Intelligent prediction method for parts demand in automotive collaborative supply chain based on spatiotemporal graph convolution

Abstract

The present invention relates to the fields of Internet big data and new generation information technology, and specifically to a method for intelligent prediction of parts demand in an automobile collaborative supply chain based on spatiotemporal graph convolution, comprising: S1: acquiring historical parts demand time series data; S2: analyzing the supply chain structure based on the historical parts demand time series data; S3: mapping several spatial dependency relationships based on the supply chain structure and calculating several adjacency matrices; S4: inputting the historical parts demand time series data and various adjacency matrices into a trained demand forecasting model, and outputting parts demand forecast data; the present invention uses graph construction technology to perform multi-dimensional decoupling of the supply chain network and generates an independent adjacency matrix for each type of spatial dependency relationship, retaining supply chain interaction features at different semantic levels, and at the same time, by introducing a spatiotemporal graph convolutional neural network, effectively capturing the spatiotemporal dynamic associations between nodes in the automobile parts supply chain, and realizing comprehensive modeling of key nodes and the complexity of multi-source data.

Description

Technical Field

[0001] The present invention relates to the field of Internet big data and new-generation information technology, and specifically to an intelligent prediction method for parts demand in an automotive collaborative supply chain based on spatiotemporal graph convolution. Background Art

[0002] In a cloud service environment, the automotive parts supply chain is experiencing a trend of increasing complexity and diversification. With the development of the automotive parts supply chain and the increasing demand for various automotive products in modern society, the automotive parts supply chain has become a dynamic system. The relationships between nodes and the time series data change over time. Time series data consists of time series data generated by the values ​​of the same statistical indicator. Time series data from nodes have dependencies on each other, and changes in time series data directly affect the behavior and decisions of other nodes in the network. Therefore, accurately predicting product demand based on the large amount of time series data generated by corresponding nodes in the automotive parts supply chain network is key to promoting the healthy development of the supply chain. Failure to effectively predict this demand information will severely impact the collaborative efficiency of the automotive parts supply chain, hindering product maintenance and parts supply under the large-scale manufacturing model.

[0003] Graph neural networks, as an effective tool, can integrate multi-source heterogeneous information into a unified structure, revealing the complex relationships between elements in the supply chain. This allows for a comprehensive view that supports cross-node collaborative interactions. However, accurately extracting inter-node demand dependencies from multi-source heterogeneous information is fundamental to achieving accurate forecasts and is currently one of the main technical challenges. Traditional demand forecasting methods struggle to achieve ideal results in complex areas like the automotive parts supply chain, primarily due to their inability to fully leverage the deep inter-node relationships within the supply chain.

[0004] Existing demand forecasting methods mostly rely on historical time series data, extracting temporal trends from historical time series data and then forecasting future demand. This approach ignores the interrelationships between multi-source data, which can easily lead to biased or unstable forecast results. To preserve the correlations between multi-source data, subsequent research has proposed training methods based on multi-task learning or multimodal data fusion. However, most existing methods cannot properly handle scenarios where data is sparse or unevenly distributed, and can lead to significant errors or overfitting. Demand forecasting in the automotive parts supply chain is multi-layered and multi-node, with significant spatial correlation and dynamic temporal characteristics. In a supply chain, different nodes belong to both logistics and transaction networks, and the relationships between nodes are complex and dynamic. Under uncertain events (such as sudden market fluctuations or supply chain disruptions), complex nonlinear dynamic characteristics are exhibited. Existing methods cannot effectively capture the spatial correlations, node interactions, and dynamic changes in automotive parts supply chain information, resulting in limited forecast accuracy and adaptability.

[0005] In demand forecasting tasks within the automotive parts supply chain, the relationships between nodes are often determined by contextual information rather than the target entity itself. Incorporating this relationship information into demand forecasting can effectively reduce data interference unrelated to node demand, lower prediction errors, and improve prediction efficiency. Furthermore, existing models are mostly based on single-dimensional or independent node predictions, lacking the ability to globally model inter-node relationships. Model evaluations typically assess the prediction accuracy of individual nodes, rather than focusing on the synergies and overall forecasting performance of the entire supply chain network. These models are unrealistic for application in the actual automotive parts supply chain because they fail to fully account for the complexity of the dynamic relationships across multiple nodes, dimensions, and time and space within the supply chain, nor do they meet the real-time and global optimization requirements of real-world scenarios. Summary of the Invention

[0006] In view of the above-mentioned deficiencies of the existing technology, the technical problem to be solved by the present invention is: how to provide an intelligent prediction method for parts demand in the collaborative automotive supply chain based on spatiotemporal graph convolution, decouple the supply chain network in multiple dimensions through graph construction technology and generate an independent adjacency matrix for each type of spatial dependency, retain the supply chain interaction characteristics at different semantic levels, and at the same time effectively capture the spatiotemporal dynamic correlation between nodes in the automotive parts supply chain by introducing a spatiotemporal graph convolutional neural network, and realize comprehensive modeling of the complexity of key nodes and multi-source data, thereby improving the accuracy and reliability of the prediction of parts demand in the collaborative automotive supply chain.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] An intelligent prediction method for parts demand in the automotive collaborative supply chain based on spatiotemporal graph convolution includes:

[0009] S1: Obtain historical parts demand time series data of the automotive supply chain;

[0010] S2: Analyze supply chain structure based on historical spare parts demand time series data;

[0011] S3: Build graphs of several spatial dependencies based on the supply chain structure, and calculate several adjacency matrices based on the graphs corresponding to various spatial dependencies;

[0012] S4: Input the historical spare parts demand time series data and various adjacency matrices into the trained demand forecasting model and output the corresponding spare parts demand forecast data;

[0013] The training steps of the demand forecasting model are as follows:

[0014] S401: Taking the historical spare parts demand time series data and various adjacency matrices as training data as model input;

[0015] S402: Using a spatiotemporal attention mechanism network, perform temporal and spatial weighting on the historical spare parts demand time series data and each adjacency matrix to obtain spatiotemporal weighted time series data for each adjacency matrix;

[0016] S403: Using the LSTM network, the time-series nonlinear dynamic changes of the spatiotemporal weighted time series data of each adjacency matrix are captured to obtain the complete time series features of each adjacency matrix;

[0017] S404: Using the GCN network, the spatial correlation features of each adjacency matrix are captured based on the complete time series features of the adjacency matrix to obtain an aggregated feature matrix of each adjacency matrix; the aggregated feature matrices of all adjacency matrices are fused to obtain a spatiotemporal comprehensive feature;

[0018] S405: extracting high-order features from the spatiotemporal comprehensive features to obtain parts demand forecast data;

[0019] S406: Calculate the loss function based on the historical spare parts demand time series data and the spare parts demand forecast data, and reverse the model parameters;

[0020] S407: Repeat steps S401 to S406 to iteratively train the model until the model converges or the maximum number of iterations is reached;

[0021] S5: The parts demand forecast data output by the demand forecast model is used as the collaborative demand forecast result of the automotive supply chain.

