for the next prediction period based on business feedback data specifically involves: Parse the demand forecast results package and, based on predefined replenishment strategies and inventory rules, convert the parsed forecast demand records into a formatted business instruction set; The formatted business instruction set is published to the power materials business chain to drive the execution of procurement and warehousing operations; Collect business feedback data generated after the execution of procurement and warehousing operations; The graph sequence joint input package for the next forecast period is updated using business feedback data.

8. The deep learning-based power resource demand forecasting method as described in claim 7, characterized in that: The update of the graph sequence joint input package for the next forecast period is based on business feedback data. The material demand sequence in the graph sequence joint input package for the next forecast period is replaced, and the supply network relationship of the warehouse supply map is corrected, and the updated graph sequence joint input package is output.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based power material demand forecasting method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based power material demand forecasting method according to any one of claims 1 to 8.

Images