[0022] Preferably, step S3 specifically includes the following steps:

[0023] S301: Graph construction is used to map the three spatial dependencies of the supply chain structure, namely, flow intensity, functional relevance, and transaction amount impact, to obtain a flow intensity graph, a functional similarity graph, and a transaction amount impact graph;

[0024] S302: Calculate the adjacency matrix A of the flow intensity graph, functional similarity graph, and transaction amount influence graph flow 、A sim and A tran .

[0025] Preferably, in step S302, the adjacency matrix A of the flow intensity graph is calculated by the following formula: flow Elements in:

[0026]

[0027] Where: I ij Represents the flow intensity between node i and node j, which is used to construct the adjacency matrix A of the flow intensity graph flow ;T ij represents the number of transactions between nodes i and j; N represents the total number of nodes.

[0028] Preferably, in step S302, the adjacency matrix A of the functional similarity graph is calculated by the following formula: sim Elements in:

[0029]

[0030] Where: S ij Represents the functional similarity between node i and node j, which is used to construct the adjacency matrix A of the functional similarity graph sim POI K (i) and POI K (j) represents the value of the Kth attribute of node i and node j respectively.

[0031] Preferably, in step S302, the adjacency matrix A of the transaction amount influence graph is calculated by the following formula: tran Elements in:

[0032]

[0033] Where: A ij Represents the transaction amount influence between node i and node j, which is used to construct the adjacency matrix A of the transaction amount influence graph tran ; C ij represents the transaction amount between node i and node j; N represents the total number of nodes.

[0034] Preferably, in step S402, the processing steps of the spatiotemporal attention mechanism network include:

[0035] S4021: The adjacency matrix A (x) And historical spare parts demand time series data X as input;

[0036] Among them A (x) ={A flow ,A sim ,A tran}, A flow 、A sim 、A tran The adjacency matrices representing the flow intensity graph, functional similarity graph, and transaction amount influence graph, respectively;

[0037] S4022: Perform time series pooling on the historical spare parts demand time series data X to obtain the feature sum X at each time step. pool ;

[0038] The formula is:

[0039]

[0040] Where: X pool Represents the sum of features over time steps; X represents the three-dimensional tensor of historical spare parts demand time series data; X[:,:,t] represents extracting the node features corresponding to time step t; X[:,:,:] represents retaining the node features of all time steps;

[0041] S4023: Transpose the historical spare parts demand time series data X to obtain X gcn_input ; Through the GCN network layer to X gcn_input and the adjacency matrix A (x) Sum along the node dimension to get the adjacency matrix A (x) Features right Sum the graph features over the time steps to obtain the adjacency matrix A (x) The spatial characteristics X gcn_pool ;

[0042] The formula is:

[0043] X gcn_input =Transpose(X,prem=[0,2,1]);

[0044]

[0045] Where: Represents the adjacency matrix A (x) Transpose represents the transposition operation; prem represents the transformed dimension order; X represents the three-dimensional tensor of the historical spare parts demand time series data; X gcn_inputRepresents a three-dimensional tensor of transposed time series data; GCN represents a graph convolution operation;

[0046] S4024: Sum the features of the time step X pool With the adjacency matrix A (x) The spatial characteristics X gcn_pool Perform fusion and normalization to obtain the adjacency matrix A (x) The spatiotemporal characteristics of z (x) ;

[0047] The formula is:

[0048]

[0049] Where: z (x) Represents the adjacency matrix A (x) The spatiotemporal characteristics of; N represents the number of nodes;

[0050] S4025: Based on the adjacency matrix A through the multi-head attention mechanism (x) The spatiotemporal characteristics of z (x) Calculating attention scores And the attention score Weighted with the historical spare parts demand time series data X, we get the adjacency matrix A (x) Spatiotemporal weighted time series data

[0051] The formula is:

[0052]

[0053] Where: Represents the adjacency matrix A (x) The spatiotemporal weighted time series data is ; h∈{1,2,…,H} represents the number of multi-head attention mechanisms; b1 and b2 represent bias vectors; ReLU and σ represent activation functions; B represents the batch size; T represents the number of time steps; and s represents the output of all attention heads aggregated.

[0054] Preferably, in step S403, the processing steps of the LSTM network include:

[0055] S4031: Use exponential smoothing to calculate the adjacency matrix A (x) Spatiotemporal weighted time series data Perform preprocessing to obtain a preprocessed sequence

[0056] The formula is:

[0057] X (t) =X reweight ;

[0058]

[0059] Where: Represents the preprocessing sequence at the current moment; represents the preprocessing sequence of the previous moment; α represents the smoothing coefficient; X (t) Represents the spatiotemporal weighted time series data at the current moment;

[0060] S4032: Preprocess the sequence Perform node segmentation to obtain the time series of each node i

[0061] The formula is:

[0062]

[0063] Where: Represents the time series of the first node;

[0064] S4033: The time series of each node i Input into the LSTM network and get the output of each node i Output for all nodes Perform splicing to obtain the adjacency matrix A (x) The complete time series characteristics H (T) ;

[0065] The formula is:

[0066]

[0067]

[0068] Where: H (T) Represents the adjacency matrix A (x) The complete time series characteristics of represents the time series of node i; D represents the dimension of the hidden state in LSTM.

[0069] Preferably, in step S404, the processing steps of the GCN network include:

[0070] S4041: Adjacency matrix A (x) The complete time series characteristics H (T) Transpose to obtain the transposed time feature H0;

[0071] The formula is:

[0072]

[0073] Where: H0 represents the adjacency matrix A (x) The transposed time characteristics of

[0074] S4042: Through the adjacency matrix A (x) The transposed time feature H0 of the corresponding adjacency matrix A (x) Perform the first step of graph convolution to obtain the adjacency matrix A (x) The feature matrix of the first propagation step

[0075] The formula is:

[0076]

[0077] Where: Represents the adjacency matrix A (x) The feature matrix of the first propagation step;

[0078] S4043: Multi-step propagation of adjacency matrix A through graph convolution (x) The feature matrix of is weighted summed up, and the information from the neighboring nodes is gradually aggregated to obtain the adjacency matrix A (x) The feature matrix propagated at each step;

[0079] The formula is:

[0080]

[0081] Where: Represents the adjacency matrix A (x) The characteristic matrix of the k-th propagation step; k∈2,3,...,T max Represents the number of propagations in the GCN network. After each propagation step, the feature matrix is ​​recalculated and updated;

[0082] S4044: Adjacency matrix A (x) The feature matrices of all propagation steps are concatenated to obtain the adjacency matrix A (x) The aggregate feature matrix The aggregate feature matrix corresponding to the adjacency matrix of the fusion flow intensity graph, functional similarity graph and transaction amount influence graph is obtained to obtain the fusion feature matrix H' GCN ; For the fusion feature matrix H' GCN Perform linear transformation and nonlinear transformation activation to obtain the spatiotemporal comprehensive feature H GCN ;

[0083] The formula is:

[0084]

[0085] H GCN =σ(H' GCN W3+b3);

[0086] Where: They represent the aggregated feature matrices corresponding to the adjacency matrices of the flow intensity graph, functional similarity graph, and transaction amount influence graph, respectively; concat represents the concatenation operation; σ represents the activation function; W3 represents the learnable weight parameter; and b3 represents the bias.

[0087] Preferably, in step S405, high-order features in the spatiotemporal comprehensive features are extracted through a fully connected neural network to obtain the parts demand forecast data.

[0088] Preferably, in step S406, the loss function is calculated using the following formula:

[0089]

[0090] Where: N represents the number of samples in a training batch; y i Represents historical spare parts demand time series data; Represents the spare parts demand forecast data output by the demand forecast model.

[0091] Compared with the existing technology, the intelligent prediction method of automobile collaborative supply chain parts demand based on spatiotemporal graph convolution in this invention has the following beneficial effects:

[0092] By analyzing the dynamic correlation of historical parts demand time series data, the present invention can accurately depict the relationship between each node in the automotive supply chain (such as suppliers, manufacturers, distributors, etc.), thereby forming an accurate supply chain structure. This can overcome the limitations of traditional methods that rely on expert experience to define network structure, and significantly improve the ability to capture the dynamic evolution characteristics of the supply chain. At the same time, through graph construction technology, the supply chain network is decoupled in multiple dimensions and an independent adjacency matrix is ​​generated for each type of spatial dependency, avoiding the oversimplification of complex spatial relationships by a single adjacency matrix. The supply chain interaction characteristics at different semantic levels are retained through a multimodal adjacency matrix, and different types of adjacency matrices reflect the spatial dependencies of different dimensions in the supply chain. Inputting this information into the demand forecasting model can significantly improve the model's adaptability to complex supply chain environments, thereby helping to improve the accuracy of subsequent automotive collaborative supply chain parts demand forecasts.

[0093] The present invention inputs the historical parts demand time series data and various adjacency matrices into the trained demand forecasting model (AutoTG-LSTM) to obtain the corresponding parts demand forecast data. By introducing the spatiotemporal graph convolutional neural network, it effectively captures the spatiotemporal dynamic correlation between nodes in the automotive parts supply chain, and realizes comprehensive modeling of the complexity of key nodes and multi-source data. First, the spatiotemporal attention mechanism network of the AutoTG-LSTM model can perform temporal and spatial weighting on the historical parts demand time series data and each adjacency matrix through a dual weighting mechanism, highlighting key time points and spatial positions, thereby efficiently extracting useful spatiotemporal features. Secondly, the LSTM network of the AutoTG-LSTM model is good at capturing the nonlinear dynamic changes of time series. By processing the spatiotemporal weighted time series data of each adjacency matrix, it can accurately predict the time series trend of future parts demand, and model the periodicity, trend and random disturbance components of parts demand through a gating mechanism, solving the problem of insufficient fitting of traditional linear models to complex time series patterns. The AutoTG-LSTM model's GCN network then mines the spatial correlation features between nodes in the supply chain based on the complete time series characteristics of each adjacency matrix, and implements cross-node feature propagation. Graph convolution operations, based on multiple adjacency matrices, aggregate the temporal features of a single node with the spatial states of its upstream and downstream nodes, breaking through the "information island" effect caused by isolated predictions and further improving the accuracy of parts demand forecasts in the collaborative automotive supply chain. Finally, by fusing the aggregated feature matrix of all adjacency matrices to obtain comprehensive spatiotemporal features and further extracting high-order features, the demand forecasting model can comprehensively consider the complex relationships between time, space, and nodes in the supply chain, significantly improving the accuracy and reliability of forecasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] In order to make the purpose, technical solutions and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0095] Figure 1 This is the network structure diagram of the demand forecasting model (AutoTG-LSTM).

[0096] Figure 2 This is the demand graph data initially constructed using neo4j.

[0097] Figure 3 This is a comparison chart of model training effects.

[0098] Figure 4 A comparison chart of the impact of the number of layers and k value on model performance. DETAILED DESCRIPTION

[0099] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0100] It should be noted that similar reference numerals and letters denote similar items in the following figures. Therefore, once an item is defined in one figure, it does not require further definition or explanation in subsequent figures. In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" indicate positions or relationships based on the positions or relationships shown in the figures, or the positions or relationships in which the inventive product is typically placed when in use. These terms are intended solely to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation, and are therefore not to be construed as limiting the present invention. Furthermore, the terms "first," "second," and "third," etc., are used solely to distinguish descriptions and are not to be construed as indicating or implying relative importance. Furthermore, terms such as "horizontal" and "vertical" do not imply that a component is absolutely horizontal or overhanging, but rather may be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but rather may be slightly tilted. In the description of the present invention, it should also be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0101] The following is a further detailed description through specific implementation methods:

[0102] Example:

[0103] This embodiment discloses a method for intelligently predicting parts demand in an automotive collaborative supply chain based on spatiotemporal graph convolution.

[0104] An intelligent prediction method for parts demand in the automotive collaborative supply chain based on spatiotemporal graph convolution includes:

[0105] S1: Obtain historical parts demand time series data of the automotive supply chain;

[0106] S2: Analyze the supply chain structure based on historical time series data of parts demand; the supply chain structure consists of nodes and edges, where a node refers to a seller / supplier / manufacturer and contains information and data about related parts, and an edge refers to the relationship between sellers / suppliers / manufacturers.

[0107] S3: Build graphs of several spatial dependencies based on the supply chain structure, and calculate several adjacency matrices based on the graphs corresponding to various spatial dependencies;

[0108] S4: Input the historical spare parts demand time series data and various adjacency matrices into the trained demand forecasting model (hereinafter referred to as AutoTG-LSTM) and output the corresponding spare parts demand forecast data;

[0109] Combine Figure 1 As shown in Figure 2, the training steps of the demand forecasting model are as follows:

[0110] S401: Taking the historical spare parts demand time series data and various adjacency matrices as training data as model input;

[0111] S402: Using a spatiotemporal attention mechanism network, perform temporal and spatial weighting on the historical spare parts demand time series data and each adjacency matrix to obtain spatiotemporal weighted time series data for each adjacency matrix;

[0112] S403: Using an LSTM (Long Short-Term Memory) network, the time-series nonlinear dynamic changes of the spatiotemporal weighted time series data of each adjacency matrix are captured to obtain the complete time series features of each adjacency matrix;

[0113] S404: Using a GCN (graph convolutional network), the spatial correlation features of each adjacency matrix are captured based on the complete time series features of the adjacency matrix to obtain an aggregated feature matrix for each adjacency matrix; the aggregated feature matrices of all adjacency matrices are fused to obtain a spatiotemporal comprehensive feature;

[0114] S405: extracting high-order features from the spatiotemporal comprehensive features through a fully connected neural network to obtain parts demand forecast data;

[0115] S406: Calculate the loss function based on the historical spare parts demand time series data and the spare parts demand forecast data, and reverse the model parameters;

[0116] S407: Repeat steps S401 to S406 to iteratively train the model until the model converges or the maximum number of iterations is reached;

[0117] S5: The parts demand forecast data output by the demand forecast model is used as the collaborative demand forecast result of the automotive supply chain.

[0118] By analyzing the dynamic correlation of historical parts demand time series data, the present invention can accurately depict the relationship between each node in the automotive supply chain (such as suppliers, manufacturers, distributors, etc.), thereby forming an accurate supply chain structure. This can overcome the limitations of traditional methods that rely on expert experience to define network structure, and significantly improve the ability to capture the dynamic evolution characteristics of the supply chain. At the same time, through graph construction technology, the supply chain network is decoupled in multiple dimensions and an independent adjacency matrix is ​​generated for each type of spatial dependency, avoiding the oversimplification of complex spatial relationships by a single adjacency matrix. The supply chain interaction characteristics at different semantic levels are retained through a multimodal adjacency matrix, and different types of adjacency matrices reflect the spatial dependencies of different dimensions in the supply chain. Inputting this information into the demand forecasting model can significantly improve the model's adaptability to complex supply chain environments, thereby helping to improve the accuracy of subsequent automotive collaborative supply chain parts demand forecasts.

[0119] The present invention inputs the historical parts demand time series data and various adjacency matrices into the trained demand forecasting model (AutoTG-LSTM) to obtain the corresponding parts demand forecast data. By introducing the spatiotemporal graph convolutional neural network, it effectively captures the spatiotemporal dynamic correlation between nodes in the automotive parts supply chain, and realizes comprehensive modeling of the complexity of key nodes and multi-source data. First, the spatiotemporal attention mechanism network of the AutoTG-LSTM model can perform temporal and spatial weighting on the historical parts demand time series data and each adjacency matrix through a dual weighting mechanism, highlighting key time points and spatial positions, thereby efficiently extracting useful spatiotemporal features. Secondly, the LSTM network of the AutoTG-LSTM model is good at capturing the nonlinear dynamic changes of time series. By processing the spatiotemporal weighted time series data of each adjacency matrix, it can accurately predict the time series trend of future parts demand, and model the periodicity, trend and random disturbance components of parts demand through a gating mechanism, solving the problem of insufficient fitting of traditional linear models to complex time series patterns. The AutoTG-LSTM model's GCN network then mines the spatial correlation features between nodes in the supply chain based on the complete time series characteristics of each adjacency matrix, and implements cross-node feature propagation. Graph convolution operations, based on multiple adjacency matrices, aggregate the temporal features of a single node with the spatial states of its upstream and downstream nodes, breaking through the "information island" effect caused by isolated predictions and further improving the accuracy of parts demand forecasts in the collaborative automotive supply chain. Finally, by fusing the aggregated feature matrix of all adjacency matrices to obtain comprehensive spatiotemporal features and further extracting high-order features, the demand forecasting model can comprehensively consider the complex relationships between time, space, and nodes in the supply chain, significantly improving the accuracy and reliability of forecasts.

[0120] In order to better introduce the technical solution of the present invention, this embodiment is described through the following parts.

[0121] 1. Adjacency Matrix

[0122] In this embodiment, the adjacency matrix is ​​calculated through the following steps.

[0123] S301: Graph construction is used to map the three spatial dependencies of the supply chain structure, namely, flow intensity, functional relevance, and transaction amount impact, to obtain a flow intensity graph, a functional similarity graph, and a transaction amount impact graph;

[0124] In this embodiment, the flow intensity graph encodes the flow volume and interaction intensity between nodes. The functional similarity graph encodes the functional and attribute similarity of nodes. The transaction amount impact graph encodes the flow intensity of the impact of transaction amounts generated between nodes on other nodes.

[0125] S302: Calculate the adjacency matrix A of the flow intensity graph, functional similarity graph, and transaction amount influence graphflow 、A sim and A tran .

[0126] 1. Flow Intensity: Transaction orders between nodes can be viewed as the flow of parts. Flow intensity represents the influence of one node on another. All nodes in the supply chain experience flow, but nodes with greater flow intensity deserve more attention because they have a greater impact on the entire supply chain. Therefore, nodes with strong flow should be given greater weight in the model.

[0127] The adjacency matrix A of the flow intensity graph is calculated by the following formula flow Elements in:

[0128]

[0129] Where: I ij Represents the flow intensity between node i and node j, which is used to construct the adjacency matrix A of the flow intensity graph flow ;T ij represents the number of transactions between node i and node j; ∑ N T ij Represents the total flow of node i; N represents the total number of nodes. By normalizing the transaction volume, we can get a matrix representing the relative flow intensity between nodes. Each element I in this matrix is ij The value of is between [0,1].

[0130] 2. Functional Relevance: In a supply chain network, each node (e.g., supplier, distributor, retailer) may possess distinct attributes that determine its functional role within the supply chain. For example, some nodes may compete due to similar functional attributes, while others may collaborate due to participating in the same order. This functional similarity between nodes can be measured using attribute similarity, which is crucial for improving the model's predictive performance.

[0131] The adjacency matrix A of the functional similarity graph is calculated by the following formula sim Elements in:

[0132]

[0133] Where: S ij Represents the functional similarity between node i and node j, which is used to construct the adjacency matrix A of the functional similarity graph sim POI K (i) and POI K(j) represents the value of the Kth attribute for node i and node j, respectively. By calculating the cosine similarity between node attributes, we can quantify the functional similarity between nodes. Similarity values ​​range from [0, 1], with values ​​closer to 1 indicating greater functional similarity between nodes.

[0134] 3. Transaction Amount Influence: In supply chain networks, transaction amount is a key indicator of inter-node influence. Larger transaction amounts generally indicate a node's greater economic importance within the supply chain and its greater influence on other nodes. Therefore, the influence of transaction amounts between nodes should not be neglected when modeling.

[0135] The adjacency matrix A of the transaction amount influence graph is calculated using the following formula: tran Elements in:

[0136]

[0137] Where: A ij Represents the transaction amount influence between node i and node j, which is used to construct the adjacency matrix A of the transaction amount influence graph tran ; C ij Represents the transaction amount between node i and node j; ∑ N C ij represents the total transaction amount of node i; N represents the total number of nodes. This normalization method calculates the transaction amount influence between each pair of nodes and forms an influence matrix with values ​​ranging from [0, 1]. A higher transaction amount influence means a larger transaction amount between nodes, resulting in a greater influence on other nodes.

[0138] 2. Spatial-Temporal Attention Mechanism Network

[0139] In this embodiment, the processing steps of the spatiotemporal attention mechanism network include:

[0140] S4021: The adjacency matrix A (x) And historical spare parts demand time series data X as input;

[0141] Among them A (x) ={A flow ,A sim ,A tran}, A flow 、A sim 、A tran The adjacency matrices representing the flow intensity graph, functional similarity graph, and transaction amount influence graph, respectively;

[0142] S4022: Perform time series pooling on the historical spare parts demand time series data X to obtain the feature sum X at each time step. pool ;

[0143] The formula is:

[0144]

[0145] Where: X pool represents the sum of features over time steps; X represents the three-dimensional tensor of historical spare parts demand time series data; X[:,:,t] represents extracting the node features corresponding to time step t; X[:,:,:] represents retaining the node features of all time steps; the result of this pooling operation is the sum of the node features of all time steps, which is used to represent the global features in the time dimension.

[0146] S4023: Transpose the historical spare parts demand time series data X to adapt to the subsequent graph convolution operation and obtain X gcn_input ; Through the GCN network layer to X gcn_input and the adjacency matrix A (x) Sum along the node dimension to get the adjacency matrix A (x) Features (the sum of graph features at each time step); Sum the graph features over the time steps to obtain the adjacency matrix A (x) The spatial characteristics X gcn_pool ;

[0147] The formula is:

[0148] X gcn_input =Transpose(X,prem=[0,2,1]);

[0149]

[0150] Where: Represents the adjacency matrix A (x) Transpose represents the transposition operation; prem represents the transformed dimension order (originally 012, now 021); X represents the three-dimensional tensor of the historical spare parts demand time series data; X gcn_input Represents a three-dimensional tensor of transposed time series data; GCN represents a graph convolution operation;

[0151] S4024: Sum the features of the time step X pool With the adjacency matrix A (x) The spatial characteristics X gcn_pool Perform fusion and normalization to obtain the adjacency matrix A (x) The spatiotemporal characteristics of z (x) ;

[0152] The formula is:

[0153]

[0154] Where: z (x) Represents the adjacency matrix A (x) The spatiotemporal characteristics of; N represents the number of nodes;

[0155] S4025: Based on the adjacency matrix A through the multi-head attention mechanism (x) The spatiotemporal characteristics of z (x) Calculating attention scores And the attention score Weighted with the historical spare parts demand time series data X, we get the adjacency matrix A (x) Spatiotemporal weighted time series data

[0156] The formula is:

[0157]

[0158] Where: Represents the adjacency matrix A (x) The spatiotemporal weighted time series data is ; h∈{1,2,…,H} represents the number of multi-head attention mechanisms; b1 and b2 represent bias vectors; ReLU and σ represent activation functions; B represents the batch size; T represents the number of time steps; and s represents the output of all attention heads aggregated.

[0159] 3. LSTM Network

[0160] In this embodiment, the processing steps of the LSTM network include:

[0161] S4031: Use exponential smoothing to calculate the adjacency matrix A (x) Spatiotemporal weighted time series data Perform preprocessing to obtain a preprocessed sequence

[0162] The formula is:

[0163] X (t) =X reweight ;

[0164]

[0165] Where: Represents the preprocessing sequence at the current moment; represents the preprocessing sequence of the previous moment; α represents the smoothing coefficient; X (t) Represents the spatiotemporal weighted time series data at the current moment;

[0166] S4032: Preprocess the sequence Perform node segmentation to obtain the time series of each node i

[0167] The formula is:

[0168]

[0169] Where: Represents the time series of the first node;

[0170] S4033: The time series of each node i Input into the LSTM network and get the output of each node i Output for all nodes Perform splicing to obtain the adjacency matrix A (x) The complete time series characteristics H (T) ;

[0171] The formula is:

[0172]

[0173] Where: H (T) Represents the adjacency matrix A (x) The complete time series characteristics of represents the time series of node i; D represents the dimension of the hidden state in LSTM.

[0174] 4. GCN Network

[0175] In this embodiment, the processing steps of the GCN network include:

[0176] S4041: Adjacency matrix A (x) The complete time series characteristics H (T) Transpose to obtain the transposed time feature H0;

[0177] The formula is:

[0178]

[0179] Where: H0 represents the adjacency matrix A (x) The transposed time characteristics of

[0180] S4042: Through the adjacency matrix A (x) The transposed time feature H0 of the corresponding adjacency matrix A (x) Perform the first step of graph convolution to obtain the adjacency matrix A (x) The feature matrix of the first propagation step

[0181] The formula is:

[0182]

[0183] Where: Represents the adjacency matrix A (x) The feature matrix of the first propagation step;

[0184] S4043: Multi-step propagation of adjacency matrix A through graph convolution (x) The feature matrix of is weighted summed up, and the information from the neighboring nodes is gradually aggregated to obtain the adjacency matrix A (x) The feature matrix propagated at each step;

[0185] The formula is:

[0186]

[0187] Where: Represents the adjacency matrix A (x) The characteristic matrix of the k-th propagation step; k∈2,3,...,T max Represents the number of propagations in the GCN network. After each propagation step, the feature matrix is ​​recalculated and updated;

[0188] S4044: Adjacency matrix A (x) The feature matrices of all propagation steps are concatenated to obtain the adjacency matrix A (x) The aggregate feature matrix The aggregate feature matrix corresponding to the adjacency matrix of the fusion flow intensity graph, functional similarity graph and transaction amount influence graph is obtained to obtain the fusion feature matrix H' GCN ; For the fusion feature matrix H' GCN Perform linear transformation and nonlinear transformation activation to obtain the spatiotemporal comprehensive feature H GCN ;

[0189] The formula is:

[0190]

[0191] H GCN =σ(H' GCN W3+b3);

[0192] Where: They represent the aggregated feature matrices corresponding to the adjacency matrices of the flow intensity graph, functional similarity graph, and transaction amount influence graph, respectively; concat represents the concatenation operation; σ represents the activation function; W3 represents the learnable weight parameter; and b3 represents the bias.

[0193] 5. Loss Function

[0194] In this embodiment, the loss function is calculated using the following formula:

[0195]

[0196] Where: N represents the number of samples in a training batch; y i Represents historical spare parts demand time series data; Represents the spare parts demand forecast data output by the demand forecast model.

[0197] 6. Experimental Description

[0198] In order to better illustrate the advantages of the technical solution of the present invention, this embodiment discloses the following experiments.

[0199] This experiment used the actual auto parts supply chain information of the project partner Dongfeng Motor Corporation to test the demand forecasting model (AutoTG-LSTM) proposed in this invention.

[0200] 1. Dataset

[0201] We experimented with the proposed AutoTG-LSTM model on a Dongfeng Motor automotive parts demand dataset. This dataset consists of orders generated from key automotive parts operations and maintenance between January 1, 2023, and December 31, 2023. For data segmentation, we used data from January 1, 2023, to July 31, 2023, for training; data from August 1, 2023, to September 30, 2023, for validation; and data from October 1, 2023, to December 31, 2023, for testing.

[0202] 2. Training details and parameter settings

[0203] The dataset contains the parts order number, supplier ID, seller ID, manufacturer ID, sales date, price, etc., as shown in Table 1. The automotive parts supply chain operation and maintenance demand data is processed through neo4j to form a graph structure that can be used for graph embedding analysis. Figure 2 As shown in Figure 2, by performing graph embedding operations on it, we obtain the corresponding three adjacency matrix weight graphs. The weight graphs represent the dependencies between nodes in the three different relationships.

[0204] Table 1 Partial dataset of automobile parts

[0205]

[0206] By performing graph embedding operations on it, the corresponding three adjacency matrix weight graphs are obtained.

[0207] 3. Experimental results

[0208] 1) Model training results

[0209] We use ReLU as the activation function in each layer of the graph convolutional network to enhance the network's ability to express nonlinear features. The entire model adopts the AutoTG-LSTM architecture, in which the learning rate is set to 0.001, and L2 regularization is introduced to reduce the risk of overfitting. The data is randomly sampled to generate batches during the training process to ensure the stability and diversity of the model training. All neural network-based methods are implemented using the TensorFlow framework, and the Adam optimizer is used to update the network parameters. In order to ensure efficient convergence and robustness of the model, we built a custom spatiotemporal graph convolution prediction model based on the pre-trained graph convolution model, and used its default configuration as the initial parameters. During the training of the model, the following training parameters were set: batch size is 32, the maximum length of the sequence is 12, the learning rate is set to 0.001, the input and output dimensions are both set to 40, and the number of layers of the spatiotemporal graph convolutional neural network is set to 2. The early stopping mechanism was not used during the training process. The selection of other hyperparameters was based on the default configuration and the performance of the model on the training set. The training results are shown in Figure 2. Figure 3 shown.

[0210] Experimental results demonstrate that our model demonstrates excellent performance for automotive parts supply chain demand information. The model is able to track the changing trends of the ground truth well. The predicted values ​​are consistent with the ground truth in multiple rising and falling intervals, demonstrating the model's strong ability to capture the changing trends of time series data. In the mid- to high-value regions of the data distribution (e.g., data points 400 to 600), the model's predictions are highly consistent with the ground truth, demonstrating the model's stable performance across these key numerical ranges. However, certain outliers or short-term fluctuations result in deviations between the predicted and ground truth values ​​(RMSE = 2.094) around data points 200 to 300 and 700 to 800, but these deviations are within a reasonable range. Therefore, our method maintains high accuracy when handling this type of demand information forecasting task, validating the effectiveness of our proposed method for the current automotive collaborative supply chain parts demand forecasting task.

[0211] 2) Ablation experiment

[0212] To investigate the contribution of each module to the overall performance of the AutoTG-LSTM model, we designed a series of ablation experiments, gradually removing or modifying key components of the model. By comparing the prediction results across different experiments, we quantified the effectiveness of each module and verified the rationality of the model design. We tested the following model variants:

[0213] (1) Using only a single adjacency matrix, without graph convolution and spatiotemporal graph attention mechanism. To study the impact of spatiotemporal dependency model, we evaluate the following variants of ST-MGCN by removing different components from the model, including: (1) flow intensity graph, (2) functional similarity graph, and (3) transaction amount intensity graph. The results are shown in Table 2. Removing any graph component leads to a significant increase in error, which demonstrates the importance of each type of relationship. These graphs encode important prior knowledge, namely regional correlations, which can be used for more accurate predictions.

[0214] (2) Remove the spatiotemporal graph attention mechanism (TGA) and only use graph convolutional GCN and LSTM for prediction.

[0215] (3) Remove the graph convolutional network (GCN) and only use the spatiotemporal graph attention mechanism and LSTM.

[0216] Table 2 shows the performance comparison of these models under the demand information of the automotive parts supply chain.

[0217] Table 2 Performance comparison of different auto parts demand forecasting model methods

[0218]

[0219] The ablation results in Table 2 show that when GCN is removed, the lack of spatial feature fusion leads to a significant decline in model prediction performance, with both RMSE and MSE significantly higher than those of AutoTG-LSTM. This demonstrates that graph convolution plays a crucial role in capturing spatiotemporal dependencies between nodes. Similarly, when TGA is removed, although the graph convolutional network is still used, the model lacks the performance of AutoTG-LSTM. This suggests that the spatiotemporal graph attention mechanism can further enhance the model's ability to learn complex dependencies between nodes, thereby improving prediction accuracy. The most direct comparison method is the method using a single adjacency matrix. However, this method fails to effectively capture the full spatiotemporal relationships, resulting in reduced prediction accuracy. Consequently, both RMSE and MSE are significantly higher than those of other methods. In summary, the graph convolutional neural network and the spatiotemporal attention mechanism significantly enhance performance in demand forecasting.

[0220] 3) Performance comparison of different spatiotemporal graph convolutional neural network layers

[0221] To study the impact of different hyperparameters of the proposed model, we evaluated the model on the original dataset and varied the two most important hyperparameters, namely the number of layers and the value of k.

[0222] from Figure 4The data shown here show that the impact of the k value on model performance fluctuates with the increase in the number of layers. When the number of layers is low (e.g., layer = 1 and layer = 3), k = 2 and k = 3 perform better, while when the number of layers is high (e.g., layer = 4 and layer = 5), k = 4 and k = 2 perform more stably. This variation indicates that model performance depends not only on the k value but also on the number of layers. This may be because increasing the number of layers allows the model to better capture complex spatiotemporal relationships, but also brings a certain risk of overfitting, especially at higher k values.

[0223] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the technical solutions. Those skilled in the art should understand that modifications or equivalent replacements of the technical solutions of the present invention that do not depart from the purpose and scope of the technical solutions of the present invention should be included in the scope of the claims of the present invention.

Claims

1. An intelligent prediction method for parts demand in the automotive collaborative supply chain based on spatiotemporal graph convolution, characterized by: include: S1: Obtain historical parts demand time series data of the automotive supply chain; S2: Analyze supply chain structure based on historical spare parts demand time series data; S3: Build graphs of several spatial dependencies based on the supply chain structure, and calculate several adjacency matrices based on the graphs corresponding to various spatial dependencies; S4: Input the historical spare parts demand time series data and various adjacency matrices into the trained demand forecasting model and output the corresponding spare parts demand forecast data; The training steps of the demand forecasting model are as follows: S401: Taking the historical spare parts demand time series data and various adjacency matrices as training data as model input; S402: Using a spatiotemporal attention mechanism network, perform temporal and spatial weighting on the historical spare parts demand time series data and each adjacency matrix to obtain spatiotemporal weighted time series data for each adjacency matrix; S403: Using the LSTM network, the time-series nonlinear dynamic changes of the spatiotemporal weighted time series data of each adjacency matrix are captured to obtain the complete time series features of each adjacency matrix; S404: Using the GCN network, the spatial correlation features of each adjacency matrix are captured based on the complete time series features of the adjacency matrix to obtain an aggregated feature matrix of each adjacency matrix; the aggregated feature matrices of all adjacency matrices are fused to obtain a spatiotemporal comprehensive feature; S405: extracting high-order features from the spatiotemporal comprehensive features to obtain parts demand forecast data; S406: Calculating a loss function based on the historical spare parts demand time series data and the spare parts demand forecast data, and reversely optimizing the model parameters; S407: Repeat steps S401 to S406 to iteratively train the model until the model converges or the maximum number of iterations is reached; S5: The parts demand forecast data output by the demand forecast model is used as the collaborative demand forecast result of the automotive supply chain.

2. The method for intelligent prediction of parts demand in a collaborative automotive supply chain based on spatiotemporal graph convolution as claimed in claim 1, characterized in that: Step S3 specifically includes the following steps: S301: Graph construction is used to map the three spatial dependencies of the supply chain structure, namely, flow intensity, functional relevance, and transaction amount impact, to obtain a flow intensity graph, a functional similarity graph, and a transaction amount impact graph; S302: Calculate the adjacency matrix A of the flow intensity graph, functional similarity graph, and transaction amount influence graph flow 、A sim and A tran .

3. The method for intelligent prediction of parts demand in a collaborative automotive supply chain based on spatiotemporal graph convolution as claimed in claim 2, characterized in that: In step S302, the adjacency matrix A of the flow intensity graph is calculated using the following formula: flow Elements in: Where: I ij Represents the flow intensity between node i and node j, which is used to construct the adjacency matrix A of the flow intensity graph flow ;T ij represents the number of transactions between nodes i and j; N represents the total number of nodes.

4. The method for intelligent prediction of parts demand in a collaborative automotive supply chain based on spatiotemporal graph convolution as claimed in claim 2, characterized in that: In step S302, the adjacency matrix A of the functional similarity graph is calculated using the following formula: sim Elements in: Where: S ij Represents the functional similarity between node i and node j, which is used to construct the adjacency matrix A of the functional similarity graph sim ; POI K (i) and POI K (j) represents the value of the Kth attribute of node i and node j respectively.

5. The method for intelligent prediction of parts demand in a collaborative automotive supply chain based on spatiotemporal graph convolution as claimed in claim 2, characterized in that: In step S302, the adjacency matrix A of the transaction amount influence graph is calculated using the following formula: tran Elements in: Where: A ij Represents the transaction amount influence between node i and node j, which is used to construct the adjacency matrix A of the transaction amount influence graph tran ; C ij represents the transaction amount between node i and node j; N represents the total number of nodes.

6. The method for intelligent prediction of parts demand in a collaborative automotive supply chain based on spatiotemporal graph convolution as claimed in claim 2, characterized in that: In step S402, the processing steps of the spatiotemporal attention mechanism network include: S4021: The adjacency matrix A (x) And historical spare parts demand time series data X as input; Among them A (x) ={A flow ,A sim ,A tran }, A flow 、A sim 、A tran The adjacency matrices representing the flow intensity graph, functional similarity graph, and transaction amount influence graph, respectively; S4022: Perform time series pooling on the historical spare parts demand time series data X to obtain the feature sum X at each time step. pool ; The formula is: Where: X pool represents the sum of features over the time step; X represents the three-dimensional tensor of the historical spare parts demand time series data; X[:,:,t] represents the extraction of the node features corresponding to the t time step; S4023: Transpose the historical spare parts demand time series data X to obtain X gcn_input ; Through the GCN network layer to X gcn_input and the adjacency matrix A (x) Sum along the node dimension to get the adjacency matrix A (x) Features right Sum the graph features over the time steps to obtain the adjacency matrix A (x) The spatial characteristics X gcn_pool ; The formula is: X gcn_input =Transpose(X,prem=[0,2,1]); Where: Represents the adjacency matrix A (x) Transpose represents the transposition operation; prem represents the transformed dimension order; X represents the three-dimensional tensor of the historical spare parts demand time series data; X gcn_input Represents a three-dimensional tensor of transposed time series data; GCN represents a graph convolution operation; S4024: Sum the features of the time step X pool With the adjacency matrix A (x) The spatial characteristics X gcn_pool Perform fusion and normalization to obtain the adjacency matrix A (x) The spatiotemporal characteristics of z (x) ; The formula is: Where: z (x) Represents the adjacency matrix A (x) The spatiotemporal characteristics of; N represents the number of nodes; S4025: Based on the adjacency matrix A through the multi-head attention mechanism (x) The spatiotemporal characteristics of z (x) Calculating attention scores And the attention score Weighted with the historical spare parts demand time series data X, we get the adjacency matrix A (x) Spatiotemporal weighted time series data The formula is: Where: Represents the adjacency matrix A (x) The spatiotemporal weighted time series data is ; h∈{1,2,…,H} represents the number of multi-head attention mechanisms; b1 and b2 represent bias vectors; ReLU and σ represent activation functions; B represents the batch size; T represents the number of time steps; and s represents the output of all attention heads aggregated.

7. The method for intelligent prediction of parts demand in a collaborative automotive supply chain based on spatiotemporal graph convolution as claimed in claim 6, characterized in that: In step S403, the processing steps of the LSTM network include: S4031: Use exponential smoothing to calculate the adjacency matrix A (x) Spatiotemporal weighted time series data Perform preprocessing to obtain a preprocessed sequence The formula is: X (t) =X reweight ; Where: Represents the preprocessing sequence at the current moment; represents the preprocessing sequence of the previous moment; α represents the smoothing coefficient; X (t) Represents the spatiotemporal weighted time series data at the current moment; S4032: Preprocess the sequence Perform node segmentation to obtain the time series of each node i The formula is: Where: Represents the time series of the first node; S4033: The time series of each node i Input into the LSTM network and get the output of each node i Output for all nodes Perform splicing to obtain the adjacency matrix A (x) The complete time series characteristics H (T) ; The formula is: Where: H (T) Represents the adjacency matrix A (x) The complete time series characteristics of represents the time series of node i; D represents the dimension of the hidden state in LSTM.

8. The method for intelligent prediction of parts demand in a collaborative automotive supply chain based on spatiotemporal graph convolution as claimed in claim 7, characterized in that: In step S404, the processing steps of the GCN network include: S4041: Adjacency matrix A (x) The complete time series characteristics H (t) Transpose to obtain the transposed time feature H0; The formula is: Where: H0 represents the adjacency matrix A (x) The transposed time characteristics of S4042: Through the adjacency matrix A (x) The transposed time feature H0 of the corresponding adjacency matrix A (x) Perform the first step of graph convolution to obtain the adjacency matrix A (x) The feature matrix of the first propagation step The formula is: Where: Represents the adjacency matrix A (x) The feature matrix of the first propagation step; S4043: Multi-step propagation of adjacency matrix A through graph convolution (x) The feature matrix of is weighted summed up, and the information from the neighboring nodes is gradually aggregated to obtain the adjacency matrix A (x) The feature matrix propagated at each step; The formula is: Where: Represents the adjacency matrix A (x) The characteristic matrix of the k-th propagation step; k∈2,3,...,T max Represents the number of propagations in the GCN network. After each propagation step, the feature matrix is ​​recalculated and updated; S4044: Adjacency matrix A (x) The feature matrices of all propagation steps are concatenated to obtain the adjacency matrix A (x) The aggregate feature matrix The aggregate feature matrix corresponding to the adjacency matrix of the fusion flow intensity graph, functional similarity graph, and transaction amount influence graph is obtained to obtain the fusion feature matrix H′ GCN ; For the fusion feature matrix H′ GCN Perform linear transformation and nonlinear transformation activation to obtain the spatiotemporal comprehensive feature H GCN ; The formula is: H GCN =σ(H′ GCN W3+b3); Where: They represent the aggregated feature matrices corresponding to the adjacency matrices of the flow intensity graph, functional similarity graph, and transaction amount influence graph, respectively; concat represents the concatenation operation; σ represents the activation function; W3 represents the learnable weight parameter; and b3 represents the bias.

9. The method for intelligent prediction of auto parts demand in collaborative supply chain based on spatiotemporal graph convolution according to claim 1, characterized in that: In step S405, high-order features in the spatiotemporal comprehensive features are extracted through a fully connected neural network to obtain parts demand forecast data.

10. The method for intelligent prediction of auto parts demand in collaborative supply chain based on spatiotemporal graph convolution according to claim 1, characterized in that: In step S406, the loss function is calculated using the following formula: Where: N represents the number of samples in a training batch; y i Represents historical spare parts demand time series data; Represents the spare parts demand forecast data output by the demand forecast model.

Images